Machine learning algorithms to predict epidural-related maternal fever: a retrospective study.

Introduction: The epidural-related maternal fever (ERMF) induced by patient-controlled epidural analgesia (PCEA) remains unpredictable. Our objective is to develop ERMF prediction models using real-world data, aiming to identify pertinent contributing factors and support obstetricians in making personalized clinical decisions. Methods: Women who used patient-controlled epidural analgesia between October 2021 and March 2023 at a tertiary hospital in Jiangsu Province were retrospectively documented. The primary outcome was the occurrence of maternal fever associated with epidural use. We developed six machine learning (ML) models and assessed the area under curve (AUC) for characteristics of subjects' performance, calibration curves, and decision curve analyses. Results: A total of 1,492 women were enrolled, with 24.3% experiencing ERMF (362 cases). The AUC ratios between the logistic regression (LR) model and the stochastic gradient descent (SGD) models showed statistical significance (p < 0.05), while the differences between the other models were not statistically significant. In comparison to the SVM model, the LR model exhibited better calibration (Brier score: 0.193; calibration slope: 0.715; calibration intercept: 0.062). Consequently, the LR model was selected as the prediction model. Furthermore, the LR-based nomogram identified eight significant predictors of ERMF, including neutrophil percentage, first stage of labor, amniotic fluid contamination during membrane rupture, artificial rupture of membranes, chorioamnionitis, post-analgesic antimicrobials, pre-analgesic oxytocin, post-analgesic oxytocin, and dinoprostone suppositories. Conclusion: Optimally applying logistic regression models can enable rapid and straightforward identification of ERMF risk and the implementation of rational therapeutic measures, in contrast to machine learning models.