Prognostic Prediction Models for Postoperative Patients with Stage I to III Colorectal Cancer: A Retrospective Study Based on Machine Learning Methods

Abstract Purpose To utilize the patient, tumor, and treatment features and compare the performance of machine learning algorithms, develop and validate models to predict overall, disease-free, recurrence-free, and distant metastasis-free survival, and screen important variables to improve the prognosis of patients in clinical settings.Methods More than 1000 colorectal cancer patients who underwent curative resection were grouped according to 4 endpoints and divided into testing sets and training sets (9:1). We applied 4 machine learning algorithms to predict 1-, 3-, and 5-year survival times. The area under the receiver operating characteristic curve (AUC) and average precision (AP) were our accuracy indicators. Vital parameters were screened by multivariate regression models. To achieve better prediction of longitudinal oncological outcomes, we performed 10-fold cross-validation except for the recurrence-free survival model (3-fold cross-validation). We iterated 3000 times after hyperparameter optimization and assessed the internal testing sets.Results The best AP values were greater than 80%, except for the overall survival model (69.5%). The best AUCs were all greater than 0.70 except for the recurrence free survival model (0.61). The models performed well. Variables that were widely correlated with prognoses, such as the TNM stage, were selected as important features; however, indirectly related indicators, such as Ki-67 level, were also selected.Conclusion We constructed an independent, high-accuracy "white-box" machine learning system for predicting survival times. This system may help in determining managing strategies for colorectal cancer patients and has future utility in personalized medicine and monitoring.