PREDICTION AND ANALYSIS OF POSTPARTUM DEPRESSION WITH
CHRONIC DISEASES AS RISK FACTORS
by
Febina Muhamed Rajeesh
M.C.A., Cochin University of Science and Technology, 2014

THESIS SUBMITTED IN PARTIAL FULFILMENT OF
THE REQUIREMENTS FOR THE DEGREE OF
MASTER OF SCIENCE
IN
COMPUTER SCIENCE

UNIVERSITY OF NORTHERN BRITISH COLUMBIA
March 2025
© Febina Muhamed Rajeesh, 2025

Abstract
Postpartum Depression (PPD) is a mental health condition that is a leading cause of annually
reported infanticide incidence. Many cases go underdiagnosed due to unawareness, and
prolonged untreated conditions can lead to psychosis, causing harm to themselves and the
infant. Hence, identifying the PPD risk has become crucial and has been widely studied in the
context of traditional risk factors. Only limited research has been conducted addressing chronic
diseases as the risk factor.

Predicting PPD by utilizing the power of Machine Learning (ML) algorithms can lead to timely
intervention and management of the condition. Data obtained from the Center for Disease
Control and Prevention – Pregnancy Risk Assessment Monitoring System (CDC-PRAMS) was
used for this thesis to identify the risk factors and forecast the likelihood of depression for
mothers who suffer from one or more chronic diseases using ML models. The performance
evaluation of the selected machine learning models—Support Vector Machines (SVM),
Random Forest (RF), Logistic Regression (LR), and Neural Network (NN) was assessed using
accuracy and F1-score, which ranged from 76% and 77% for NN to 89% and 88% for LR. The
impact of each key predictor identified in the SHAP analysis demonstrated close alignment
across all models and highlighted the significance of chronic disease. The results also highlight
how chronic diseases potentially interact with other common risk factors to increase the
likelihood of PPD. An interactive dashboard is created to visualize and present preprocessed
data using charts and graphs. Also, a diagnostic screening tool developed based on the trained
models demonstrates the potential of ML as a screening tool to improve diagnostic precision
and support personalized care for enhanced quality of life.
ii

Table of Contents

Abstract .................................................................................................................................................. ii
Table of Contents .................................................................................................................................. iii
List of Tables......................................................................................................................................... vi
List of Figures ...................................................................................................................................... vii
List of Equations ................................................................................................................................... ix
List of Abbreviations.............................................................................................................................. x
Acknowledgment .................................................................................................................................. xi
Chapter 1 ................................................................................................................................................ 1
Introduction ............................................................................................................................................ 1
1.1

Utilizing ML in Healthcare: Addressing PPD........................................................................ 3

1.1.1

Postpartum Depression (PPD) ...........................................................................................4

1.1.2

Predictive Modeling for PPD Detection ............................................................................6

1.1.3

PPD Screening Tools.........................................................................................................8

1.2

Problem Statement ................................................................................................................. 9

1.3

Research Methodology......................................................................................................... 10

1.4

Data Source .......................................................................................................................... 11

1.5

Contributions ........................................................................................................................ 13

1.6

Organization of Thesis ......................................................................................................... 14

Chapter 2 .............................................................................................................................................. 15
Literature Review ................................................................................................................................. 15
2.1

PPD Predictive Analysis ...................................................................................................... 15

2.1.1

PPD Risk Factors Analysis..............................................................................................21

2.1.2

Summary .........................................................................................................................29

iii

2.2

Postpartum Depression Risk Factors: Chronic Illness ......................................................... 30

2.2.1
2.3

Summary .........................................................................................................................34

CDC-PRAMS Data PPD Analysis ....................................................................................... 34

2.3.1

Summary .........................................................................................................................38

2.4

PPD Tool .............................................................................................................................. 38

2.5

Summary .............................................................................................................................. 40

Chapter 3 .............................................................................................................................................. 42
Methodology ........................................................................................................................................ 42
3.1

Data Collection..................................................................................................................... 43

3.1.1
3.2

Ethics Approval ...............................................................................................................44

Data Preparation ................................................................................................................... 44

3.2.1

Data Collection ................................................................................................................45

3.2.2

Data Preprocessing ..........................................................................................................45

3.2.3

Feature Selection .............................................................................................................49

3.2.4

Feature Engineering and Derived Variables ....................................................................50

3.2.5

Data Encoding and Scaling .............................................................................................51

3.2.6

Challenges .......................................................................................................................52

3.3

SPSS Modeler ...................................................................................................................... 53

3.4

Predictive Model .................................................................................................................. 54

3.4.1

Artificial Neural Network................................................................................................56

3.4.2

Logistic Regression .........................................................................................................58

3.4.3

Random Forest Tree ........................................................................................................60

3.4.4

Support Vector Machine..................................................................................................62

3.5

Dashboard ............................................................................................................................ 63

3.6

Summary .............................................................................................................................. 64

Chapter 4 .............................................................................................................................................. 65

iv

Experiment and Results........................................................................................................................ 65
4.1

PPD Dashboard .................................................................................................................... 65

4.1.1

Summary .........................................................................................................................70

4.2

Evaluation Metrics: Accuracy and F1 Score ........................................................................ 71

4.3

Predictive Modeling: An Analysis of Four Models to Predict PPD..................................... 72

4.4

Model Comparison ............................................................................................................... 73

4.4.1
4.5

Performance Metrics: Discussion ....................................................................................74

Post-Training Analysis: Identify Key Predictors ................................................................. 76

4.5.1

Summary: Post-Training Analysis ..................................................................................88

4.6

Identifying and Analyzing Interactions Among Key Predictors .......................................... 90

4.7

Summary: Interaction Analysis .......................................................................................... 100

4.8

PD Screening Tool ............................................................................................................. 102

4.9

Conclusion.......................................................................................................................... 104

Chapter 5 ............................................................................................................................................ 106
Conclusion and Future Work ............................................................................................................. 106
5.1

Future Work ....................................................................................................................... 108

v

List of Tables
Table 1. Predictor Variables ................................................................................................................. 12
Table 2. Precision, Recall, and F1 Score.............................................................................................. 73
Table 3. Key Features and Their Predictive Strength by Model .......................................................... 89
Table 4. Top Common Key Interactive Features Identified by the Selected Models ........................ 101

vi

List of Figures
Figure 1. Projected Growth of the Global Mental Health Apps Market (2024-2029) [5]...................... 2
Figure 2. Spectrum of PPD [14]. ............................................................................................................ 5
Figure 3. PPD Rate Trend in 2014 -2018 [46]. .................................................................................... 21
Figure 4. PPD and Pre-Existing Health Conditions [46]...................................................................... 22
Figure 5. Proposed Research Model. ................................................................................................... 42
Figure 6. Data Preparation Process [70]............................................................................................... 45
Figure 7. a) ANN (b)Structure of ANN for a Neuron in a Hidden Layer with One Output [76]. ....... 57
Figure 8. The Sigmoid Curve in LR [78] ............................................................................................. 59
Figure 9. RF: An Ensemble Learning Algorithm [80]. ........................................................................ 61
Figure 10. Linear SVM [82]................................................................................................................. 62
Figure 11. PPD Dashboard ................................................................................................................... 66
Figure 12. PPD Key Influencers .......................................................................................................... 67
Figure 13. PPD and Maternal Age Group ............................................................................................ 68
Figure 14. PPD by Maternal Education Level and Marital Status (Married) ....................................... 68
Figure 15. PPD by Disease Factors ...................................................................................................... 69
Figure 16. PPD by Prenatal Factors. .................................................................................................... 70
Figure 17. Model Comparison: Accuracy and F1 Score. ..................................................................... 73
Figure 18. Feature Significance Using SHAP on NN Model............................................................... 78
Figure 19. SHAP Summary: Visualizing Feature Impact and Distribution Using NN Model ............ 79
Figure 20. Feature Significance Using SHAP on LR Model ............................................................... 80
Figure 21. SHAP Summary: Visualizing Feature Impact and Distribution Using LR Model ............. 81
Figure 22. Feature Significance Using SHAP on RF Model ............................................................... 83
Figure 23. SHAP Summary: Visualizing Feature Impact and Distribution Using RF Model ............. 84
Figure 24. Feature Significance Using SHAP on the SVM Model ...................................................... 86

vii

Figure 25. SHAP Summary: Visualizing Feature Impact and Distribution Using SVM Model ......... 87
Figure 26. SPSS Feature Selection Node ............................................................................................. 91
Figure 27. Feature Selection Results .................................................................................................... 92
Figure 28. Interaction Feature Significance Using SHAP on NN Model ............................................ 94
Figure 29. Interaction Feature Significance Using SHAP on LR Model ............................................. 96
Figure 30. Interaction Feature Significance Using SHAP on RF Model ............................................. 98
Figure 31. Interaction Feature Significance Using SHAP on the SVM Model.................................... 99
Figure 32. PPD Screening Tool - Front End ...................................................................................... 103
Figure 33. PPD Screening Tool - Front End ...................................................................................... 103
Figure 34. PPD Screening Tool - Prediction ...................................................................................... 104

viii

List of Equations
Equation 1. Output of the Neuron ........................................................................................................ 57
Equation 2. Sigmoid Function.............................................................................................................. 59
Equation 3. Accuracy by Percentage.................................................................................................... 71
Equation 4. F1 Score Formula. ............................................................................................................ 71
Equation 5. Precision Formula ............................................................................................................. 72
Equation 6. Recall Formula.................................................................................................................. 72

ix

List of Abbreviations
AUC

Area Under Curve

BI

Business Intelligence

CDC

Centre for Disease and Control

EHR

Electronic Health Record

GBM

Gradient Boosting Machine

LR

Logistic Regression

ML

Machine Learning

NN

Neural Network

OR

Odds Ratio

PPD

Postpartum Depression

PRAMS

Pregnancy Risk Assessment Monitoring System

RF

Random Forest

SVM

Support Vector Machine

x

Acknowledgment
I would like to begin by expressing my heartfelt gratitude to my family, especially my husband,
Muhamed Rajeesh, our children, and our parents, whose unwavering love, support, and
encouragement have been my greatest source of strength throughout this journey. Next, I
acknowledge the traditional and unceded territory of Lheidli Tenneh on which UNBC is
located. I am grateful for the opportunity to learn, work, and conduct research on this land.

I extend my sincere thanks to my supervisor, Dr. Waqar Haque, for his invaluable guidance,
motivation, and support. His guidance has been essential in shaping my research and academic
growth. I would like to extend my deepest gratitude to my committee members, Dr. Fan Jiang
and Dr. Kafui Monu, for their thoughtful feedback, constructive suggestions, and
encouragement.

Special thanks to the PRAMS Working Group, which includes the PRAMS Team, Division of
Reproductive Health, CDC, and the PRAMS sites for their role in conducting PRAMS
surveillance and allowing the use of their data. I want to express my heartfelt gratitude to Dr.
Rahul G.P. Sukumariamma for his invaluable assistance in reviewing the data. I am also
thankful for the support of my friends in Prince George and India. Additionally, I would like
to extend special thanks to my colleagues at BIRG LAB for their ongoing encouragement and
motivation throughout this journey.

Thank you all for being part of this significant chapter in my life.

Febina Muhamed Rajeesh
xi

Chapter 1
Introduction
Processing large volumes of complex health data with traditional analytics methods often takes
too long to reveal valuable insights. As a branch of Artificial Intelligence (AI), Machine
Learning (ML) techniques like Deep Neural Network (DNN) models have gained increased
attention in extracting hidden patterns, finding relationships within datasets, and other valuable
information from the massive quantity of health data. ML can enhance focused patient care by
learning from the patterns and making predictions to improve speed and precision [1]. In
addition to healthcare, it provides valuable benefits to other industries, such as finance, to
improve trading strategies and various other sectors to adopt automated systems [2].

The significance of AI in the medical field has grown as ML techniques are increasingly used
to diagnose, forecast, and treat various health conditions, including mental illnesses [3]. With
the help of vast patient datasets, these systems can aid clinicians in making more accurate
assessments and individualized treatment plans, thereby improving the overall quality of
mental healthcare delivery [4]. While prior ML research has focused on the benefits of
improving mental health conditions such as depression, Alzheimer's, and schizophrenia, the
current research in this area is proving to be increasingly vital in advancing broader mental
health care and showcasing its impact in the field. Therefore, the impact of ML and big data
applications is invaluable in mental health care and management advancements for improving
treatment options and patient support, thus contributing to public health initiatives.

1

Figure 1 presents the global market for mental health apps, expected to grow from $7.03 billion
in 2024 to $15.95 billion by 2029, with a compound annual growth rate (CAGR) of 18.7% [5].
This increase reflects the growing role of ML and digital technologies in mental health,
enabling more personalized care, enhanced support, and improved accessibility through
relevant applications. Together, these areas showcase the potential of ML to enhance mental
health care and address critical challenges in this field, such as identifying risk factors and
developing individualized treatment plans.

Figure 1. Projected Growth of the Global Mental Health Apps Market (2024-2029) [5].

Maternal mental health poses a considerable global challenge. Nearly one in five women will
encounter a mental health condition during pregnancy or within the year following childbirth;
of those with perinatal mental health conditions, 20% will also be grappling with suicidal
thoughts or engaging in acts of self-harm [6]. Postpartum psychiatric disorders typically fall
2

into three categories: baby blues or postpartum blues, postpartum depression (PPD also known
as postnatal depression), and postpartum psychosis. Unfortunately, these conditions are often
disregarded and left untreated. Mothers grappling with mental health challenges may also
experience exacerbated physical health issues, which could potentially impact the well-being
of their infants. The formation of attachment between mother and infant may be affected by
the cognitive and emotional challenges faced by mothers with PPD [7]. Timely intervention is
important to prevent complications.

Numerous studies have investigated the use of ML techniques for predicting PPD and
effectively contributing to mental healthcare practices [8], [9], [10]. As machine and deep
learning methods gain popularity, analyzing existing work to better guide future research
becomes necessary. A comprehensive literature review is provided in Chapter 2.

1.1 Utilizing ML in Healthcare: Addressing PPD
ML has been applied to various areas of mental health, including clinical psychology and
psychiatry, enhancing research methods, treatments, public health studies, and analysis of
support community behaviors. It aids in offering automated solutions to enhance predictive
accuracy and reveal complex patterns in data. The ability of ML to process diverse data types,
including clinical records, has made it a powerful tool in mental health research. Analyzing
complex data, predicting risk factors, providing patient-specific treatment approaches, and
monitoring symptoms enhance the quality of mental health care. Early intervention and
targeted support could be facilitated for at-risk mothers by this predictive model for PPD,
potentially resulting in improved outcomes for both mothers and their infants. Overall, ML
3

demonstrates the potential to improve the efficiency of clinical and research processes and
generate new insights into maternal mental health and well-being.

1.1.1 Postpartum Depression (PPD)
PPD is characterized by symptoms such as profound sadness, fatigue, anxiety, frequent crying,
irritability, and alterations in sleeping or eating habits [11] and is highly prevalent among
women in the first year of delivery. It is influenced by environmental and genetic factors,
making it a complex, multifaceted condition. While its exact cause remains unknown, it is
believed to result from physical, emotional, genetic, and social factors. It is also the most
underdiagnosed obstetric complication globally, including in the United States [12]. Although
the feelings associated with PPD are often temporary and seen as mild conditions by health
professionals, they highlight the crucial need for support during the postpartum period to
manage emotional challenges and maintain well-being. PPD does not go away on its own, and
it is one of the leading causes of the annually reported infanticide incidence [11].

Figure 2 illustrates the spectrum of postpartum mood disorders, highlighting postpartum blues,
PPD, and postpartum psychosis, with their prevalence rates, key symptoms, and severity levels.
Timely diagnosis and proper identification are expected to largely cure this condition. A very
different disorder with psychotic symptoms associated with the postpartum period is
postpartum psychosis. Unknown and untreated diseases lead to prolonged psychosis in severe
cases. Approximately 1-2 cases per 1000 births result in postpartum psychosis, requiring
emergency medical attention. As a severe condition, there is an increased risk of suicide and

4

risk of harm to the baby [13]. Hospitalization, psychotherapy, and medication are the usual
treatments.

Figure 2. Spectrum of PPD [14].

Biological elements like reproductive hormones, stress hormones, and thyroid hormones, in
conjunction with psychological factors, significantly contribute to the onset of PP. Recognizing
key risk factors enables healthcare providers to identify women who may be at high risk of
developing PPD, ensuring they receive medical interventions to lower the persistence during
motherhood [15].

5

Common risk factors for PPD [11] are:
•

A change in hormone levels after childbirth.

•

Previous experience with depression or anxiety.

•

Family or personal history of depression or mental illness.

•

Stresses due to caring for a newborn and managing new life changes.

•

Having a challenging baby who cries more than usual is hard to comfort or whose sleep
and hunger needs are irregular and hard to predict.

•

Having a baby with special needs (premature birth, medical complications, illness).

•

First-time motherhood, very young motherhood, or older motherhood.

•

Other emotional stressors or stress events, such as the death of a loved one or family
problems.

•

Financial or employment problems.

•

Isolation and lack of social support.

Clinical organizations can identify trends and patterns and support personalized interventions
using data mining, predictive analytics, and ML techniques such as Microsoft Power BI and
IBM SPSS Modeller. Business Intelligence (BI) platforms provide tools for analysis and
visualization, transforming data into strategic information to answer specific queries and guide
decisions and planning [16].

1.1.2 Predictive Modeling for PPD Detection
Predictive modeling anticipates future outcomes by utilizing historical and current data and
applying advanced statistical techniques and ML algorithms to generate accurate forecasts and
6

practical recommendations. It identifies patterns and relationships between independent and
predicted variables and provides meaningful insights into data [17]. Prediction models play an
important role in clinical decision-making, helping to improve personalized and targeted
treatments [18].

ML models, such as neural networks (NN), logistic regression (LR), and decision trees, are
widely used for predictive analysis in mental health, especially for screening and diagnosing
depression. It has also been proven as an effective approach to analyzing PPD, focusing on
prediction, optimizing risk assessment, and improving clinical outcomes [19], [20]. These
models use factors such as demographics, medical history, and psychosocial variables to
improve PPD management [9], [21]. A recent review evaluated the risk of bias and
applicability of PPD prediction models commonly used in screening and diagnosis [21]. An
innovative model using deep reinforcement learning to improve the process of identifying PPD
was proposed in [22]. A scoping review [23] on ML methods for predicting PPD highlights
the need for further clinical research to fine-tune ML algorithms for improved prediction and
treatment outcomes.

The work presented in this thesis examines the relationship between independent variables,
including chronic disease features, and the target variable PPD using Support Vector Machines
(SVM), Random Forest (RF), Logistic Regression (LR), and Neural Network (NN) models.
This approach has proven to be very effective for analyzing PPD by focusing on prediction,
optimizing risk assessment, and improving clinical decision-making [8], [20], [24], [25].

7

1.1.3 PPD Screening Tools
Primary screening tools widely used for assessing depressive symptoms postpartum include
the Postpartum Depression Screening Scale (PDSS) and the Edinburgh Postnatal Depression
Scale (EPDS). EPDS [26] is a 10-item self-reporting questionnaire designed to screen for PPD.
Among other variables, it assesses mood, anxiety, and anhedonia experienced by mothers in
the postpartum period. Each item is scored on a point scale, resulting in a total score ranging
from 0 to 30. A cut-off score of ≥13 is commonly used to indicate the need for further
evaluation, while some guidelines recommend ≥10 in primary care settings for broader
assessment. Likewise, the PDSS [27] is a 35-item self-reporting tool specifically developed to
identify PPD and assess its severity across seven dimensions: anxiety/insecurity, emotional
lability, cognitive impairment, loss of self, guilt/shame, sleep/appetite changes, and suicidal
thoughts. A PDSS cut-off score of ≥9 suggests the need for a formal diagnostic assessment.
There are web-based tools that utilize screening scales such as EPDS and PDSS. These tools
allow users to complete the questionnaire digitally, providing immediate scoring and feedback
to support early detection and further evaluation by healthcare providers.

The BabyCenter [28] postpartum depression quiz is an online screening tool based on the
EPDS, featuring 10 multiple-choice questions to assess emotional well-being, mood, and
anxiety over the past week. Healthy Mom [29] offers a similar online platform with a test based
on validated tools like the EPDS, allowing new mothers to complete screenings digitally with
immediate results and guidance. Mental Health America (MHA) [29] provides another online
test using similar tools to assess PPD risk. This includes questions on general health conditions,
such as heart disease, reproductive health concerns (e.g., PCOS, endometriosis), diabetes,
8

asthma, and neurological conditions. These screenings combine standard scale-based items
with contextual questions about health and life factors, providing a score out of 30. These tools
raise awareness about postpartum mental health, provide quick insights, and encourage users
to seek professional help if the results indicate potential concerns. Besides EPDS and PDSS, a
few other innovative tools are explained in Section 2.4.

1.2 Problem Statement
PPD is a complex and prevalent mental health condition, often challenging to diagnose and
treat effectively. Traditional methods in healthcare often fall short of efficiently analyzing
complex data related to mental health [30], leading to delayed detection. Without detection
tools and personalized treatment plans, individuals with depression face prolonged struggles,
affecting their quality of life.

In 2019, almost one-quarter (23%) of mothers who recently gave birth in Canada reported
conditions consistent with either PPD or an anxiety disorder [31]. Although the risk factors for
PPD have been studied, many potentially critical clinical variables remain unexplored [32].
Previous research on mental illness has demonstrated that depression frequently co-occurs with
chronic diseases [33], [34]. Another study highlighted chronic physical illness as a significant
predictor of PPD but examined it broadly without focusing on specific types or conditions [35].
Further evidence suggests that women with chronic conditions such as diabetes, hypertension,
heart disease, migraines, and other neurological disorders are at an increased risk for
peripartum mental illness [36]. Chronic illness may also reduce appetite, leading to amino acid
deficiencies that hinder serotonin production, a critical factor in depression [35]. Overall,
9

studies specifically investigating the relationship between chronic diseases and PPD are
limited.

Following the above insights, the current study presumes that chronic diseases serve as a
potential risk factor for PPD, warranting further exploration. This research hypothesizes that
chronic diseases are significant risk factors for PPD and attempts to find evidence for this
hypothesis using ML models. Further, when chronic disease is identified as a risk factor, it is
important to know the specific disease that increases the likelihood of PPD.

1.3 Research Methodology
A comprehensive literature review established the foundation for this research. The research
progressed by determining relevant data sources, suitable tools, and modeling techniques to
ensure compatibility with the dataset and research objectives.

The ML-based classification was used to build predictive models. This approach helps find
correlations among variables to analyze the PPD risk factors to aid healthcare providers
working with chronic disease patient groups. The source data was integrated into a database
using SQL Server Management Studio (SSMS) 16.0, and relevant variables were selected
using SPSS Modeler 18.4 in consultation with medical personnel. The preprocessed data was
used to develop the models and generate results in Python 3.11.11 [37]. The models underwent
training, testing, and validation stages to ensure their accuracy and reliability in predicting
outcomes related to the research objectives.

10

Furthermore, an interactive dashboard was developed using Power BI [38] to visually present
valuable insights in a user-friendly dashboard. Flask 2.3.3 [39] integrated a backend framework
into the dashboard to offer personalized risk assessment based on user input. Chapter 3
provides a detailed description of the research methodology.

1.4

Data Source

The primary data source for this research is the Centre for Disease Control and Prevention –
Pregnancy Risk Assessment Monitoring System (CDC-PRAMS) [40], a population-based
surveillance system that gathers data on maternal experiences surrounding pregnancy to
improve maternal and infant health. Sampling from birth certificate registries, PRAMS surveys
new mothers monthly, providing valuable data for research purposes. This research uses data
from 202,745 mothers surveyed across all states from 2016 to 2020. The categories of predictor
variables used by this study are summarized in Table 1.

11

Table 1. Predictor Variables

Indicator
Demographic Variables

Disease Variables

Pregnancy and Childbirth Factors
Psychosocial Variables
Lifestyle and Behavioral Factors

Variable
Age, gender, race/ethnicity, marital status, education level,
and income.
Medical history and family history of mental health
conditions.
Gestational age, delivery mode, prenatal care utilization,
and complications during pregnancy and childbirth.
Stress levels during pregnancy and postpartum, as well as
social support networks and relationship status.
Substance use (e.g., smoking) and physical activity during
pregnancy.

Postpartum Factors

Breastfeeding, sleep patterns, and infant care.

Geographic Factors

Geographic factors, such as regional variations.

Income and Socioeconomic
Status
Infant Health and Development

Income levels and socioeconomic status.
Infant health and development, such as birth weight,
immunizations, and infant feeding practices.

Parenting and Infant Care

Parenting practices, infant sleep routines, and childcare

Practices

arrangements.
Healthcare utilization during pregnancy and the

Healthcare Utilization

postpartum period, including the number of prenatal care
visits, healthcare-provider interactions, and access to
healthcare services

Social Determinants of Health,
Communities, and Neighborhood
Characteristics

Social determinants of health, such as housing stability,
food security, and access to transportation.

12

1.5 Contributions
This research focuses on enhancing PPD care for mothers by developing predictive models,
creating a diagnostic screening tool, and designing a multi-level dashboard. First, developing
ML models to forecast the likelihood of depression incorporating multiple risk factors,
including chronic diseases, provides the following contributions:

•

Detection of key risk factors of PPD to help identify high-risk individuals, enabling
early interventions, tailored treatment strategies, and improved preventive care.

•

Identification of specific chronic diseases and their interaction with other risk factors
that impact the onset of PPD in vulnerable populations.

Second, the developed dashboard visually highlights relationships and patterns in the data,
which offers the following contributions:

•

Insights into PPD statistics, including prevalence, correlations, and trends.

•

Helps identify gaps in PPD care, enabling experts to create targeted care plans and
improve support for vulnerable postpartum populations.

Third, developing a diagnostic screening tool integrating predictive models into an ensemble
framework to estimate depression probability based on user input demonstrates the potential
of ML for effective PPD management, timely identification, and customized preventive care.

13

1.6 Organization of Thesis
This thesis focuses on PPD risk factor identification and prediction, comprising five chapters
focusing on prediction through various ML methodologies. The first chapter introduces PPD
and its connection with chronic diseases and explains why this research is important. Chapter
two reviews existing studies on PPD. Chapter three explains the design of the prediction
models. Chapter four describes the experimental setup and presents the findings and results of
the models. Finally, Chapter five provides insights and recommendations for future research.

14

Chapter 2
Literature Review
The consequences of untreated or late-detected PPD can be severe, potentially leading to severe
issues and effects on both the mother and the child or may escalate to postpartum psychosis.
Thus, prior studies in the field of PPD were primarily aimed at identifying risk factors and
developing predictive models to facilitate timely intervention and effective management of
PPD. However, the research on the impact of chronic diseases on PPD is limited. This chapter
provides an overview of the current knowledge, gaps, and advancements in predicting PPD,
highlighting the significance of identifying at-risk individuals.

2.1 PPD Predictive Analysis
Statistical algorithms and ML techniques have been used in predictive analytics to reveal
complex patterns and predict future trends based on historical information. This approach is
especially important in the context of PPD where developing predictive models can assist
healthcare professionals to identify individuals who are at heightened risk of developing this
condition and enable proactive measures to be taken. A brief literature review on PPD
predictive analysis is provided in this section.

A two-year longitudinal study [41] was conducted with 301 pregnant participants to validate
the Postpartum Depression Predictors Inventory-Revised (PDPI-R) tool in Turkish women and
assess its predictive validity for PPD using two standards: the EPDS and the Structured Clinical
15

Interview. Eligible participants for the study were over 18, in gestational weeks 28-32, carrying
a single fetus, and without psychiatric disorders or medical complications for the baby.
Statistical analyses conducted in SPSS confirmed the predictive validity of both versions with
specific cut-off points established for accurate prediction of PPD using EPDS. The study
identified various predictors of PPD, which included unwanted pregnancy, previous history of
depression, life stress, self-esteem, social support, and marital satisfaction. The tool examined
both prenatal and postnatal versions, with accuracies ranging from 64% to 81%. While the
findings support routine screening with the tool to assess women's susceptibility to PPD,
limitations in generalizability and interpretation should be addressed through additional
research.

A study aimed to utilize ML to develop a predictive model for identifying women at a higher
risk of PPD [42]. Data from 686 mothers in Sri Lanka collected within six months of childbirth
was used, and the risk levels for PPD were assessed using the Sri Lankan version of EPDS.
The questionnaire included multiple factors such as physical disabilities, drug abuse history,
life stressors, social support, socioeconomic status, family problems, marital status, babyrelated issues, education level, and number of children. The EPDS scores were categorized into
mild, moderate, severe, and profound PPD. Exploratory data analysis and Spearman
correlation were performed to identify relevant features for predicting PPD. Four models were
built using an Adaptive Neuro Fuzzy Inference System with Genetic Algorithm (ANFIS-GA),
ANFIS, and Feed Forward Artificial Neural Network (FFANN), SVM, and RF. Among these
models, FFANN with 25 epochs exhibited superior performance compared to other models,
achieving high accuracy in testing, training, and validating datasets. Researchers identified
ANFIS-GA performance in running the Genetic Algorithm (GA) for five hundred generations,
16

resulting in the lowest loss and optimal parameters. Despite the promising findings, the study
did not provide a detailed discussion on the selected features for the predictive models and
present any identified risk factors for PPD, which could have provided valuable insights into
the underlying causes of PPD.

A study in 2021 employed data mining algorithms, specifically J48, Random Tree, RF, and
Reduce Error Pruning (REP) Tree, in combination with ensemble techniques such as Adaptive
Boosting and Bagging considering social and demographic factors and identifying common
symptoms, utilizing non-clinical information and observations [43]. The analysis aim was to
identify the conditions under which women are more likely to experience PPD by surveying
ninety-six participants and implementing the algorithms to analyze and predict the risk level
of PPD. The study examined women who had given birth within the past year, irrespective of
their depression history, and considered both urban and rural areas to assess the influence of
quality of life on PPD probability. The collected information focused on social, mental, and
quality of life aspects as well as 31 attributes of EPDS score to assess different levels of risk.
The models achieved accuracy in the range of 84 - 94%. The findings highlighted associations
between unplanned pregnancy, spousal support, delivery method (Caesarean vs. vaginal), and
urban residence with PPD risk. The study revealed a concerning trend where most women
facing PPD go unnoticed or undiagnosed. Also, women living in rural areas have had a lower
prevalence of PPD, indicating a positive impact on lifestyle and living conditions. The
limitations of this work include a small participant pool and a limited sampling area. This was
also acknowledged by the authors.

17

A regional study conducted in five Arab countries (Jordan, Palestine, Lebanon, Saudi Arabia,
and Bahrain) during the COVID-19 pandemic aimed to develop ML models for predicting
maternal depression and anxiety among pregnant and postpartum women [25]. The data was
collected through a web-based questionnaire. The dataset included 3569 pregnant women over
18 who had normal pregnancies during the COVID-19 pandemic. Exclusions applied to those
conceiving during the pandemic and those with risk factors like miscarriage and chronic health
issues. ANOVA test was implemented to assess risk factor distributions and the ML models,
including Gradient Boosting (GB), Random Forest (RF), Extreme Randomized Forest (XRT),
Naïve Bayes (NB), Support Vector Machine (SVM), Multilayer Neural Network (MNN), and
Decision Tree (DT) were evaluated for their ability to predict mental health outcomes. Higher
depression rates were associated with factors such as family problems, inadequate sleep,
psychological issues, financial challenges, and a COVID-19 diagnosis. Women, with an
average age of 28.5 years, experienced varying levels of depression and anxiety influenced by
various factors. The Gradient Boosting and RF models accurately predicted depression and
anxiety symptoms. In general, all models performed similarly in terms of accuracy.

PPD risk prediction models were developed using electronic health record (EHR) datasets from
single and multiple sites [44]. Data from Weill Cornell Medicine and New York-Presbyterian
Hospital in New York City, USA, focusing on individuals aged 18 to 45, was utilized to
develop the models. An end-to-end framework was developed to extract features and process
the data, which included demographics, clinical diagnoses, medication prescriptions,
laboratory results, and unstructured clinical notes. A PPD diagnosis was defined as using an
antidepressant within one year after childbirth, based on the Systematized Nomenclature of
Medicine. ML models, including RF, Decision Trees, Extreme Gradient boosting, Regularized
18

Logistic Regression, and Multilayer Perceptron, were used to train and optimize the model
using Sequential forward selection separately for women with or without a mental disorder
history. In the PPD group, higher emergency room visit rates, Caesarean delivery, medication
prescriptions, past or current mental health conditions, being a single parent, and pregnancy
complications (e.g., palpitation, diarrhea, vomiting, abdominal pain) were associated with
increased risk of PPD. While this study incorporated clinical diagnoses, medication
prescriptions, and laboratory results, it did not consider pre-pregnancy health conditions.

The need for a more reliable comprehensive tool for identifying mothers at risk of PPD was
recognized and implemented in a study by Elday Hochman et al. [19]. Longitudinal EHR data
over a period of eight years from Clalit Health Services (CHS) was utilized to develop a PPD
risk prediction model. The data excluded non-singleton births, stillbirths, and births ending
with infant death. After primary and secondary exclusions, the final cohort consisted of
214,359 individuals, from which the training and validation sets were derived. Out of the initial
156 (sociodemographic, clinical, and obstetric) potential predictor features in the model
development, 65 features were identified as having a significant impact on model performance
due to their nonzero SHAP values. The model was built using XGBoost using the SHAP mean
absolute values to measure how each feature contributed to predicting PPD. A calibration plot
showed consistency between predicted and observed risks across different risk scores. The
model's accuracy was measured using AUROC, and 95% confidence intervals were calculated
using bootstrapping. The model consistently performed well for age and first-time births but
less accurately for women without a psychiatric history. Additionally, the birth cohort was
categorized into "low-risk" and "high-risk" based on PPD risk scores, with the "high-risk"
group characterized by factors such as prior psychiatric disorders, increased third-trimester
19

healthcare visits, higher BMI, smoking, lower birthweight, and gestational age. This group also
experienced more pregnancy complications and showed distinct blood test patterns, more tests
during pregnancy, and earlier gestational tests compared to the "low risk" group. The study
suggests using symptoms and predictive modeling together makes it more accurate at spotting
those at risk of or experiencing PPD.

A PPD detection and prevention study in expectant mothers suggested perinatal (or peripartum)
interventions [45]. This pilot study used six ML algorithms featuring longitudinal clinical
information and patients' socio-demographic characteristics. The statistically significant
variables associated with PPD were selected (71 variables were diagnoses, and 22 were
medications). This study evaluated each model's performance using the area under the receiveroperator curve in 10-fold cross-validation by differentiating variables into demographic,
diagnoses, and medication variables. The SVM model showed the highest AUC whereas
Decision Tree with least AUC. Race, threatened abortion, prior or prenatal depression disorder,
anxiety, backache in the 3rd trimester, and muscle pain in the 2nd trimester were significant
associated factors. Also, antidepressant use across three trimesters was found to be a strong
predictor of PPD, consistent with previous studies. Though medication variables were included
in the research, none of the chronic illnesses were included.

A study was done to diagnose PPD in women and develop appropriate treatment plans using
demographic, depression, and pregnancy survey data [24]. The ML-based algorithm, including
decision trees, Naive Bayes, and SVM, and a functional-gradient boosting algorithm were
implemented to predict PPD. A final classification tree was made, including factors like being
a first-time mother, finances, age, education, other illnesses, sleeping trouble, and
20

relationships. As this study focused on survey data, the clinical variables were not used in
predicting PPD. This research examined several stressors before pregnancy and found that
women with depression before getting pregnant were four times more likely to have PPD.
Furthermore, the highest prevalence of PPD was observed in women who were younger,
unmarried, had a lower income, had low education, or had a baby born less than 37 weeks.

2.1.1 PPD Risk Factors Analysis
Risk factors refer to identifiable characteristics or conditions that research has shown to elevate
the chances of experiencing PPD. It is crucial to identify the factors associated with PPD. It
has been identified that women within the age range of 18-24 and those with pre-existing
anxiety health conditions are particularly susceptible to receiving PPD diagnoses (Figure 3 and
Figure 4) [46]. A literature review of studies focusing on identifying risk factors for predicting
PPD is provided in this section.

Figure 3. PPD Rate Trend in 2014 -2018 [46].

21

Figure 4. PPD and Pre-Existing Health Conditions [46].

A randomized controlled trial (RCT) examining the effectiveness of tranexamic acid in
preventing postpartum hemorrhage (PPH) after vaginal delivery was conducted as part of the
Tranexamic Acid for Preventing Postpartum Hemorrhage After Vaginal Delivery (TRAAP)
[47]. The study was carried out in 15 maternity units in France and included women who were
18 years or older and planning to give birth vaginally to a single live fetus at 35 or more weeks
of gestation. Three types of exposure variables were investigated: pre-pregnancy
characteristics (including psychiatric history), pregnancy and delivery factors, and postpartum
variables (like neonatal complications and childbirth memories). Factors such as age,
psychiatric history, and experiences during labor and delivery were found to be linked to PPD.
PPD was found to be associated with negative childbirth memories, as well as induced labor
and operative vaginal delivery, with the latter two being significant risk factors. Pre-pregnancy
characteristics and obstetric events, emphasizing the need for targeted screening and
intervention strategies, were identified as risk factors. Although the findings provide valuable

22

insights into postpartum care, potential limitations are underestimation and reliance on selfreported screening.

The effects of different factors on PPD and post-traumatic stress disorder (PTSD) among
women in Slovakia who experienced traumatic childbirth were determined in [48]. A multistep approach was employed to investigate the impact of sociodemographic factors, maternal
and child health complications, and childbirth experiences on postpartum well-being. The
study investigated the postpartum experiences of 437 females, including childbirth satisfaction,
maternal and child health complications, and overall preparedness. Standardized
questionnaires assessed PTSD symptoms, PPD risk, birth satisfaction, and perceptions of
trauma and respect during childbirth. This study found an association between primiparity,
maternal and child health issues, previous trauma, and PPD/PTSD, with lower birth satisfaction
levels correlating with higher depression and PTSD symptoms, emphasizing the significance
of subjective birth trauma perception and perceived birth respect as contributors. The study
acknowledged limitations in sample representation and the complexity of understanding the
factors involved while highlighting the high reliability of data collection and analysis which
enhanced the validity of the research findings.

In 2021, an online survey involving [49] 3,523 postpartum women in Brazil, South Korea,
Taiwan, Thailand, and the United Kingdom was conducted to investigate factors related to
PPD symptoms during the COVID-19 pandemic. The criteria for this study included
individuals aged 18 to 49 who are within the first six months postpartum. Descriptive
statistics were performed, including calculating frequencies and percentages for categorical
variables and means and standard deviations for continuous variables. Chi-square tests and t23

tests were used to detect associations with PPD symptoms, where PPD was defined as an
EPDS score exceeding 13. A binary logistic regression model was employed to pinpoint the
factors explaining the association with PPD. This included calculating crude and adjusted
odds ratios (OR and AOR) and their corresponding 95% confidence intervals. The
independent variables that exhibited a p-value below 0.05 in the simple linear regression
analysis were utilized in the multiple linear regression model. Results indicate that both
Taiwan and Thailand exhibited a reduced risk of experiencing PPD symptoms compared to
individuals in Brazil. Furthermore, 29.3% of women exhibited PPD symptoms linked to
younger age. High-risk factors for PPD include younger women experiencing health
problems during pregnancy, delivery, or postpartum, no improvement or worsening of food
security during COVID-19, feeding babies with expressed human milk or complementary
food, receiving low or medium support, and having low social support. However, this
research identified health problems as a contributing factor to PPD but did not delve into
specific details regarding the types or nature of these health problems.

Several factors were examined, including socio-demographics (age, education, occupation),
substance use, reproductive characteristics (pregnancy index, inter-delivery interval, live
children), and maternal social factors (violence, support, gender satisfaction, decision-making)
in a study conducted in Ethiopia [50]. This study included 479 postpartum mothers, all within
the first year after childbirth and aged at least 18 years old. Both bivariate and multivariable
analyses were conducted using a binary logistic regression model. Variables with p < 0.25 in
the initial analysis were included in a multivariable model to control for confounding, and those
with p < 0.05 were considered statistically significant. In the results, a prevalence of 23% was
observed for late PPD, alongside factors such as husband's Khat use, partner dissatisfaction
24

with the baby's gender, difficulty in meeting the husband's sexual needs, postpartum intimate
partner violence, and low social support. It is noted that being cross-sectional does not
necessarily establish cause-and-effect relationships.

A cross-sectional study was conducted involving 1028 women aiming to investigate the link
between dietary patterns and elevated PPD symptoms among participants in the initial phase
of the Maternal and Child Health cohort study [51]. The study employed the Food Frequency
Questionnaire (FFQ) to collect dietary data. Four distinct dietary patterns—prudent, sweet and
dessert, junk food, and western—were identified using exploratory factor analysis (EFA) and
EPDS. Data analysis involved various statistical tests, including the chi-square test, Fisher's
exact test, independent sample t-test, and multiple logistic regression (MLR). The study
utilized a multiple logistic regression model to examine how high PPD symptoms relate to
dietary patterns. The study found that 24% of participants experienced high PPD symptoms.
Compared with the Prudent dietary pattern, high adherence to the Western dietary pattern was
linked to an increased risk of PPD symptoms. However, no significant association was
observed between sweet, dessert, or junk food patterns and high PPD symptoms. When
comparing the two groups based on income, smoking, history of chronic disease, gestational
diabetes, dietary supplement uses during pregnancy, types of childbirth, and the gender of the
baby, no significant differences were found. A limitation of this study is that it excluded women
with prior mental illness or depression, epilepsy, nephropathy, or cancer, and abnormal
pregnancies.

In 2019, Smorti et al. [52] analyzed PPD in 161 Italian nulliparous low-risk women to
investigate multiple risk factors and relate them to labor and birth experience. Socio25

demographic and clinical measures included age, educational level, work status, marital status,
information about the number of years of their relationship, and information about a planned
or unplanned pregnancy. Psychological measures included levels of depression and anxiety
during pregnancy. Relational measures included the relationship with the mother, father, and
partner, as well as prenatal attachment to the child. In addition, labor measures and mode of
delivery were also included. Age, induced labor, prenatal anxiety, and depression were found
to be significant risk factors for PPD. On the other hand, the length of the relationship with the
partner, level of education, employment status, and planned pregnancy were not significant.
The study's limitation lies in its inclusion criteria, as it only encompassed nulliparous women
aged between 18 and 42 with no risk during pregnancy, excluding those with twins, previous
miscarriage, or abortion experiences.

In 2017, G. Palumbo et al. [53] established screening procedures in Italy to identify women at
risk of PPD as part of the "Prevention and Early Intervention for Risk of PPD" project. A total
of 1,558 women were screened, with an average age of 32.5 years. The Italian version of the
EPDS was used to assess symptoms of depression. Information collected included delivery,
maternal health problems, pre-delivery and pregnancy details, stressful events in the past year,
mental health conditions before pregnancy, perceived family, and sociodemographic variables.
Using SPSS, the sociodemographic and psychosocial characteristics of women with and
without current PPD symptoms were summarized. LR was used to identify the PPD. A
significant proportion of women (7.1%) in the study tested positive for PPD, and those with
positive scores showed higher rates of self-harm thoughts, recent stressful events, depressive
and anxiety symptoms, psychotropic drug use, lower social support, and relationship
dissatisfaction. The analysis revealed that loss of interest during pregnancy and lack of
26

psychological support from family and friends were the most significant predictors of PPD
symptoms. Other significant predictors included anxiety or depressive symptoms during
pregnancy, job loss or change, personal health problems after childbirth, sleep disorders,
difficult infant temperament, and history of a mental health condition. The study acknowledges
certain limitations, including using a non-random sample, and it may introduce bias and limit
the generalizability of the findings, as this group might have had experiences different from
the broader population.

A web-based survey using a cross-sectional design analyzed PPD risk factors in U.S. women
who gave birth to a live infant after COVID-19 was declared a public health emergency [54].
Out of the 670 postpartum patients in the study, one in three screened positive for PPD, and
one in five showed major depressive symptoms. Formula-feeding was associated with a 92%
higher likelihood of PPD and a 73% higher likelihood of major depressive symptoms compared
to breastfeeding. In addition, PPD risk increased by 74% with infant NICU admission, 4% per
postpartum week, and 71% due to COVID-19-related worry. Due to reduced healthcare access,
changes in peripartum care, and limited early telehealth options, infection concerns for
themselves or their infants during the COVID-19 pandemic also increased the likelihood of
screening positive for PPD. This study has limitations due to its reliance on a social mediabased convenience sample, which may not fully reflect the diversity of the postpartum
population, and it did not consider all potential PPD factors or regional infection rate
differences.

A study assessed the probability and the prevalence of diagnosed PPD occurrence among firsttime mothers (primiparous women) and examined how maternal factors associated with PPD
27

differ between adolescent mothers (ages 12–19) and adult mothers (ages 20 and older) [55].
An observational design was utilized to analyze data from 61,226 first-time mothers who had
at least one non-emergency visit within ten years prior to their first live birth. Data were
obtained from three sources: (1) electronic health records (EHR) from Intermountain
Healthcare, (2) electronic health records (EHR) from the University of Utah Health, and (3)
the Utah Population Database (UPDB), which encompassed a wide range of socioeconomic,
demographic, psychological, and clinical factors. Using multivariable logistic regression, the
study found that a history of depression or anxiety is a significant factor. A higher rate of PPD
was diagnosed among adolescents (6.1%) compared to adults (3.8%), and maternal
demographic factors associated with PPD were more prevalent among adolescents.
Adolescents also had a more considerable proportion of potential PPD risk factors, such as
non-White race, Hispanic ethnicity, and various health and lifestyle factors. Moreover,
adolescent mothers exhibited a greater tendency to postpone prenatal care, with a higher
occurrence of inadequate prenatal care and a prevalence of tobacco use and infections in
comparison to adult mothers. It was found that the youngest and oldest individuals among the
participants had a higher risk. The limitation of the study included a potential misclassification
of PPD and associated maternal factors due to reliance on diagnostic codes, which may
underestimate mild PPD cases and the unavailability of clinical records from outside the study's
health systems.

A cohort study of all women with singleton births over 11 years in Sweden found that the risk
of PPD increased with a history of depression, advanced maternal age, and gestational diabetes
[56]. Poisson regression models estimated the relative risk of PPD using incidence rate ratios.
Women with a history of depression have a 20-fold higher risk of PPD compared to those with
28

no prior depression. For women with a history of depression, higher PPD risk was associated
with advanced maternal age (>30 years) and pregestational diabetes. In contrast, among women
without a history of depression, PPD risk increased with young age (<24 years), older age (>35
years), and Caesarean or instrument-assisted delivery. Also, moderate preterm delivery (32–
36 weeks) increased PPD risk regardless of depression history. Gestational diabetes raised PPD
risk for all women, while pregestational diabetes posed an additional risk only for women with
a history of depression. A shorter gestational age increases PPD risk, particularly for preterm
births. Although 11 years of data were analyzed, the predictors were limited to obstetric,
gestational, psychiatric, and demographic factors, excluding other disease factors.

2.1.2 Summary
Research analyzing the risk factors for PPD has found that a person's history of depressive
disorders, such as depression or anxiety, is a significant predictor [4], [53], [55]. Previous
studies have primarily looked at factors related to maternal and paternal demographics when
considering potential predictors. Other key factors associated with PPD include health
problems during pregnancy [49], lack of social support, experiences of violence after giving
birth, and dissatisfaction in the partner relationship [50]. Some studies have considered
gestational diabetes as a separate factor, but there is no statistical evidence to confirm a
significant link [51]. Age is a unique risk factor, with younger age being associated with a
higher risk [56] and its impact on predicting PPD can vary depending on different combinations
of factors.

29

In PPD research, statistical methods and ML techniques have been employed to forecast future
outcomes. By examining historical data, the research seeks to reveal concealed patterns and
comprehend the factors influencing the onset and severity of PPD. This approach enables early
detection and precise intervention, enhancing the welfare of postpartum individuals and their
families. With the utilization of predictive analytics, researchers are refining their capacity to
identify PPD risk factors and tailor treatments to individual needs. It was found that psychiatric
history, stress, gestation, and demographic-related variables are reliable predictors of PPD
[19], [24].

2.2 Postpartum Depression Risk Factors: Chronic Illness
People with chronic illnesses often have a higher prevalence of depression and anxiety [57].
This section presents a concise overview of the literature on the relationship between chronic
diseases and PPD.

R. D. Björvang et al. [58] aimed to address existing gaps in the literature concerning the
relationship between Diabetes Mellitus in Pregnancy (DMP) and perinatal depression,
specifically examining antepartum and PPD separately. They conducted a longitudinal
investigation within a large prospective cohort study. By including multiple assessment
methods for depression and adjusting for relevant confounders, the study provided a
comprehensive understanding of the association between DMP and perinatal depression. The
findings revealed a significant positive association between DMP and PPD. However, the study
did not find a statistically significant association between DMP and antepartum depression
once confounders were considered. Sub-analyses examined specific types of Diabetes Mellitus
30

(pre-gestational, gestational, and unspecified) and found varying associations with perinatal
depressive symptoms. The study highlights the complexities of metabolic factors, hormonal
changes, and psychological well-being during this period. The acknowledged limitations were
low prevalence of DMP which may impact generalizability, and possible biases due to sample
characteristics such as higher education levels. Also, future research is suggested to investigate
how different types of DMP are related to depression over time and explore underlying
biological pathways.

A systematic review and meta-analysis [59] of longitudinal studies investigated the association
between thyroid autoimmunity during pregnancy and the weeks after childbirth and the risk of
developing PPD. Women with and without anti-thioperoxides antibody-Thyroid Peroxidase
Antibodies (TPOAb) positivity were specifically focused. The analysis involved five studies,
encompassing data from 449 TPOAb-positive women and 2483 TPOAb-negative women. It
revealed a significantly higher risk of developing PPD in the TPOAb-positive group, although
there was some heterogeneity. The study indicates that thyroid autoimmunity during pregnancy
and postpartum is linked to a higher risk of PPD. However, more rigorously designed studies
are needed to verify this association and to better understand the underlying pathophysiological
mechanisms, which was a limitation of this research. Additionally, only one specific disease
was explored in this study.

A multinational study was conducted during the COVID-19 pandemic [60] aimed to assess the
mental health of pregnant and breastfeeding women (9041) in Ireland, Norway, Switzerland,
the Netherlands, and the UK. Descriptive statistics were used to analyze women's
characteristics and mental health scores. Associations between depressive symptoms, anxiety,
31

and stress were assessed using logistic and linear regression models, with significant covariates
retained in the final adjusted model, including country, maternal age, professional status,
smoking, chronic somatic and mental illness, planned pregnancy, and breastfeeding status.
Among the pregnant women, 15% exhibited major depressive symptoms, while 11% had
moderate to severe generalized anxiety symptoms. In the breastfeeding cohort, 13% had major
depressive symptoms, and 10% experienced significant generalized anxiety. Pregnant and
breastfeeding women who reported chronic mental illness, smoked, or lived in the UK or
Ireland were likelier to experience major depressive symptoms. These symptoms were also
linked to not breastfeeding, having an unplanned pregnancy, and having a chronic physical
illness after childbirth. This study has a limitation related to potential selection bias, as the
survey was distributed via social media and did not assess whether depressive symptoms were
linked to postpartum status.

The prevalence of PPD was high in an institutional-based cross-sectional study conducted in
Ethiopia [61]. All mothers who had delivered within the last 12 months and followed up for
PPD were selected for this study. The PPD outcome variable was measured by using the EPDS.
This study uses the SPSS statistical package, and bivariate analysis was done to find the
association between the independent and dependent variables. Among the 408 participants,
7.4% had chronic medical illnesses. The result demonstrated that about 33.82% had PPD, and
those who had other chronic physical illnesses were 7.7 times more likely to develop PPD. In
addition, unstable marital status, infant death incident, unplanned pregnancy, and age greater
than 30 had increased probability for PPD. Even though this research explored chronic physical
illnesses, the prediction of specific chronic illnesses was not determined.

32

A review examined the prevalence of PPD in three populations: women who use substances,
women with current or experience of abuse, and women with chronic illness [62]. No matter
the direct or indirect effect, this review found that PPD increased among women with substance
use and those who experienced abuse, and these women are more likely to seek treatment.
Additionally, this review found an increased risk of PPD in only one of the four reviewed
studies involving women with chronic illness. The combined effect of multiple chronic health
problems over a single or specific chronic illness increases the risk of developing PPD. The
main limitation acknowledged in this review was the failure to confound depression in
pregnancy in most of the studies.

A study in Lebanon analyzed the determinants and prevalence of PPD [63] in 369 women from
two locations, Beirut and Bekaa Valley. The chosen women were screened using EPDS, and
scores above a threshold of 12/13 were considered as PPD. Bivariate analyses were performed
using chi-square tests and identified covariates were adjusted to predict the PPD using multiple
LR analyses. In both areas, the study determined the significant factors of PPD as the lack of
support and prenatal depression. Also, women with prenatal depression and multiple chronic
health problems were more likely to develop PPD. The prevalence of PPD was significantly
higher in Bekaa Valley (21%) than in others. At the same time, the pattern of having more than
one chronic illness was seen as a significant predictor for PPD in both regions. However, this
study did not reveal or classify the chronic diseases that act as risk factors for PPD. In addition,
the limitation mentioned in this study was that rather than psychiatric evaluation or other
objective measures the potential risk factors were based on self-reports and retrospective
information.

33

2.2.1 Summary
Considering chronic diseases as the risk factor for developing PPD, only limited studies have
been conducted. Among those, researchers found out that women with chronic illnesses had a
higher chance of having PPD, and those having more than one chronic physical illness had an
increased risk for PPD. In addition, the data used in those studies was on a minimal population
or a concise period. Moreover, none of the studies had determined the specific chronic illnesses
impacting PPD.

2.3 CDC-PRAMS Data PPD Analysis
Researchers have previously utilized the PRAMS dataset for studies exploring various aspects
of PPD, including its prevalence, significant risk factors, and the impact of health, maternal
behaviors, and experiences on postpartum mental health. Below is a brief overview of selected
research articles that have utilized PRAMS data to analyze PPD, shedding light on significant
findings and their implications for maternal and infant health.

A. Paul et al. [8] used 28,755 records of PRAMS data and utilized different ML algorithms,
including RF and SVM, to predict PPD using metrics such as AUC values, accuracy, precision,
recall, F1 score, specificity, and sensitivity. Health factors such as pre-pregnancy BMI,
nutrition status, and health problems while ignoring demographic and socioeconomic factors
were considered. The analysis of different models revealed that SVM performed the best
overall, achieving an accuracy of 74% and highest precision and F1 score. TabNet showed the
highest AUC value, while the Gradient Boosting Machine (GBM) had the highest specificity,
34

and KNN had the highest sensitivity. While integrating ML models into remote monitoring
systems for early detection of PPD was suggested, the dataset's state-specific nature and limited
diversity may pose challenges to maintaining accuracy across varied populations. Also, no risk
factors were analyzed since this research focuses on the ML model evaluation.

PPD risk factors and prevalence were assessed using general-purpose statistical software
(STATA) on PRAMS data [64]. This study highlighted the need for evaluating the
epidemiology of PPD detection, prevention, and control strategies. A Chi-square test was
conducted to assess the difference between the PPD and non-PPD groups, and LR models were
created to analyze the risk factors. The evaluation showed that 12% of the 4,022 mothers had
PPD, with prevalence fluctuating over the years. PPD was also common among unmarried
mothers, those with lower education, low income, preterm births, low birth weight babies, and
those receiving WIC during pregnancy. Additionally, mothers with prior depression or
hypertension had the highest prevalence, while no significant differences were observed based
on delivery type. Sociodemographic factors such as younger age, being Black, unmarried, or
having a lower income were associated with higher odds of PPD. Depression before pregnancy,
abuse before or during pregnancy, job loss, extended time away from husband/partner due to
work-related travel, husband/partner not wanting the pregnancy, or arguing more than usual
with a husband/partner were the key risk factors determined. Clinical factors like hypertension,
unintended pregnancies, preterm births, and low birth weight babies also increased the risk.
Additional stressors, such as divorce, homelessness, financial difficulties, substance abuse in
someone close, and the death of a loved one, further elevated the risk of PPD. The sample data
for this study was limited in size and was chosen only from a single state.

35

In 2014, Sundaram S et al. [65] investigated the relationship between maternal health issues
and PPD using PRAMS data. A Hierarchical Linear Modeling (HLM) approach was used to
explore various maternal morbidities and their potential association with PPD. A total of 13
medical and obstetric issues were examined that can occur during pregnancy within two
distinct samples: women from all states and women specifically from Alaska and Maine.
Approximately 13.8% of participants showed symptoms of PPD, while 7.6% had a formal
diagnosis of PPD. Five statistically strong significant maternal morbidities identified with both
PPD symptoms and PPD diagnosis are vaginal bleeding, kidney/bladder infection, nausea,
preterm labor, and bed rest. Hypertension and blood transfusion emerged as strong predictors
for the diagnosis of PPD. The main findings reveal that both PPD symptoms and diagnosis are
associated with preterm labor and nausea in Alaska and Maine. Additionally, vaginal bleeding,
kidney/bladder infection, blood transfusion, and bed rest were associated with PPD symptoms
only, whereas pre-existing diabetes was associated with PPD diagnosis only. The study
acknowledges several limitations, including the potential for recall bias and underreporting of
PPD symptoms due to the self-reported nature of PRAMS data. Discrepancies between selfreported symptoms and PPD diagnosed by a healthcare provider were also noted. Limited
generalizability was highlighted due to the state-specific nature of the data. Furthermore, the
study points out the potential for bias due to uncontrolled variables and the exclusion of
observations with missing data.

The prevalence of PPD using PRAMS data was analyzed considering the effects of different
stressful life events [66]. The overall prevalence of self-reported PPD was 14.8%, which was
higher in teen mothers, unmarried mothers, and low-income mothers, lower in mothers with
graduate degrees, and among those with stressful life events. There was no significant
36

difference in prevalence by race or parity. This study conducted Chi-square tests, t-tests, and
logistic regression analysis by categorizing 13 stressors into four groups: financial, relational,
trauma-related, and emotional. The results found the highest probability of PPD in mothers
who experienced severe relationship problems with low financial and high trauma-related
stresses. Interaction effects showed the highest PPD likelihood with high relational and trauma
stresses but low financial stress. Married teens were also more likely to experience PPD than
married mothers aged 30 or older. The limitations highlighted include the inability to establish
causal inference due to the cross-sectional design and potential biases from the self-reported
nature of the data. The generalizability of the findings was restricted to states with similar
demographics, such as higher proportions of low-income, less educated, and unmarried
mothers. Additionally, key factors like pregnancy complications, birth outcomes, stress
perception, past depression history, biological factors, and paternal support were not examined.

Shin et al. [20] developed predictive models utilizing the PRAMS 2012-2013 dataset. Nine
ML algorithms, including k-nearest neighbor (kNN), recursive partitioning (RPART; a
decision tree-based method), support vector machine (SVM), stochastic GBM, Random Forest
(RF), Neural Network (NN), Naïve Bayes (NB), Logistic Regression (LR), and AdaBoost were
used in this study. The study revealed that among those tested, RF emerged as the most
effective in predicting PPD, followed by SVM. Random Down-Sampling and The Synthetic
Minority Over-Sampling Technique (SMOTE) were used to address the imbalance, and a 10fold cross-validation strategy was used to evaluate the classification models and the area AUC
as the performance metrics. The study results showed that women of younger age, lower
education, small-for-gestational-age infants, smoking reduction during pregnancy, and
unmarried status are more likely to have PPD. Other important key predictors were stress
37

during pregnancy, prior depression, weeks of breastfeeding, income, dental hygiene, and infant
gender. Self-reporting bias and the likelihood of limiting generalizability due to missing data
were recognized as limitations.

2.3.1 Summary
The research on CDC-PRAMS data primarily focused on maternal, relationship, and
psychiatric variables to predict PPD and its risk factors. However, current studies in this field
often rely on data from specific locations (states) or limited time periods. Additionally, no
comprehensive study has considered chronic illnesses as potential predictors of PPD. SVM
was identified as a high-performing model [8], [20]. Key factors consistently identified across
studies include a history of depression [20], [64], [67], younger age [20], [64], [66] and stress
[20], [64], [66]. The predictors identified in these studies were consistent with findings from
other research using different datasets.

2.4 PPD Tool
Section 1.1.3 discussed standard PPD screening tools, including EPDS and PDSS.
Advancements in PPD screening have introduced new tools that assess a broader range of
factors, such as psychosocial and cognitive risks. Details of these tools are outlined below.

AJ Bjertrup et al. [68] examine how negative thinking about infant distress during pregnancy
relates to PPD and provide a proof of concept to test the practicality of measuring negative
neurocognitive bias using an online risk screening tool, REDCap. Eighty-seven participants in

38

their second or third trimester completed the online screening via a webpage on infant distress
reactivity and psychosocial risk questionnaires. Bivariate associations were analyzed, followed
by LR and multiple LR for PPD and post-birth depressive symptoms as outcome variables.
The predictor variables included Affective Neurocognitive Processing of Infant Stimuli
(emotional reaction to an infant distress video) and psychosocial factors. Other factors included
age, depressive symptoms during pregnancy, education level, pregnancy week, parental
bonding, attachment to the unborn child, job status, financial worries, income level, social
support, family history of psychiatric disorders, personal history of psychological illness, loss
of a parent, and childhood trauma. Key findings showed that negative emotional reactivity to
an infant distress video and negative interpretations of infant cries during pregnancy led to
PPD and symptom severity. The practicality of the online screening tool was less than ideal.
The low response rate suggests the need for a shorter, more optimized screening tool. As
mentioned in the paper, the key limitation indicated that the tool is not feasible in its current
form, and validation is required on large-scale data.

In 2012 McDonald S et al. [69] developed and validated a prenatal screening tool to identify
women at risk of postpartum distress. A longitudinal cohort data from the All Our Babies
(AOB) study was used to develop the tool based on regression coefficients. The independent
variables selected to derive the screening tool are depression in late pregnancy/4 months
postpartum, anxiety in late pregnancy/4 months postpartum, stress in late pregnancy, history
of depression, history of abuse, and social support in late pregnancy, and relationship tension.
The tool demonstrated higher sensitivity and broader focus compared to the EPDS, effectively
identifying women at risk of both depression and anxiety. However, it showed lower specificity
and performed better in women with middle to high socioeconomic status groups. Limitations
39

include reliance on self-reported measures instead of standard psychiatric diagnoses, missing
data, and underrepresentation of ethnic minorities and low-income households. Despite these
challenges, the tool provides a comprehensive approach to psychosocial risk assessment during
pregnancy.

In summary, beyond traditional scales like the EPDS, innovative tools provide a broader
perspective by incorporating emotional reactivity, psychosocial variables, and personal
histories. While these tools show promise, low response rates and limited generalizability
highlight the need for optimization to create more reliable and effective solutions. Additionally,
there are limited studies focused on developing new tools for screening or calculating the
likelihood of PPD, emphasizing the need for further research in this area.

2.5 Summary
In the field of developing predictive models for PPD, research commonly considers a range of
variables related to mental health, such as psychiatric history (during pregnancy, before
pregnancy, history of PPD, other depression or psychiatric illnesses, and family history of
depression). Furthermore, identified risk factors for PPD encompass financial aspects (such as
low income), age, mode of delivery, gestational medical conditions or illnesses (including
gestational diabetes), unplanned pregnancies, preterm births, and lower levels of education.
These models also incorporate factors associated with the gestational period, maternal factors,
and variables related to the family environment. However, it is essential to acknowledge that
there has been a relatively limited focus on investigating the potential impact of pre-gestational
health, disease, and specific conditions as predictive factors for PPD in the current literature.
40

While research has shed some light on the relationship between chronic illnesses and PPD, it
remains a relatively understudied area. Significantly, the existing literature has not identified
specific diseases that can be definitively classified as predictors for the onset of PPD. It has
been identified depression as a potential risk factor for developing chronic diseases,
particularly cardiovascular disease [12]. Also, Thyroid autoimmunity during pregnancy was
identified as a risk factor for PPD [59]. Further exploration and investigation are needed to
better understand the potential role of these diseases in predicting PPD. It is crucial to
emphasize that the understanding of predictive models integrating chronic diseases as potential
risk factors for PPD prediction is still in its early stages and lacks comprehensive insight.

41

Chapter 3
Methodology
This chapter provides a comprehensive overview of the research methodology. The proposed
framework includes analytical methods for predicting PPD and identifying risk factors, a
dashboard for enabling data exploration, and a web-based tool to find the likelihood of PPD
based on user input. The research methodology comprises several components ranging from
data collection to the visualization and final web-based tool (Figure 5). These components are
explained in this section.

Figure 5. Proposed Research Model.

42

3.1 Data Collection
The primary data source, PRAMS has undergone several phases of questionnaires, starting
from Phase 1 (1988–1989) and continuing to Phase 9 (2023–present). The questionnaire is
periodically updated and includes three parts:

•

Core questions: Used by all sites, covering topics like maternal pre-pregnancy status,
prenatal care, smoking, and alcohol use.

•

Standardized questions: Optional questions for sites to choose from.

•

Site-developed questions: Unique questions used by individual sites.

The PRAMS contains a standard set of variables in five categories, as given below [40]:

•

Birth Certificate: Provides information primarily on maternal and infant demographics.

•

Operational Variables: Indicate the mode used to complete the questionnaire, such as
by mail or phone.

•

Weighting Variables: Account for the PRAMS survey design and ensure statistical data
weighting.

•

Questionnaire Data: Represents the responses collected from the PRAMS survey.

•

Analytic Variables: Pre-calculated variables derived by combining different variables
within the dataset.

The anonymized dataset obtained for this study comprised 950 fields which includes maternal
and paternal demographics, infant- and family-related factors, chronic diseases, postpartum,
43

gestational, and obstetric variables. The collected records had a valid PPD symptom variable
‘PP_DEPRESS1’ with values 1 and 2 (False and True). The data was provided as a SAS file
containing detailed records, which was then exported to an SSMS database for processing.
SSMS was chosen for its robust data management capabilities.

3.1.1 Ethics Approval
The Research Ethics Board (REB) of the University of Northern British Columbia (UNBC)
reviewed the ethics approval application and determined that REB approval was not required.

3.2 Data Preparation
Data preparation (Figure 6) is a critical step in the data analysis process, involving procedures
that make raw data suitable for consumption by ML models [70]. These procedures include
cleaning, transforming, and organizing data. Adequate data preparation ensures optimal
outcomes by enhancing the efficiency and reliability of ML models.

1 PP_DEPRESS refers to the target variable "Post-Partum Depression," which was used to predict the likelihood

of post-partum depression in mothers.

44

Figure 6. Data Preparation Process [70].

3.2.1 Data Collection
The primary focus of this research was to identify potential high-risk chronic diseases
contributing to PPD. So, in addition to the risk factors described in Section 1.1.1, diseaserelated variables were included in the input features. Specifically, variables such as diabetes
and asthma were added to address the research question.

3.2.2 Data Preprocessing
Data preprocessing involves broader tasks such as handling missing values, feature
engineering, and handling imbalances in labeled data to improve data quality and relevance,
making the analysis more effective and valuable [71]. Proper execution of these tasks ensures
reliable results from predictive models.

45

Identifying the relevant variables that affect the outcome of interest poses many challenges.
Sometimes, unrelated variables indirectly influence the outcome, making it crucial to consider
all potential variables carefully. The initial dataset obtained was overly complex, leading to
model overfitting. This required extensive preprocessing to make the data suitable for analysis.
Identifying and handling issues in the dataset includes several tasks, which are explained
below.

Missing value treatment

Handling missing values is vital to reduce the impact of poor data quality on the model’s
performance. Issues such as biased predictions, reduced model reliability, and compromised
decision-making often stem from unclean data [71]. Dealing with missing data primarily
involves two major methods: imputation and row deletion. The underlying factors for missing
values in PRAMS data include item-non-response (explained in Section 3.2.6) and skipping
patterns. Item non-response occurs when individuals choose not to answer certain questions. It
significantly increases the risk of developing bias, leading to inaccurate or unreliable
predictions. Since PRAMS data is survey-based, records with missing data were removed to
address item non-response bias. Imputation was applied in cases involving skip patterns. For
example, in PRAMS’ Phase 5 survey, if a mother reported that her infant was still in the
hospital, she was instructed to skip the question about whether her infant was alive. It is
reflected in the variable INFLIVE5, where responses were coded as missing for mothers whose
infants were still hospitalized. Missing values for INFLIVE5 were imputed by referencing the
variable LTH_HOSP, where a response of "7" indicated that the baby was still in the hospital.

46

Challenges arose when analyzing subgroups within the PRAMS dataset. The primary issue
was sample size constraints, with limited entries for certain features in specific states.
Additionally, when state-specific features were excluded from surveys, the results became
skewed, reducing the reliability of specific analyses.

Irrelevant and redundant information

Knowledge discovery becomes more complicated when there are a lot of irrelevant and
duplicative details or noisy and unreliable data. This difficulty arises for the following reasons:

•

Multifaceted questioning: The survey may ask similar or nearly identical questions in
different sections. It can lead to repeated or conflicting information if participants
answer these questions again.

•

Incomplete updates: Updating the survey may change or remove questions, but their
old versions might remain. It can result in redundant questions in the survey.

For instance, the dataset comprised MH_PPDPR, MH_PPINT (these questions are adapted
from PHQ-2 - Phase 2 Questionnaire), and PP_DEPRESS variables which are defined as
follows:

MH_PPDPR: Since your new baby was born, how often have you felt down, depressed, or
hopeless?

MH_PPINT: Since your new baby was born, how often have you had little interest or pleasure
in doing things you usually enjoy?
47

PP_DEPRESS: Coded as “yes” for PPD symptoms if mother responded “always or often” to
MH_PPDPR or MH_PPINT and coded as “no” if mother responded “sometimes, rarely, or
never” for MH_PPDPR and MH_PPINT.

The PP_DEPRESS is an analytic variable composite of two variables, MH_PPDPR and
MH_PPINT. In the context of composite target variable PP_DEPRESS, MH_PPDPR and
MH_PPINT primarily serve as building blocks and there was no specific need to analyze them
separately. Here, removing the variables MH_PPDPR and MH_PPINT was considered.
Similarly, instances of redundancy can arise, such as the features INFQ_AGE and
INFQ_AGE_MOD, both representing infant age in days whereas the latter was identified as
being the cleaned version.

In addition, capturing different phases of the questionnaire in varying formats may lead to
similar variables. For example, starting from Phase 5, the INFLIVE5 variable reflects a change
in the survey skip pattern, where mothers with infants still in the hospital were instructed to
skip the question, "Is your baby alive now?" This case is coded as missing in INFLIVE5. In
earlier phases (Phase 4 or before), mothers were required to respond to this question, and their
responses were recorded in the INF_LIVE variable for comparability. While both variables
record the same frequencies for deceased infants, INFLIVE5 has fewer entries for alive infants
due to the enforced skip pattern. Additionally, the INFLIVE5_RAW variable captures
responses from mothers who did not follow the skip instructions and answered the infant alive
question. This response was coded as missing in INFLIVE5 to align with the enforced skip
pattern. Since this research focuses on Phase 5 to Phase 8 survey data, INFLIVE5 is selected,

48

and other similar variables, such as INF_LIVE and INFLIVE5_RAW, are removed during
preprocessing.

3.2.3 Feature Selection
Feature selection is necessary to choose the most relevant features (input variables) from the
original set of features for building a predictive model. Its purpose is to boost the performance
of the model, mitigate overfitting, and enhance the interpretability of the model [72].

Variable inclusion and exclusion

Input variables in predictive models were carefully selected to include only relevant factors,
ensuring model simplicity and improved accuracy. All the chronic diseases and health-related
variables available in the dataset were added to the input features to address the research
question. Variables deemed irrelevant to the predictive goals were removed to minimize errors
caused by unnecessary information. For instance, the variable "PRE_STI," which captures
discussions about sexually transmitted infections (STIs) such as chlamydia, gonorrhea, or
syphilis occurring before pregnancy was determined to be irrelevant and was removed. This
step ensured that the model focused solely on meaningful predictors and excluded irrelevant
features.

Another example of irrelevant variables is the PRE_CKD, which refers to standard question
L26, (Did you do any of the following things at any time during the 12 months before you got
pregnant with your new baby?) and (option d: A health care worker checked me for diabetes).

49

This question asks whether a healthcare provider conducted tests to check for diabetes,
regardless of the mother’s diabetes status, were removed, recognizing its irrelevancy.

In addition, the inclusion and exclusion of variables were based on data completeness and
relevance to the thesis objective. Only eight states with complete and reliable data for key
chronic disease variables were retained. Participant records were filtered to include mothers
aged 15–55, focusing on the most relevant demographic. Variables critical to the study were
prioritized, while redundant or less impactful features were removed using the feature selection
algorithm, which is explained in Section 0.

3.2.4 Feature Engineering and Derived Variables

Feature engineering involves transforming and creating features to make data suitable for
modeling. Correcting irregularities, enhancing feature relevance,

and meaningful

transformations of raw data are the main processes involved in feature engineering [71].

Feature engineering was conducted to create new variables such as ABUSE and CHRONIC
(Flag type). Derived variables from existing features were introduced to the dataset by
transforming or combining existing variables. Thus, additional features, such as MOM_AGE
from MDOB_YR4 and ID and YEAR from ID, were created. Creating these new features helps
reveal hidden patterns in the data, enhancing the accuracy and performance of the model's
prediction. All these processes resulted in a final dataset comprising 23,855 records and 42
features. A nurse practitioner reviewed and confirmed the selection of these final variables,
ensuring the clinical relevance and validity of the data used for analysis.
50

3.2.5 Data Encoding and Scaling
Data transformations and scaling are vital to address issues arising from disparate data ranges
and categorical representations, which can otherwise result in biased predictions or
inefficiencies in model performance [71]. Encoding categorical data and scaling numerical
values makes the data more suitable for analysis, fostering accurate and reliable predictive
outcomes.

Encoding such as one-hot or label encoding involves converting categorical variables into
numerical formats which can be interpreted by the model. Tree-based models, such as Random
Forests (RF), natively handle such data through their tree-splitting mechanism allowing them
to evaluate categories directly without assuming specific relationships between data types, thus
eliminating the need for encoding. Pre-processed data was scaled for features such as maternal
and paternal race and month of birth to be utilized in the LR, NN, and SVM. These ML models
are further explained in Section 3.4.

Scaling ensures that numerical variables are standardized to a consistent range or distribution,
improving model convergence and performance. Variables such as infant age were scaled for
NN and SVM because of their sensitivity to feature magnitude or distribution. Similar to
encoding, tree-based models such as RF do not require scaling.

51

3.2.6 Challenges
Unbalanced data

Conducting thorough exploratory data analysis (EDA) is essential to identify imbalances in
datasets, which can severely affect the performance of ML models. Unbalanced data can skew
the model predictions toward the majority class. SMOTE was applied to handle imbalances in
data. SMOTE generates synthetic samples for the minority class to balance the dataset and
improve the model's generalization ability across classes.

Item non-response

When individuals choose not to respond to survey questions, it can result in a biased survey.
PRAMS relies on surveys that are sent to new mothers. Not all mothers who receive the
questionnaires opt to participate and provide their responses due to a variety of reasons, such
as time constraints, privacy concerns, or a lack of interest. This non-response can introduce
bias, as non-respondents may have different characteristics or experiences than respondents,
posing a significant challenge. It affects the quality and representativeness of the data. With
this bias, the data accurately reflects only part of the population, impacting the validity of
research findings and decisions based on the data. To address this, biased records were
removed, and subgroup analyses were conducted to assess whether non-response bias varied
across different demographic groups.

52

3.3 SPSS Modeler
Although previous research on PRAMS data has relied heavily on SAS (Statistical Analysis
System), SPSS (Statistical Package for Social Science), and STATA as the primary analysis
software, data mining and ML techniques have gained attention. As a result, some researchers
have started to explore open-source software such as IBM's SPSS Modeler, which offers robust
analytical capabilities. The reason behind choosing this tool is its following capabilities:

1. The software itself finds the appropriate model for given data.
2. Automated feature selection and classification model selection.
3. Easy to use with a simple drag-and-drop interface.
4. Easy to add source data from multiple sources.

A node in SPSS Modeler represents a specific operation or analysis and is interconnected to
other nodes to form a stream that sequences data analysis process steps from input to output.
This tool was selected for this thesis for specific tasks such as feature selection and selecting
the best predictive models for classification using nodes explained below [73]:

Feature Selection Node:

The most significant predictors for a given analysis are identified using this node. It evaluates
input fields relative to a specified target, ranking them based on importance. This process
involves:

53

•

Screening: Eliminate inputs such as those with excessive missing values or
minimal variation that are unimportant or outliers.

•

Ranking: Sort and assign ranks.

•

Selecting: identifies the subset of features to utilize in the selected models.

Auto Classifier Node:

It compares multiple predictive models for nominal or binary targets and selects suitable
models for the given dataset and target. It allows users to:

•

Select Algorithms: Choose from various modeling methods, such as decision trees,
neural networks, or support vector machines.

•

Experiment with Options: Test multiple combinations of algorithm parameters to
identify the optimal configuration.

•

Rank Models: Evaluate and rank each candidate model based on criteria such as
accuracy or F1 score.

3.4 Predictive Model
Once the data is prepared and cleansed, the next step is to use it for predictive modeling. The
four models used for this study are NN, LR, RF, and SVM. The rationale for selecting these
models is based on their demonstrated superior performance in predicting PPD, as identified
in the literature. Model-specific reasons are provided in Section 4.3.

54

Input data

The preprocessed dataset was imported from the SQL server to Jupyter Notebook in the Visual
Studio environment and read using pandas' data frame. Considering the target variable
'PP_DEPRESS,' the dataset was partitioned into training and testing sets. Better results are
achieved when 20–30% of the data is allocated for testing, with 70–80% used for training [74].
Therefore, a 70:30 split ratio was selected for training and testing to ensure sufficient data for
model training while reserving enough data to evaluate the trained model's performance on
unseen data.

Model Training (Learning)

The ML models were trained on the prepared dataset. In this phase, the models learned the
underlying patterns and associations between the input features and the target variable
'PP_DEPRESS.' The models were retrained by tuning hyperparameters, which are configurable
settings influencing the learning process and model performance. A stratified fold crossvalidation was applied to ensure that each fold maintains the same proportion of each class as
the original dataset, which is important given the imbalanced nature of the data.

Model Testing (Decision-Making)

After training, the model's effectiveness was assessed using a separate, smaller test dataset
without the target variable. Accuracy and F1 scores were calculated from test results to evaluate
how well the models learned to predict PPD. This is described in detail in Section 4.2.

55

3.4.1 Artificial Neural Network
Artificial Neural Networks (ANNs) are advanced ML models inspired by the human brain's
neural structure. Models learn and recognize the patterns by designing the network to process
the data in layers [75]. The structure of these layers can solve complex, non-linear problems.
Mainly, an ANN consists of three components:

Input Layer: Accepts raw input data and forwards it to the neurons of another layer in
the network for processing.

Hidden Layers: This layer applies weights, biases, and activation functions to uncover
relationships within the data. An activation function introduces non-linearity to an NN and
learns complex patterns and relationships in data. Deeper networks use multiple hidden
layers.

Output Layer: Final predictions or classifications are produced in the output layer.

The data flows from the input layer to the output layer through the hidden layer, and each
layer's neurons are interconnected. ANNs learn by adjusting weights during training using
backpropagation to compute error gradients. Optimization methods such as gradient descent
utilize the gradients to reduce prediction errors and enhance model performance. Figure 7
shows the schematic diagrams of ANN.

56

n
Figure 7. a) ANN (b)Structure of ANN for a Neuron in a Hidden Layer with One Output [76].

The output of a neuron in an artificial NN is evaluated as
n

hj = f (∑ wij xi + bj )
i=1

Equation 1. Output of the Neuron

where

xi : Input values

wij : Weight connecting input xi to neuron hj
57

bj : Bias term for neuron hj

f: Activation function

hj : Output of the j-th neuron

The versatile characteristics of ANN make it suitable in fields such as image recognition,
language processing, predictive modeling, handwriting recognition, fraud detection, and
speech-to-text systems. A Multi-Layer Perceptron (MLP) NN classifier was chosen due to its
ability to balance complexity and flexibility, making it well-suited for binary classification
tasks.

3.4.2 Logistic Regression
Logistic regression (LR) is the most commonly used regression model that creates a linear
relationship between the independent variables and the logarithm of the odds for the target
variable. This method is designed for binary classification problems where the target variable
has two possible outcomes (e.g., success/failure, yes/no, or 0/1). Its straightforward
implementation and efficiency make it a reliable choice for classification tasks, predicting the
probability of a target class using the sigmoid function, which outputs values between 0 and 1
[77]

The logarithmic odds are transformed into probabilities using the sigmoid function. A
threshold (commonly 0.5) is applied to these probabilities to classify the outcomes. Each
58

coefficient in the model reflects the predictor's impact on the log-odds of the target, aiding in
understanding its effect on the outcome. Figure 8 is a visualization of the sigmoid function
used in LR, illustrating how LR transforms input features into a probability distribution suitable
for binary classification.

Figure 8. The Sigmoid Curve in LR [78]

The sigmoid function, also known as the logistic function, is a mathematical function that maps
any real-valued number into a value between 0 and 1, producing an S-shaped curve. This
characteristic makes it particularly useful in LR for modeling probabilities. The function is
defined as:

σ(x) =

1
1 + e−x

Equation 2. Sigmoid Function
59

In this equation, e represents the base of the natural logarithm, and x is the input value. The
sigmoid function effectively compresses input values into a range between 0 and 1, which can
be interpreted as probabilities in the context of binary classification. LR is widely recognized
for its practical healthcare, social sciences, and finance applications [77].

3.4.3 Random Forest Tree
RF is a powerful ensemble learning model used for classification and regression tasks. It
constructs multiple decision trees during training and combines them to make predictions to
improve accuracy and reduce overfitting. Unique trees are built by selecting random data and
features. Robust and accurate models are created with a reduced likelihood of overfitting, with
the diverse nature of the randomly constructed trees in the forest [79]. This makes it a popular
choice for ML tasks for its robustness and reliability of results. Figure 9 shows the RF
algorithm, which combines the results of multiple decision trees by majority voting or
averaging to produce robust and accurate predictions. Structure and Functioning of RF:

1) Ensemble of Decision Trees: Each decision tree is trained on multiple random subsets
of the data (with replacement, known as bootstrapping).

2) Feature Selection: A random selection of attributes is made at each division in a tree,
promoting variety among the trees.

60

3) Prediction Aggregation: For classification tasks, the final prediction is determined by
a majority vote, where the class predicted by most trees is chosen. For regression tasks,
the final prediction is the average of all individual tree predictions.

Figure 9. RF: An Ensemble Learning Algorithm [80].

These ensemble characteristics reduce the overfitting that generalizes better than individual
decision trees, making it suitable for applications such as medical diagnosis, fraud detection,
and customer segmentation.

61

3.4.4 Support Vector Machine
SVM is one of the most robust prediction methods and a powerful supervised ML models
primarily used for classification and regression tasks. It identifies support vectors, which are
the closest data points to the hyperplane, and uses them to maximize the margin between
classes and ensure the best possible separation of classes. This makes SVM improve
generalization, reduce misclassification, and handle high-dimensional spaces and complex
datasets [81].

SVM can effectively handle non-linear relationships by transforming data into a higherdimensional space using kernel functions (e.g., linear, polynomial, radial basis function). For
example, a radial basis function kernel allows SVM to create flexible decision boundaries for
complex datasets. Figure 10 presents a simple SVM with a linear kernel.

Figure 10. Linear SVM [82].

62

SVM is widely used in applications such as image recognition, spam detection, and
bioinformatics [82]. It is robust to overfitting, especially in cases where the number of
dimensions exceeds the number of data points.

3.5 Dashboard
A business intelligence (BI) dashboard allows users to analyze and report on key performance
indicators and other metrics. Many health organizations have online visualization tools
available for the public to explore trends of incidence, mortality, demographics, and other
statistics etc., to intuitively visualize vast amounts of data. BI dashboards are effective as they
reduce the time needed to visualize data in charts, graphs, and maps.

A dashboard was built using Power BI for several reasons such as its capability to connect to
various data sources, built-in insights with drag-and-drop features, and the creation of a
shareable and customizable wide range of visuals. Power BI also offers a free standalone
desktop application. It offers a user-friendly platform that integrates data cleaning, statistical
analysis, and interactive visualizations, enabling seamless collaboration, cost-effective
solutions, and powerful insights backed by Microsoft’s robust support and extensive ecosystem
[83].

63

3.6 Summary
This chapter outlined the methodology used in this research, describing various components
ranging from data sources to visualization techniques. The experiments and outcomes are
explained in the following chapter.

64

Chapter 4
Experiment and Results
This chapter outlines the data visualization, evaluation of selected ML models that predict the
risk of PPD, and a PPD web-based tool. The post-training analysis section identifies key
predictors, focusing on the relationship between chronic diseases and the risk of PPD. The
predictive modeling sub-section presents results from various ML models and provides a
comparative performance evaluation. The post-training analysis examines the selected chronic
diseases to demonstrate the validity of the hypothesis. Finally, the functionality of a tool that
predicts PPD based on user inputs is presented.

4.1 PPD Dashboard
A dashboard created using Power BI is presented in this section, which supports EDA in
exploring the data to gain insights and identify trends along multiple key performance
indicators (KPIs) (Figure 11). The dashboard showcases data segmented across various
dimensions, including maternal age groups, PPD prevalence, disease variables, maternal
education groups, prenatal variables, and infant factors over the years and states. Of the 23,855
individuals in the cleaned dataset, 3,259 (13%) were identified as having PPD, and 11,000
were reported to have chronic diseases.

65

Figure 11. PPD Dashboard

The Key Influencers

The Key Influencers functionality in Power BI uses statistical methods along with odds ratios
to identify and rank factors that most influence the target outcome. Figure 12 reveals the factors
that significantly influence the likelihood of depression during postpartum. Individuals having
a history of depression in specific maternal age groups are 2.2 times more likely to experience
PPD. Being married and belonging to paternal race group 1 (Asian, excluding Indian, Japanese,
Filipino, or Chinese) increases the likelihood of PPD by 1.75 times, whereas having a chronic
disease raises the probability by 1.74 times. Additional factors such as having a history of

66

hypertension increase the probability by 1.51 times, and Type 1 or Type 2 diabetes diagnosis
raises the likelihood by 1.23 times. These insights provide valuable information for
understanding the drivers of PPD in the population.

Figure 12. PPD Key Influencers

Demographic Features

Figure 13 and Figure 14 highlight the prevalence of PPD across maternal age groups and
education levels. The 25-29 age group (978 cases) has the highest number of cases, followed
by the 30-34 (820 cases), and the lowest in the 15-18 (53 cases) group. PPD prevalence also
increased with education levels, with the highest cases in women with 13-15 years of education
(1,071 cases) and the lowest in those with 0-8 years (66 cases). Marital status shows more

67

married individuals with PPD in higher education categories, particularly among those with
12-15 years of education (1,165 married).

Figure 13. PPD and Maternal Age Group

Figure 14. PPD by Maternal Education Level and Marital Status (Married)

68

Disease Factors

Figure 15. PPD by Disease Factors

Figure 15 depicts the trends in various health conditions among patients from 2016 to 2020,
including heart disease, thyroid dysfunction, asthma, seizures, chronic disease, anemia, and
PCOS. Anemia and Asthma exhibit moderate increases over the years, while the remaining
diseases remain relatively stable with minimal fluctuations. All features with PPD show similar
trends in their increase, which could have an impact on PPD.

Prenatal Factors

Figure 16 illustrates the trends in the prevalence of PPD, with gestational diabetes, depression,
and high blood pressure (HBP) from 2016 to 2020. Though gestation diabetes shows the

69

highest and depression the lowest, all factors show a steady increase throughout the period.
These patterns indicate a potential relationship between prenatal health conditions and PPD.

Figure 16. PPD by Prenatal Factors.

4.1.1 Summary
The dashboard provides a concise overview of PPD prevalence. The Key Influencers tool
identifies critical factors impacting PPD, such as history of depression, maternal age, chronic
disease, and marital status, enabling data-driven insights that are consistent with the important
predictors identified by ML models. Visualization helps ensure accurate insights into patterns
and relationships between features and the target variable and helps identify imbalances in the
data.

70

4.2 Evaluation Metrics: Accuracy and F1 Score
The accuracy of the model is determined by dividing the number of instances where the
predicted values match the actual values by the total number of predictions [84],

Number of Correct Predictions
) × 100
Accuracy (%) = (
Total Number of Predictions

Equation 3. Accuracy by Percentage

In addition, the F1 score, the harmonic mean of precision and recall, provides a more detailed
evaluation, which is especially useful in imbalanced datasets. It helps to balance the trade-off
between false positives and false negatives, ensuring a thorough assessment of the model's
predictive capabilities. Accuracy and F1 [87] scores provide a complete view of the model's
effectiveness in accurately classifying observations.

F1 Score = 2 × (

Precision × Recall
)
Precision + Recall

Equation 4. F1 Score Formula.

where
Precision is the ratio of correctly predicted positive observations to the total predicted
positives, given by [85]:

71

Precision =

True Positives
True Positives + False Positives

Equation 5. Precision Formula

Recall (or Sensitivity) is the ratio of correctly predicted positive observations to all the actual
positives, given by [85]:

Recall =

True Positives
True Positives + False Negatives

Equation 6. Recall Formula

4.3 Predictive Modeling: An Analysis of Four Models to Predict PPD
During the training phase, the performance of the models was improved by tuning the
hyperparameters. This section presents the selected parameters for each model and compares
the corresponding performance metrics.

Hyperparameter optimization was conducted on NN and SVM using RandomizedSearchCV to
efficiently explore a wide range of hyperparameter combinations, while LR and RF employed
GridSearchCV to search for the best configurations. The search was performed in parallel to
enhance computational efficiency. Each model's decision threshold was optimized to improve
the balance between precision and recall. The models were trained using sample weights to
account for class imbalances.

72

4.4 Model Comparison
Figure 17 illustrates the performance of models on selected features, showing their influence
on the probability of PPD. The models were evaluated based on their accuracy and F1 scores
defined earlier.

Figure 17. Model Comparison: Accuracy and F1 Score.

Table 2. Precision, Recall, and F1 Score
Model

Precision

Recall

F1 Score

NN

0.75

0.79

0.77

LR

0.90

0.86

0.88

RF

0.79

0.85

0.82

SVM

0.79

0.81

0.80

73

The model comparison reveals that LR attained 89% accuracy and an 88% F1 score. The RF
model achieved 83% accuracy and 82% F1 score, followed by the SVM model with 83%
accuracy and 80% F1 score. The NN model demonstrated 76% accuracy and 77% F1 score.
The close similarity between accuracy and F1 scores across models suggests balanced
performance across the majority and minority classes. This balance indicates that the models
are not biased toward one class, effectively managing the trade-off between precision and
recall, as presented in Table 2.

4.4.1 Performance Metrics: Discussion
This section discusses the performance metrics observed during the model evaluation, with a
particular focus on the influence of parameter tuning. The results were analyzed to understand
how different tuning approaches influenced the selected models' overall accuracy and F1 score.
A stratified cross-validation was applied to test if the model was biased or overfitted.

Neural Network
o An initial 75% F1 score was achieved using Python's MLPclassifier with default
parameters.
o Tuning the parameters to two layers with a ReLU (Rectified Linear Unit) activation
function, an Adam optimizer, and L2 regularization set at 0.001, along with a
sigmoid (output layer), improved the F1 score to 77%.

74

Logistic Regression
o Tuning the solver parameter for liblinear, saga, and lbfgs resulted in varying
accuracy scores. The saga solver achieved around 65% accuracy, while the other
two solvers—liblinear and lbfgs (Limited-memory Broyden-Fletcher-GoldfarbShanno)—each achieved approximately 89% accuracy, finding the optimal
threshold value of 46 instead of 50.
o LR achieved 89% and 88% accuracy and F1 score, respectively, using
GridSearchCV with a maximum of 5000 iterations, an L2 penalty, liblinear solver,
and a regularization strength (C) of 0.01.
Random Forest
o The basic RandomForestClassifier configuration with default parameters—100
trees, a minimum of 2 samples required to split an internal node, no maximum tree
depth, and at least 1 sample per leaf—achieved an F1 score of 80%.
o Tuning hyperparameters with GridSearchCV, the RF model was configured with
300 trees, a minimum of 2 samples required to split a node, no maximum tree depth,
and max_features set to 'sqrt' yielded an F1 score of 82%.
Support Vector Machine
o The SVM model, initially configured with parameters such as enabling probability
estimation and setting the random state to 42, achieved an F1 score of 77%.
o Tuning the 0.01 RBF (Kernel Coefficient) and adjusting the C (regularization
parameter) improved the F1 score to 80%.

75

The performance achieved by LR is likely due to its ability to effectively model linear
relationships between features and the binary target, as well as the use of optimized
hyperparameters that balance precision and recall. The ensemble learning approach in RF
effectively captures complex nonlinear relationships in the data by creating multiple decision
trees, likely resulting in robust classification performance. Likewise, capturing complex
nonlinear relationships through the RBF kernel in SVM might have resulted in enhanced
predictions. Stratification ensures that each fold reflects the class imbalance in the dataset,
forcing the model to learn from both classes, which may slightly reduce accuracy but improve
generalization.

4.5 Post-Training Analysis: Identify Key Predictors
This section identifies key predictors, focusing on the relationship between chronic diseases
and the risk of PPD. SHAP (Shapley Additive exPlanations) [86], a game theoretic approach,
was utilized in this post-training model analysis. Other model-explaining techniques, including
the Permutation Importance and Partial Dependence plots, allow us to understand the
importance of features but lack local interpretability. Since the SHAP can be model-specific
or agnostic, it has the flexibility to adapt to various ranges of models.

The SHAP value was calculated to interpret the impact of input features on the target
PP_DEPRESS. The 'SHAP' Python package enabled the creation of various plots to understand
the key predictors and how the values of each predictor impact the target. Plots were based on
the mean absolute SHAP values of each critical feature. The Beeswarm plots in Figure 19,
Figure 21, Figure 23, and Figure 25 show a deeper insight into each value of every significant
76

predictor, whether it has a positive or negative impact. Each dot corresponds to an individual
prediction, shifting the color gradient from blue for lower feature values to red for higher ones.
The positive SHAP value for a feature value indicates that it has a strong positive contribution
to predicting PPD, and the negative value indicates the reverse. Table 3 shows the key features
and their predictive strength by selected models with label information. The SHAP values for
significant features using each model are presented below.

Neural Network

Figure 18 demonstrates the key features identified by the NN model using SHAP. The bars
show the absolute SHAP value, meaning a more critical feature with a higher SHAP value on
the top. The feature CHRONIC was on the top, having a SHAP value of 1.33, followed by
other selected chronic diseases, anxiety, anemia, and asthma, which are consistent (except
anxiety) with other models described later.

Key chronic predictors of PPD identified are anxiety, anemia, heart disease, PCOS, epilepsy,
thyroid dysfunction, and pre-pregnancy diabetes. Other features such as length of stay at the
hospital, infant sex, high blood pressure, both father's and mother's race, paternal education,
and Kessner Index2 are also underscored as key predictors.

2

The Kessner Index measures the adequacy of prenatal care by evaluating when prenatal care began,
the total number of visits, and the length of the pregnancy [87].

77

Figure 18. Feature Significance Using SHAP on NN Model

Figure 19 displays the contribution of the feature CHRONIC towards PPD with a higher value
(1- presence). Chronic diseases such as heart disease, anxiety, anemia, asthma, PCOS,
78

epilepsy, thyroid dysfunction, and pre-pregnancy diabetes had lower SHAP value for their
presence. The other vital features identified are infant sex and place of birth. Having high blood
pressure during pregnancy had a significant influence on predicting PPD.

Figure 19. SHAP Summary: Visualizing Feature Impact and Distribution Using NN Model

79

Logistic Regression
Figure 20 shows the key predictors of PPD. The feature CHRONIC comes after paternal and
maternal race and the infant’s birth month. CHRONIC and MH_PGDX8 (Health problems
during pregnancy – Depression) got the same SHAP value of .29.

Figure 20. Feature Significance Using SHAP on LR Model

80

The SHAP analysis across each feature is shown in Figure 21. Considering the feature
CHRONIC, a red marker (1-presence of chronic disease) falls in the positive SHAP value
region, indicating that the mother with chronic disease had a higher chance of getting PPD.

Figure 21. SHAP Summary: Visualizing Feature Impact and Distribution Using LR Model

81

In addition to the presence of chronic disease, depression during pregnancy, younger age of
mother, lower Kessner Index, having high stress, "Other" marital status, history of depression,
smoking behavior, and lower paternal education level have been identified as the key factors
in predicting risk of depression. When considering specific chronic diseases, epilepsy emerges
as a strong contributor, while the presence of asthma, anxiety, and anemia show a lesser impact.

Random Forest
The key features identified by the RF model using SHAP are shown in Figure 22. Similar to
LR and NN, the feature CHRONIC is indicated as one of the key predictors of PPD. In addition,
other similar features such as Kessner Index, depression during pregnancy, paternal and
maternal race, paternal education, asthma, and anemia are found in the top predictors.

82

Figure 22. Feature Significance Using SHAP on RF Model

The SHAP plot in Figure 23 shows the detailed effect of each feature value in predicting PPD,
highlighting that the feature CHRONIC has a positive contribution, consistent with both the
LR and SVM models.

83

Figure 23. SHAP Summary: Visualizing Feature Impact and Distribution Using RF Model

It is noted from the above chart that depression during pregnancy, younger maternal age, highstress levels, older infant age, being of a race such as Chinese, Japanese, Filipino, or Hawaiian,
and lower paternal and maternal education levels are strong contributors to predicting PPD
risk. Additionally, diseases such as asthma and anemia show lower SHAP values, which are
consistent with LR and SVM. Having depression (prenatal and before pregnancy), younger

84

maternal age group, "Other" marital status, lesser prenatal care visit, and a lesser Kessner Index
aligns with LR results. Not being breastfed emerges as an additional vital contributor to the
feature importance analysis.

Support Vector Machine

Figure 24 illustrates the key features identified by the SVM model using SHAP. The feature
CHRONIC was at the top, with a SHAP value of more than 0.90, followed by other selected
chronic diseases, such as anxiety, anemia, and asthma. Further down, thyroid dysfunction is
also identified as an essential feature in predicting PPD.

85

Figure 24. Feature Significance Using SHAP on the SVM Model

The beeswarm plot in Figure 25 visualizes a detailed overview of how individual feature values
affect the target PP_DEPRESS. The distribution reveals that when the binary feature
CHRONIC is 1 (presence), it gives a positive SHAP value, thus contributing positively to the
PPD prediction. Consistent with the LR model, the mother's depression during pregnancy
(MH_PGDX8) holds similar significance in predicting PPD. In contrast to LR, marital status

86

(MARRIED) shows mixed values in both directions, making it difficult to determine which
category of this feature is a vital contributor.

Figure 25. SHAP Summary: Visualizing Feature Impact and Distribution Using SVM Model

Features like maternal age group, stress level, and maternal education level exhibit effects
similar to LR. The SHAP value indicates that mothers in the younger age group or with lower

87

education are more strongly linked to PPD. Other key chronic disease predictors, such as
anxiety, anemia, asthma, and thyroid dysfunction, exhibit lower SHAP values, indicating a
diminished influence on predicting PPD. In comparison with RF, the SHAP plot for
breastfeeding shows a similar pattern, with non-breastfeeding being denser in the positive
direction.

4.5.1 Summary: Post-Training Analysis
Table 3 presents key features that strongly contribute to predicting PPD across selected ML
models. Features such as CHRONIC, asthma, anemia, maternal race, paternal education, and
history of depression are consistently presented as key features with comparable SHAP values.
Additionally, stress level, maternal age, maternal education, paternal race, marital status,
Kessner Index, depression during pregnancy, infant age, and anxiety were identified as key
features among at least three out of the four models. Diabetes before pregnancy is common in
both RF and NN, breastfeeding is common in both RF and SVM, and length of stay in the
hospital is common in both NN and SVM. Only a few additional vital features emerged
uniquely in each model. Notably, LR and RF show a presence of a history of depression, while
SVM shows an absence of a history of depression, indicating variability in its influence across
models.

88

Table 3. Key Features and Their Predictive Strength by Model
Feature

Label

NN

LR

RF

SVM

CHRONIC

Chronic disease

Presence

Presence

Presence

Presence

MH_PGDX8

Depression during
pregnancy (binary)

NA

Presence

Presence

Presence

KESSNER

Kessner Index –
number of prenatal
care (ordinal)

Uneven
effect

Lower

Lower

NA

STRS_T_G

Stress level group
(ordinal)

NA

Higher

Higher

Higher

INFQ_AGE_MOD

Infant age (days)

Uneven
effect

NA

older

older

MAT_AGE_NAPHSIS

Maternal age group
(ordinal)

NA

Younger

Younger

Younger

MAT_ED

Maternal education
group (ordinal)

NA

MediumLower

Medium

MediumLower

PAT_ED

Paternal education
group (ordinal)

Uneven
effect

MediumLower

Medium

Uneven effect

HTH_ASMA

Asthma (binary)

Absence

Absence

Absence

Absence

HTH_IRON

Anemia (binary)

Absence

Absence

Absence

Absence

HTH_ANX

Anxiety (binary)

Absence

Absence

NA

Absence

MARRIED

Marital status
(binary)

NA

Married

Married

Uneven effect

Uneven
effect

Uneven
effect

Other
Asian,
White,
Black,
Indian, and
Chinese

Uneven effect

Other
Asian,
White,
Black,
Indian, and
Chinese

NA

Presence

Absence

MAT_RACE

Maternal race
(nominal)

PAT_RACE

Paternal race
(nominal)

Uneven
effect

Other Asian,
White, Black,
Indian,
Chinese,
Japanese, and
Filipino

BPG_DEPRS8

History of
depression (binary)

Uneven
effect

Presence

89

The identified key predictors contributing to PPD are greatly consistent across all models.
Moreover, deeper individual feature value analysis revealed that the strength of these features
in predicting PPD is also closely aligned across the models. Any minor differences between
the models are likely due to the handling of nonlinearities and interactions.

Additionally, the feature CHRONIC was consistently identified in all models and showed
similar strength in predicting PPD. However, the SHAP values for the presence of individual
chronic diseases were low, indicating a need for further interaction analysis. Such patterns
often occur when models see these features as part of broader interactions rather than as strong
individual predictors.

4.6 Identifying and Analyzing Interactions Among Key Predictors
The low SHAP values for individual chronic diseases indicated the likelihood for these features
to contribute more significantly through interactions rather than as standalone predictors. New
features were generated to assess the combined effect of features identified as important
through SHAP analysis. The following steps involve extracting the combination and analyzing
the interaction for all selected models.

90

Step 1: Generating Interaction Features

Interaction features were developed by combining selected input features. Since the nominal
features were encoded, the selected features had Boolean and ordinal datatypes. A bitwise
"AND" operation was performed between feature pairs. Each new feature was given a
meaningful name and added to the original dataset as a new column.

Step 2: Data Preparation and Modeling with Interaction Features

After creating interaction features, they were filtered to retain only those identified as
significant in the feature selection node in SPSS Modeler [73]. Figure 26 shows the feature
selection node for interaction features.

Figure 26. SPSS Feature Selection Node

91

Figure 27. Feature Selection Results

The generated features were loaded using the Excel data source node of the SPSS Modeler. A
Merge node was used to merge the data, which was connected to a Type node. The Type node
helped define each feature, including its type and target. The measurement column describes
92

the datatypes, and the least key features were ignored due to their large single category. These
results were then used to filter and create a final dataset with relevant input features. Finally,
the dataset used in the model to identify the significant interaction.

Step 3: Post-Training SHAP Analysis of Interactions

SHAP analysis was conducted to identify the most impactful interactions and provided insights
into how feature combinations influence the predictions. The following results of SHAP values
highlight significant feature interactions for each model.

Neural Network

Figure 28 illustrates the key interaction features of the NN model identified using SHAP.
Maternal education is found at the top as a significant feature, along with other features, such
as ‘No medical risk factors,’ high blood pressure and gestational diabetes, number of perinatal
care visits, chronic diseases, anemia, and asthma. Similarly, the number of prenatal care visits,
along with factors such as the absence of medical risk factors, high blood pressure, and diabetes
during pregnancy ranks top on the chart.

93

Figure 28. Interaction Feature Significance Using SHAP on NN Model

The feature Chronic emerges as a key predictor for PPD when interacting with factors such as
maternal education level, gestational diabetes, hypertension during pregnancy, Number of
perinatal care visits, and paternal racial background. Also, two chronic diseases, asthma, and
anemia, interacted similarly with the maternal education level.
94

Logistic Regression

Figure 29 highlights the key interaction predictors for PPD using LR. Similar to NN, maternal
education is shown at the top of the chart. Maternal age groups are the most common significant
interaction features with maternal and paternal education and marital status. Additionally, the
number of perinatal care visits interacts with the history of depression, abuse, maternal
education, and the Kessner Index.

95

Figure 29. Interaction Feature Significance Using SHAP on LR Model

Maternal and paternal education features were identified with a SHAP value of 0.18.
Considering the chronic feature, the analysis found a moderate interaction with the maternal
age group (SHAP value 0.017) and paternal education (SHAP value .012). No other relevant
interactions were found between specific chronic diseases and any other feature. Other key
interaction predictors include paternal and maternal race, abuse and maternal education, and
96

the Kessner Index with the number of perinatal care visits, paternal education, and number of
perinatal care visits with a history of depression, maternal education, and abuse.

Random Forest

Figure 30 shows the key interaction features identified by SHAP in the RF model. The
significance of paternal education and maternal education are found to resemble the LR model.
In addition, other similar inactive features include paternal education with ‘No medical risk
factors,’ maternal age, maternal education, asthma and stress level. The interaction of Chronic
disease with paternal education and maternal age were identified as key predictors of PPD,
which is similar to the LR model. In contrast to NN and LR, maternal education with ‘No
medical risk factors’ was found to be the top predictor of PPD. Additionally, the feature
Chronic shows interactions with the infant age, history of depression, paternal and maternal
education, and stress level. Finally, one of the chronic diseases, asthma, interacts with the
maternal age group, paternal education, and infant age.

97

Figure 30. Interaction Feature Significance Using SHAP on RF Model

Support Vector Machine
The key interaction features for predicting PPD using the SVM are presented in Figure 31. In
alignment with NN, paternal education is found at the top of the chart, with diabetes, high
98

blood pressure during pregnancy, and the absence of medical risk factors. Similarly, the
maternal age group with maternal education, ‘No medical risk factors,’ and paternal education
are aligned with RF. Paternal education and maternal education are similar to those of the LR
model.

Figure 31. Interaction Feature Significance Using SHAP on the SVM Model

99

Key interaction predictor features identified in the SHAP plot include paternal education
interacting with the number of perinatal care visits, infant sex, Kessner Index, and abuse.
Kessner Index, together with high blood pressure and diabetes during pregnancy, paternal
education, infant sex, and the absence of medical risk factors, also show significant
interactions. Hospital length of stay also interacts notably with high blood pressure, the absence
of medical risk factors, and maternal education in predicting PPD. No relevant interactions
were found with the feature Chronic.

4.7 Summary: Interaction Analysis
Rather than examining features individually, interaction analysis captures the combined
influence of multiple features. This approach enhanced the model's interpretability in the
prediction process. Table 4 presents common key interaction features identified in the selected
models.

Each model captures distinct interaction features for predicting PPD. In the SVM model,
features such as infant sex and length of hospital stay showed interactions with other variables
and were uniquely significant to this model. Similarly, in the NN model, interactions involving
anemia were uniquely significant.

Features such as high BP, gestational diabetes, maternal age group, ‘No medical risk factors,’
number of perinatal care visits, chronic disease, and maternal and paternal education frequently
interact with other predictors, emphasizing their roles in PPD prediction modeling. Notably,
paternal education and maternal education appear across all models, while chronic disease
100

appears in three models: NN, LR, and RF.

Table 4. Top Common Key Interactive Features Identified by the Selected Models
Feature

Models

Chronic_AND_Maternal_Education

NN and RF

Chronic_AND_Paternal_Education

LR and RF

Chronic_AND_Maternal_Age_Group

LR and RF

No_Med_Risk_Factor_AND_Maternal_Education

NN and RF

No_Med_Risk_Factor_AND_ #Perinatal_Care_Visit

NN and SVM

No_Med_Risk_Factor_AND_ Maternal_Age_Grpd

RF and SVM

No_Med_Risk_Factor_AND_Paternal_Education

RF and SVM

Preg_Diab_AND_ #Perinatal_Care_Visit_ Grpd

NN and SVM

Preg_High_BP_AND_ #Perinatal_Care_Visit_ Grpd

NN and SVM

Maternal_Age_Grpd_AND_Paternal_Education

LR, RF, and SVM

Maternal_Age_Grpd_AND_Maternal_Education

LR, RF, and SVM

Maternal_Education_AND_Paternal_Education

LR, RF, and SVM

101

4.8 PD Screening Tool
An interactive machine-learning-based tool was developed to predict the likelihood of PPD for
mental health assessment. The backend of this tool incorporates pre-trained models and
associated scaling and encoding mechanisms necessary for preprocessing the user input. The
front end, shown in Figure 32 and Figure 33, allows users to input values for a reduced set of
critical features identified during feature selection, while non-critical features are automatically
assigned default values based on insights from post-training analysis.

Once the user submits their input, the backend processes the data, applying the necessary
transformations based on the trained models, and predicts the probability of PPD using multiple
machine learning models. An ensemble model aggregates predictions from the individual
models, with weights assigned based on the F1 scores of each model to ensure accurate and
reliable results. The final output, expressed as a percentage likelihood of PPD, is sent back to
the front end for user interpretation. For instance, a 74.90 % likelihood of PPD was identified
for the sample inputs (Figure 34). This result is consistent with the patient’s disease history
and related factors observed in the training dataset. This personalized assessment of PPD risk
will aid healthcare professionals in assessing and predicting PPD risks more effectively,
showcasing the practical use of ML techniques in improving PPD care.

102

Figure 32. PPD Screening Tool - Front End

Figure 33. PPD Screening Tool - Front End

103

Figure 34. PPD Screening Tool - Prediction

4.9 Conclusion
The ML models exhibited a range of performance scores, from NN achieving 77% accuracy
and 76% F1 score to LR reaching 89% accuracy and 88% F1 score. Overall, the results
underscore differences in how each model captures and interprets the underlying patterns in
the data. LR's optimized parameters would have enabled effective classification of the target,
balancing precision, and recall. Additionally, RF's robust ensemble learning and SVM's
effective margin optimization with the RBF kernel likely contributed to their performance
scores. The NN might have faced challenges in capturing complex data patterns.

In conclusion, all models demonstrate a close alignment in their SHAP analysis in identifying
key predictors. The SHAP summary plots reveal similar positive and negative SHAP values
for multiple key features across the models. Disease features such as CHRONIC, asthma,
104

anemia, and a history of depression were consistently identified as key predictors across all
models, with comparable SHAP values. Demographic features such as younger maternal age,
lower education, race, infant age, and being married were found to be strong predictors. In
addition, variables such as depression during pregnancy and higher stress levels contributed to
the prediction of PPD.

When identifying the key features, CHRONIC is consistently presented among the top four
predictors across all models, supporting the research hypothesis that CHRONIC is a significant
predictor. Furthermore, the positive SHAP values across all models confirm that the presence
of CHRONIC is a vital contributor to PPD.

Integrating interaction features offered valuable insights into influential factors of PPD. When
considered together, the SHAP analysis highlights that certain demographic, health, and
situational factors significantly predict PPD risk factors. A few interaction features appeared
as the key predictors across all models, while a few other features were distinct. These results
indicate that chronic disease, paternal and maternal education, maternal age group, high BP,
gestational diabetes, ‘No medical risk factors,’ and number of perinatal care visits are key
predictors. This consistency highlights their significance in predicting PPD.

Chronic disease was once again identified as an important predictor in three models (NN, LR,
and RF), frequently interacting with other features and highlighting its vital role in predicting
PPD. Identifying independent chronic diseases, asthma, and anemia are repeatedly found as
key predictors, along with other features in the analysis using interaction features, reinforcing
the hypothesis. Thus, results show that chronic diseases play a significant role in PPD risk.
105

Chapter 5
Conclusion and Future Work
PPD is a significant global concern that affects both new mothers and their infants. Despite its
prevalence, awareness and understanding of PPD remain limited, leading to many cases being
undiagnosed and untreated. When left unaddressed, this condition can escalate into severe
psychosis over time. Early detection is crucial, as untreated PPD can lead to self-harm.
Recognizing key risk factors allows healthcare professionals to identify women who are at a
higher risk of developing PPD. This research presents a comprehensive process designed to
prepare, analyze, visualize, and calculate PPD probabilities to identify PPD risk factors,
specifically when chronic diseases are present.

Five years of data (over 200,000 records) obtained from CDC_PRAMS was used to
demonstrate the predictive power of ML models. The research methodology started with
implementing various preprocessing techniques such as encoding, scaling, and balancing data
using SMOTE on the raw data. Subsequently, the most relevant variables, such as asthma,
heart disease, diabetes, and thyroid dysfunction, were selected to address the hypothesis. Power
BI was utilized to visualize the prepared data along various dimensions. The dashboard
highlights a higher prevalence of postpartum depression among young mothers aged 25-29 and
those with higher education levels (13-15 years of education). Key Influencer functionality on
the dashboard highlights factors such as history of depression, being married, paternal race,
and having a chronic disease that increases the likelihood of PPD. These insights offer a
valuable understanding of the population, feature relationships, and underlying patterns.
106

Four predictive models, NN, LR, RF, and SVM, were implemented to predict the risk factors
of PPD. The evaluation was based on accuracy and F1 scores, ensuring the model predicts
accurately and handles class distributions appropriately. All models achieved over 75%
accuracy and F1 scores, demonstrating satisfactory predictive performance across the test
dataset. The metrics revealed that accuracy and F1 scores were closely aligned, indicating that
the model performs well across both majority and minority classes. The predicted risk factors
are mostly consistent across all models, and the feature CHRONIC was highlighted as a
significant predictor, thereby supporting the research hypothesis. Also, key predictors such as
depression (during and before pregnancy), higher stress, and maternal age are consistent with
other findings [20], [24], [42], [60], [64]. Additionally, the SHAP analysis provided an indepth understanding of each contributing factor of PPD. It further indicated the potential for
feature interactions, particularly with those that showed a low SHAP value. The interaction
analysis revealed factors such as high blood pressure, gestational diabetes, maternal age group,
number of prenatal care visits, chronic disease, and maternal and paternal education interacting
with other predictors frequently.

A web-based tool utilized the developed model as the backend for generating individualized
predictions, allowing users to input relevant data and calculate the probability of PPD. This
tool provides proof of concept for PPD screening in real-world scenarios.

In summary, the objective of this thesis was to identify the risk factors, particularly focusing
on chronic diseases. These findings aim to support timely identification and intervention,
reduce the risk of prolonged mental health challenges for mothers, and enhance their overall
well-being.
107

5.1 Future Work
Although the PRAMS dataset provides valuable insights, its state-specific focus may limit the
applicability of the model to more diverse populations. Also, a deeper investigation of
interaction features, and incorporation of additional disease variables and clinical features is
recommended to enhance the understanding of predictor impacts. Developing interaction
features and including them in ML models has shown the importance of how input features
interact. A detailed analysis of these interactions could help optimize the models to improve
their performance and make them more valuable as a backend for the PPD prediction tool,
providing more reliable screening results.

Also, the use of a broader dataset could improve the evaluation of predictive models and PPD
calculation tool. Furthermore, the study can apply a longitudinal approach to identify how risk
factors and symptoms change over time. Time-series data from longitudinal studies could
capture these trends and make more accurate predictions.

108

References

[1]

M. Javaid, A. Haleem, R. P. Singh, R. Suman, and S. Rab, “Significance of machine learning
in healthcare: Features, pillars and applications,” Int. J. Intell. Netw, vol. 3, pp. 58–73, 2022,
doi: 10.1016/j.ijin.2022.05.002.

[2]

N. L. Rane, M. Paramesha, S. P. Choudhary, and J. Rane, “Machine learning and deep
learning for big data analytics: A review of methods and applications,” Partners Univ. Int.
Innov. J., vol. 2, no. 3, pp. 172–197, Jun. 2024, doi: 10.5281/zenodo.12271006.

[3]

H. Byeon, “Advances in machine learning and explainable artificial intelligence for depression
prediction,” Int. J. Adv. Comput. Sci. Appl., vol. 14, no. 6, 2023, doi:
10.14569/IJACSA.2023.0140656.

[4]

N. K. Iyortsuun, S. H. Kim, M. Jhon, H. J. Yang, and S. Pant, “A review of machine learning
and deep learning approaches to mental health diagnosis,” Feb. 01, 2023, MDPI. doi:
10.3390/healthcare11030285.

[5]

The Business Research Company, “Mental Health Apps Global Market Report.” [Online].
Available: https://www.thebusinessresearchcompany.com/report/mental-health-apps-globalmarketreport#:~:text=Mental%20Health%20Apps%20Market%20Size,(CAGR)%20of%2015.4%25.
[Accessed: Nov. 09, 2024].

[6]

World Health Organization, “Launch of the WHO guide for integration of perinatal mental
health in maternal and child health services.” [Online]. Available:
https://www.who.int/news/item/19-09-2022-launch-of-the-who-guide-for-integration-ofperinatal-mental-health. [Accessed: Mar. 20, 2024].

[7]

J. F. Raisa, M. S. Kaiser, and M. Mahmud, “A machine learning approach to early detection of
postpartum depression in Bangladesh,” in Lect. Notes Comput. Sci. (including subser. Lect.
Notes Artif. Intell. Lect. Notes Bioinform.,) Springer Science and Business Media Deutschland
GmbH, 2022, pp. 241–252. doi: 10.1007/978-3-031-15037-1_20.

[8]

A. Paul, S. D. Pragada, D. N. Murthy, M. L. J. Shruthi, and S. Gurugopinath, “Performance
comparison of machine learning techniques for early detection of postpartum depression using
the PRAMS dataset,” in 2023 IEEE 15th Int. Conf. Comput. Intell. Commun. Netw. (CICN),
2023, pp. 310–315. doi: 10.1109/CICN59264.2023.10402321.

[9]

M. Zhong, H. Zhang, C. Yu, J. Jiang, and X. Duan, “Application of machine learning in
predicting the risk of postpartum depression: A systematic review,” Dec. 01, 2022, Elsevier
B.V. doi: 10.1016/j.jad.2022.08.070.

[10]

S. Graham et al., “Artificial intelligence for mental health and mental illnesses: An
overview.Nov. 01, 2019, Current Medicine Group LLC 1, doi: 10.1007/s11920-019-1094-0.

[11]

Wikipedia, “Postpartum depression.” [Online]. Available:
https://en.wikipedia.org/wiki/Postpartum_depression. [Accessed: Aug. 23, 2022].

109

[12]

M. Lampariello, “Gaps Identified in Postpartum Depression Screening,” Clinical Advisor, Jul.
2020, [Online]. Available: https://www.clinicaladvisor.com/home/topics/mood-disorderinformation-center/gaps-identified-postpartum-depression-screening/. [Accessed: Aug. 04,
2022].

[13]

S. E. Jacob, “Postpartum-Depression,” Simhaee OBGYN & Cosmetic. [Online]. Available:
https://simhaeeobgyn.com/content/postpartum-depression. [Accessed: Sep. 12, 2024].

[14]

“Postpartum Depression,” UBC Wiki. [Online]. Available:
https://wiki.ubc.ca/Postpartum_Depression. [Accessed: Aug. 19, 2022.

[15]

A. Šebela, J. Hanka, and P. Mohr, “Etiology, risk factors, and methods of postpartum
depression prevention,” Ceska Gynekol, vol. 83, no. 6, pp. 468–473, 2018, [Online].
Available: https://pubmed.ncbi.nlm.nih.gov/30848154/. [Accessed: Jan. 11, 2025].

[16]

“Business intelligence: A complete overview,” Tableau. [Online]. Available:
https://www.tableau.com/learn/articles/business-intelligence. [Accessed: Aug. 05, 2024].

[17]

V. Kumar and M. G. L., “Predictive Analytics: A Review of Trends and Techniques,” Int J
Comput Appl, vol. 182, pp. 31–37, Jul. 2018, doi: 10.5120/ijca2018917434.

[18]

Y. Park et al., “Comparison of Methods to Reduce Bias from Clinical Prediction Models of
Postpartum Depression,” JAMA Netw Open, vol. 4, no. 4, p. e213909, Apr. 2021, doi:
10.1001/jamanetworkopen.2021.3909.

[19]

E. Hochman et al., “Development and validation of a machine learning-based postpartum
depression prediction model: A nationwide cohort study,” Depress Anxiety, vol. 38, no. 4, pp.
400–411, Apr. 2021, doi: 10.1002/da.23123.

[20]

D. Shin, K. J. Lee, T. Adeluwa, and J. Hur, “Machine learning-based predictive modeling of
postpartum depression,” J Clin Med, vol. 9, no. 9, pp. 1–14, Sep. 2020, doi:
10.3390/jcm9092899.

[21]

W. Qi et al., “Predictive models for predicting the risk of maternal postpartum depression: A
systematic review and evaluation,” J Affect Disord, vol. 333, pp. 107–120, 2023, doi:
10.1016/j.jad.2023.04.026.

[22]

S. Shen, S. Qi, and H. Luo, “Automatic model for postpartum depression identification using
deep reinforcement learning and the differential evolution algorithm,” Int. J. Adv. Comput. Sci.
Appl. (IJACSA), vol. 14, no. 11, p. 2023, doi: 10.14569/IJACSA.2023.0141115.

[23]

K. Saqib, A. F. Khan, and Z. A. Butt, “Machine Learning Methods for Predicting Postpartum
Depression: Scoping Review”, JMIR Ment. Health., vol. 8, no. 11, p. e29838, doi:
10.2196/29838.

[24]

S. Natarajan, A. Prabhakar, N. Ramanan, A. Bagilone, K. Siek, and K. Connelly, “Boosting
for postpartum depression prediction,” in Proc. IEEE/ACM Int. Conf. Connected Health:
Appl., Syst. Eng. Technol. (CHASE), 2017, pp. 232–240. doi: 10.1109/CHASE.2017.82.

[25]

R. Qasrawi et al., “Machine learning techniques for predicting depression and anxiety in
pregnant and postpartum women during the COVID-19 pandemic: A cross-sectional regional
study,” F1000Res, vol. 11:390, 2022, doi: 10.12688/f1000research.110090.1.

110

[26]

J. L. Cox, J. M. Holden, and R. Sagovsky, “Detection of postnatal depression: Development of
the 10-item Edinburgh postnatal depression scale,” Br. J. Psychiatry, vol. 150, no. 6, pp. 782–
786, 1987, doi: 10.1192/bjp.150.6.782.

[27]

C. T. Beck and R. K. Gable, “Postpartum depression screening scale: Development and
psychometric testing,” Nurs Res, vol. 49, no. 5, pp. 272–282, 2000, [Online]. Available:
https://journals.lww.com/nursingresearchonline/fulltext/2000/09000/postpartum_depression_s
creening_scale__development.6.aspx. [Accessed: Mar. 20, 2023].

[28]

“Postpartum depression screening for new moms,” Baby Center. [Online]. Available:
https://www.babycenter.com/baby/postpartum-health/postpartum-depression-quiz_20000530.
[Accessed: Mar. 28, 2024].

[29]

“Find out if you have Postpartum Depression,” HealthyMom. [Online]. Available:
https://www.healthymom.org/postpartum-depression-test. [Accessed: Oct. 28, 2024].

[30]

N. Ramesh, Y. Ghodsi, and H. Bolhasani, “A survey on the role of artificial intelligence in the
prediction and diagnosis of schizophrenia,” arXiv, May 2023, doi: 10.48550/arXiv.2305.14370.

[31]

“Maternal Mental Health in Canada, 2018/2019,” Statistics Canada, [Online]. Available:
https://www150.statcan.gc.ca/n1/daily-quotidien/190624/dq190624b-eng.htm. [Accessed: Sep.
14, 2022].

[32]

L. E. Ross and C.-L. Dennis, “The prevalence of postpartum depression among women with
substance use, an abuse history, or chronic illness: A systematic review,” J. Womens Health
(Larchmt), vol. 18, no. 4, pp. 475–486, 2009, doi: 10.1089/jwh.2008.0953.

[33]

F. Hamed, E. Supriyanto, S. Osman, and T. A. E. K. Ali, “Risk prediction of major depressive
disorder using artificial neural network,” in Proc. 2020 3rd Int. Sem. Res. Inf. Technol. Intell.
Syst. (ISRITI), Institute of Electrical and Electronics Engineers Inc., Dec. 2020, pp. 68–72. doi:
10.1109/ISRITI51436.2020.9315463.

[34]

R. Chen et al., “3-year risk prediction of coronary heart disease in hypertension patients: A
preliminary study,” in 2017 39th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. (EMBC), 2017,
pp. 1182–1185. doi: 10.1109/EMBC.2017.8037041.

[35]

S. Kerie, M. Menberu, and W. Niguse, “Prevalence and associated factors of postpartum
depression in Southwest, Ethiopia, 2017: A cross-sectional study,” BMC Res Notes, vol. 11,
no. 1, Aug. 2018, doi: 10.1186/s13104-018-3730-x.

[36]

H. K. Brown, A. Qazilbash, N. Rahim, C. L. Dennis, and S. N. Vigod, “Chronic medical
conditions and peripartum mental illness: A systematic review and meta-analysis,” Am J
Epidemiol, vol. 187, no. 9, pp. 2060–2068, Sep. 2018, doi: 10.1093/aje/kwy080.

[37]

“Python 3.11.11 documentation,” Python. [Online]. Available: https://docs.python.org/3.11/.
[Accessed: Jan. 05, 2023].

[38]

“Power BI,” Microsoft. [Online]. Available: https://www.microsoft.com/en-us/powerplatform/products/power-bi/. [Accessed: Dec. 09, 2023].

[39]

“Flask,” Pallets Projects. [Online]. Available: https://flask.palletsprojects.com/en/stable/.
[Accessed: Sep. 05, 2024].

111

[40]

“PRAMS Data,” Centers for Disease Control and Prevention (CDC). [Online]. Available:
https://www.cdc.gov/prams/php/data-research/index.html. [Accessed: Apr. 14, 2022].

[41]

F. Ayhan, A. Akalin, H. Balsak, and A. Erden, “Psychometric testing and the predictive
validity of the postpartum depression predictors inventory-revised (PDPI-R): A longitudinal
study with Turkish women,” Sexual & Reproductive Healthcare, vol. 40, p. 100965, 2024,
doi: 10.1016/j.srhc.2024.100965.

[42]

T. H. K. R. Prabhashwaree and N. M. Wagarachchi, “Towards machine learning approaches
for predicting risk level of postpartum depression,” in 2022 6th SLAAI Int. Conf. Artif. Intell.
(SLAAI-ICAI), 2022, pp. 1–6. doi: 10.1109/SLAAI-ICAI56923.2022.10002477.

[43]

P. Mazumder and S. Baruah, “A community based study for early detection of postpartum
depression using improved data mining techniques,” in 2021 IEEE Int. Conf. Comput. Syst.
Inf. Technol. Sustain. Solut. (CSITSS), 2021, pp. 1–7. doi:
10.1109/CSITSS54238.2021.9682941.

[44]

Y. Zhang, S. Wang, A. Hermann, R. Joly, and J. Pathak, “Development and validation of a
machine learning algorithm for predicting the risk of postpartum depression among pregnant
women,” J Affect Disord, vol. 279, pp. 1–8, 2021, doi: 10.1016/j.jad.2020.09.113.

[45]

S. Wang, J. Pathak, and Y. Zhang, “Using electronic health records and machine learning to
predict postpartum depression,” in Stud. Health Technol. Inform., 2019. doi:
10.3233/SHTI190351.

[46]

“Trends in pregnancy and childbirth complications in the U.S.-Postpartum Depression,”
[Online]. Available: https://blog.bcbsks.com/2020/08/12/trends-in-pregnancy-and-childbirthcompilations-in-the-u-s-postpartum-depression/. [Accessed: Nov. 14, 2022].

[47]

A. Froeliger, C. Deneux-Tharaux, L. Loussert, H. Bouchghoul, H. Madar, and L. Sentilhes,
“Prevalence and risk factors for postpartum depression 2 months after a vaginal delivery: A
prospective multicenter study,” Am J Obstet Gynecol, vol. 230, no. 3, Supplement, pp. S1128S1137.6, 2024, doi: 10.1016/j.ajog.2023.08.026.

[48]

B. Ďuríčeková, Z. Škodová, and M. Bašková, “Risk factors associated with postpartum
depression and PTSD after birth in a sample of Slovak women,” Heliyon, vol. 10, no. 1, 2024,
doi: 10.1016/j.heliyon.2023.e23560.

[49]

K. P. Coca, L. Y. Chien, E. Y. Lee, A. C. de P. Souza, S. A. Hong, and Y. S. Chang, “Factors
associated with postpartum depression symptoms among postpartum women in five countries
during the COVID-19 pandemic: an online cross-sectional study,” BMC Psychiatry, vol. 23,
no. 1, Dec. 2023, doi: 10.1186/s12888-023-04607-0.

[50]

L. F. Wedajo, S. S. Alemu, M. H. Jarso, A. M. Golge, and D. E. Dirirsa, “Late postpartum
depression and associated factors: community-based cross-sectional study,” BMC Womens
Health, vol. 23, no. 1, Dec. 2023, doi: 10.1186/s12905-023-02444-7.

[51]

S. Dehghan-Banadaki, M. Hosseinzadeh, F. Madadizadeh, and H. Mozaffari-Khosravi,
“Empirically derived dietary patterns and postpartum depression symptoms in a large sample
of Iranian women,” BMC Psychiatry, vol. 23, no. 1, Dec. 2023, doi: 10.1186/s12888-02304910-w.

112

[52]

M. Smorti, L. Ponti, and F. Pancetti, “A comprehensive analysis of post-partum depression
risk factors: The role of socio-demographic, individual, relational, and delivery
characteristics,” Front Public Health, vol. 7, Oct. 2019, doi: 10.3389/fpubh.2019.00295.

[53]

G. Palumbo, F. Mirabella, and A. Gigantesco, “Positive screening and risk factors for
postpartum depression,” European Psychiatry, vol. 42, pp. 77–85, May 2017, doi:
10.1016/j.eurpsy.2016.11.009.

[54]

C. J. Shuman et al., “Postpartum depression and associated risk factors during the COVID-19
pandemic,” BMC Res Notes, vol. 15, no. 1, Dec. 2022, doi: 10.1186/s13104-022-05991-8.

[55]

N. Jeon, J. Kent‐Marvick, J. N. Sanders, H. Hanson, and S. E. Simonsen, “Comparing
maternal factors associated with postpartum depression between primiparous adolescents and
adults: A large retrospective cohort study,” Birth, vol. 51, no. 1, pp. 218–228, 2024, doi:
10.1111/birt.12785.

[56]

M. E. Silverman et al., “The risk factors for postpartum depression: A population-based
study,” Depress. Anxiety, vol. 34, no. 2, pp. 178–187, Feb. 2017, doi: 10.1002/da.22597.

[57]

M. Chaponda, N. Aldhouse, M. Kroes, L. Wild, C. Robinson, and A. Smith, “Systematic
review of the prevalence of psychiatric illness and sleep disturbance as co-morbidities of HIV
infection in the UK,” Int J STD AIDS, vol. 29, no. 7, pp. 704–713, 2018, doi:
10.1177/0956462417750708.

[58]

R. D. Björvang, I. Liakea, B. Carpentsier, Z. Kozinszky, A. Skalkidou, and E. Fransson,
“Association of Diabetes Mellitus in Pregnancy and Perinatal Depression,” Psychosom. Med,
vol. 86, no. 1, 2024, doi: 10.1097/PSY.0000000000001261.

[59]

E. Minaldi et al., “Thyroid autoimmunity and risk of post-partum depression: a systematic
review and meta-analysis of longitudinal studies,” J Endocrinol Invest, vol. 43, no. 3, pp. 271–
277, 2020, doi: 10.1007/s40618-019-01120-8.

[60]

M. Ceulemans et al., “Mental health status of pregnant and breastfeeding women during the
COVID-19 pandemic—A multinational cross-sectional study,” Acta Obstet Gynecol Scand,
vol. 100, no. 7, pp. 1219–1229, Jul. 2021, doi: 10.1111/aogs.14092.

[61]

S. Kerie, M. Menberu, and W. Niguse, “Prevalence and associated factors of postpartum
depression in Southwest, Ethiopia, 2017: A cross-sectional study,” BMC Res Notes, vol. 11,
no. 1, Aug. 2018, doi: 10.1186/s13104-018-3730-x.

[62]

S. N. Vigod, L. Villegas, C. L. Dennis, and L. E. Ross, “Prevalence and risk factors for
postpartum depression among women with preterm and low-birth-weight infants: a systematic
review,” BJOG, vol. 117, no. 5, 2010, doi: 10.1111/j.1471-0528.2009.02493.x.

[63]

M. Chaaya, O. M. R. Campbell, F. El Kak, D. Shaar, H. Harb, and A. Kaddour, “Postpartum
depression: Prevalence and determinants in Lebanon,” Arch. Women’s Ment. Health, vol. 5,
no. 2, 2002, doi: 10.1007/s00737-002-0140-8.

[64]

N. Luong, S. Thuma, A. Santore, Z. Ma, S. Watkins, and E. McCarty, “Prevalence and
associated risk factors of postpartum depression among mothers in Pennsylvania, United
States: An analysis of the pregnancy risk assessment monitoring system (PRAMS) data, 20122015,” Int. J. Transl. Med. Res. Public Health, vol. 5, no. 2, 2021, doi: 10.21106/ijtmrph.386.

113

[65]

S. Sundaram, J. S. Harman, and R. L. Cook, “Maternal morbidities and postpartum depression:
An analysis using the 2007 and 2008 pregnancy risk assessment monitoring system,” Women’s
Health Issues, vol. 24, no. 4, 2014, doi: 10.1016/j.whi.2014.05.001.

[66]

M. Qobadi, C. Collier, and L. Zhang, “The effect of stressful life events on postpartum
depression: Findings from the 2009–2011 Mississippi Pregnancy Risk Assessment Monitoring
System,” Matern. Child Health J, vol. 20, pp. 164–172, Nov. 2016, doi: 10.1007/s10995-0162028-7.

[67]

J. J. Gifford, J. R. Pluchino, R. Della Valle, and J. M. Schwarz, “Regional differences in
various risk factors for postpartum depression: applying mixed models to the PRAMS
dataset,” Front Glob Womens Health, vol. 2, 2021, doi: 10.3389/fgwh.2021.726422.

[68]

A. J. Bjertrup, M. S. Væver, and K. W. Miskowiak, “Prediction of postpartum depression with
an online neurocognitive risk screening tool for pregnant women,” European
Neuropsychopharmacology, vol. 73, pp. 36–47, 2023, doi: 10.1016/j.euroneuro.2023.04.014.

[69]

S. McDonald et al., “Development of a prenatal psychosocial screening tool for post-partum
depression and anxiety,” Paediatr Perinat Epidemiol, vol. 26, no. 4, pp. 316–327, Jul. 2012,
doi: 10.1111/j.1365-3016.2012.01286.x.

[70]

A. Opidi, “Solving data challenges in machine learning with automated tools,” TOPBOTS.
[Online]. Available: https://www.topbots.com/data-preparation-for-machine-learning/.
[Accessed: Oct. 31, 2023].

[71]

D. Pyle, “Data preparation for data mining,” Elsevier Science, 1999. [Online]. Available:
https://books.google.ca/books?hl=en&lr=&id=hhdVr9FJfAC&oi=fnd&pg=PR17&dq=Dorian+Pyle,+Data+Preparation+for+Data+Mining.+Elsevier+
Science,+1999.&ots=6icUeRJAat&sig=0UAV6G8kycN6f-nPAMDj8kl018#v=onepage&q=Dorian%20Pyle%2C%20Data%20Preparation%20for%20Data%20
Mining.%20Elsevier%20Science%2C%201999.&f=false. [Accessed: Dec. 12, 2024].

[72]

J. Cai, J. Luo, S. Wang, and S. Yang, “Feature selection in machine learning: A new
perspective,” Neurocomputing, vol. 300, Mar. 2018, doi: 10.1016/j.neucom.2017.11.077.

[73]

“IBM SPSS Modeler,” IBM. [Online]. Available: https://www.ibm.com/products/spssmodeler. [Accessed: Jun. 09, 2022].

[74]

A. Gholamy, V. Kreinovich, and O. Kosheleva, “Why 70/30 or 80/20 relation between
training and testing sets: A pedagogical explanation,” Int. J. Intell. Technol. Appl. Stat, vol. 11,
no. 2, pp. 105–111, 2018, doi: 10.6148/IJITAS.201806_11(2).0003.

[75]

R. Qamar and B. Zardari, “Artificial Neural Networks: An Overview,” Mesopot. J. Comput.
Sci., vol. 2023, pp. 130–139, Aug. 2023, doi: 10.58496/MJCSC/2023/015.

[76]

M. Razavipour, J.-G. Legoux, D. Poirier, B. Guerreiro, J. Giallonardo, and B. Jodoin,
“Artificial Neural Networks Approach for Hardness Prediction of Copper Cold Spray Laser
Heat Treated Coatings,” J. Therm. Spray Technol., Jan. 2022, doi: 10.1007/s11666-02101311-x.

[77]

D. W. Hosmer Jr, S. Lemeshow, and R. X. Sturdivant, “Applied logistic regression,” Appl.
Logist. Regress. John Wiley & Sons, 2013. [Online]. Available:

114

https://books.google.ca/books?id=bRoxQBIZRd4C&dq=Logistic+regression&lr=&source=gb
s_navlinks_s. [Accessed: Dec. 12, 2024].
[78]

“Machine learning internship presentation,” Slideshare. [Online]. Available:
https://www.slideshare.net/slideshow/machine-learning-internship-presentation/258122105.
[Accessed: Oct. 10, 2024].

[79]

G. Biau and E. Scornet, “A random forest guided tour,” TEST, vol. 25, no. 2, pp. 197–227,
2016, doi: 10.1007/s11749-016-0481-7.

[80]

A. Brital, “Random forest algorithm explained,” Blog. [Online]. Available:
https://anasbrital98.github.io/blog/2021/Random-Forest/. [Accessed: Oct. 09, 2024].

[81]

C. Cortes and V. Vapnik, “Support-vector networks,” Mach Learn, vol. 20, no. 3, pp. 273–
297, 1995, doi: 10.1007/BF00994018.

[82]

DataFlair Team, “Real-life applications of SVM (Support Vector Machines),” DataFlair.
[Online]. Available: https://data-flair.training/blogs/applications-of-svm/. [Accessed: Jul. 06,
2024].

[83]

L. T. Becker and E. M. Gould, “Microsoft Power BI: Extending Excel to manipulate, analyze,
and visualize diverse data,” Serials Review, vol. 45, no. 3, pp. 184–188, Jul. 2019, doi:
10.1080/00987913.2019.1644891.

[84]

S. R. Priest, M.-P. Austin, B. B. Barnett, and A. Buist, “A psychosocial risk assessment model
(PRAM) for use with pregnant and postpartum women in primary care settings,” Arch.
Women’s Ment. Health, vol. 11, no. 5, pp. 307–317, 2008, doi: 10.1007/s00737-008-0028-3.

[85]

L. A. Jeni, J. F. Cohn, and F. De La Torre, “Facing imbalanced data–recommendations for the
use of performance metrics,” in 2013 Humaine Assoc. Conf. Affect. Comput. Intell. Interact.,
2013, pp. 245–251. doi: 10.1109/ACII.2013.47.

[86]

L. S. Shapley, “Notes on the N-Person Game — II: The Value of an N-Person Game,” 1951.
[Online]. Available: https://policycommons.net/artifacts/4837582/notes-on-the-n-persongame-ii/5674259/. [Accessed: Mar. 13, 2024].

[87]

J. R. Bloch, K. Dawley, and P. D. Suplee, “Application of the Kessner and Kotelchuck
Prenatal Care Adequacy Indices in a Preterm Birth Population,” Public Health Nurs., vol. 26,
no. 5, pp. 449–459, Sep. 2009, doi: 10.1111/j.1525-1446.2009.00803.x.

115