American College of Surgeons survival calculator for biliary tract cancers: using machine learning to individualize predictions.

Background: Although cancer prognosis is most commonly estimated by tumor stage, survival is multifactorial. Our objective was to develop an American College of Surgeons "Biliary Tract Cancer Survival Calculator" prototype using machine learning to generate personalized survival estimates based on patient, tumor, and treatment factors.

Methods: The National Cancer Database was used to identify all patients with biliary tract malignancies between 2010 and 2017 including intrahepatic bile duct, extrahepatic bile duct, and gallbladder cancers. Included variables were determined based on random forest algorithms and review by subject matter experts. Data were split into 80% training and 20% test data sets. Extreme gradient boosting with survival embeddings, a machine learning class, generated 3-year survival curves. Internal 5-fold cross validation was evaluated through concordance statistics (c-index), Brier scores, distant calibration, and time-dependent area under the curve.

Results: Overall, 62,877 patients were included. Metastatic disease, age at diagnosis, and lack of surgical treatment were identified as most influential on worse survival outcomes via random forest. The final model included patient (age, sex, race and ethnicity, comorbidities), tumor (clinical TNM stage, disease site, grade), and treatment (surgery, chemotherapy, radiation) factors. Accurate model discrimination, calibration, and performance was demonstrated on internal validation (c-index: 0.74, Brier score: 0.14, distant calibration: P < .001, area under the curve: 0.83). These metrics were notably improved compared to a model based solely on stage (c-index: 0.64, Brier score: 0.18, distant calibration: P < .001, time-dependent area under the curve: 0.68).

Conclusion: This "Biliary Tract Cancer Survival Calculator" represents a highly accurate and comprehensive prognostic tool to estimate individualized survival estimates in real time.