Unsupervised phenotype clustering of non-ischemic dilated cardiomyopathy with AI-assisted T1 mapping cardiac MR

Aims: Prognostic stratification and individual management are essential in the heterogeneous population of non-ischemic dilated cardiomyopathy (NIDCM). We applied unsupervised machine learning (ML) clustering in NIDCM cohorts, using semi-automated artificial intelligence (AI)-based cardiac magnetic resonance imaging (CMR) measurements with multimodal data to identify distinct phenotypes, characterize echocardiographic remodeling trajectories, and evaluate prognostic significance. Methods and results: We analyzed 347 patients with NIDCM from two tertiary centers who underwent CMR and echocardiography at baseline, with follow-up echocardiography at a median 12 months. The cohort was randomly divided into derivation (n=242) and validation (n=105) sets using stratification by the composite outcome. Remodeling trajectories were evaluated using follow-up echocardiographic changes, and associations with outcomes were assessed by multivariable Cox regression adjusted for age and sex. Using eleven comprehensive clinical, laboratory, echocardiographic, and CMR-derived variables, partitioning around medoids clustering identified three phenotypes: (i) a younger, male-predominant preserved phenotype; (ii) a metabolic, fibrotic-remodeling phenotype; and (iii) an atrial fibrillation-predominant biventricular dysfunction phenotype. Cluster 1 showed the most favorable prognosis, whereas Cluster 3 had the highest risk of the composite outcome. Although LV reverse remodeling occurred across all clusters, Cluster 3 was characterized by attenuated LA reverse remodeling, suggesting persistent LA dysfunction. Conclusion: Unsupervised ML-based clustering of NIDCM patients, integrating AI-derived CMR parameters with multimodal data, identified three clusters exhibiting distinct patterns in longitudinal echocardiographic trajectories and outcomes. This strategy may enable more individualized management in heterogeneous NIDCM.