Artificial Intelligence-Enhanced Electrocardiography for Detection and Prediction of Hypertrophic Cardiomyopathy across Monogenic and Polygenic Susceptibility

Among 1,095 carriers of pathogenic or likely pathogenic sarcomere variants, the validated AI-ECG model detected the HCM phenotype at first clinical assessment (P+) with a pooled AUROC of 0.91 (95% CI 0.89–0.93), and at the prespecified threshold of 0.15 sensitivity was 0.78, specificity 0.89, PPV 0.95, and NPV 0.59. Prior AI-ECG work focused largely on cross-sectional HCM detection and was constrained by low positive predictive value in unselected populations; this study applied the same model (an EfficientNet-B3 convolutional neural network, used without retraining) in a genetically at-risk population and validated it across three independent international centers. Multicenter pooled cohort of 1,095 individuals, with site-specific AUROCs of 0.88, 0.92, and 0.90, a prespecified threshold, DeLong confidence intervals, and consistent results across sex, age, and gene subgroups.

Among genotype-positive/phenotype-negative (G+/P−) individuals at baseline, a higher AI-ECG score predicted development of HCM during follow-up: unadjusted HR 1.55 (95% CI 1.28–1.88; p<0.001) and adjusted HR 1.38 (95% CI 1.11–1.71; p=0.004) after adjustment for age, sex, and site, with a pooled Harrell C-index of 0.73. Individualized risk stratification for G+/P− individuals has been limited, leaving clinicians reliant on repeated imaging; this analysis extends the AI-ECG score from cross-sectional phenotype detection to prediction of incident disease, including an age-based time scale (phenotype-free survival at age 60 of 11.6% versus 61.3%). Analytic cohort of 239 individuals with median follow-up of 6.6 years and 55 events (23.0%); hazard ratios were directionally consistent across sites with no evidence of heterogeneity (I² = 0%, Cochran's Q = 0.19, p = 0.91), though event counts at individual sites were small.

In 57,007 UK Biobank participants, the AI-ECG score and a polygenic risk score (PRS) provided complementary information: HCM prevalence rose from 0.02% (9/41,721) in the both-low group to 1.48% (17/1,148) in the both-high group, and the age- and sex-adjusted odds ratio for HCM was 60.2 (95% CI 26.5–137.2) in the both-high group versus 15.0 for high AI-ECG alone and 4.1 for high PRS alone, with no significant interaction (p=0.69). This analysis is described as the first to demonstrate complementarity between an image-based AI-ECG phenotypic marker and polygenic risk for HCM in the same population, linking monogenic and polygenic susceptibility pathways. Large sample (57,007 participants) but only 51 HCM cases, limiting precision of stratum-specific estimates; PRS was based on the PGS_MTAG reported by Zheng et al., and the analysis was restricted to a genetically homogeneous European-ancestry group.

The findings apply to carriers of pathogenic or likely pathogenic sarcomere variants undergoing cascade screening and long-term imaging surveillance, and to risk stratification in general populations with available ECG and whole-genome sequencing data; the model takes a 300×300-pixel 12-lead ECG image as input and outputs a continuous score between 0 and 1, and can be deployed locally from DICOM or PDF/JPEG files without sharing patient data, which suits multicenter collaboration and low-cost scaling. The authors state explicitly that whether a negative AI-ECG result can safely defer imaging requires prospective evaluation and should not be inferred from this study.

This is a preprint that has not been certified by peer review, and the authors note the results should not be used to guide clinical practice. The relatively small number of incident HCM cases limits statistical power for site-specific and gene-stratified subgroup analyses; index ECG selection differed across sites (Yale used a 90-day window around imaging, whereas Erasmus MC and Motol used the first available ECG), which may explain differences in site-specific discrimination; pathogenic and likely pathogenic variants were not distinguished; and UK Biobank is a volunteer cohort that is predominantly of European ancestry with a small absolute number of HCM cases, limiting precision of stratum-specific estimates. In addition, this reading is based on the full text, and the specific values in figures and supplementary tables were not individually verified, so exact stratum-specific numbers should be confirmed against the original figures and tables.