关键词: 回肠造口, 脱水, 机器学习, 预测模型

Abstract: Objective To construct a dehydration risk prediction model for ileostomy patients based on machine learning algorithms and conduct internal validation.Methods A retrospective study was conducted including 520 ileostomy surgery patients admitted to our hospital from June 2020 to June 2023 as study subjects.Patients were divided into modeling set and testing set at a 1∶1 ratio,and postoperative dehydration was statistically observed.Based on clinical data from the modeling set,univariate and multivariate Logistic regression analyses were used to identify independent influencing factors for dehydration after ileostomy surgery.Random forest model,XGBoost model,and decision tree model were constructed to predict postoperative dehydration in patients.The predictive efficacy of the models was evaluated,the optimal model was selected,and internal validation was performed.Results In the total cohort,74 patients (14.23%) developed dehydration within 30 days postoperatively,with 38 patients (14.62%) in the modeling set and 36 patients (13.85%) in the validation set developing postoperative dehydration.The dehydration group had lower body mass index than the non-dehydration group,while age,diabetes,American Society of Anesthesiologists (ASA) classification,diuretic use,postoperative chemotherapy proportion,and postoperative day 1 C-reactive protein levels were higher than the non-dehydration group (P<0.05).Multivariate Logistic regression model showed that age,diabetes,diuretic use,postoperative chemotherapy,and postoperative day 1 C-reactive protein were all risk factors for postoperative dehydration in ileostomy surgery patients (P<0.05).Receiver operating characteristic (ROC) curve analysis showed that the predictive AUCs of the random forest model,XGBoost model,and decision tree model were 0.913,0.905,and 0.832,respectively,with the random forest model having the highest predictive efficacy.Calibration curve and decision curve analysis showed that the random forest model had good consistency and clinical net benefit.Random forest feature importance ranking showed that the importance ranking of independent influencing factors for postoperative dehydration in patients was postoperative day 1 C-reactive protein,age,diuretic use,diabetes,and postoperative chemotherapy,in that order.Additionally,internal validation showed that the predictive sensitivity,specificity,and accuracy of the random forest model were 86.11%,87.05%,and 86.92%,respectively.Conclusion The random forest model constructed based on age,diabetes,diuretic use,postoperative chemotherapy,and postoperative day 1 C-reactive protein has the best performance in predicting postoperative dehydration in ileostomy surgery patients.Importance ranking based on this model can provide support for formulating targeted clinical intervention strategies.

Key words: ileostomy, dehydration, machine learning, prediction model