# Analyzing Key Predictors of Postoperative Delirium Following Coronary Artery Bypass Grafting and Aortic Valve Replacement: A Machine Learning Perspective

## TL;DR

This study uses machine learning to identify key factors that predict delirium after heart surgery, aiming to improve early detection and prevention.

## Contribution

The novel contribution is applying interpretable machine learning to uncover combined clinical and biochemical predictors of postoperative delirium.

## Key findings

- Sedation during mechanical ventilation was the strongest predictor of postoperative delirium.
- XGBoost achieved high accuracy (97.6%) in predicting delirium using clinical and biochemical variables.
- SHAP analysis revealed interactions between sedation, ventilation, and biomarkers like fibrinogen and troponin I.

## Abstract

Background and Objectives: Postoperative delirium (POD) is a frequent and severe complication following cardiac surgery, particularly in high-risk patients undergoing coronary artery bypass grafting (CABG) and aortic valve replacement (AVR). Despite extensive research, predicting POD remains challenging due to the multifactorial and often non-linear nature of its risk factors. This study aimed to improve POD prediction using an interpretable machine learning approach and to explore the combined effects of clinical, biochemical, and perioperative variables. Materials and Methods: This study included 131 patients who underwent CABG or AVR. POD occurrence was assessed using standard diagnostic criteria. Clinical, biochemical, and perioperative variables were collected, including patient age, sedation type, and mechanical ventilation status. Machine learning analysis was performed using an XGBoost classifier, with model interpretation achieved through SHapley Additive exPlanations (SHAP). Univariate logistic regression was applied to identify significant predictors, while SHAP analysis revealed variable interactions. Results: POD occurred in 34.3% of patients (n = 45). Patients who developed POD were significantly older (67.7 ± 6.5 vs. 64.5 ± 8.7 years, p = 0.020). Sedation with mechanical ventilation and the type of sedative used were strongly associated with POD (both p < 0.001). Sedation during mechanical ventilation showed the strongest association (OR = 2520.0; 95% CI: 80.9–78,506.7; p < 0.00001). XGBoost classifier achieved excellent performance (AUC = 0.998, accuracy = 97.6%, F1 score = 0.976). SHAP analysis identified sedation, mechanical ventilation, and their interactions with fibrinogen, troponin I, leukocyte parameters, and lung infection as key predictors. Conclusions: This study demonstrates that an interpretable machine learning approach can enhance POD prediction, providing insights into the combined impact of multiple clinical, biochemical, and perioperative factors. Integration of such models into perioperative workflows may enable early identification of high-risk patients and support individualized preventive strategies.