Abstract
BACKGROUND: Machine learning (ML) models in healthcare are crucial for predicting clinical outcomes, and their effectiveness can be significantly enhanced through improvements in accuracy, generalisability, and interpretability. To achieve widespread adoption in clinical practice, risk factors identified by these models must be validated in diverse populations. METHODS: In this cohort study, 86 870 community-dwelling older adults ≥65 years from the UK Biobank database were used to train and test three ML models to predict 30-day emergency hospitalisation. The three ML models, Random Forest (RF), XGBoost (XGB), and Logistic Regression (LR), utilised all extracted variables, consisting of demographic and geriatric syndromes, comorbidities, and the Drug Burden Index (DBI), a measure of potentially inappropriate polypharmacy, which quantifies exposure to medications with anticholinergic and sedative properties. 30-day emergency hospitalisation was defined as any hospitalisation related to any clinical event within 30 days of the index date. The model performance metrics included the area under the receiver operating characteristics curve (AUC-ROC) and the F1 score. RESULTS: The AUC-ROC for the RF, XGB and LR models was 0.78, 0.86 and 0.61, respectively, signifying good discriminatory power. The DBI, mobility, fractures, falls, hazardous alcohol drinking and smoking were validated as important variables in predicting 30-day emergency hospitalisation. CONCLUSIONS: This study validated important risk factors for predicting 30-day emergency hospitalisation. The validation of important risk factors will inform the development of future ML studies in geriatrics. Future research should prioritise the development of targeted interventions to address the risk factors validated in this study, ultimately improving patient outcomes and alleviating healthcare burdens.