Predictive modeling for early detection of refractory esophageal stricture following esophageal atresia surgery: insight from a machine learning study.

backgroundRefractory esophageal stricture (RES) presents a challenging complication after esophageal atresia (EA) repair. Earlier identification of patients with RES could help clinical decision-making. However, there are currently limited articles to predict the occurrence of RES after EA repair.

methodsAll patients diagnosed with esophageal strictures (ES) who underwent primary endoscopic balloon dilations (EBD) after EA repair in our institution between January 2017 to December 2022 were retrospectively reviewed. They were divided into two groups according to the presence or absence of suffering RES to analyze the risk factors for developing RES after EA repair and construct a predictive model using the machine learning (ML) method, based on the baseline characteristics, early postoperative complications, and ES features at the first EBD.

results92 patients were included in our study, including the RES group (n = 59) and the non-RES group (n = 33). In terms of patient demographics, multivariate analysis revealed that LGEA (p = 0.021), and anastomosis leak (p = 0.016) were the risk factors for developing RES after EA repair. Further analysis of the stricture characteristics after EA repair between the two groups revealed that esophageal stricture at a two-thirds lower level (p = 0.006), eccentric esophageal stenosis (p = 0.048), and a higher upper pouch stricture index (U-SI) (p = 0.011) emerged as significant risk factors for developing RES following EA repair. Then, we constructed ML models for predicting RES following LEAP based on patient demographics and stricture characteristics. Among all the ML models, the decision tree (DET), random forest (RF), and multilayer perceptron (MLP) models demonstrated perfect predictive value in the validation set. Among them, the RF model showed the highest average AUC in the training set (AUC = 0.756, 95% CI: 0.584-0.928), indicating better reproducibility. Subsequently, we used Shapley additive explanation (SHAP) to interpret the most important features and found that U-SI, rTEF, stricture level, and anastomotic leakage were identified as the most important predictive features. Furthermore, an RF-based clinical decision support system was deployed, providing dynamic visualization for predicting RES.

conclusionLGEA, anastomosis leak, lower stricture level, eccentric stricture, and higher U-SI value were risk factors for RES after EA repair. By utilizing the ML predictive model, we could efficiently and accurately predict the occurrence of RES.