Integrating B-mode ultrasound radiomics and clinical variables to predict 26-week progression-free survival in advanced hepatocellular carcinoma under targeted immunotherapy.

Authors

Abstract

Hepatocellular carcinoma (HCC) is the third leading cause of cancer-related mortality globally, with over 905,700 new cases and 830,200 deaths in 2022, and its incidence is projected to rise due to increasing metabolic risk factors. Targeted immunotherapy has revolutionized treatment for advanced HCC, yet response rates remain low (∼30%), necessitating reliable predictive tools to identify non-responders early and personalize care. While radiomics from CT/MRI has shown promise, ultrasound-based multimodal models are underexplored despite ultrasound's accessibility. To develop and validate a multimodal model integrating clinical features, ultrasound characteristics, and radiomics for predicting 26-week progression-free survival (PFS) in advanced HCC patients receiving targeted immunotherapy. In this retrospective cohort study at Quanzhou First Hospital, 222 patients with advanced HCC (BCLC stage B/C) were split 7:3 into training (n=155) and validation (n=67) cohorts. Clinical variables (e.g., tumor diameter, margins, pseudocapsule, GGT, CA19-9) and ultrasound radiomics features were extracted from baseline B-mode images using PyRadiomics. Seventeen machine learning algorithms were benchmarked; Random Forest was selected to build clinical, radiomics, and combined models. Performance was evaluated via ROC curves (AUC), calibration plots, decision curve analysis (DCA), net reclassification improvement (NRI), integrated discrimination improvement (IDI), and Kaplan-Meier survival analysis with log-rank tests. The combined model achieved AUCs of 0.979 (training) and 0.978 (validation), superior to clinical (0.870/0.810) and radiomics (0.770/0.852) models. SHAP analysis identified tumor diameter, obscure margins, pseudocapsule, IBIL, CA19-9, and radiomics textures (e.g., entropy, kurtosis) as top predictors. Calibration was excellent (Brier score < 0.1), and DCA showed highest net benefit. NRI/IDI confirmed added value (negative for unimodal models). Risk stratification (0.50 cutoff) separated high- vs. low-risk groups with significant OS differences (log-rank p<0.0001; median OS ∼200/250 days high-risk vs. unreached low-risk). The combined clinical-ultrasound-radiomics model robustly predicts 26-week OS in advanced HCC under targeted immunotherapy, outperforming unimodal approaches and offering a cost-effective tool for personalized management. Prospective multicenter validation is warranted.

Tags

Topics