Machine learning to evaluate the effects of non-clinical social determinant features in predicting colorectal Cancer mortality in a medically underserved Appalachian population.

Colorectal cancer (CRC) is the 2nd leading cause of cancer death in the United States (US). Rural Appalachia suffers the highest CRC incidence and mortality rates. There are several non-clinical health-related social determinant factors (SDOH) associated with cancer mortality. This study describes novel predictive modeling that uses demographic, clinical, and SDOH features from health records data from Appalachian community cancer centers to predict 5-year CRC survival. We trained, validated, and tested four gradient-boosted tree ensemble (XGBoost) machine learning models which were developed using selected combinations of available features. The area under the receiver operating characteristic curve was greatest in the model that included SDOH features with demographic and clinical features (0.79; P < 0.0001). Feature stratification showed rurality as the top SDOH feature. It is demonstrated that the ML model performs better when SDOH features are included, and that rurality significantly impacts CRC survival in Appalachia. The study provides preliminary indications that further data collection and evaluation of SDOH factors would strengthen our understanding of their impact on cancer survival in Appalachia and other underserved populations and improve development of strategies for care delivery.