Improving Delirium Prediction in Surgery: Enhanced Machine Learning Models vs. Traditional Risk Scores

Postoperative delirium (POD) is a prevalent complication among older surgical patients; however, accurately predicting its occurrence using traditional point-based risk scores remains challenging. This study aimed to develop a delirium prediction model using a contemporary dataset and to compare the performance of modern machine learning (ML) algorithms with that of the Inouye and Marcantonio risk indices.

A retrospective cohort of 283 surgical patients aged ≥ 50 years (POD incidence = 8.8%) was analyzed. Candidate predictors included demographic characteristics, frailty markers, comorbidities, surgical category, and planned intensive care unit (ICU) admission. Class imbalance was addressed using the Synthetic Minority Over-sampling Technique. Five classifiers-ridge logistic regression (baseline), random forest, XGBoost, CatBoost, and a stacking ensemblewere trained using 5-fold stratified cross-validation. Model discrimination, assessed by the area under the receiver operating characteristic curve (AUC), along with F1 score, precision, and recall, was averaged across folds. For comparison, the Inouye 4-factor and Marcantonio 7-factor scoring systems were recalculated for each patient.

The mean patient age was 65.7 ± 8.6 years, and 40.6% were female. The incidence of POD varied significantly across surgical specialties, ranging from 18.4% in hepatobiliary procedures to 3.7% in general and vascular surgeries. Univariate analysis identified older age, functional dependency, planned ICU admission, and high-risk surgery as significant risk factors. The AUCs of the ML models ranged from 0.57 (logistic regression) to 0.78 (stacking ensemble). Among individual models, CatBoost achieved the best overall balance (AUC = 0.77). In contrast, traditional risk scores demonstrated poor performance (Inouye AUC ≈ 0.50; Marcantonio ≈ 0.56).

Contemporary ML models outperformed conventional delirium risk scores in predicting POD, although their overall discrimination remained moderate. Key predictors included age, frailty markers, and surgical complexity. Integration of ML-based screening into electronic health records may enable earlier identification of high-risk patients and facilitate targeted delirium prevention strategies. External validation in larger, multi-center cohorts is warranted to confirm these findings.