Abstract / Summary
Abstract Background Frailty is highly prevalent among elderly emergency department (ED) patients and is associated with mortality, readmission, and functional decline. Current screening tools provide static assessments without dynamic risk prediction capability. Objective To develop and externally validate a multi-modal machine learning model integrating Clinical Frailty Scale (CFS) data with electronic medical records (EMR) for predicting frailty risk in elderly ED patients. Methods This retrospective cohort study screened 521 consecutive patients aged > = 60 years presenting to the ED of a tertiary hospital in Shandong Province between June 2024 and March 2025; 474 patients met eligibility criteria. Seven features were selected: age, respiratory rate, SpO2, C-reactive protein (CRP), albumin, CRP/albumin ratio, and Charlson Comorbidity Index (CCI). A support vector machine (SVM) with radial basis function kernel was trained with 5-fold stratified cross-validation and evaluated on a held-out test set (n = 143, 30%). External validation was performed on 22,670 patients from the MIMIC-IV database (version 2.2) using a HistGradientBoostingClassifier with 19 available clinical features, targeting 28-day in-hospital mortality. Results In the internal validation cohort (n = 474, mean age 72.9 ± 7.9 years, 50.7% female), the SVM model achieved test set AUC = 0.940, accuracy = 0.853, sensitivity = 0.851, specificity = 0.855, PPV = 0.838, NPV = 0.867, and F1 = 0.844; 5-fold cross-validation mean AUC was 0.902 ± 0.036. In the MIMIC-IV external validation cohort (n = 6801), the HistGradientBoosting model achieved AUC = 0.788 (95% CI: 0.750–0.828), with sensitivity = 0.741 and specificity = 0.688 at the Youden-optimal threshold. The CRP/albumin ratio and inflammatory-nutritional markers were dominant predictors in both cohorts. Conclusion The multi-modal SVM model demonstrated good discriminative ability (AUC = 0.940) for frailty risk prediction in elderly ED patients. External validation on MIMIC-IV confirmed clinically meaningful predictive performance for 28-day mortality (AUC = 0.789), with inflammatory-nutritional and organ function markers emerging as dominant predictors across both cohorts.