Electrocardiographic sex index: a continuous representation of sex.

Abstract

Clinical risk calculators consider sex as a binary variable. However, sex is a complex trait with anatomic, physiologic, and metabolic attributes that are not easily summarized in this manner [1]. We propose a continuous representation of sex, the ECG Sex Index (ESI), derived via artificial intelligence analyses of electrocardiograms (ECG-AI).We used an ECG repository at Wake Forest Baptist Health (Winston-Salem, NC) to develop a convolutional neural network-based ECG-AI model to detect sex from standard 12-lead ECGs. We utilized a rank-ordered transformation of the outcomes of ECG-AI to create the ESI. We also created a sex discordance index (SDI) from the ESI and assessed its utility in 1-year risk prediction for all-cause mortality, heart failure, and kidney failure.The Wake Forest cohort included 3,573,844 ECGs and electronic health record data from 754,761 patients; 75% were White, 17% were Black, and 51% were female, with a mean age (SD) of 61 (17) years. The PhysioNet external validation cohort included 45,152 ECGs from 10,646 patients from two hospitals in China. The PhysioNet cohort was 100% Asian, 43.6% female, and had a mean age (SD) of 59 (20) years. ECG-AI provided a holdout area under the curve of 0.95 and an external validation area under the curve of 0.92. Lower ESI scores in males and higher ESI scores in females were associated with a greater risk for clinical outcomes. The ESI and SDI demonstrated comparable accuracy to binary sex in logistic regression analyses and outperformed binary sex in predicting clinical outcomes, highlighting their value as predictors in risk calculators for all-cause mortality, heart failure, and kidney failure.