She Code Africa ML/AI Challenge – Postpartum Depression Prediction

📌 Project Overview

This project was developed for the She Code Africa ML/AI Hackathon 2025, with the challenge theme:
“Postpartum Depression Prediction using Supervised Learning.”

The goal was to build a machine learning model that predicts the Hamilton Depression Rating at 6 months (hamd_6m) using demographic data, medical history, birth complications, and social support scores.

Beyond just accuracy, we aimed to:

• Understand the factors contributing to postpartum depression
• Build a model that is both reliable and explainable
• Document our approach so that others can easily follow and improve on it

📂 Dataset

• Source: Provided by hackathon organizers
• Target variable: hamd_6m
• Features:
  • Demographic details (age, education, employment, etc.)
  • Birth-related data (complications, delivery type)
  • Social support scores
  • Medical history indicators

📑 Dataset schema: View here

🔍 Methodology

1. Data Understanding & EDA

• Inspected missing values → imputed them and created flags where useful.
• Checked for duplicates and inconsistencies.
• Explored distributions → applied log transformation to skewed target variable (hamd_6m).
• Looked for correlations between features and hamd_6m to guide feature selection.

2. Feature Engineering

• Encoded categorical variables (One-Hot Encoding for models).
• Added clinically meaningful features:
  • is_first_pregnancy (based on first_child and kids_no)
  • total_trauma (sum of abortion, child death, stillbirth)
  • Interaction features (age_x_ses, support_x_financial, baselineDep_x_childloss)
  • Binary flags (childloss_flag, abortion_flag)
• Standardized continuous features for linear models.

3. Modeling

We experimented with:

• Baseline Models: OLS (Linear Regression), Lasso → performed poorly.
• Tree-Based Models: Random Forest (best single model), XGBoost.
• Final Choice: Stacking Ensemble (Random Forest + XGBoost) with tuned hyperparameters, which delivered the best performance.

4. Validation Strategy

• Used 5-Fold Cross Validation to ensure robust evaluation.
• Metrics:
  • RMSE (Root Mean Squared Error) – measures how far predictions are from actual.
  • MAE (Mean Absolute Error).
  • R² (explained variance).

---

📊 Results

Model	RMSE	MAE	R²
OLS	~3.69	~2.75	0.60
Lasso	~3.69	~2.76	0.61
Random Forest	~2.76	~1.72	0.78
XGBoost	~2.86	~1.79	0.76
Stacked Model	0.74	0.46	0.97

The final stacked model with engineered features reduced RMSE from 2.76 → 0.74, meaning predictions are within ~1 point of the actual HAMD score - a huge improvement over earlier models.

1. Clone the repository

   git clone <your-repo-link>
   cd sca-ppd-prediction

2. Install requirements

   pip install -r requirements.txt

3. Run training

   python train.ipynb

4. Generate predictions

   python predict.ipynb