Development of a machine learning-based risk prediction model for perioperative neurocognitive disorders

Hengjun Wan, Qing Zhong, Ana Kowark, Mark Coburn, Yuling Tang, Yiyun Li, Xiaobin Wang, Qiuran Zheng, Xiaoxia Duan,
Development of a machine learning-based risk prediction model for perioperative neurocognitive disorders,
Journal of Clinical Anesthesia,
Volume 107,
2025,
112016,
ISSN 0952-8180,
https://doi.org/10.1016/j.jclinane.2025.112016.
(https://www.sciencedirect.com/science/article/pii/S0952818025002776)
Abstract: Background
Perioperative neurocognitive disorder (PND) is a common complication that significantly increases patient mortality and healthcare burden. Existing predictive models lack standardisation and personalisation, especially for elderly patients undergoing non-cardiac elective surgery.
Methods
This study first identified 13 key feature variables through LASSO regression and then constructed ten machine learning prediction models based on this subset of variables. Model performance was validated via ROC/AUC and decision curve analysis. SHAP interpreted the optimal model, enabling development of a clinical risk assessment tool. Kaplan-Meier analysis examined the association between risk factors and PND onset timing.
Results
The incidence of PND was 12.5 % (255/2042). The AUC values across the ten machine learning models ranged from 0.615 to 0.877. Among these, the neural network model demonstrated the optimal predictive performance (AUC = 0.877, 95 % CI: 0.839–0.916). SHAP analysis identified hyperlipidaemia (highest SHAP value), smoking, ASA classification III, and low education level as key risk factors. Survival analysis showed that smoking, ASA classification III, and hypertension were associated with earlier onset of PND (log-rank test, P < 0.05).
Conclusion
This study systematically identified core risk factors for PND in non-cardiac surgical patients using machine learning, and developed both logistic regression-based nomograms and online tools that prioritize interpretability and practicality to support clinical decision-making. The primary modifiable factors include hyperlipidaemia, smoking, and ASA classification. Survival analysis revealed that smokers and hypertensive patients experienced earlier onset of perioperative neurocognitive disorder (PND). However, multicentre validation is warranted, alongside the development of individualised strategies informed by risk stratification.
Keywords: Perioperative neurocognitive disorder; Machine learning; Risk prediction; Surgery; Hyperlipidaemia