Abstract
AIMS: Dilated cardiomyopathy (DCM) has a highly variable presentation and disease course. Current stratification strategies are complex and require multimodality evaluation. Using machine learning (ML) on a large dataset obtained at first cardiological evaluation, this study aims to identify specific DCM subgroups. METHODS AND RESULTS: In a retrospective cohort of DCM patients, baseline clinical, genetic, and outcome data were collected. Unsupervised clustering was performed and then simplified to identify patient subgroups. The subgroups were characterized based on outcomes, including all-cause mortality/heart transplantation (HT)/left ventricular assist device implantation (LVAD), sudden cardiac death/major ventricular arrhythmias (SCD/MVA) and heart failure-related death/HT/LVAD. These findings were then validated in an external population. In the derivation cohort of 409 patients (mean age 46 ± 14 years, 71% male), two cluster-subgroups were identified: CL1 (82%) and CL2 (18%), mainly differentiated by electrocardiogram (ECG) characteristics. A lower yield of pathogenic/likely pathogenic variants was found in CL2 versus CL1 (15% vs. 47%, p < 0.001). A simplified clustering using only three variables (QRS duration, presence of left bundle branch block, intrinsicoid deflection >50 ms) was equally effective and validated in the external cohort of 160 patients (mean age 54 ± 13 years, 68% male). A lower risk for SCD/MVA events was observed for CL2 in the primary (hazard ratio 0.29, 95% confidence interval 0.13-0.67) and validation cohort (p = 0.017). CONCLUSIONS: Using ML, baseline ECG variables were found to effectively identify two DCM subgroups differing in disease progression and genetic background. This approach could serve as a valuable tool for improving risk stratification of DCM patients upon their initial evaluation.