Using ECG Machine Learning for Detection of Cardiovascular Disease in African American Men and Women: the Jackson Heart Study

Background: Almost half of African American (AA) men and women have cardiovascular disease (CVD). Detection of prevalent CVD in barbershops would facilitate secondary prevention of CVD. We sought to investigate the cross-sectional association of prevalent CVD and sex with global electrical heterogeneity (GEH) and develop a tool for CVD detection. Methods: Participants from the Jackson Heart Study (JHS) with analyzable ECGs (n=3,679; age, 62 {+/-} 12 years; 36% men) were included. QRS, T, and spatial ventricular gradient (SVG) vectors magnitude and direction, and traditional metrics were measured on 12-lead ECG. Linear regression and mixed linear models with random intercept were adjusted for cardiovascular risk factors, sociodemographic and anthropometric characteristics, type of median beat, and mean RR intervals. Random forests, convolutional neural network, and lasso models were developed in 80%, and validated in 20% samples. Results: In fully adjusted models, women had a smaller spatial QRS-T angle (-12.2(-19.4 to-5.1) {degrees} ; P=0.001), SAI QRST (-29.8(-39.3 to -20.3) mV*ms; P<0.0001), and SVG elevation (-4.5(-7.5 to -1.4) {degrees} ; P=0.004) than men, but larger SVG azimuth (+16.2(10.5-21.9) {degrees} ; P<0.0001), with a significant random effect between families (+20.8(8.2-33.5) {degrees} ; P=0.001). SAI QRST was larger in women with CVD as compared to CVD-free women or men (+15.1(3.8-26.4) mV*ms; P=0.009). Men with CVD had smaller T area [by 5.1 (95%CI 1.2-9.0) mV*ms] than CVD-free men, but there were no differences when comparing women with CVD to CVD-free women. Machine-learning detected CVD with ROC AUC 0.69-0.74; plug-in-based model included only age and QRS-T angle. Conclusions: GEH varies by sex. Sex modifies an association of GEH with CVD. Automated CVD detection is feasible.