Combining deep neural network feature extraction from 1,553 echocardiographic video clips with over 100 clinical variables from 156 HCM patients, HCM-PhenotypeNet used UMAP dimensionality reduction and Deep Embedded Clustering to resolve four distinct phenogroups: an Advanced Obstructive/Metabolic subtype, an Early-onset Genotype-Positive subtype, a Late-onset Mild Non-obstructive subtype, and a Hypertrophic Heart Failure-predominant subtype. Cluster separation was strong, achieving a silhouette score of approximately 0.89, though no mortality differences were detected across groups.
HCM affects roughly 1 in 500 adults and remains notoriously difficult to risk-stratify because its morphological and symptomatic heterogeneity resists simple classification. Current guidelines rely on discrete thresholds — septal wall thickness, outflow tract gradient, genetic status — that miss the condition's continuum of severity. Multimodal AI phenotyping, as attempted here, represents a logical evolution toward data-driven subtyping that could eventually personalize surveillance intervals, device implantation decisions, and emerging myosin inhibitor therapy (mavacamten, aficamten). The trend toward higher heart failure and atrial fibrillation burden in the obstructive and heart failure-predominant clusters is clinically plausible and internally consistent.
However, the cohort of only 156 patients is a critical limitation — clusters derived from small samples are statistically fragile and prone to overfitting. The absence of a mortality signal likely reflects underpowering. As a preprint not yet peer-reviewed, these findings should be treated as hypothesis-generating. External validation in multi-center HCM registries is essential before clinical adoption.