心不全（駆出率保持型）患者における1年死亡率の予測因子

Heart failure with preserved ejection fraction (HFpEF) is increasingly common. While widely used prognostic scores often rely on limited variables and linear assumptions, they are likely to miss complex risk patterns. We aimed to develop and internally validate prediction models for 1-year all-cause mortality after first hospitalization for decompensated HFpEF and to compare them with standard survival models.

We performed a retrospective cohort study (2010-2020) using electronic medical records from a large academic health system. Adults with HFpEF (EF≥50%) admitted for a first time heart failure exacerbation were included. Variables spanned demographics, comorbidities, laboratory tests, echocardiography, discharge medications, and outcomes. Data were split into training (80%) and test (20%) sets with stratification by outcome. Missing values were handled with multiple imputation by chained equations. Two tree-based classifiers (Extreme Gradient Boosting and Light Gradient Boosting) were tuned with cross-validation and evaluated by area under the receiver operating characteristic curve (AUROC) and calibration. Time-to-event models included Cox proportional hazards, random survival forest (RSF), and gradient boosting survival (GBS) with concordance index and calibration assessment. Global and local (patient-level) explainability was extracted from each model, with cross-model predictor ranking and comparison.

We analyzed 7,840 index admissions; mean age was 78 years with 55.6% women. One-year mortality was 31.5% . Test-set AUROC was 0.751 for Extreme Gradient Boosting and 0.749 for Light Gradient Boosting with acceptable calibration. GBS achieved the highest concordance index (0.718), followed by RSF (0.711) and Cox (0.704). Lower serum albumin, older age, higher N-terminal pro B-type natriuretic peptide, renal dysfunction, and lower hemoglobin were the most consistent risk signals.

Diverse modeling families produced similar discrimination and coherent predictors. A transparent risk tool using routinely available admission data appears feasible, allowing for patient-level risk assessment for precision health.