AI-Enhanced Handheld ECGs to Screen for Prior Myocardial Infarction in Rural India

Efforts to move care out of hospitals and into the community face a hard constraint: while most medications can be delivered anywhere, diagnosis is still mostly trapped inside the formal health system—clinics, hospitals, and the people and machines inside them. This is especially binding in low- and middle-income countries due to distance, cost, and limited health system capacity. Low-cost data collection devices (e.g., handheld electrocardiograms (ECGs), retinal cameras) are promising, but do not solve the bottleneck: their outputs still require expert interpretation. Artificial intelligence (AI) could turn these raw signals into actionable outputs at scale, but most medical AI is trained in high-income settings and deteriorates in different populations and settings. We assemble a new dataset in rural Tamil Nadu, India that links low-cost device signals to hospital-grade reference testing. We use it to build a proof-of-concept screening tool: an algorithm that predicts the result of a hospital-based reference-standard test (regional wall motion abnormality on echocardiography), using only data from a handheld ECG device (cost: $60). This would allow targeting of secondary prevention medications—high-dose statins, antiplatelet agents, and beta blockers—proven to be cost-effective in both high- and low-resource settings. We evaluate predictive performance and cost-effectiveness of such screening, benchmarking against common clinical risk scores (including the American Heart Association 10-year risk score). In a held-out test set, our model identifies subgroups with rates high enough to make screening very cost-effective: the top 2.5% have a 9.3% rate of evidence consistent with prior heart attack—about six times the base rate—delivering benefits at US$1,999 per life-year saved. The model also flags high-yield patients who would be missed by standard clinical scores, underscoring the limits of transporting Western-derived risk tools to low- and middle-income settings. To enable extensions to other conditions in the dataset, we make all data freely available for research.