43P - Machine learning radiomics based on CT to predict response to lenvatinib plus tislelizumab based therapy for unresectable hepatocellular carcinoma

Background

Lenvatinib with PD-1 inhibitors showed promising efficacy in treating unresectable hepatocellular carcinoma (uHCC) but predicting treatment response (TR) remains challenging due to tumor heterogeneity. We aimed to develop models based on clinical and radiomics features using machine learning (ML) algorithms to predict the efficacy of lenvatinib (LEN) combined with tislelizumab (TIS) based therapy in uHCC patients (pts).

Methods

110 uHCC pts treated with LEN combined with TIS based therapy between 2019 and 2024 were recruited. Among them, 30 pts received LEN plus TIS, while 80 received the combination of LEN, TIS, and TACE. Clinical data and contrast-enhanced CT images were collected. TR was assessed by mRECIST. Five machine learning algorithms, including logistic regression, random forest (RF), neural networks, support vector machines and naive Bayes, were applied to develop prediction models based on clinical and radiomic features. The predictive performance of these models was evaluated using ROC curves, accuracy (ACC), calibration curves (CC), and decision curve analysis (DCA).

Results

Of 110 eligible pts, ORR was 33.6% with all partial response, and DCR was 75.9%. The cohort was randomly divided into a training set of 88 pts and a validation set of 22 pts. A total of 240 radiomic features were extracted, and the top 15 significant features were selected using the mRMR algorithm for further analysis. Prediction models based on 20 clinical variables and 15 radiomic features were developed using five different machine learning algorithms. The RF model showed the best predictive performance, with average area under the curve (AUC) of 0.99 and 0.72, and ACC values of 0.99 and 0.66 for the training and validation sets, respectively. CC and DCA further demonstrated the robust performance of the RF model in predicting treatment response.

Conclusions

Machine learning models based on clinical and radiomics features were successfully constructed to predict the efficacy of LEN and TIS based therapy in uHCC pts. These models offer potential guidance for personalized treatment strategies.

Clinical trial identification

NCT05543304.

Legal entity responsible for the study

The First Affiliated Hospital of Wenzhou Medical University.

Funding

Zhejiang Provincial Health Commission (WKJ-ZJ-2438).

Disclosure

All authors have declared no conflicts of interest.