Survival machine learning model of T1 colorectal postoperative recurrence after endoscopic resection and surgical operation: a retrospective cohort study.

objectiveTo construct a postoperative recurrence prediction model for patients with T1 colorectal cancer after endoscopic resection and surgical operation via survival machine learning algorithms.

methodsBased on two tertiary-level affiliated hospitals, case data of 580 patients with T1 colorectal cancer treated by endoscopic resection and surgery were obtained, and patients' personal information, treatment modalities, and pathology-related information were extracted. After Boruta's algorithmic feature selection, predictors with significant contributions were identified. The patients were divided into a train set and a test set at a ratio of 7:3, and five survival machine learning models were subsequently built, namely, Randomized Survival Forest (RSF), Gradient Boosting (GB), Survival Tree (ST), CoxPH and Coxnet. Interpretability analysis of the model is based on the SHAP algorithm.

resultsPatients at high risk of lymph node metastasis have a poor prognosis, but different treatment modalities do not significantly affect the prognosis of patients with recurrence. The Random Survival Forest model shows better performance, with a C-index and Integrated Brier Score of 0.848 and 0.098 in the test set, respectively, and its time-dependent AUC is 0.918. The interpretability analysis of the model revealed that submucosal invasion depth < 1000 μm, tumor budding grade of BD1, lymphovascular invasion and perineural invasion are absent, well differentiated cancer cells, and tumor size < 20 mm have positive effects on the model, lts negative gain characteristics are a contributing factor to patient relapse.

conclusionsThe prognostic model constructed via survival machine learning for patients with T1 colorectal cancer has good performance, and can provide accurate individualized predictions.