Developing a Machine Learning Based Automated Screening Tool to Diagnose Silent Heart Attacks in Resource-Constrained Settings

Abstract

Aging populations worldwide pose significant financial and social challenges for low- and middle-income countries, particularly in supporting elderly individuals with chronic illnesses. These challenges are exacerbated by a shortage of high-fidelity diagnostic technology. While point-of-care tests offer a low-cost and mobile solution to limited diagnostic access, they lack the precision of more expensive tests and specialized medical expertise. This thesis develops a supervised machine learning-based diagnostic tool for silent heart attacks using both point-of-care and gold standard data from elders in Tamil Nadu, India, as part of a proposed solution to bridge the diagnostic gap. The research explores whether point-of-care data can be used to reliably identify risk, transforming low-cost inputs into predictions with signal. We also investigate how to operationalize these predictions in a referral pipeline. The results demonstrate that single-lead ECG inputs can effectively detect signals indicative of silent heart attacks. Based on model predictions and a cost-benefit analysis, we suggest risk score thresholds for classifying high-risk individuals for whom it is cost-effective to refer to escalated care. Additionally, integrating single-lead and 12-lead ECG data enhances diagnostic accuracy and supports the development of an operational referral pipeline.