[Machine learning and radiomics-based model for predicting response to lenvatinib combined with TACE in patients with unresectable hepatocellular carcinoma].

Authors

Abstract

<b>Objective:</b> To evaluate the predictive value of machine learning combined with radiomics for treatment response to lenvatinib combined with transarterial chemoembolization (TACE) in patients with unresectable hepatocellular carcinoma (uHCC). <b>Methods:</b> This retrospective study enrolled 117 uHCC patients treated with lenvatinib-TACE at the First Affiliated Hospital of Wenzhou Medical University from January 2020 to December 2023. Patients were randomly divided into a training set (<i>n</i>=81) and an internal validation set (<i>n</i>=36) at a 7∶3 ratio. An additional 24 uHCC patients treated with lenvatinib-TACE between January and December 2024 were included as an external validation set. The follow-up period concluded in September 2025. Patients were stratified into a non-response group (PD+SD) and a response group (PR) based on therapeutic outcomes. Overall survival (OS) and progression-free survival (PFS) were estimated using the Kaplan-Meier method, with differences between groups compared via the log-rank test. Additionally, abdominal contrast-enhanced CT features were extracted from uHCC patients, and the top 25 most representative features were selected using the maximum relevance minimum redundancy (mRMR) algorithm. Six machine learning algorithms-multilayer perceptron (MLP), decision tree (DT), random forest (RF), logistic regression (LR), extreme gradient boosting (XGBoost), and support vector machine (SVM)-were employed to develop predictive models based on enhanced CT radiomic features after feature selection and dimensionality reduction. Model performance was evaluated using receiver operating characteristic (ROC) curves and area under the curve (AUC), and the differences in AUC were compared using Delong test. Calibration curves and decision curve analysis (DCA) were used to validate model accuracy and clinical utility, respectively. <b>Results:</b> The response group consisted of 57 patients [mean age: (60.5±9.8) years; 50 males], while the non-response group comprised 84 patients [mean age: (58.9±13.8) years; 74 males]. The overall median follow-up time was 31.7(25.3,44.8) months. Notably, both OS and PFS were significantly superior in the response group compared to the non-response group (both <i>P</i><0.05). The MLP model yielded an AUC of 0.856 (95%<i>CI</i>: 0.805-0.906) in the training set, 0.790 (95%<i>CI</i>: 0.641-0.927) in the internal validation set, and 0.836 (95%<i>CI</i>: 0.585-1.000) in the external validation set. In terms of model comparison, the MLP significantly outperformed the DT, RF, and XGBoost models in the training set, and surpassed the DT model in the internal validation set (all <i>P</i><0.05); however, no statistically significant differences were observed against other models in the external validation set. Notably, the MLP was the only model to demonstrate consistent calibration across all three datasets, and DCA confirmed its superior clinical net benefit. <b>Conclusion:</b> This study developed six machine learning prediction models based on imaging data, among which the MLP model demonstrated optimal performance in predicting treatment response to the combination therapy of lenvatinib and TACE, enabling effective efficacy prediction and clinical decision-making guidance for uHCC patients.

Tags

Topics