Clinical value of machine learning model combining test data and enhanced CT images to predict primary hepatocellular carcinoma

Objective To investigate the clinical value of machine learning methods for the detection of primary hepatocellular carcinoma (HCC) using laboratory data and enhanced CT imaging. Methods The enhanced CT images and test data of 654 patients with liver diseases admitted in Shiyan Taihe Hospital from 2013 to 2021 were retrospectively analyzed. These consisted of 199 patients with primary HCC, 223 patients with hepatitis, and 232 patients with liver cirrhosis. They were randomly allocated to the training set and testing set at 7∶3. Five machine learning algorithms (including logistic regression, support vector machine, random forest, decision tree, and AdaBoost) were implemented to train the model using test data, CT enhanced images, and combined test data/CT images. Receiver operating characteristic curve was used to calculate the area under curve (AUC), accuracy, sensitivity and specificity to verify the model. Results From the analysis of data modality, the use of a multimodal method combining test data and enhanced CT images significantly improved classification accuracy of primary HCC compared to the unimodal modality. From the analysis of the machine learning models: (1) The detection efficiency of primary HCC in the test dataset was higher when using the random forest algorithm, with an accuracy rate of 84.77%, specificity of 89.96%, sensitivity of 86.11% and AUC value of 0.889; (2) The detection efficiency of primary HCC was higher when applying the decision tree algorithm to the arterial phase enhanced CT group, with an accuracy rate of 76.14%, specificity of 77.84%, sensitivity of 61.90%, and AUC value of 0.789; (3) The AdaBoost algorithm applied to the combined data set of test data and arterial phase enhanced CT data demonstrated higher detection efficiency for primary HCC, with an accuracy rate of 87.31%, specificity of 88.28%, sensitivity of 84.62%, and AUC value of 0.918. Conclusion The construction of models using test data and enhanced CT images with multimodal data offers enhanced prediction of primary HCC. Notably, the constructed AdaBoost model based on test data and arterial phase enhanced CT radiomics features can exhibit greater predictive accuracy.