To leverage perioperative indicators in developing an explainable machine learning (ML) model for postoperative delirium (POD) prediction, discover distinct data-driven POD subtypes through hierarchical clustering analysis, and enhance personalized risk stratification to inform targeted clinical interventions.

This is a secondary analysis of several prospective observational studies, including 1106 patients who had non-cardiac surgery. Univariate analysis and the least absolute shrinkage and selection operator (LASSO) regression was used to screen essential features associated with POD. We compared six algorithms: adaptive boosting with classification trees, random forest (RF), neural networks, support vector machines, extreme gradient boosting with classification trees and logistic regression. SHapley Additive exPlanations (SHAP). was used to interpret the best one and to externally validate it in another large tertiary hospital. Among patients who developed POD, we conducted hierarchical clustering analysis on the risk factors (identified through univariate screening in the prediction model) to delineate distinct subtypes. We then compared the length of postoperative hospital stay and mortality rates (at 1, 3, 6, and 12 months postoperatively) between the identified clusters.

We identified 14 POD risk factors to develop ML models. The RF model performed best among the six ML models (area under the curve [AUC] of 0.85, 95 % confidence interval [CI], 0.78–0.91). SHAP analysis highlighted surgery duration, preoperative mini-mental state examination score, and Edmonton Frail Scale as the top predictors of POD. Hierarchical clustering identified three distinct POD subtypes: Subtype 1 (high-risk profile with significant comorbidity and inflammatory dysregulation, longest hospitalization: 21.5 days ([interquartile range (IQR) 19–28]; p   <  0.001), Subtype 2 (resilient majority with optimal survival; Log-rank p   <  0.001), and Subtype 3 (advanced age, frailty and low cognitive reserve, shortest hospitalization: 5 days [IQR 4–8]). Kaplan-Meier analysis showed significant 12-month survival differences among the subtypes (Subtype 2  >  Subtype 3  >  Subtype 1; p   <  0.001).

Our study validated the utility of ML models, particularly RF, in predicting POD and identified three novel data-driven subtypes with distinct clinical characteristics.