Beyond TNM staging: Machine learning of individualized colon cancer outcomes.

Abstract

Anatomic TNM staging lacks patient information that can influence cancer survival. Using machine learning, specifically extreme gradient boosting with survival embeddings, a model can be developed to improve predictive accuracy of patient survival by including additional clinical information. A retrospective study of 382,531 colon cancer cases diagnosed between 2018 and 2021 using the National Cancer Database. A random forest analysis was performed followed by training a machine learning model to develop survival predictions. Model 1 used only TNM staging to predict survival, whereas model 2 used TNM and the additional variables to predict survival. Predictive efficacy was compared using integrated Brier scores (a measure that predicts error rates), Harrell concordant indices (a measure that evaluates the discriminative ability of survival models), and time-dependent areas under the curve (a measure that assesses the performance of survival models). Predictive accuracy metrics demonstrated improvements in all areas including a 4-year Brier score of 0.19 for the TNM-only model, and 0.14 for the TNM with predictors model; a Harell concordance index for TNM-only of 0.73 and 0.83 for TNM with predictors. Finally, the average time-dependent area under the curve for TNM only was 0.75, and the average time-dependent area under the curve for TNM with predictors was 0.87. Machine learning can be used to develop models of cancer survival that include clinical parameters resulting in more accurate predictions of patient outcomes compared with traditional TNM anatomic staging.