Machine Learning Improves Prediction Over Logistic Regression on Resected Colon Cancer Patients

Introduction: Despite advances, readmission and mortality rates for surgical patients with colon cancer remain high. Prediction models using regression techniques allows for risk stratification to aid periprocedural care. Technological advances have enabled large data to be analyzed using machine learning (ML) algorithms. A national database of colon cancer patients was selected to determine whether ML methods better predict outcomes following surgery compared to conventional methods. Methods: Surgical colon cancer patients were identified using the 2013 National Cancer Database (NCDB). The negative outcome was defined as a composite of 30-d unplanned readmission and 30- and 90-d mortality. ML models, including Random Forest and XGBoost, were built and compared with conventional logistic regression. For the accounting of unbalanced outcomes, a synthetic minority oversampling technique (SMOTE) was implemented and applied using XGBoost. Results: Analysis included 528,060 patients. The negative outcome occurred in 11.6% of patients. Model building utilized 30 variables. The primary metric for model comparison was area under the curve (AUC). In comparison to logistic regression (AUC 0.730, 95% CI: 0.725-0.735), AUC's for ML algorithms ranged between 0.748 and 0.757, with the Random Forest model (AUC 0.757, 95% CI: 0.752-0.762) outperforming XGBoost (AUC 0.756, 95% CI: 0.751-0.761) and XGBoost using SMOTE data (AUC 0.748, 95% CI: 0.743-0.753). Conclusions: We show that a large registry of surgical colon cancer patients can be utilized to build ML models to improve outcome prediction with differential discriminative ability. These results reveal the potential of these methods to enhance risk prediction, leading to improved strategies to mitigate those risks.