Development and Internal Validation of an Electronic Medical Record–Based Machine Learning Model for Frailty Prediction in Hospitalized Patients: Retrospective Cohort Study

Abstract Background Frailty is prevalent among hospitalized patients and is associated with increased risk of complications, prolonged hospitalization, higher mortality, and health care costs. Existing frailty screening tools, including the Clinical Frailty Scale (CFS) and Hospital Frailty Risk Score (HFRS), have limitations related to subjectivity, resource intensity, or reliance on retrospective data, limiting their use in routine acute care. Early identification of frailty is critical to support targeted interventions and resource allocation, yet scalable approaches for predicting frailty early during hospital admission remain limited. Objective This paper evaluates whether routinely collected electronic medical record (EMR) data available within the first 72 hours of admission can be used to predict frailty in hospitalized patients. Methods A retrospective cohort study was conducted using routinely collected EMR data from the Royal Adelaide Hospital (RAH), including overnight admissions with a length of stay of at least 72 hours between July 2023 and December 2024. Frailty was defined using the HFRS (score >4). Predictors were derived from EMR data available within 72 hours of admission, including demographics, medications, laboratory results, nursing assessments, and medical and allied health consult orders. Data were split into training (21,436/26,824, 80%) and test (5388/26,824, 20%) sets. Multiple machine learning models were trained, with the F 2 -score as the primary optimization metric. Model performance was evaluated using the F 2 -score, recall, area under the receiver operating characteristic curve (AUROC), calibration, and clinical utility analyses. Results This study included 26,824 admissions, of whom 9427 (35%) were classified as frail. The final XGBoost (XGB) model achieved an F 2 -score of 0.755 (95% CI 0.741-0.768), an AUROC of 0.777 (95% CI 0.764-0.790), and a recall of 0.873 (95% CI 0.857-0.887) on the test dataset, with good calibration and an expected-to-observed ratio close to 1 (0.996, 95% CI 0.963-1.028). Key predictors included polypharmacy, laboratory results (including creatinine, albumin, and sodium), allied health consult orders, and nursing assessment data. Model performance was lower at 24 hours postadmission and improved by discharge. Conclusions An EMR-based frailty prediction model using data available at 72 hours postadmission can identify patients with frailty early in their hospital stay, supporting targeted interventions and operational planning. Reduced performance in Indigenous patients highlights the need for culturally validated frailty indices and ongoing evaluation of equity in predictive modeling.