Application value of machine learning algorithms for preoperative prediction of microvascular invasion in hepatocellular carcinoma
10.3760/cma.j.issn.1673-9752.2020.02.008
- VernacularTitle:机器学习算法在肝细胞癌微血管侵犯术前预测中的应用价值
- Author:
Hongzhi LIU
1
;
Haitao LIN
;
Zhaowang LIN
;
Jun FU
;
Zongren DING
;
Pengfei GUO
;
Jingfeng LIU
Author Information
1. 福建医科大学孟超肝胆医院东南肝胆健康大数据研究所 福建省肝胆医学中心,福州 350025
- From:
Chinese Journal of Digestive Surgery
2020;19(2):156-165
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To investigate the application value of machine learning algorithms for preoperative prediction of microvascular invasion (MVI) in hepatocellular carcinoma (HCC).Methods:The retrospective and descriptive study was conducted. The clinicopathological data of 277 patients with HCC who were admitted to Mengchao Hepatobiliary Hospital of Fujian Medical University between May 2015 and December 2018 were collected. There were 235 males and 42 females, aged (56±10)years, with a range from 33 to 80 years. Patients underwent preoperative magnetic resonance imaging examination. According to the random numbers showed in the computer, all the 277 HCC patients were divided into training dataset consisting of 193 and validation dataset consisting of 84, with a ratio of 7∶3. Machine learning algorithms, including logistic regression nomogram, support vector machine (SVM), random forest (RF), artificial neutral network (ANN) and light gradient boosting machine (LightGBM), were used to develop models for preoperative prediction of MVI. Observation indicators: (1) analysis of clinicopathological data of patients in the training dataset and validation dataset; (2) analysis of risk factors for tumor MVI of the training dataset; (3) construction of machine learning algorithm prediction models and comparison of their accuracy of preoperative tumor MVI prediction. Measurement data with normal distribution were represented as Mean± SD, and comparison between groups was analyzed using the paired t test. Count data were described as absolute numbers, and comparison between groups was analyzed using the chi-square test. Univariate and multivariate analyses were performed using the Logistic regression model. Results:(1) Analysis of clinicopathological data of patients in the training dataset and validation dataset: there were 157 males and 36 females in the training dataset, 78 males and 6 females in the validation dataset, showing a significant difference in the sex between the training dataset and validation dataset ( χ2=6.028, P<0.05). (2) Analysis of risk factors for tumor MVI of the training dataset: of the 193 patients, 108 had positive MVI, and 85 had negative MVI. Results of univariate analysis showed that age, the number of tumors, tumor diameter, satellite lesions, tumor margin, alpha fetaprotein (AFP), alkaline phosphatase (ALP), fibrinogen were related factors for tumor MVI [ odds ratio ( OR)=0.971, 2.449, 1.368, 4.050, 2.956, 4.083, 2.532, 1.996, 95% confidence interval ( CI): 0.943-1.000, 1.169-5.130, 1.180-1.585, 1.316-12.465, 1.310-6.670, 2.214-7.532, 1.016-6.311, 1.323-3.012, P<0.05]. Results of multivariate analysis showed that AFP>20 μg/L, multiple tumors, larger tumor diameter, unsmooth tumor margin were independent risk factors for tumor MVI ( OR=3.680, 3.100, 1.438, 3.628, 95% CI: 1.842-7.351, 1.334-7.203, 1.201-1.721, 1.438-9.150, P<0.05). Larger age was associated with lower risk of preoperative tumor MVI ( OR=0.958, 95% CI: 0.923-0.994, P<0.05). (3) Construction of machine learning algorithm prediction models and comparison of their accuracy of preoperative tumor MVI prediction: ①machine learning algorithm prediction models involving logistic regression nomogram, SVM, RF, ANN and LightGBM were constructed based on results of multivariate analysis including age, AFP, the number of tumors, tumor diameter, tumor margin, and consistency analysis of the logistic regression nomogram prediction model showed a good stability. For the training dataset and validation dataset, the area under curve (AUC) of logistic regression nomogram model, SVM model, RF model, ANN model, LightGBM model was 0.812, 0.794, 0.807, 0.814, 0.810 and 0.784, 0.793, 0.783, 0.803, 0.815, respectively, showing no significant difference between SVM model and logistic regression nomogram model, between RF model and logistic regression nomogram model, between ANN model and logistic regression nomogram model, between LightGBM model and logistic regression nomogram model [(95% CI: 0.731-0.849, 0.744-0.860, 0.752-0.867, 0.747-0.862, Z=0.995, 0.245, 0.130, 0.102, P>0.05) and (95% CI: 0.690-0.873, 0.679-0.865, 0.702-0.882, 0.715-0.891, Z=0.325, 0.026, 0.744, 0.803, P>0.05)]. ② Clinicopathological factors were selected using RF, LightGBM machine learning algorithm to construct corresponding prediction models. According to importance scale of factors to prediction models, factors with importance scale>0.01 were selected to construct RF model, including age, tumor diameter, AFP, white blood cell, platelet, total bilirubin, aspartate transaminase, γ-glutamyl transpeptidase, ALP, and fibrinogen. Factors with importance scale>5.0 were selected to construct LightGBM model, including age, tumor diameter, AFP, white blood cell, ALP, and fibrinogen. Due to lack of factor selection ability, factors based on results of univariate analysis were secected to construct SVM model and ANN model, including age, the number of tumors, tumor diameter, satellite lesions, tumor margin, AFP, ALP, and fibrinogen. For the training dataset and validation dataset, the AUC of SVM model, RF model, ANN model, LightGBM model was 0.803, 0.838, 0.793, 0.847 and 0.810, 0.802, 0.802, 0.836, respectively, showing no significant difference between SVM model and logistic regression nomogram model, between RF model and logistic regression nomogram model, between ANN model and logistic regression nomogram model, between LightGBM model and logistic regression nomogram model [(95% CI: 0.740-0.857, 0.779-0.887, 0.729-0.848, 0.789-0.895, Z=0.421, 0.119, 0.689, 1.517, P>0.05) and (95% CI: 0.710-0.888, 0.700-0.881, 0.701-0.881, 0.740-0.908, Z=0.856, 0.458, 0.532, 1.306, P>0.05)]. Conclusion:Machine learning algorithms can predict MVI of HCC preoperatively, but its application value needs to be further verified by large sample data from multi centers.