HCM-PhenotypeNet: Deep Multimodal Phenotyping of Hypertrophic Cardiomyopathy from Echocardiographic Video and Clinical Data

Abstract Background: Hypertrophic cardiomyopathy (HCM) is a heterogeneous condition with variable morphology and symptoms, leading to adverse outcomes such as atrial fibrillation, heart failure and sudden cardiac death. Data-driven phenotyping may help refine risk stratification and guide management. Objectives: To propose an automated machine learning framework (HCM-PhenotypeNet) that integrates echocardiographic video and clinical data to identify clinically meaningful HCM phenotypes. Methods: We developed a multimodal pipeline using a deep neural network to extract features from 1,553 echocardiography video clips, combined with >100 clinical variables from 156 HCM patients. Patientlevel vectors were reduced using UMAP and clustered using Deep Embedded Clustering (DEC). Cluster quality was assessed using internal validation metrics, and clusters were characterized through statistical comparisons and association rule mining. Results: The cohort included 156 HCM patients (mean age 53.6 +/- 13.9 years, 40.4% female) with preserved systolic function and diverse racial/ethnic backgrounds. The optimal pipeline (DEC with 4 clusters) yielded phenogroups with strong separation (silhouette {approx} 0.89): (A) Advanced Obstructive/Metabolic phenotype; (B) Early-onset, Genotype Positive phenotype; (C) Late-onset, mild non-obstructive phenotype; and (D) Hypertrophic Heart Failure-predominant phenotype. Whilst there was no mortality difference, there was a trend towards increased heart failure and atrial fibrillation burden in the Advanced Obstructive/Metabolic and Heart Failure-predominant subgroups. Conclusions: HCM-PhenotypeNet enabled automated identification of four clinically relevant HCM subgroups. These phenogroups highlight the heterogeneity of HCM and may support personalized risk stratification and management but require validation in larger cohorts. Keywords: Cardiomyopathy Subtypes, Echocardiography, Phenotyping, Video Analysis, Clustering