Integrating the Hospital Frailty Risk Score into Explainable Machine Learning to Predict Mortality in Older Adults...

Background/Objectives: Community-acquired pneumonia (CAP) is a leading cause of mortality in older adults. Traditional prognostic scores may underestimate risk in frail patients by assuming linear relationships between predictors and outcomes. This study aimed to develop and validate explainable machine learning models integrating the administrative Hospital Frailty Risk Score (HFRS) to predict in-hospital mortality in a nationwide cohort of older adults in Chile. Methods: A retrospective cohort study was conducted using anonymized hospital discharge records from the Chilean National Health Fund (FONASA), including 58,306 hospitalization episodes of adults aged ≥60 years across 72 public hospitals. Fourteen supervised machine learning algorithms were trained using five routinely collected predictors: age, sex, HFRS, Charlson Comorbidity Index, and length of stay. Model performance was evaluated on an independent test set using AUC-ROC. SHAP (SHapley Additive exPlanations) values were calculated to assess global and individual predictor contributions. Results: The Extra Trees classifier achieved the highest discriminative performance (AUC-ROC 0.862), outperforming logistic regression (0.642) and other linear models. SHAP analyses identified HFRS as the most influential predictor (mean |SHAP| = 0.66), followed by length of stay, age, and comorbidities. Conclusions: Ensemble tree-based models incorporating administrative frailty measures provide superior mortality prediction compared to traditional linear approaches. Frailty emerged as the primary driver of risk, supporting scalable early stratification using routinely available hospital data.