P1.064. Limitations of Preoperative Prediction Models for Complications After Esophagectomy: A Multi-Center Analysis

Abstract

Esophageal Cancer: Surgical Treatment of Esophageal Cancer Preoperative risk stratification for esophagectomy complications relies on clinical prediction models; however, their discriminative performance in multi-institutional settings remains poorly defined. We hypothesized that standard preoperative variables would demonstrate limited predictive validity across heterogeneous surgical cohorts. We analyzed 2,490 patients undergoing esophagectomy across four institutions in Asia (total n=2,490; individual center n range 75–1,012). Four outcomes were studied: recurrent laryngeal nerve palsy (RLNP), anastomotic leak (AL), pulmonary complications (PC), and vocal cord palsy (VCP). Logistic regression models with bootstrap-validated odds ratios (1,000 iterations) were evaluated by 5-fold cross-validated AUC. SHAP (SHapley Additive exPlanations) via Gradient Boosting Machines quantified variable importance. Decision curve analysis (DCA) assessed net clinical benefit across threshold probabilities 2–70%. Association between tumor location and each complication was assessed using chi-squared tests. All prediction models demonstrated poor-to-fair discrimination: RLNP AUC 0.533, AL AUC 0.586, PC AUC 0.676, and VCP AUC 0.556. Tumor location was the only statistically significant categorical predictor of RLNP—upper/cervical location was associated with higher RLNP incidence compared to middle thoracic tumors (39.1% vs. 28.0%; OR 1.22, 95%CI 1.05–1.42; p=0.009). No significant association was observed between tumor location and AL, PC, or VCP. DCA demonstrated negligible clinical net benefit for RLNP and AL models; only the PC model provided modest benefit (max net benefit gain +0.057) at threshold probabilities of 5–20%. SHAP analysis identified FEV1%, PNI score, and BMI as the highest-importance variables for RLNP prediction, with tumor location ranking sixth—indicating that location contributes a statistically real but clinically modest signal. Standard preoperative variables are insufficient for individualized risk stratification of RLNP, anastomotic leak, or vocal cord palsy after esophagectomy. Statistical significance (p=0.009 for location–RLNP association) does not translate to clinically meaningful predictive power (AUC 0.533). Tumor location should be incorporated into RLNP preoperative counseling. Improved prediction will require prospective integration of real-time intraoperative data. Pulmonary complication risk approaches clinically actionable prediction (AUC 0.676) and may guide respiratory prehabilitation targeting.