Construction and evaluation of a preoperative prediction model for hepatocellular carcinoma with microvascular invasion based on machine learning algorithm
10.3760/cma.j.cn113884-20230703-00193
- VernacularTitle:基于机器学习算法肝细胞癌伴微血管侵犯术前预测模型的构建与评估
- Author:
Yubo ZHANG
1
;
Peng LEI
;
Yang BO
;
Gang YANG
;
Wei ZHANG
;
Danyang ZHANG
Author Information
1. 宁夏医科大学研究生院，银川 750004
- Keywords:
Carcinoma, hepatocellular;
Machine learning;
Microvascular invasion;
Predictive model
- From:
Chinese Journal of Hepatobiliary Surgery
2023;29(11):801-807
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To screen preoperative microvascular invasion (MVI)-related indicators in patients with hepatocellular carcinoma by machine learning, and to construct a predictive model for predicting MVI and evaluate it.Methods:The clinical data of hepatocellular carcinoma patients who underwent radical resection from January 2018 to March 2023 in General Hospital of Ningxia Medical University were retrospectively analyzed. A total of 437 patients were enrolled, including 325 males and 112 females, aged (56.3±13.6) years. The 437 patients were divided into a training set ( n=305) and a test set ( n=132) by computer-generated random numbers on a 7∶3 basis; the training set was used to construct the predictive model as well as to internally validate it by the five-fold cross-validation method, and the test set was used to externally validate the model. Two machine learning Boruta algorithm and LASSO regression, were used to screen MVI characteristic variables and construct multifactorial logistic regression prediction models. Receiver operating characteristic (ROC) curve, calibration curves, and decision curve were evaluated for predictive modeling, applying Shapley's additive explanatory analysis (SHAP) of the significance of key variables. Results:The intersection (5 variables) of 8 characteristic variables selected by Boruta algorithm and 8 variables selected by LASSO regression were selected: aspartate aminotransferase/lymphocyte ratio (ALR), tumor margin, intratumbral necrosis, tumor number and tumor maximum diameter, and the logistic regression model was constructed. The area under ROC curve for predicting the MVI were 0.77 (95% CI: 0.70-0.82) (training set), 0.76 (95% CI: 0.63-0.87) (validation set), and 0.84 (95% CI: 0.78-0.91) (test set). The prediction results of calibration curve logistic regression model were close to those of reagent, and the analysis of decision curve indicates that the model had good clinical application value. According to the mean absolute SHAP value, the order of importance was tumor margin, tumor maximum diameter, tumor number, ALR, and intratumoral necrosis. Conclusion:Tumor margin, tumor maximum diameter, tumor number, ALR and intratumoral necrosis were independent influencing factors for hepatocellular carcinoma associated with MVI, and the logistic regression model based on these factors was effective in predicting MVI.
