Identifying patterns of high intraoperative blood pressure variability in noncardiac surgery using explainable machine learning: a retrospective cohort study

Abstract

Background High intraoperative blood pressure variability (HIBPV) is significantly associated with postoperative adverse complications. However, practical tools to characterize perioperative factors associated with HIBPV remain limited. This study aimed to develop explainable supervised machine learning (ML) models to classify patients with HIBPV and to identify structural perioperative patterns associated with HIBPV through model interpretation.Materials and Methods This retrospective cohort study analyzed 47,520 noncardiac surgery cases from Beijing Tsinghua Changgung Hospital. We applied four ML algorithms—Extreme Gradient Boosting (XGBoost), Random Forest (RF), Light Gradient Boosting Machine (LightGBM), and Logistic Regression (LR)—to classify patients with or without HIBPV. The overall population and each age subgroup (pediatric, adult, elderly) underwent independent 70/30 train-test splits for model development. Model performance was assessed using the area under the receiver operating characteristic curve (AUROC). SHapley Additive exPlanations (SHAP) values were used to interpret model outputs and assess feature importance.Results Among 47,520 noncardiac surgeries, 1,996 (4.2%) were classified as HIBPV. XGBoost and RF achieved the best performance, with AUROC values of 0.85 (95% confidence intervals (CI): 0.84–0.86) and 0.84 (95% CI: 0.82–0.85). Intraoperative average heart rate (HR) and bispectral index (BIS) were the most influential variables. In patients aged 50 ∼ 70, higher sevoflurane dosage was associated with reduced HIBPV risk. Among hypertensive patients, elevated intraoperative blood calcium (>1.10 mmol/L) was associated with increased HIBPV risk.Conclusion The models enabled accurate classification of HIBPV cases and highlighted key discriminative perioperative variables through SHAP-based interpretation. Intraoperative HR and BIS were significant contributing factors. Moreover, interactions between sevoflurane and age and between hypertension and calcium levels may inform individualized hemodynamic management strategies.

Keywords

• Machine learning
• blood pressure variability
• Anesthesia management