Specificity of artificial intelligence-enhanced electrocardiography for diagnosis and prediction of cardiovascular disorders

Background Artificial intelligence (AI) applications for electrocardiograms (ECGs) have been proposed for the detection and prediction of a range of specific structural and functional cardiac abnormalities. To better define the clinical utility as diagnostic and predictive tools, we sought to explore the specificity of the cross-sectional and longitudinal phenotypic associations of several AI-ECG tools. Purpose To systematically evaluate the cross-sectional and longitudinal phenotypic associations of 6 AI-ECG models across a US-based tertiary care hospital, 4 community hospitals, an outpatient medical network, and the UK Biobank. Methods We deployed 6 AI-ECG models on ECG images, including five validated models for the detection of left ventricular systolic dysfunction (LVSD), aortic stenosis (AS), mitral regurgitation (MR), left ventricular hypertrophy (LVH), a composite model for structural heart disease (SHD), and a negative control AI-ECG model for biological sex. Diagnosis codes from the electronic health records were transformed into phenotype codes and a phenome-wide association study (PheWAS) was conducted. We assessed the association of AI-ECG-probabilities with clinical phenotypes, (i) cross-sectionally using age/sex-adjusted logistic regression, and (ii) longitudinally for new-onset CV diseases in age/sex-adjusted Cox regression. Results The study included 265,187 individuals (mean age 59±18 years, 146,090 [55%] women) across sites, with one random ECG per person. Each of the 5 AI-ECG models had differentially stronger association with cardiovascular phenotypes compared with other phenotype groups, which was not observed for the AI-ECG model for sex, which was most strongly associated with non-CV phenotypes (Figure 1). Each of the AI-ECG models was significantly associated with their target phenotype, but they also exhibited similar or stronger associations with a broad range of other cardiovascular phenotypes. For instance, the AI-ECG model for AS was more strongly associated with heart failure NOS (OR 3.2, p <10⁻³⁰⁰) than with aortic valve disease (OR 2.7, p <10⁻²⁵⁹). Each of the models had similar strong cross-phenotype associations (Figure 2A). For predicting future disease, AI-ECG models had strong non-specific associations with a broad range of CV phenotypes, spanning both intended and related phenotypes (Figure 2B). These findings were consistent across models and cohorts. Conclusion Despite AI-ECG being developed to detect specific cardiovascular conditions, they are non-specific and detect a range of CV abnormalities and predict the occurrence of a range of adverse CV outcomes. These findings suggest that several AI-ECG models best serve as general biomarkers of CV health rather than dichotomous diagnostic or predictive tools.