Scalable AI for Diagnosis of Silent Heart Attacks in Tamil Nadu, India

AI could dramatically increase access to diagnostic health screenings in low-resource settings by predicting health outcomes using low-cost, mobile tests. This project will develop and evaluate the impact of an AI-based referral tool that predicts risk of silent heart attacks in Tamil Nadu, India.

In prior work with the government partner, the researchers developed a machine learning model that predicts likelihood of silent heart attack using handheld electrocardiogram data. Initial results are promising, with the model identifying high-risk individuals with reasonably high fidelity. This project will support final preparation and implementation of a randomized evaluation. First, they will collect an additional 6,000 data points to improve model precision and gather demographic data needed for fairness audits prior to deployment. Second, they will evaluate the tool in a randomized evaluation embedded within an existing door-to-door healthcare program, where frontline government volunteers use the tool to assess risk during household visits and refer high-risk patients to health facilities for further testing and treatment.

The evaluation will compare the AI-based referral tool to a standard risk score based on demographics and blood pressure. The researchers will measure effects on silent heart attack diagnosis rates and diagnostic test yield. To understand mechanisms, they will decompose effects into (1) an informational effect – shifting who gets referred; and (2) a behavioral effect – changing willingness to follow up upon receiving a referral due to the novelty of the tool. They will also assess cost-effectiveness and conduct subgroup fairness analysis to inform safe, equitable deployment in similar settings.