nguivo/ML-multimodal-heart-disease-prediction-africa

A machine learning-based multimodal heart disease prediction framework I built as part of MSc. Eng desertation. It combines ECG data and clinical data to predict cardiovascular disease, with Africa as a case study. # Multimodal Heart Disease Prediction — African Context MSc thesis pipeline: machine learning-based multimodal data integration (ECG + clinical/demographic data) for heart disease prediction, with a focus on the African epidemiological context. **Stack:** Python · scikit-learn · XGBoost · PyTorch · NeuroKit2 · SDV (CTGAN) · SHAP ## Overview Cardiovascular disease prediction models are almost always trained on datasets collected in Europe and North America, yet the epidemiological profile of heart disease in Sub-Saharan Africa differs substantially (hypertensive and rheumatic heart disease are disproportionately common, and patients tend to present younger). This project asks what happens to a model trained on Western data when it meets that different reality, and whether training on Africa-representative data closes the gap. Since no freely available African ECG dataset exists (at the time of this study), the African arm of this study is a simulation: African-context ECG and clinical records are synthesized from a European baseline (PTB-XL) using class-conditional feature shifts and CTGAN, grounded in published African cardiovascular registries (THESUS-HF, INTERHEART Africa, REMEDY, Ataklte et al. 2015). The full pipeline (feature extraction, synthesis, and three candidate models) runs on CPU, since GPU access wasn't available for this project. ## Three-stage experimental design | Stage | Trained on | Evaluated on | Question it answers | |---|---|---|---| | 1. The European baseline | PTB-XL ECG (5-fold CV) | PTB-XL ECG | Does the model work at all on real clinical ECG data? | | 2. Domain gap | *(no retraining, reuses Stage 1's XGBoost)* | Synthetic African ECG | How much does performance drop when a European-trained model meets an African-context population? | | 3. African multimodal | Synthetic African ECG + clinical (5-fold CV) | Synthetic African ECG + clinical | Does training on African-context, multimodal data recover the lost performance? | ## Resul …