Objective To explore the application value of machine learning prediction model for neural invasion in gallbladder cancer based on enhanced computed tomography (CT) and clinical characteristics.

Methods The retrospective cohort study was conducted. The clinical and imaging data of 502 patients with gallbladder cancer who were admitted to The First Affiliated Hospital of Xinxiang Medical University from January 2010 to June 2024 were collected. There were 171 males and 331 females, aged 65(range, 35‒91)years. All patients underwent preoperative abdominal enhanced CT and radical resection. The 502 patients were randomly divided into a training set of 351 cases and a test set of 151 cases at a 7:3 ratio. The training set was used to construct prediction model, and the test set was used to validate prediction model. Observation indicators: (1)neural invasion in gallbladder cancer and influencing factor analysis; (2) construction and validation of machine learning prediction models for neural invasion in gallbladder cancer. Comparison of count data between groups was conducted using the chi-square test. Comparison of ordinal data between groups was conducted using the Mann-Whitney U test. Logistic regression model was performed for univariate and multivariate analyses. Independent influencing factors were incor-porated to construct machine learning models using the standard library modules based on Python 3.9. Receiver operating characteristic (ROC) curves were plotted, and the accuracy, sensitivity, specificity, area under the curve (AUC), precision, F1 score, positive predictive value, negative predic-tive value, and Kappa value were calculated to evaluate the predictive performance of the models. The Delong test was used to assess the differences in AUC among different models in the test set. The Hosmer-Lemeshow test and Brier score were used to evaluate the calibration of the models.

Results (1) Neural invasion in gallbladder cancer and influencing factor analysis. Of the 502 patients with gallbladder cancer, 131 cases had neural invasion, and 371 cases had no neural invasion. Results of multivariate analysis showed that total bilirubin, carcinoembryonic antigen, CA199, CA125, neutrophil-lymphocyte ratio, liver invasion detected by CT, vascular invasion detected by CT, hilar or retroperi-toneal lymph node metastasis detected by CT, and tumor stages T3 and T4 were independent influencing factors for neural invasion in patients with gallbladder cancer odds ratios=3.747, 2.395, 3.917, 3.596, 2.805, 2.377, 3.523, 2.774, 5.080, 6.809, 95% confidence interval (CI) as 1.890‒7.430, 1.154‒4.971, 2.054‒7.472, 1.807‒7.155, 1.506‒5.225, 1.241‒4.553, 1.666‒7.449, 1.483‒5.189, 2.050‒12.589, 2.552‒18.168, P<0.05. (2) Construction and validation of machine learning predic-tion models for neural invasion in gallbladder cancer. Based on the independent influencing factors, seven machine learning models were constructed, including logistic regression, K-nearest neighbors, support vector machine, random forest, decision tree, back-propagation neural network, and gradient boosting machine. The ROC curves of seven machine learning models in the test set were plotted, and the AUC were 0.900(95%CI as 0.851‒0.948), 0.741(95%CI as 0.646‒0.829), 0.836(95%CI as 0.762‒0.895), 0.782(95%CI as 0.701‒0.855), 0.839(95%CI as 0.770‒0.901), 0.817(95%CI as 0.738‒0.887), 0.843(95%CI as 0.770‒0.909), respectively. Results of Delong test showed that the logistic regression model had the highest AUC. The sensitivity and specificity of the logistic regression model were 0.868 and 0.805 respectively, indicating the best balance. Results of Hosmer-Lemeshow test showed that the logistic regression model had a good goodness-of-fit (χ²=5.320, P>0.05). The Brier score of the logistic regression model was relatively low, as 0.168, which verified its calibration advantage.

Conclusion Total bilirubin, carcinoembryonic antigen, CA199, CA125, neutrophil-to-lymphocyte ratio, liver invasion detected by enhanced CT, vascular invasion detected by enhanced CT, hilar or retroperitoneal lymph node metastasis detected by enhanced CT, and tumor stages T3 and T4 are independent influencing factors for nerve invasion in patients with gallbladder cancer. Seven machine learning models are constructed based on enhanced CT and clinical characteristics to predict neural invasion in gallbladder cancer, of which the logistic regression model demonstrates good predictive performance.