Evaluating non-linear thresholds and synergistic interactions in the comorbidity of preoperative insomnia and emergence agitation: a SHAP-based machine learning study.

BACKGROUND

Preoperative insomnia disorder (PID) and emergence agitation (EA) often present as a complex perioperative comorbidity. Traditional linear models may not adequately capture their underlying interactive mechanisms.

METHODS

This retrospective, single-center study analyzed 542 adult surgical patients. Machine learning algorithms, particularly XGBoost, were developed using stratified random splitting (70% training, 30% testing) to predict the PID-EA co-occurrence. The Shapley Additive exPlanations (SHAP) framework was utilized to interpret statistical associations and non-linear interactions among clinical variables.

RESULTS

The XGBoost model demonstrated superior discriminative performance (AUC = 0.874 [95% CI: 0.812-0.936], AUPRC = 0.883, Brier score = 0.150) compared to logistic regression (AUC = 0.749 [95% CI: 0.675-0.823]). SHAP analysis indicated that preoperative cortisol and melatonin were the primary features associated with the model's predictions. Partial dependence plots suggested potential non-linear risk thresholds for neuroendocrine markers. Furthermore, interaction analyses illustrated that severe postoperative pain may synergistically amplify the risk associated with preoperative insomnia.

CONCLUSION

The integration of XGBoost and SHAP offers an interpretable approach to evaluating the PID-EA comorbidity network. These findings highlight potential clinical thresholds and interaction patterns, warranting further validation through multi-center prospective studies.