Research Spotlight: Leveraging AI to Detect ‘Hidden Signals’ in Electrocardiograms

Doctors record your heart’s electrical activity with a simple test called an electrocardiogram, or ECG. In this study, our team wanted to explore a surprising question: can artificial intelligence (AI) “guess” whether someone identifies as Black or White just from that heart tracing, and what might that mean for health equity? We also wanted to determine the origin of these potential differences, which may have profound implications for how we diagnose and treat cardiovascular disease (CVD).

We trained computer models on a very large trove of ECGs collected at Massachusetts hospitals over decades, carefully filtering to only include people without heart or brain disease and balancing the data for fairness during training.

We found that everyday medical data carries “hidden signals” linked to social realities. If left unchecked, those signals could reinforce inequities — for example, if an AI tool behaves differently across communities. While broader studies are needed, the takeaway is clear: to build fair medical AI, we must recognize and correct for the social fingerprints that our bodies, and our data, inevitably bear.

Q: What methods or approach did you use?

We obtained results from about 10 million standard ECG tests from a clinical database, corresponding to about 1.76 million people. This data was collected at various Mass General Brigham sites between 1971 and 2021 and included patients living across Massachusetts. We then developed, optimized and tested an AI model called a convolutional neural network-based classifier to analyze the heart traces.

Q: What did you find?

Overall, we discovered that deep learning models can accurately differentiate between self-declared Black and White healthy individuals from ECG signals alone. The best-performing model we developed was able to correctly identify whether an individual was Black or White from their ECG more than 86% of the time, performing consistently across sexes.

We also identified that the portion of an ECG signal that represents when the heart’s ventricles contract to pump out blood, known as the QRS complex, played a significant role in race classification.

Interestingly, the physiological patterns the model focused on to predict race looked similar in the ECGs of Blacks and Whites at birth but changed as people grew older. This points to the idea that social and environmental factors, as opposed to genetics, may influence self-declared racial signatures on ECGs. Indeed, we observed that high-income groups have better classifier performance than low-income groups due to misclassification of Whites as Blacks in the low-income group. This indicates that the performance of the classifier is susceptible to the demographics of various socio-economic subgroups, which emphasizes the significance of addressing socio-economic disparities when developing fair AI models.

Q: What are the implications?

While there is a lack of clinical data regarding racial groups, there is an increasing belief that epigenetic factors, including social determinants of health and access to health care, play a major role in the manifestation of disease in genotype-negative individuals.

Removal of race from risk prediction, and inclusion of a measure of place-based social disadvantage supports a more equitable approach to CVD prevention, indicating that absolute risk assessment for total CVD supports a more comprehensive clinician-patient risk communication and preventive decision-making process, offering support for a holistic approach to screening, risk assessment, and prevention of CVD among patients.

The strong capacity of models to recognize self-declared race in ECG signals could lead to patient harm. In other words, AI models could potentially not only predict a patient’s race from ECG signals but make use of this capability to produce different health outcomes for members of different racial groups.

Q: What are the next steps?

Though much progress has been made in understanding race-dependent disparities in CVD, significant policy improvements are needed to promote equitable care and improve patient outcomes. Adopting the revised American Heart Association policy of employing a “social deprivation index” to address social determinants of health — reflecting socioeconomic status and other factors affecting health, instead of race — will likely result in improved outcomes for patients with CVD.