Machine-learning clustering uncovers divergent response pathways to mavacamten in obstructive HCM: a portuguese multicentre perspective

Background Response to mavacamten in obstructive hypertrophic cardiomyopathy (oHCM) is heterogeneous due to differences in clinical features, biomarkers, and cardiac remodeling. Traditional subgroup analyses may miss multidimensional phenotypes influencing treatment effects. Unsupervised clustering offers a data-driven approach to identify latent response profiles, potentially improving patient stratification and supporting personalized therapy. Objective To identify distinct mavacamten response phenotypes using unsupervised clustering of clinical, biomarker, and imaging parameters in patients with oHCM, based on real-world data from three Portuguese centers. Methods Consecutive mavacamten-treated oHCM patients from three Portuguese centers were analyzed using K-means unsupervised clustering. Input variables captured multidimensional changes from baseline to optimal dose, including LVOT gradients, maximal wall thickness, indexed LV mass, left atrial diameter (LAd), E/e′, and NYHA class, along with baseline genotype category and CYP2C19 metabolizer status. Missing values were imputed with medians and all features were standardized (z-scores) before model training. Silhouette analysis across k=2 to k=6 supported a two-cluster model, and clinical and echocardiographic profiles of each cluster were examined to derive clinically interpretable response phenotypes. Results A total of 83 patients were included (mean age 65±12 years; 68.7% female). Two response phenotypes emerged as the optimal cluster solution. - Cluster 0 (“hyper-responders”, n=20) included younger patients (66.3±10.6 vs 60.4±14.3 years, p=0.096) with markedly higher baseline obstruction (rest LVOT gradient 101±37 vs 46±29mmHg, p<0.001; provoked 143±45 vs 85±36mmHg, p<0.001), slightly larger LAd (46±6 vs 43±6mm, p=0.06), and higher E/e′ (17.7±6.0 vs 15.2±5.2, p=0.08). This group had pronounced haemodynamic and diastolic improvement, with reductions in LVOT gradients (rest -98±34 vs -32±28mmHg, p<0.001; provoked -135±45 vs -66±35mmHg, p<0.001) and E/e′ (-6.7±3.5 vs -2.8±2.7, p=0.01). - Cluster 1 (“typical responders”, n=63) consisted of older patients with less severe baseline gradients and exhibited a moderate but consistent response across all remodeling domains. Reductions in LAd (-4.1±4.5 vs -3.3±4.1mm, p=0.42) and indexed LV mass (-32.7±41 vs -17.5±39 g/m², p=0.20) were observed in both clusters. Genotype distribution and CYP2C19 metabolizer class did not significantly differ. Conclusions Data-driven phenogrouping identified two mavacamten response pathways: a younger, highly obstructive hyper-responder group with striking reverse remodeling, and a milder obstructive group with moderate improvement. Hemodynamic load, but not genotype, drives treatment benefit. These phenogroups may support more personalized mavacamten use in real-world oHCM.