Various machine learning prediction models for short-term postoperative prognosis of colorectal cancer based on intraoperative hypothermia.

This retrospective study aimed to investigate inadvertent intraoperative hypothermia (IIH) in patients undergoing laparoscopic radical resection for colorectal cancer (CRC) and to compare predictive models for short-term prognosis. Among 296 CRC patients (January 2022 to June 2024), univariate and multivariate logistic regression identified independent risk factors. Based on these, five models (nomogram, neural network, decision tree, random forest, gradient boosting machine) were constructed and evaluated using ROC, calibration curves, decision curves, and confusion matrices, with an additional 102 patients (July 2024 to July 2025) for clinical validation. Of the 296 patients, 72 (24.32%) had poor short-term prognosis, and 102 (34.46%) experienced IIH. Restricted cubic spline analysis showed IIH duration >55 minutes significantly increased poor prognosis risk. IIH was an independent risk factor (OR=4.498, P<0.001), remaining significant after adjusting for tumor location, TNM stage, and CEA (OR=4.245, P<0.001). Rectal tumor location, TNM stage III, and CEA ≥5 ng/mL were also risk factors. Except for the decision tree, the four other models showed good predictive performance in both sets. In the independent validation (102 cases), prediction accuracies were 83.33%, 78.43%, 78.43%, 77.45%, and 75.49%, respectively. In conclusion, IIH is an independent risk factor for poor short-term prognosis after laparoscopic CRC resection, with significantly increased risk when duration exceeds 55 minutes. Logistic regression, neural network, random forest, and gradient boosting machine all demonstrate excellent predictive performance.