Predicting Depression Trajectories Following Cognitive Behavioral Therapy Using Machine Learning

This repository contains code and documentation for a machine learning project aimed at predicting depression trajectories (remission, relapse, and recurrence) following initiation of Cognitive Behavioral Therapy (CBT)-based treatment for perinatal depression, using data derived from a cluster randomized trial (CRT) conducted in Pakistan.

Background

Perinatal depression affects approximately 25% of women in low- and middle-income countries and poses significant risks to maternal and child health. CBT has demonstrated short-term efficacy in reducing depressive symptoms; however, there is uncertainty about its long-term effectiveness, particularly regarding relapse and recurrence. The "Thinking Healthy" intervention, a CBT-based program implemented by minimally trained community health workers in Pakistan, significantly reduced depression at 6 months and 1 year postpartum. However, this effect diminished by the 7-year follow-up.

This project applies advanced machine learning techniques to identify predictors of long-term remission, relapse, and recurrence using longitudinal data from the original trial. The ultimate goal is to inform personalized, scalable strategies for maintaining mental health outcomes.

Project Objectives

Develop and evaluate predictive models using logistic regression, random forest, XGBoost, and neural networks.

Perform nested cross-validation with cluster stratification to robustly estimate model performance.

Utilize SHAP feature importance analysis to identify influential predictors and enhance interpretability.

Create hypothetical patient "phenotype profiles" to enhance clinical utility.

Repository Structure

data/

• THP_clean.dta (original study data)
• dat_ml_6m_imputed.csv
• dat_ml_1y_imputed.csv
• dat_ml_7y_imputed.csv

scripts/

• data_cleaning.Rmd
• modeling.ipynb

README.md

data/: Contains raw and processed datasets.

scripts/: R Markdown file for cleaning and deriving outcomes using the original study data; Jupyter notebook for preprocessing, imputation, modeling, and evaluation.

Usage TBD

Key Results TBD

Original Studies

Baranov V, Bhalotra S, Biroli P, Maselko J. Maternal Depression, Women’s Empowerment, and Parental Investment: Evidence from a Randomized Controlled Trial. American Economic Review. 2020;110(3):824-59.

Rahman A, Malik A, Sikander S, Roberts C, Creed F. Cognitive behaviour therapy-based intervention by community health workers for mothers with depression and their infants in rural Pakistan: a cluster-randomised controlled trial. Lancet. 2008;372(9642):902-9.