Improving patient flow through hospitals with machine learning based discharge prediction

Abstract

AbstractBackgroundAdvanced analytics, underpinned by large-scale electronic health record (EHR) data, have the potential to transform the efficiency of healthcare delivery.MethodsWe used data from 4 hospitals in Oxfordshire, UK (01-February-2017 to 31-January-2020), to develop machine learning models for predicting hospital discharge in the next 24 hours, conditional on the duration of hospitalisation to date. We fitted separate XGBoost models for patients in hospital after planned and emergency admissions and for patients with different lengths of stay since admission. Models were trained and tested using data from 126,054 emergency and 45,609 planned admissions.FindingsIn held-out test data (01-February-2019 to 31-January-2020), individual patient discharge was predicted with positive and negative predictive values of 87·5% and 94·1% respectively in emergency patients (sensitivity 82·3%, specificity 98·5%, AUPRC 87·6%, AUC 87·5%) and 95·6% and 97·8% after a planned admission (sensitivity 83·5%, specificity 98·1%, AUPRC 90·5%, AUC 90·1%). Combining individual discharge probabilities allowed accurate estimates of the total number of discharges from hospital, mean average error 5·8% for emergency and 3·1% for planned patients. Previous hospital exposure, procedures, antibiotic use, and comorbidities were the most predictive features, but the most important features differed between planned and emergency patients, and over the course of admission. Performance was best in short-stay and planned admission patients, but generally robust over time, across subgroups, and different model training strategies.InterpretationOur approach could help improve flow of patients through hospitals, resulting in faster care delivery and patient recovery, and better use of healthcare resources.FundingNational Institute for Health Research.Research in ContextEvidence before this studyThere is unprecedented demand for healthcare services, and therefore significant interest in improving the efficiency of healthcare delivery. Prediction of hospital discharge could potentially facilitate more efficient and rapid delivery of care. We searched PubMed and Google Scholar using the terms (“patient flow” OR “hospital discharge” OR “bed state”) AND (“machine learning” OR “decision support”) to 31 March 2023. Despite successful applications of machine learning elsewhere in healthcare, there are relatively few implementations of machine learning based tools in operational management of healthcare. Previous studies are often limited to a specific context, e.g., one specialty, and often fail to demonstrate adequate levels of performance to be operationally useful.Added value of this studyWe describe an approach for accurately predicting hospital discharge within the next 24 hours that performs well across the whole hospital. We use established machine learning methods, but a novel modelling framework considering emergency and planned admissions separately, and fitting separate models conditional on the prior length of stay in the current admission. We provide accurate predictions of discharge within 24 hours for emergency patients, positive and negative predictive values of 87·5% and 94·1%, respectively, and 95·6% and 97·8% for planned admissions, in held-out test data. Combining individual discharge probabilities, we achieve highly accurate estimation of the total number of discharges from hospital over the next 24 hours, with mean average error of 5·8% and 3·1% for emergency and planned patients, respectively. Our method outperforms previous studies. We also provide detailed analyses exploring the optimal quantity and recency of training data needed, variations in predictive power throughout the day, performance across subgroups, and details of the key factors driving our predictions and how they vary across patient groups and during a hospital admission.Implications of all the available evidenceAdoption of our approach has the potential to improve the flow of patients through hospitals, leading to more efficient care delivery, better patient experiences and enhanced utilization of healthcare resources.