Highlight
– AI-enhanced electrocardiogram (AI-ECG) models detect not only specific cardiac conditions but also broadly associate with numerous cardiovascular phenotypes.
– Phenotypic associations of AI-ECG models are highly correlated across different cardiovascular disease targets, indicating shared detection patterns.
– These models predict new-onset cardiovascular diseases, questioning their current role as binary diagnostic tools and advocating their utility as comprehensive cardiovascular risk biomarkers.
Study Background and Disease Burden
Cardiovascular disease (CVD) remains a leading cause of morbidity and mortality worldwide. Early detection and risk stratification are critical for timely intervention and improved outcomes. Electrocardiography (ECG) is a ubiquitous, non-invasive tool that reflects electrophysiological cardiac activity, commonly used for diagnosing various cardiac abnormalities.
With advances in artificial intelligence (AI), AI-enhanced ECG models have been developed to identify specific anatomical and functional cardiac abnormalities such as left ventricular systolic dysfunction (LVSD), aortic stenosis (AS), mitral regurgitation (MR), and left ventricular hypertrophy (LVH). While these models boast high diagnostic accuracy for their intended targets, their phenotypic selectivity—whether they function strictly as condition-specific classifiers or broader cardiovascular risk markers—remains unclear. Understanding this selectivity is vital because clinical utility depends on precise interpretation of what these AI outputs represent.
Study Design
This large-scale observational study incorporated four distinct populations derived from electronic health records (EHR) and prospective cohort studies, aggregating data from 233,689 individuals (mean age 59±18 years; 56% women). Only one random ECG per individual was analyzed.
Six AI models were deployed, including five validated image-based AI-ECG classifiers trained to detect LVSD, AS, MR, LVH, and a composite structural heart disease (SHD) model. Additionally, a negative control model trained for biological sex prediction was included, alongside six novel models developed for non-cardiovascular conditions.
Diagnosis codes from EHR and cohort databases were transformed into interpretable phenotypes through a phenome-wide association study (PheWAS) framework. Associations between AI-ECG-derived probabilities and cross-sectional phenotypes were evaluated by logistic regression, while prospective associations with incident cardiovascular diseases were assessed using Cox proportional hazards regression. Pearson correlation coefficients of phenotypic signatures across models were calculated to assess similarity.
Key Findings
All five validated AI-ECG models showed strong, statistically significant associations with their respective target phenotypes (e.g., LVSD model with LVSD diagnosis), confirming their diagnostic validity (p<10⁻⁶). More importantly, these models also demonstrated similar or more pronounced associations with a wide array of other cardiovascular phenotypes beyond their original targets.
Odds ratios for association with cardiovascular versus non-cardiovascular phenotypes ranged from 2.16 to 4.41, indicating a preferential detection of cardiovascular pathology. In contrast, the sex prediction model did not exhibit such cardiovascular specificity.
Phenotypic association patterns between AI-ECG models trained on different conditions were remarkably similar, with Pearson correlation coefficients ranging from 0.67 to 0.96. This high concordance was absent in non-cardiovascular models. These findings were consistently reproducible across external datasets and in both cross-sectional and prospective analyses.
Prospective analysis demonstrated that AI-ECG-predicted probabilities significantly anticipated the new onset of diverse cardiovascular conditions, suggesting these models capture underlying cardiovascular risk signals rather than isolated phenotypic footprints.
Expert Commentary
These results challenge the prevailing notion that AI-ECG models function as narrow binary classifiers tailored to specific cardiac diseases. Instead, the high phenotypic selectivity overlap and broad cardiovascular associations suggest these models detect convergent physiological and pathophysiological cardiac alterations present across multiple disease states.
This multifunctional detection capability opens new avenues for AI-ECG use as integrated cardiovascular biomarkers, offering a holistic risk evaluation rather than isolated disease screening. Clinicians should thus interpret positive AI-ECG results in a broader cardiovascular context, prompting comprehensive cardiovascular risk assessment and monitoring.
Nevertheless, limitations include the reliance on diagnosis codes that may imperfectly capture true clinical phenotypes and the varying prevalence of conditions across populations. Prospective validation in diverse clinical settings and exploration of underlying mechanistic signals identified by AI-ECG will further refine clinical application.
Conclusion
Artificial intelligence-enhanced ECG models, although initially designed as specific diagnostic classifiers, demonstrate broad cardiovascular phenotypic associations and predict future cardiovascular disease development. This phenotypic non-selectivity supports a paradigm shift from binary diagnostic tools to versatile cardiovascular biomarkers that can enhance early detection, risk stratification, and potential precision medicine strategies.
The integration of AI-ECG into clinical workflows should leverage their holistic cardiovascular risk insights, supplementing rather than replacing traditional diagnostic frameworks. Future research should focus on mechanistic elucidation of AI-derived features, prospective clinical utility studies, and model refinement to optimize cardiovascular care impact.
References
Croon PM, Dhingra LS, Biswas D, Oikonomou EK, Khera R. Phenotypic Selectivity of Artificial Intelligence-enhanced Electrocardiography in Cardiovascular Diagnosis and Risk Prediction. Circulation. 2025 Sep 1. doi: 10.1161/CIRCULATIONAHA.125.076279. Epub ahead of print. PMID: 40888124.
Attia ZI, Noseworthy PA, Lopez-Jimenez F, et al. An artificial intelligence-enabled ECG algorithm for the identification of patients with atrial fibrillation during sinus rhythm: a retrospective analysis of outcome prediction. Lancet. 2019;394(10201):861-867. doi:10.1016/S0140-6736(19)31721-0
Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019 Jan;25(1):44-56. doi: 10.1038/s41591-018-0300-7.
