Artificial Intelligence Enabled Phenogrouping of Heart Failure with Preserved Ejection Fraction Depicts Early and End-Stage Trajectories

BackgroundHeart failure with preserved ejection fraction is challenging to diagnose, precluding the initiation of prognostic medications. A deeper understanding of HFpEF phenotypes and trajectories could address this, however previous studies analyzed trial cohorts, limiting generalizability. Here we apply artificial intelligence (AI) algorithms to longitudinal electronic health record data (EHR) to characterize reproducible real-world HFpEF phenogroups. MethodsWe applied natural language processing (NLP) to extract structured and unstructured data from EHRs at a multi-site hospital network, including comorbidities, echocardiographic measurements, and all-cause mortality outcomes. HFpEF cases were identified by applying contemporaneous guideline-directed criteria. Phenogroups were derived at the first documented mention of heart failure using latent class analysis, then reassigned annually to quantify the frequency and directionality of transitions. External validation was performed at an independent hospital network. ResultsWe identified 2,223 patients (median age 75 years; 60% female) meeting HFpEF diagnostic criteria, with a median follow-up of 4 years. A clinician-recorded HFpEF diagnosis was present for only 10.4%. Four phenogroups were identified: Elderly-Atrial Dysfunction (32%), Cardio-Renal-Metabolic (23%), Obesity-Predominant (24%), and Young-Low Comorbidity (22%). The Young-Low Comorbidity group had the lowest rate of clinician-assigned HFpEF diagnosis (3%), and more than half of patients transitioned phenogroups within five years, suggesting this may represent an early-stage HFpEF phenotype. By contrast, Elderly-Atrial Dysfunction and Cardio-Renal-Metabolic groups displayed stable group membership (<10% transitioned), with high mortality rates (adjusted hazard ratio = 1.30 [1.02-1.67] and 1.49 [1.18-1.87] respectively). Phenogroup reproducibility was validated externally (n=3,349), demonstrating consistent clinical features and survival trends. ConclusionAI-enabled phenotyping of real-world data identified reproducible, prognostically significant HFpEF phenogroups. The identification of an underrecognized, early-stage subgroup underscores a critical opportunity for earlier diagnosis and targeted intervention to alter the clinical trajectory. Clinical perspectivesO_ST_ABSWhat is new?C_ST_ABSHeart failure with preserved ejection fraction (HFpEF), as a heterogenous syndrome, can be characterized into four reproducible phenogroups, including an early-HFpEF phenotype. What are the clinical implications?Future research could build on implementation of scalable, EHR-enabled diagnostic strategies for early HFpEF detection. Longitudinal studies examining the impact of intervention on early phenogroups may provide valuable insights into modifying disease progression.