Machine-learning phenotyping and exploratory medication-outcome associations in heart failure with preserved ejection fraction: a multicohort retrospective study

Background The marked clinical and pathophysiological heterogeneity of HFpEF has repeatedly undermined uniform treatment strategies. Consequently, identifying distinct phenotypes and medication-associated patterns across phenotypes may inform future precision-medicine research. Methods In this multicohort retrospective study, patients with HFpEF were included from MIMIC-IV (derivation cohort, n = 2,511), the MIMIC-III CareVue subset (internal validation cohort, n = 1,524), and the Second Hospital of Hebei Medical University (external application cohort, n = 340). Unsupervised K-prototypes clustering was used to derive phenotypic groups, with K = 2 evaluated as a lower-resolution sensitivity solution. All phenotype-stratified medication analyses were exploratory. In the derivation and internal validation cohorts, overlap weighting was the primary weighted analysis and stabilized inverse probability of treatment weighting was a sensitivity analysis; external medication estimates were reported descriptively because adequate within-group covariate balance was not achieved. A TabPFN phenotype-assignment classifier was developed after recursive feature elimination, evaluated in internal validation against independently reclustered labels aligned by prespecified one-to-one Hungarian assignment on cluster-profile distances, and applied without refitting in the external cohort. Results The primary three-cluster solution identified clinically interpretable but partially overlapping cardiorenal-metabolic, hypertensive-pulmonary, and low-blood-pressure/arrhythmia profiles. K = 2 had a higher mean silhouette width than K = 3 (0.083 versus 0.060), whereas both solutions showed high median resampling stability (ARI, 0.940 versus 0.924). A graded 365-day mortality difference was observed across the three clusters in the derivation cohort, but between-cluster survival differences were not statistically significant in the internal validation or external application cohorts. In the derivation and internal validation cohorts, no weighted cluster-specific medication association remained statistically significant after FDR correction, and all medication-by-cluster interaction tests were nonsignificant. External medication estimates were descriptive and were not used for confirmatory inference. The fixed TabPFN classifier achieved one-versus-rest AUCs of 0.951–0.969 in internal validation (10-bin multiclass ECE = 0.028). Conclusion Routinely collected ICU data supported an exploratory framework for describing overlapping HFpEF phenotypes and reproducibly assigning machine-derived labels. The analyses did not establish phenotype-specific medication effects, and the classifier should be regarded as a research tool pending prospective multicenter validation.