Comparative diagnostic performance of machine learning models and traditional scores for HFpEF in older adults

AIMS: Diagnosing heart failure with preserved ejection fraction (HFpEF) remains challenging, particularly in older individuals. We hypothesized that machine learning (ML) approaches could improve diagnostic accuracy compared with HFpEF scores. METHODS: We evaluated the diagnostic performance of four supervised ML algorithms (random forest [RF], extreme gradient boosting [XGBoost], support vector machines, and decision trees) to identify HFpEF in individuals aged 60 to 80 years. The models were trained on three derivation cohorts (N = 1474; HFpEF: KaRen, MEDIA cohorts; community-based without HF: Malmö Preventive Project) and validated in two independent cohorts (N = 542; HFpEF: HF-Nancy cohort; community-based without HF: STANISLAS cohort). Performance metrics included accuracy, F-measure, area under the receiver operating characteristic curve (AUC), and C-index. ML models were also compared with HFA-PEFF, H2FPEF, and HFpEF-ABA scores. RESULTS: Among 2017 participants, RF and XGBoost demonstrated the highest diagnostic value, outperforming traditional HFpEF scores (AUC: RF, 0.98; XGBoost, 0.96; HFA-PEFF, 0.86; H2FPEF, 0.79). RF and XGBoost also showed the greatest gain in discriminative capacity among ML algorithms when compared with H2FPEF (ΔC-index: RF +0.20, XGBoost +0.18), HFA-PEFF (ΔC-index: RF +0.12, XGBoost +0.10), and HFpEF-ABA score (ΔC-index: RF +0.17, XGBoost +0.15). Elevated natriuretic peptides were by far the most influential feature in both RF and XGBoost models (36% of model explainability). CONCLUSIONS: Machine learning algorithms, particularly RF and XGBoost, demonstrated superior diagnostic accuracy compared to established HFpEF scoring systems. These findings support the potential integration of ML-based tools into clinical workflows to facilitate earlier identification of HFpEF and prompt initiation of guideline-recommended therapies.