E239: A National Cancer Survival Calculator for Hepatobiliary Cancers using Machine Learning

While cancer prognosis has traditionally been estimated by tumor stage, survival is multifactorial. Our objective was to develop a national “Cancer Survival Calculator” using the National Cancer Database (NCDB) and machine learning to better tailor survival estimates for patients with hepatobiliary cancers.

Methods:

From the NCDB, we identified all patients with primary liver, intrahepatic/perihilar/distal bile duct, and gallbladder (“hepatobiliary”) cancers (2010-2017). Factors included in the model were selected via random forest algorithms to rank order variables most influential on survival as well as clinical review by site specific experts. Data were split into 75% training and 25% test datasets. Extreme gradient boosting with survival embeddings, a machine learning class, was used to estimate prognosis and generate a 3-year survival curve with the training dataset. Internal model validation using the test dataset was assessed with concordance statistics (c-index) and Brier scores (1.0 and 0.0 representing perfect accuracy, respectively).

Results:

Of 182,766 patients, 69.3% had primary liver, 25.3% bile duct, and 5.4% gallbladder cancer. In rank order, metastatic disease, elevated bilirubin, clinical T category, age at diagnosis, undergoing surgery, receipt of chemotherapy, and histology were most influential on survival via random forest. The final model included tumor (clinical TNM stage, histology, grade, tumor markers), patient (age, sex, race/ethnicity, comorbidities) and treatment (surgery, chemotherapy, radiation) specific factors. Accurate model discrimination and performance was demonstrated, estimating 3-year survival with a c-index of 0.72 and Brier score of 0.15 compared to TNM stage alone (c-index 0.65, Brier score 0.19). Patient specific survival curves using a web application were able to be generated; for example, a 66 year old male with T2N0M0 intrahepatic cholangiocarcinoma 3.0 cm in size who undergoes partial hepatectomy with adjuvant chemotherapy has an estimated 1-year survival rate of 86% (95%CI: 80%-90%) and 3-year survival rate of 60% (95%CI: 53%-68%, Figure).

Conclusions:

This “Cancer Survival Calculator” provides a comprehensive prognostic tool for patients with hepatobiliary cancers incorporating patient, tumor, and treatment specific factors. A web-based platform under development will enable clinicians to input these factors and calculate personalized survival estimates in real time. Future work involves pilot testing across accredited cancer programs and generating models for all primary cancers in the NCDB.