Development of risk prediction model for delirium using routinely collected electronic health records

Background: Delirium, an acute confusional state in hospitalised older adults, is common and associated with poor outcomes. We developed and internally validated a machine learning model for early identification of patients at risk of delirium using routinely collected electronic health record data. Setting and Methods: Risk prediction model was developed using electronic health data from all adults over 50 years in a large metropolitan publicly funded hospital in Victoria, Australia. Candidate predictors, informed by the literature, were sourced from the first 48 hours of admission. Delirium diagnosed during admission was the outcome. Logistic regression (LR) and Random Forest (RF) models were developed using 26 variables, with 10-fold stratified nested cross-validation and multiple imputation, and internal validity estimated by Harrell bootstrap optimism correction over 200 repetitions. Model performance was assessed using area under curve (AUC) and calibration plots and metrics including decision curve analysis. Results: Following pre-processing, a cohort of 28,565 admissions from 20,136 patients was obtained, with 2,284 (8.0%) delirium. The Random Forest model achieved good discrimination (apparent AUC 0.887, 95% CI 0.881-0.894; optimism-corrected AUC 0.880) with improved calibration following Platt scaling (corrected calibration slope 0.99). At a 15% risk threshold, sensitivity was 66.0% and specificity was 88.9%, identifying approximately 15.5% of admissions for targeted assessment and intervention. Conclusions: These findings support the feasibility of using routinely collected EHR data for delirium risk identification in acute hospital settings. External validation is needed to confirm generalizability before implementation in other healthcare systems, and reliance on coded delirium diagnoses may under-ascertain cases, particularly hypoactive delirium.