Inflammation-guided timing of thoracoscopic repair for recurrent tracheoesophageal fistula: predictive value of preoperative chest CT and a transfer learning model.

BACKGROUND: Recurrent tracheoesophageal fistula (RTEF) is one of the most challenging long-term complications following primary repair of esophageal atresia/tracheoesophageal fistula (EA/TEF). Thoracoscopic revision has been increasingly used in patients. Anastomotic leakage (AL) remains an important postoperative complication after RTEF repair. Objective predictors for surgical timing and AL risk are limited. METHODS: We retrospectively analyzed patients undergoing thoracoscopic RTEF repair at Beijing Children’s Hospital between January 2019 and January 2025. Preoperative, intraoperative, and postoperative variables were evaluated to identify predictors of AL. Based on preoperative features, a transfer learning (TL) model was developed to predict AL risk. RESULTS: A total of 92 thoracoscopic repairs were performed in 84 patients, with AL occurring in 23/92 cases (25.0%). Univariate analysis identified significant associations between AL and preoperative CT-determined pulmonary inflammation adjacent to the anastomosis (P < 0.001), persistent upper esophageal negative-pressure suction (P = 0.008), oral feeding (P = 0.037), approach used for primary EA/TEF repair (P = 0.021), and esophageal gap > 2 cm (P < 0.001). Multivariate analysis confirmed that preoperative inflammation severity (moderate: OR = 17.07, P = 0.014; severe: OR = 50.62, P = 0.001) independently increased AL risk, whereas an esophageal gap ≤ 2 cm was protective (OR = 0.18, P = 0.014). The TL model integrating preoperative features achieved excellent discrimination (training AUROC 0.929 [95%CI: 0.850–0.986]; validation AUROC 0.863 [95%CI: 0.688–1.000]) with high accuracy (0.919). CONCLUSIONS: Thoracoscopic RTEF repair is safe and effective when guided by preoperative assessment. CT-quantified pulmonary inflammation at the T4 vertebral level is a powerful independent predictor of AL. The proposed TL-based model enables individualized, data-driven optimization of surgical timing for improved outcomes.