Abstract
Identifying postpartum depression (PPD) early is important, but it may go undetected at times as standard screening depends on mothers attending postpartum visits which many of them may miss. In this project, I built a cross-cultural machine learning workflow that can predict PPD risk. The machine learning models were trained on three datasets, namely Pregnancy Risk Assessment Monitoring System (PRAMS), Colorado, with 8,520 records and a Bangladesh maternal health dataset with 800 records, a merged dataset that had a combination of these two datasets and had 9320 records. The main contribution of this project is a PPD prediction tool powered by a model that is trained on this diversified merged dataset from two different countries. A broader idea of postpartum risk was captured as the model trained on the merged dataset which had both clinical mental health indicators like depression history and postpartum depressive symptoms like socioeconomic factors, pregnancy and delivery information and cultural and family related factors. Six models were trained on all three datasets using class weights and Synthetic Minority Over Sampling Technique (SMOTE) to handle class imbalance, tuned with GridSearchCV and 5-fold cross validation, and evaluated using AUC, recall, precision, and F1. A fairness analysis using the Disparate Impact Ratio (DIR) confirmed that models perform fairly across demographic groups. Models trained on the PRAMS dataset achieved AUC scores in the range 0.90, while Bangladesh models achieved AUC scores in the 0.80 range, and best model was FCNN on the merged dataset (AUC 0.9251). We also developed a web-based monitoring system for both clinicians and patients. For clinicians, the web-based health monitoring system provides interactive visualizations of PPD risk patterns using all 3 datasets. They can look up any single patient by ID and view their full health profile, and use a K-Nearest Neighbors based similar patient lookup feature that finds the five most similar patients from the 9,230 records in the merged dataset. For patients, the web-based health monitoring system provides a six section risk assessment form that collects health and lifestyle information and returns an instant PPD risk prediction from the trained FCNN model, along with personalized recommendations generated by Meta Llama 3 based on the patient's predicted risk level and the key risk factors identified by the trained FCNN model. This work provides a customized early intervention mechanism for PPD detection using ML models trained on culturally diversified datasets.