The AUGIS Survival Predictor: Prediction of Long-Term and Conditional Survival After Esophagectomy Using Random Survival Forests

The aim of this study was to develop a predictive model for overall survival after esophagectomy using pre/postoperative clinical data and machine learning. SUMMARY BACKGROUND DATA: For patients with esophageal cancer, accurately predicting long-term survival after esophagectomy is challenging. This study investigated survival prediction after esophagectomy using a Random Survival Forest (RSF) model derived from routine data from a large, well-curated, national dataset. METHODS: Patients diagnosed with esophageal adenocarcinoma or squamous cell carcinoma between 2012 and 2018 in England and Wales who underwent an esophagectomy were included. Prediction models for overall survival were developed using the RSF method and Cox regression from 41 patient and disease characteristics. Calibration and discrimination (time-dependent area under the curve) were validated internally using bootstrap resampling. RESULTS: The study analyzed 6399 patients, with 2625 deaths during follow-up. Median follow-up was 41 months. Overall survival was 47.1% at 5 years. The final RSF model included 14 variables and had excellent discrimination with a 5-year time-dependent area under the receiver operator curve of 83.9% [95% confidence interval (CI) 82.6%–84.9%], compared to 82.3% (95% CI 81.1%—83.3%) for the Cox model. The most important variables were lymph node involvement, pT stage, circumferential resection margin involvement (tumor at < 1 mm from cut edge) and age. There was a wide range of survival estimates even within TNM staging groups, with quintiles of prediction within Stage 3b ranging from 12.2% to 44.7% survival at 5 years. CONCLUSIONS: An RSF model for long-term survival after esophagectomy exhibited excellent discrimination and well-calibrated predictions. At a patient level, it provides more accuracy than TNM staging alone and could help in the delivery of tailored treatment and follow-up.