Interpretable Machine Learning Enables Preoperative Physiologic Risk Stratification for Dysphagia After Anti-Reflux Surgery

Postoperative dysphagia remains one of the most clinically significant complications following anti-reflux surgery, yet existing preoperative risk stratification approaches incompletely integrate esophageal motility, reflux burden, and esophagogastric junction biomechanics. We developed an interpretable machine learning framework integrating multimodal physiologic and clinical data to predict new-onset postoperative dysphagia and translate these relationships into a clinically deployable risk score. Following screening of 878 consecutive patients undergoing anti-reflux surgery at a high-volume tertiary referral center, 428 patients met inclusion criteria and underwent multimodal preoperative physiologic assessment including high-resolution manometry, EndoFLIP impedance planimetry, Bravo pH monitoring, and GERD-HRQL evaluation. Patients were divided into derivation (n = 362) and independent holdout validation (n = 66) cohorts. An ensemble machine learning framework integrating logistic regression, random forest, XGBoost, and support vector machine models was developed using engineered higher-order physiologic interaction features and subsequently translated into an interpretable point-based scoring system. The ensemble model demonstrated strong discrimination in the derivation cohort (AUC = 0.91, accuracy = 80.9%, sensitivity = 86.2%, specificity = 75.7%) with preserved performance in the independent holdout cohort (AUC = 0.72, balanced accuracy = 68.7%, sensitivity = 70.0%, specificity = 67.4%). Interaction features integrating reflux burden, esophageal contractility, distensibility, and patient-level modifiers demonstrated greater predictive utility than isolated physiologic variables alone. The resulting cumulative risk score enabled stratification into distinct postoperative dysphagia susceptibility groups across both derivation and validation cohorts and was deployed as an open-source web-based calculator for individualized risk estimation. These findings demonstrate that integrated multimodal esophageal physiology combined with interpretable machine learning enables clinically meaningful prediction of postoperative dysphagia after anti-reflux surgery and may support future personalized perioperative risk stratification and prospective multicenter validation.