Machine learning phenotyping of HefpEF :a systematic review and meta analysis of validation cohorts with external generalizability assessment
J Gohri, A R Rawal, A M Aryan, S K Kalra, A S Y Singh Yadav, N S SutariaAbstract
Background
HFpEF encompasses 3–5 mechanistically distinct phenotypes (hypertensive, metabolic, restrictive, senile) requiring differential therapeutic approaches. Traditional diagnostic scores (H2FPEF, HFA-PEFF) use categorical cutoffs, limiting precision. Machine learning (ML) phenotyping enables superior patient subtyping with prognostic stratification potential.
Objective
To evaluate ML phenotyping diagnostic accuracy versus established clinical scores, assess prognostic outcome discrimination by phenogroup, and characterize clinical implementation readiness.
Methods
Parallel search of MEDLINE, Embase, IEEE Xplore (2018–January 2026) for observational/prospective studies applying supervised ML (random forests, gradient boosting) or unsupervised clustering to HFpEF diagnosis/phenotyping. Inclusion: baseline LVEF ≥50%, N≥100, follow-up ≥6 months, reported diagnostic accuracy (AUC) or prognostic metrics (c-statistic). Primary outcomes: diagnostic AUC, phenogroup-stratified outcome discrimination (c-statistic for HF-hospitalization/mortality). QUADAS-AI bias assessment applied.
Results
Twenty-eight studies (n=8,456 HFpEF; 1,248 events; median follow-up 2.5 years) included. ML achieved diagnostic AUC 0.82–0.94 (median 0.88) versus H2FPEF 0.76–0.84 (p<0.001). Three–four distinct phenogroups identified; highest-risk demonstrated 3.0-fold increased HF-hospitalization/mortality versus lowest-risk (p<0.001). External multicenter validation c-statistic 0.75 (range 0.71–0.79), supporting real-world implementation.
Conclusions
ML phenotyping significantly enhances HFpEF diagnostic accuracy and enables robust prognostic stratification. External validation demonstrates clinically meaningful discrimination.