[Machine learning models to predict intestinal necrosis in acute mesenteric venous thrombosis]

Objective: To develop and validate machine learning-based models for predicting the risk of transmural irreversible intestinal necrosis (ITIN) in patients with acute mesenteric venous thrombosis (AMVT). Methods: A retrospective case series analysis was conducted on 156 patients with AMVT admitted to the Department of General Surgery,Tianjin Medical University General Hospital,between January 2015 and December 2024. There were 106 males and 50 females, and 62 patients aged ≥60 years and 94 patients aged <60 years. Patients were randomly assigned by the random number table method to a training set (n=115) and a validation set (n=41) at a 7∶3 ratio. No significant differences were found between the two sets regarding clinical characteristics,laboratory findings,imaging results,or the incidence of ITIN (all P>0.05). Independent risk factors for ITIN identified via multivariable logistic regression in the training set were used to construct seven machine learning models (Logistic regression,K-nearest neighbors,support vector machine,random forest, decision tree, BP neural network,and AdaBoost) using Python 3.9. The predictive performance was assessed using receiver operating characteristic (ROC) curves,sensitivity, specificity, accuracy, area under the curve (AUC), Brier score, F1 score, and Cohen's kappa coefficient. Decision curve analysis (DCA) was performed to evaluate clinical utility,and internal validation was conducted using the validation set. Results: Multivariable Logistic regression analysis identified that onset duration ≤4 d,body temperature ≥39 ℃,intestinal obstruction on contrast-enhanced CT, and serum lactate ≥2.0 mmol/L as independent risk factors for ITIN in patients with AMVT (all P<0.05), machine learning models were constructed based on these factors. In the training set, the AUC values were 0.913, 0.937, 0.921, 0.937, 0.939, 0.938, and 0.936, respectively. The Brier scores,F1 scores, kappa coefficients, and DCA demonstrated favorable predictive performance for all seven models in the training set, although performance in the validation set was slightly lower. Conclusions: Short onset duration,high fever,intestinal obstruction on contrast-enhanced CT, and elevated serum lactate are critical predictors of ITIN in AMVT patients. The machine learning models based on these indicators exhibit robust predictive performance in the training cohort,though further external validation is required to confirm their generalizability.