Machine Learning-based Predictions of Healthcare Contacts Following Emergency Hospitalisation Using Electronic Health Records

Abstract Emergency care systems are challenged by the needs of an ageing population, requiring tailored inputs facilitated by early care needs assessment. We examined the potential of Machine Learning algorithms to identify in-hospital healthcare needs in older patients after emergency admission, developed from linked electronic health record (EHR) data within South-East Scotland. Gradient-boosting (XGBoost) prediction models were trained on frailty markers and nursing risk assessments to predict healthcare contacts, adverse outcomes and requirements for specialist input between arrival and 72 hours following admission. Across 98,242 patients, the predicted contact error rate varied between 49% at point of emergency attendance and 34% at 72 hours post-admission. Area-under-the-curve reached 0.89 in predicting need for urgent geriatric services, and 0.83 for in-hospital rehabilitation. Pressure ulcer risk prominently reduced nursing and rehabilitation contact frequency. EHR data can predict granular estimates of in-hospital activity early after ED attendance, facilitating quicker allocation to appropriate urgent care pathways.