321 RANDOM SURVIVAL FOREST FOR PERSONALISED PROGNOSTICATION AFTER ESOPHAGECTOMY
Abstract For patients with esophageal cancer, producing accurate prediction models for long-term survival after esophagectomy has proved challenging. We investigated whether Random Survival Forests (RSF), a machine learning method, could produce an accurate prognostic model for overall survival after esophagectomy. Methods The study used data from the 'National Oesophago-Gastric Cancer Audit' (NOGCA) and included patients diagnosed with esophageal adenocarcinoma or squamous cell carcinoma between 2012 and 2018 in England and Wales and who underwent a curative esophagectomy with adequate lymphadenectomy (15 lymph nodes) and survived to discharge (n = 6838). Missing data was handled using multiple imputation. 15 variables were selected for inclusion using Random Forest variable importance and used to train the final model. The same variables with non-linearity transformations were used to develop a traditional Cox regression model for comparison. Results Median survival was 50 months. The final RSF model had good discrimination on internal validation with a C-index of 0.7627 (0.7625–0.7629), exceeding the cox model 0.7539 (0.7541–0.7537). At 3 years post-surgery, overall survival was 56.2%. The RSF yielded a mean predicted survival of 59.3% (IQR 33.3–87.1%) with good calibration (Figure 1) compared to 57.4% (38.4%–79.8%) for the cox model. The most influential variables were lymph node involvement and pT/ypT stage, however other variables including neoadjuvant treatment completion and surgical complications were also important. Decision curve analysis was undertaken which also showed an increased net benefit with the RSF model. Conclusion A Random Forest survival model provided better performance in predicting survival after curative esophagectomy. This will allow more personalised predictions to be delivered clinicians and patients. An online web app is provided at https://uoscancer.shinyapps.io/NOGCA/