Leveraging electronic health records for data-driven comorbidity phenotyping in heart failure: Identification of eight risk-stratified phenotypes in a Chinese population

Objective: To identify comorbidity subgroups of heart failure (HF) from electronic health record (EHR) data by computational phenotyping and evaluate their associations with mortality and healthcare utilization. Background: Multimorbidity drives HF complexity, yet conventional classifications capture it poorly and EHR-based comorbidity phenotyping in Asian HF populations remains limited. Methods: In this retrospective cohort using EHR data from 462 hospitals in Yichang, China (2018-2023), latent class analysis was applied to 17 ICD-10-coded comorbidity indicators among 95,593 adults with incident HF. Cox and negative binomial models assessed all-cause mortality and healthcare utilization, adjusted for age and sex. Sensitivity analyses examined stricter HF definitions, alternative class solutions, wider comorbidity ascertainment, and competing risk of death. Results: Eight phenotypes emerged; most prevalent were Cardiometabolic (31.4%) and Minimal Comorbidity (28.9%; low probabilities for all conditions except COPD). Versus Cardiometabolic, Malignancy-Anemia had the highest mortality (adjusted HR 5.00, 95% CI 4.69-5.32), followed by Cardiorenal-Anemia (2.35), Severe Multimorbidity (1.99) and Minimal Comorbidity (1.93, 1.85-2.01). Malignancy-Anemia had the highest hospitalization rate (incidence rate ratio 2.85, 95% CI 2.68-3.03), yet its cumulative incidence of first hospitalization did not differ from the reference once death was treated as a competing event (subdistribution HR 1.03; P = 0.287), reflecting high early mortality. Excess mortality in Minimal Comorbidity was not attenuated by capturing index-encounter diagnoses, adjusting for baseline healthcare contact, or stratifying by COPD; phenotype ordering held across alternative HF definitions and class solutions. Conclusions: Phenotypes derived entirely from routinely coded data separate HF patients into prognostically distinct groups differing up to five-fold in adjusted mortality. Low coded comorbidity does not indicate low clinical risk, with direct implications for risk tools built on administrative data. With adjustment limited to age and sex in one region, these are hypothesis-generating prognostic associations requiring external validation and prospective evaluation before clinical use.