Machine learning-based Predictive Modeling of Anastomotic stricture after Esophageal atresia repair

Background Anastomotic stricture (AS) is a common complication after esophageal atresia (EA) surgery. This study analyzed the clinical features to identify the relevant risk factors for EA and constructed a risk prediction model to guide clinical decision-making. Method: Clinical data of 162 children with EA who visited and underwent surgery at our center between June 2015 and June 2022 were collected. Based on the postoperative anastomotic condition, they were divided into the AS and non-AS groups. Univariate and multivariate analyses were performed to screen for independent risk and protective factors to construct the model. The EA cases were randomly divided into a training set and a test set (7:3), and a logistic regression (LR) prediction model was constructed using the machine learning algorithm for the training set. The model performance was tested and evaluated using the test set. Result: Feature selection analysis showed that weight (OR = 0.999, 95%CI 0.998–0.999), gap length (OR = 1.101, 95%CI 1.040–3.254), and anastomotic fistula (OR = 6.671, 95%CI 0.684–155.450) were the risk factors for AS after EA surgery (P<0.05). An LR prediction model was constructed using these three risk factors, and the AUC of the training and test set was 82.49% and 76.86%, respectively. The accuracy, precision, and recall were 76.99%, 72.73%, and 68.54% for the training set and 77.55%, 70.00%, and 66.67% for the test set. The clinical decision curve showed that the AS risk prediction model had a high clinical net benefit. Conclusion: Gap length, weight, and anastomotic fistula are the main risk factors for AS after esophageal anastomosis surgery in children with EA. The prediction model constructed using these risk factors and the machine learning algorithm had high performance in predicting AS after surgery. This model can guide targeted interventions to improve surgical efficacy and improve patient prognosis.