A Machine-Based Learning Model for Recurrence Prediction and Timing After Endoscopic Eradication Therapy for...

Tools that can predict recurrence in patients with Barrett's esophagus (BE)-related neoplasia treated with endoscopic eradication therapies (EET) to guide surveillance decisions are needed. We aimed to develop and validate a machine learning (ML)-based prediction tool to predict the risk and timing of recurrence status post EET.

Three prospective United States databases of patients who underwent EET for BE-related neoplasia and achieved complete eradication of intestinal metaplasia (CE-IM; n = 1114) were utilized to develop and internally validate a ML-based prediction tool using the Random Forest model and imputation techniques. Predictors incorporated in this model included demographics, endoscopy and pathology results, and EET details. A Cox proportional hazards model was utilized to predict the time to recurrence. External validation was performed using the United States Radiofrequency Ablation database (n = 1397).

BE recurrence occurred in 29.2% (n = 734) of patients and BE-related neoplasia recurrence in 10.6% (n = 265), with a mean time to recurrence of 21.3 months (mean follow-up, 37.7 months). The top predictors for recurrence included BE length, body mass index, age, sessions needed to achieve CE-IM, and baseline histology. The model was well-calibrated, and area under the receiver operating characteristic curve (AUC) was 0.92 (95% confidence interval [CI], 0.85-0.95) on internal validation and 0.91 (95% CI, 0.87-0.95) on external validation for BE recurrence. For BE-related neoplasia recurrence, the AUC was 0.90 (95% CI, 0.88-0.93). The model had moderate discriminative performance to predict timing of recurrence with a C-index of 0.701 at 1 year (AUC 0.71), 0.68 (AUC 0.69) at 3 years, and 0.66 (AUC 0.69) at 5 years.

This United States-based externally validated tool accurately predicts BE and BE-related neoplasia recurrence and timing post EET. This practical tool may help provide a personalized approach to surveillance strategies.