ECG Deep Learning Models' Disparities in Predicting Heart Failure

​Electrocardiogram (ECG) deep learning models have proven to be capable of predicting heart failure in patients. The technology offers a potential solution for early intervention and improved patient outcomes. However, studies have shown disparities in the accuracy of these models based on the backgrounds of the patients. As this deep learning model machine becomes integrated into clinical practice, it is vital to address these disparities, which are the result of differences in patients’ age, sex, and race.

The primary model declined with increased patient age and performed significantly worse in Black patients from birth to 40 years of age when compared to other racial groups, with it being most pronounced among young Black women.

Despite efforts to address disparities in model performance by integrating race, sex, and age, the disparities persisted. However, using a customized probability threshold tailored to an individual's race, sex, and age led to an enhanced F1 score.

According to a report on the Stanford Medicine website, Dr. Sanjiv Narayan and other medical practitioners concluded that the results of this study highlight the risk of increasing existing health disparities through the development of these models for heart failure prognosis. To minimize this risk, practitioners can explore the use of personalized probability thresholds based on individual patient factors, leading to more inclusive and efficient healthcare solutions.

Sanjiv Narayan

Shop