Abstract / Summary
Abstract Objective This study aimed to develop and validate an interpretable machine learning model for predicting delirium in adult intensive care unit (ICU) patients with constipation and to construct an online risk calculator to facilitate real-time clinical assessment. Methods This retrospective analysis included 4,010 patients from the MIMIC-IV database. Feature selection was performed using Recursive Feature Elimination (RFE), the Least Absolute Shrinkage and Selection Operator (LASSO), and the Boruta algorithm. Eight machine learning (ML) models were developed, and the optimal model was identified through comprehensive validation. Model interpretability was achieved using SHapley Additive exPlanations (SHAP), and a web-based calculator was deployed to enhance clinical usability. Results By integrating three feature selection strategies, 18 key predictors were identified for model construction. The gradient boosting machine (GBM) demonstrated acceptable calibration and moderate predictive performance in the validation cohort. SHAP feature importance indicated four major risk drivers for delirium: Glasgow Coma Scale (GCS), Sequential Organ Failure Assessment (SOFA) score, sedative use, and mechanical ventilation. The optimal model was subsequently implemented as an online risk calculator. Conclusion We developed and validated an interpretable machine learning model for delirium risk stratification among ICU patients with constipation. The integration of interpretable ML with an online calculator provides a practical tool to support early identification and timely management of delirium in this vulnerable population. The model was further evaluated using same-centre temporal validation with the MIMIC-III cohort.