Development and validation of a machine learning-based prediction model for intraoperative hypothermia in Chinese patients undergoing gastrointestinal surgery.

Abstract

Background: Intraoperative hypothermia, defined as a core temperature < 36.0 °C, is a common complication during gastrointestinal surgery, potentially affecting patient outcomes. This study aimed to develop a predictive model for intraoperative hypothermia based on clinical features.

Methods: In this retrospective, single-center study, we analyzed data from 418 patients who underwent gastrointestinal surgery between January and March 2024 at the First Hospital of Putian City, China. Core temperature was measured intraoperatively using a deep nasal probe every 15 min. Five machine learning models (logistic regression, random forest, support vector machine, XGBoost, and multilayer perceptron) were evaluated to develop a prediction model for hypothermia. Logistic regression (LR) was identified as the optimal model and used to develop a nomogram based on key features.

Results: Among 25 clinical features, 12 showed significant differences between the two groups. The LR model demonstrated the best predictive performance [accuracy = 0.808, area under the curve (AUC) = 0.872] and identified six key predictors: temperature at surgery start, surgery duration, cisatracurium use, preoperative temperature, anesthesia time, and preoperative red blood cell count. A nomogram constructed with these features exhibited excellent predictive ability (AUC = 0.877) and clinical utility, as confirmed by decision curve analysis.

Conclusion: This study highlights key predictors of intraoperative hypothermia and presents a reliable nomogram for risk prediction in patients undergoing gastrointestinal surgery. These findings can inform targeted interventions and improve perioperative care. Further validation with diverse cohorts is warranted to enhance generalizability.