Unsupervised machine learning uncovers functional recovery endotypes following valvular surgery

Background Patients undergoing valvular heart surgery show heterogeneous postoperative recovery that is not fully explained by left ventricular ejection fraction (LVEF), the cornerstone metric currently guiding postoperative management. Multidimensional phenotyping through unsupervised machine learning may identify clinically meaningful endotypes beyond conventional classifications. Purpose The PRO-FEVs study aimed to derive latent endotypes using unsupervised clustering of comprehensive baseline data and to evaluate their association with functional recovery and LVEF. Methods We enrolled 436 consecutive patients undergoing valvular surgery. Baseline demographic, clinical, echocardiographic and functional variables were collected. Latent endotypes were identified using an enhanced K-prototypes clustering algorithm for mixed data. Model selection was guided by the silhouette index computed on Gower distances. Differences across clusters were assessed using Kruskal–Wallis and Chi-square tests. Outcomes assessed at discharge from cardiac rehabilitation included length of stay (LoS), modified Barthel Index (MBI), and six-minute walk test (6MWT). Results We analysed data from 436 patients (mean age 72.8 ± 9.4 years; 57.6% male). At baseline, hypertension was present in 61.9% of participants and atrial arrhythmias in 12.2%. Left ventricular ejection fraction was overall preserved. The optimal clustering solution identified three endotypes (silhouette coefficient = 0.34; n=106, 153, and 177, respectively). The first endotype was predominantly characterised by aortic stenosis, preserved rhythm stability, and better baseline functional capacity. The second was mainly defined by aortic regurgitation and larger ventricular dimensions. The third was characterised by mitral regurgitation, a higher prevalence of atrial arrhythmias (p<0.001) and worse baseline autonomy. Importantly, LVEF did not differ across endotypes (p=0.119). Post-rehabilitation 6MWT performance differed significantly across groups (p<0.001), with greater walking capacity observed in the aortic stenosis endotype. In contrast, LoS and MBI at discharge did not differ significantly across groups. Conclusions In the PRO-FEVs study, unsupervised clustering identified three clinically coherent post–valvular surgery endotypes associated with distinct functional recovery trajectories despite similar LVEF. These findings suggest that structural valve phenotype and arrhythmic burden, rather than ejection fraction alone, may drive early rehabilitation outcomes. Endotype-driven stratification could support more tailored rehabilitation strategies and inform individualized post-operative management.