Applying unsupervised machine learning to 347 non-ischemic dilated cardiomyopathy (NIDCM) patients across two tertiary centers, AI-assisted T1-mapping cardiac MRI combined with 11 multimodal variables identified three clinically distinct phenotypes: a younger, male-predominant preserved cluster with favorable prognosis; a metabolic-fibrotic remodeling cluster; and an atrial fibrillation-predominant biventricular dysfunction cluster carrying the highest composite risk. Crucially, while left ventricular reverse remodeling occurred across all three groups over 12 months of follow-up, Cluster 3 showed attenuated left atrial reverse remodeling, suggesting persistent LA dysfunction as a mechanistic driver of poor outcomes.
NIDCM has long frustrated clinicians with its phenotypic heterogeneity — patients with identical ejection fractions can follow wildly divergent trajectories. This work advances a growing body of precision cardiology research using unsupervised clustering (similar to approaches validated in heart failure with preserved ejection fraction) to move beyond single-parameter risk scores. The integration of AI-derived T1-mapping parameters — reflecting myocardial fibrosis burden — with echocardiographic and clinical data is methodologically compelling. However, the cohort of 347 patients remains modest for a three-cluster ML model, and the two-center Korean tertiary design limits generalizability to broader populations. Cox regression adjustments were limited to age and sex, leaving residual confounding possible. As a preprint not yet peer-reviewed, these findings require independent external validation before influencing clinical protocols. If confirmed, phenotype-guided management — particularly targeting LA dysfunction in high-risk clusters — could meaningfully refine NIDCM therapy selection.