DEVELOPMENT AND VALIDATION OF A MACHINE LEARNING (ML) BASED DECISION-MAKING TOOL TO DETERMINE RISK AND TIMING OF RECURRENCE OF BARRETT’S ESOPHAGUS (BE) NEOPLASIA AFTER SUCCESSFUL ENDOSCOPIC ERADICATION THERAPY (EET)

Background:
Recent studies have described the durability of EET for BE-related neoplasia and the risk of neoplastic recurrence after achieving complete eradication of intestinal metaplasia (CE-IM). Current guidelines for endoscopic surveillance intervals post-EET are based on limited evidence. We aimed to develop and validate an ML-based decision-making tool to predict the risk of BE neoplasia recurrence and the timing of recurrence after CE-IM.

Methods:
Feature selection, model training, and validation were based on individual patient-level data from 4 prospective US registries of BE patients with dysplasia undergoing EET. Demographics, endoscopic findings, worst BE histology, EET details, and CE-IM rates were recorded. Only patients achieving CE-IM were included in this analysis. Recurrence was defined as the histologic finding of BE with or without dysplasia/neoplasia on surveillance endoscopy after CE-IM. Subjects were randomly assigned to training and test cohorts in a 4:1 ratio. Predictors of recurrence and timing of recurrence were identified using Adaptive Boosting. Missing data was addressed using mean imputation and enriched using a synthetic-minority over-sampling method. ML models were subsequently trained using these predictors and evaluated using area under the receiver operating curve (AUC) with 5-fold cross-validation and an independent test set. Time to BE recurrence was classified in intervals of 6 months, up to 5 years post-EET.

Results:
2773 BE with neoplasia patients (mean age 65 yrs, 80.5% males, 96.5% Caucasians) with a mean follow-up of 38.3±70 months were included (Figure 1). BE recurrence occurred in 760 (27.4%) patients while 282 (10.2%) had neoplastic BE recurrence. Predictors of BE recurrence and timing were identified in the ML model and included demographics, BMI, family history, hiatal hernia, duration and length of BE, grade of dysplasia, number of sessions to CE-IM, and type of EET. After 5-fold cross-validation, the discriminatory power of the BE recurrence ML model was tested on the validation set resulting in an 86% prediction accuracy (90% sensitivity, 75% specificity). The prediction accuracy for BE neoplasia recurrence was 84% (84% sensitivity, 77% specificity). Performance of the model in predicting the time to BE recurrence with 0.70 AUC for 1 year and 0.66 for 5 years. A web-based prediction and clinical decision-making tool (Figure 2) was created that provides the risk of BE recurrence and a personalized “heat map” of the time to recurrence up to 5 years post-EET.

Conclusions:
This US-based study demonstrates the feasibility of developing an ML-based decision-making BE recurrence tool that has the potential to report an objective risk and timing of recurrence, providing a personalized approach to surveillance intervals in post-EET patients. External validation will be necessary prior to clinical adoption.