Aims: Early detection of cardiovascular risk in the general population is still a challenge. The electrocardiogram (ECG) is a cost-effective screening test for large populations. We hypothesize that individuals with subclinical cardiovascular disease (CVD) might share similar ECG features that can be used for non-invasive risk stratification. This study aimed to identify distinct clusters of individuals without prevalent CVDs based on their ECG morphology using unsupervised learning, and inves-tigate their association with incident CVD risk. Methods: A median heartbeat was derived from 15-seconds resting ECGs (leads I, II and the 6 precordial leads) from 61,489 individuals without prevalent CVDs in the UK Biobank study. ECG parameters relat-ed to time intervals, amplitudes, and morphology were calculated for each lead's median heartbeat. Using k-means clustering, individuals were classified into k distinct clusters. Survival analysis assessed the association of each cluster with incident atrial fibrillation, heart failure (HF), myocardial infarction, major adverse cardiovascular events (MACE) and ventricular arrhythmias risk (9-years follow-up). Results: The model distinguished 3 clusters with varying morphologi-cal features, which significantly differed in terms of HF and MACE events rate (Table 1). Cluster 2 (N = 2,562) included the highest rate of HF (1.5%, p<0.001) and MACE events (2.4%, p<0.001). These individu-als exhibited lower QRS amplitudes (in leads II, V4 and V6), wider QRS duration, longer QTc and Tpec intervals, greater T wave morphological variations with respect to a normal reference and major variability in ST segment deviation than clusters 1 and 3.
Conclusions: Our analysis has identified in an unsupervised manner a group of individuals at risk of HF and MACE using 15-seconds ECGs enabling fast and noninvasive risk assessment in large populations.