Post-operative delirium risk estimation and assessment with machine learning: development and validation of the Open-DREAM Model

BACKGROUND: Postoperative delirium (POD) is a common and serious complication in older surgical patients, associated with increased morbidity, prolonged hospitalisation and increased healthcare costs. Existing predictive models have limited clinical adoption due to implementation costs, suboptimal accuracy and poor generalisability. OBJECTIVES: To develop an open-access, interpretable machine learning (ML) model using only preoperative data to predict the risk of POD in patients aged at least 60 years undergoing surgery. DESIGN: Observational diagnostic study using prospectively collected routine care data. SETTING: Single secondary care hospital in Switzerland, January 2023 to January 2024. PATIENTS: A total of 1425 patients were screened; 748 met the inclusion criteria (≥60 years, noncardiac, nonintracranial surgery, no preoperative delirium). INTERVENTIONS: None. MAIN OUTCOME MEASURE: POD, defined as a Nursing Delirium Screening Scale (Nu-DESC) score at least 2 at any peri-operative assessment. RESULTS: Three ML algorithms were trained and compared: Support Vector Machines; Logistic Regression and XGBoost. The calibrated XGBoost model achieved an area under the receiver operating characteristic curve (AUC) of 0.80, with a sensitivity of 0.89 at the optimal probability threshold of 0.35. Feature selection using SHapley Additive exPlanations (SHAP)-derived rankings reduced the predictor set from 55 to 15 features without loss of AUC. Type of anaesthesia (spinal vs. general) was the strongest predictor. CONCLUSIONS: An open-access ML-based model using routinely collected preoperative variables can predict POD with high sensitivity and preserved interpretability. External validation is warranted. Future research should explore causal inference for modifiable risk factors, acknowledging the limitations of observational data. This open-access application is currently intended for research and educational use only until external validation confirms its performance in independent patient groups.