S1012 Development and Validation of Predictive Models for Pathological Complete Response in Esophageal Adenocarcinoma Patients Receiving Neoadjuvant Chemoradiotherapy: A Decision Curve Analysis

Introduction: Identifying esophageal adenocarcinoma (EAC) patients likely to achieve pathological complete response (pCR) after neoadjuvant chemoradiotherapy (nCRT) could refine treatment strategies, potentially identifying candidates for organ preservation. We developed advanced predictive models for pCR and identified key predictive factors. Methods: Data from 9,065 EAC patients (Stage I-III, cM0) treated with nCRT and esophagectomy (2018-2021) were extracted from the National Cancer Database. Features included demographics, comorbidities, clinical stage, tumor grade, histology, lymphovascular invasion, radiotherapy dose, chemotherapy type, and diagnosis-to-surgery interval. Multi-Layer Perceptron (MLP), XGBoost, and logistic regression models were developed. Predictive accuracy was assessed by area uder the curve (AUC) and Brier score, and key predictive features were identified. Results: The pCR rate was 19.7% (N = 1,813 in test set). Both XGBoost and MLP models achieved an AUC of 0.715, outperforming logistic regression (AUC 0.707). XGBoost demonstrated superior calibration, as indicated by a lower Brier score (0.144), compared to MLP (Brier score 0.221) and logistic regression (Brier score 0.220). Using model-specific importance measures, lymphovascular invasion (LVI) status consistently ranked as a top predictor across all models. Other leading predictors for the MLP model included interval to surgery and positive clinical N-stage. For XGBoost, clinical T-stage and a specific insurance type were prominent. Logistic regression's top contributors, aside from LVI, included diagnostic confirmation methods, median income quartile, and urban/rural classification. Conclusion: Advanced machine learning models, especially XGBoost, accurately predict pCR in EAC patients post-nCRT and demonstrate potential clinical utility. Identified predictors such as lymphovascular invasion, clinical stage, and treatment characteristics offer insights into patient subgroups with differing response likelihoods. If prospectively validated, these models could aid in shared decision-making and tailoring therapy.