Background Targeted-immune systemic therapy (TIST) is currently the first-line systemic treatment for hepatocellular carcinoma (HCC). However, there is a lack of reliable biomarkers for predicting efficacy and identifying potential beneficiaries.

Objective To explore the feasibility of using Magnetic Resonance Imaging (MRI) radiomic features for response prediction.

Methods This study included 191 HCC patients who received targeted immunotherapy at the Department of Hepatobiliary and Pancreatic Surgery, Chinese PLA General Hospital, from January 2019 to January 2023. The patients were randomly divided into training cohort and validation cohort at a 7:3 ratio. All patients underwent dynamic contrast-enhanced MRI before treatment and during follow-up. Treatment efficacy was assessed according to RECIST1.1 criteria, with complete or partial response considered effective. Tumor regions were delineated on pre-treatment dynamic contrast-enhanced MRI images, and radiomic features were extracted. Meaningful radiomic features were selected using the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm. Various machine learning algorithms, including logistic regression, k-nearest neighbor (KNN), naive Bayes, neural network, support vector machine (SVM), decision tree, XGBoost, and random forest were used to construct radiomic prediction models. The performance of each model was evaluated using receiver operating characteristic (ROC) curves and confusion matrices. The prognostic predictive value of the models was analyzed using Cox regression and Kaplan-Meier curves.

Results Medical data of 191 patients were collected, including 163 male patients (85.3%) and 28 female patients (14.7%), with a median age of 55 (range: 24-74) years. The overall objective response rate was 30.4%. There were 133 cases in the training set and 58 cases in the validation set, with no statistically significant differences in gender and age between the two groups (P > 0.05), the objective response rates in training cohort and validation cohort were 30.1% and 31.0%, respectively. Eight ML models were successfully constructed based on the selected radiomic features. In the training cohort with 10-fold cross-validation, the AUC values for the KNN, naive Bayes, and SVM models were 0.826, 0.810, and 0.801, respectively. In the validation cohort, the AUC values were 0.830, 0.836, and 0.825, respectively. Survival analysis showed that the progression-free survival of patients predicted to have effective treatment by the above three models was significantly prolonged (P < 0.05).

Conclusion Radiomic models based on machine learning algorithms and contrast-enhanced MRI can accurately predict the efficacy of TIST in HCC patients, demonstrating potential utility in assisting clinical decision-making.