Abstract
OBJECTIVE: Machine learning (ML) models have emerged as tools to predict length of stay (LOS) and disposition decision (DD) in emergency departments (EDs) to combat overcrowding. However, site-specific ML models are not transferable to different sites. Our objective was to develop interpretable ML models to predict LOS and DD at specific time points, all while establishing a transparent data analysis framework. This framework was designed to be easily adapted by other institutions for the development of their own ML models. METHODS: We analysed data from 297 392 ED visits of patients aged 18 and above at a quaternary hospital between 30 June 2019 and 31 December 2022. Eight ML algorithms were evaluated, and ultimately, twelve lasso models built from 21 features were trained to predict four outcomes of LOS and DD at three time points post-triage. Hold-out testing and cross-validation were conducted for these models. RESULTS: The area under the curve values were 0.862/0.868/0.878 for binary LOS predictions at 10, 60 and 120-minute time points and 0.839/0.851/0.863 for binary DD predictions. The accuracies were 60.2%/60.7%/61.9% for ternary LOS predictions and 61.5%/62.3%/63.4% for ternary DD predictions. CONCLUSIONS: Interpretable ML models demonstrated outstanding performances in predicting both LOS and DD. The transparent data analysis framework can be easily adapted by other institutions.