Machine learning prediction model for delirium after heart valve replacement with cardiopulmonary bypass: A large-scale cohort study

Clinical Snapshot

Key Findings

• P Value: P < 0.001 for mortality association

• Effect Size: Random Forest model AUROC 0.854 vs traditional nomogram AUROC 0.754

• Primary Outcome: Postoperative delirium incidence 25.1% (270/1076)

• Nnt Or Sensitivity: External validation AUROC 0.793 demonstrating good generalizability

• Confidence Interval: 95% CI 0.764-0.945 for Random Forest model

Clinical Application

Requires implementation of machine learning infrastructure and standardized data collection for perioperative variables Relevant to Australian cardiac surgery centers. Could integrate with existing perioperative care pathways. May require TGA consideration for clinical decision support tools if implemented as medical device software Adult patients undergoing heart valve replacement with cardiopulmonary bypass in tertiary cardiac centers

Abstract

OBJECTIVE: Postoperative delirium (POD) is a common and severe complication following heart valve replacement (HVR) with cardiopulmonary bypass (CPB), associated with poor outcomes. Accurate early prediction is crucial for targeted prevention. This study aimed to develop and externally validate machine learning (ML) models for predicting POD in this high-risk population. METHODS: We conducted a retrospective cohort study involving 1076 adult patients who underwent HVR with CPB between January 2018 and December 2022. POD was assessed using the CAM-ICU. Perioperative factors were analyzed to develop a traditional logistic regression-based nomogram and four ML models. The best-performing model was then tested on an independent external validation cohort of 281 patients. RESULTS: The incidence of POD was 25.1% (270/1076) and was associated with significantly increased mortality (11.9% vs. 0.2%, P < 0.001) and prolonged hospitalization. In the internal test set, the Random Forest (RF) model demonstrated the highest predictive performance with an AUROC of 0.854 (95% CI, 0.764-0.945), outperforming the traditional nomogram (AUROC: 0.754). The RF model maintained robust performance in the external validation cohort, achieving an AUROC of 0.793. Model interpretation using SHAP analysis identified postoperative awakening time (PAT), postoperative mechanical ventilation time (PMVT), and postoperative reintubation (POR) as the most influential predictors. CONCLUSIONS: Machine learning models, particularly Random Forest, provide a robust and generalizable tool for the early prediction of POD after HVR. These models outperform traditional statistical approaches and offer interpretable insights, highlighting PAT as a critical, modifiable target for perioperative management to mitigate POD risk.

References

1. 1.Li, J., Xue, M., Peng, D., Han, K., Bai, R., An, R., Zhang, L., Zhou, B., Zhang, L., & Shen, X. (2026). Machine learning prediction model for delirium after heart valve replacement with cardiopulmonary bypass: A large-scale cohort study. Asian Journal of Psychiatry, 104930. https://doi.org/10.1016/j.ajp.2026.104930