Machine Learning for Predicting Postpartum Depression and Suicidal Tendencies in Bangladesh

Abstract

Postpartum depression (PPD) has become a significant mental health disorder affecting mothers across the world, with prevalence rates of 6% in high-income countries (HICs) and 20% in low middle-income countries (LMICs) leading to serious consequences for maternal and neonatal health. In countries like Bangladesh, PPD remains a critical issue, contributing to maternal suicides due to the overwhelming psychological burden. The condition often goes undiagnosed in such regions due to cultural stigma surrounding mental health and a lack of education, preventing mothers from seeking help. This thesis aims to develop a machine learning model using symptom and behavioral data, to detect suicidal ideation in women with PPD, reducing diagnostic delay for early intervention and improving access to care by recommending mental health resources. Additionally, the study investigates the influence of various feature selection methods, namely, expert-selected features, Sequential Feature Selection (SFS), and Pearson’s correlation (PCC) as well as balancing techniques such as undersampling and oversampling on model performance, developed using four machine learning algorithms; Decision Tree (DT), Logistic Regression (LR), Random Forest (RF) and Support Vector Machine (SVM). Random Forest, trained using the original unsampled dataset and expert-selected features, gave the best performance across other models when using various feature selection techniques and sampling approaches. Decision Trees are also reasonably accurate while Logistic Regression and Support Vector Machines, on the other hand, are less suitable for this prediction. While expert-selected features give the best outcome, it is also identified that PCC selected features are most beneficial for LR and SVM while tree-based classifiers are benefited more with features selected by SFS. Comparisons with existing research reveal that the developed models achieve good performance, with AUC values ranging from 0.86 to 0.98 for Random Forest and Decision Tree, outperforming the PPD and suicide prediction models from existing research.