To construct a predictive model for perineural invasion using XGBoost based on clinical information and data of colorectal cancer patients and evaluate its effectiveness.

Real-world clinical data of 178 colorectal cancer patients in our hospital were selected and subjected to data cleaning and feature engineering processing. An XGBoost model was trained to select the optimal subset of features and construct the final model. The model was interpreted through SHapley Additive Explanation. The Boruta algorithm was applied to filter features and construct five models: Linear Discriminant Analysis, Naive Bayes, K Nearest Neighbor, Classification and Regression Trees, and Random Forest, and their oversampling class balance models were then generated. The performance of different models was compared.

XGBoost applied Sequential Feature Selector to select the optimal feature subset containing 9 features and constructed the final model. Compared with other machine learning models, the performance of the XGBoost model was superior and the training speed was faster.

We have successfully constructed a predictive model for perineural invasion in colorectal cancer patients based on XGBoost. It can provide preoperative prediction of nerve invasion for clinical doctors, especially surgeons, and provide a basis for the development of comprehensive treatment plans.