DH-452784-1 UNSUPERVISED MACHINE LEARNING IDENTIFIES PROGNOSTICALLY SIGNIFICANT PHENOGROUPS FROM NEURAL NETWORK-DERIVED ECG FEATURES

Abstract

Artificial intelligence (AI)-ECG models trained for specific classification tasks learn to recognise many subtle ECG features that are not related to conventional ECG parameters or abnormalities. These novel neural-network (NN)-derived ECG features have the potential to be transferable and applicable to a range of clinically useful tasks. We aimed to use NN-derived ECG features derived from a simple diagnostic classification model to identify prognostically significant clinical phenotypes. We extracted NN-derived ECG features from an AI-ECG model trained for six rhythm/conduction diagnoses. We applied unsupervised machine learning (K-means clustering) to these features. We evaluated the clinical significance of the phenogroups using survival analysis including, uniquely, 3 external cohorts, across 2 continents. The European cohorts had very few subjects in Phenogroup C, this group was therefore removed from survival analysis in those cohorts. A phenome wide association study (PheWAS) and a genome wide association study (GWAS) were performed. In the derivation cohort (n = 1558421), 3 phenogroups were identified. As shown in the Figure, phenogroup B had a 2.57-fold higher risk of mortality compared to phenogroup A, while phenogroup C had a 15% lower risk (phenogroup B vs A, HR 2.57 95% CI 2.51-2.63, p < 0.001; phenogroup C vs A, HR 0.85 95% CI 0.82-0.87, p < 0.001). In external validation in 3 cohorts (UK Biobank n = 42386, Whitehall II = 5066, SaMi-TROP = 1631) across 2 continents covering the spectrum of healthy volunteers to patients with established cardiomyopathy, the predictive value of phenogroup B was retained. PheWAS demonstrated significant correlations in the phenogroups with non-ECG phenotypes such as chamber volume from cardiac MRI and physical measures such as blood pressure and arterial stiffness, while GWAS identified genetic correlates (SCN5A, SCN10A, CAV1) that relate to arrhythmia and conduction disease susceptibility. For the first time, ARHGAP24 was identified as a gene associated with a prognostically significant phenogroup. We describe the use of NN-derived ECG features to identify prognostically significant phenogroups from the 12-lead ECG, and show its broad applicability through external validation using a large and geographically diverse mortality-linked ECG dataset. We also identified the phenotypic and genotypic correlates of these phenogroups. NN-derived ECG features may be used for risk prediction in a wide range of settings.