Machine Learning Model Enabled with Data Optimisation for Prediction of Coronary Heart Disease

Cardiovascular disorders remain leading cause for mortality worldwide, necessitating robust early risk assessment. Although machine learning models show promise, most rely on conventional preprocessing, which lacks model portability across datasets. We propose an integrated preprocessing pipeline enhancing model generalizability. Our methodology standardises features solely based on training statistics and then transforms test data identically to prevent leakage. We handle class imbalance through synchronised oversampling, enabling consistent performance despite distribution shifts. This framework was evaluated on an open-source dataset of clinical parameters from an African cohort using classifiers like support vector machines and gradient boosting. All models achieved upto 80% accuracy. Remarkably, evaluating the identical models on five external European and Asian datasets maintains 80% - 86% accuracy. Our reproducible data conditioning strategy enables precise and transportable heart disease risk prediction, overcoming population variability. The framework provides the flexibility to readily retrain models on new data or update risk algorithms for clinical implementation in diverse locales. Our work accelerates the safe translation of machine learning to guide cardiovascular screening worldwide.