Highlight
This study demonstrates that machine learning-based radiomic models derived from pre-treatment CT imaging, combined with clinical data, outperform established clinical biomarkers such as BCLC stage and ALBI grade in predicting survival and response to atezolizumab plus bevacizumab immunotherapy in hepatocellular carcinoma (HCC). The integrated model accurately stratified patients into high- and low-risk groups with significant differences in overall survival (OS), progression-free survival (PFS), and immune checkpoint inhibitor response rates, validated across independent international cohorts.
Study Background and Disease Burden
Hepatocellular carcinoma (HCC) is a leading cause of cancer-related mortality worldwide, often diagnosed at an advanced stage, making curative treatment options limited. Atezolizumab plus bevacizumab (A/B) has emerged as a first-line immunotherapy regimen for unresectable HCC, improving clinical outcomes compared with prior standards. However, only a minority of patients respond favorably to this combination, highlighting the urgent need for robust prognostic and predictive biomarkers to guide personalized treatment decisions. Current clinical biomarkers, including BCLC stage and ALBI grade, have limited predictive accuracy for immunotherapy response and survival outcomes in this setting. Radiomics — the high-throughput extraction of quantitative imaging features — in conjunction with advanced machine learning techniques, offers a promising approach to capture tumor heterogeneity and host factors from routine imaging, potentially improving outcome prediction without additional invasive procedures.
Study Design
This retrospective multicenter study included 152 patients with unresectable HCC treated with atezolizumab plus bevacizumab at two international institutions: Imperial College London (ICL) and Assistance Publique-Hôpitaux de Paris (AP-HP). Pre-treatment computed tomography (CT) scans were used for analysis. Deep learning models were employed for automated segmentation of the whole liver, enabling standardized radiomic feature extraction. These radiomic features were combined with clinical variables to develop predictive models for 12-month mortality following immunotherapy initiation.
Multiple machine learning algorithms (seven in total) were evaluated in conjunction with thirteen feature selection methods to identify optimal predictive models in the training ICL cohort. K-means clustering stratified patients into high- and low-risk groups. Model performance was then independently validated in the AP-HP cohort. Primary endpoints included overall survival (OS), progression-free survival (PFS), and response rates to immune checkpoint inhibitors.
Key Findings
The integrated radiomic-clinical model significantly outperformed conventional clinical biomarkers. In the training ICL cohort, the model achieved an area under the receiver operating characteristic curve (AUC) of 0.89 (95% CI 0.75–0.99), compared to 0.61 for the BCLC stage and 0.48 for ALBI grade (both p < 0.001). Validation in the independent AP-HP cohort confirmed robust predictive ability with an AUC of 0.75 (95% CI 0.64–0.85).
The model-stratified high-risk group exhibited notably shorter median OS compared to the low-risk group in both cohorts (ICL: 5.6 vs. 28.2 months, AP-HP: 5.8 vs. 15.7 months, p < 0.001). Similarly, progression-free survival was reduced in high-risk patients (ICL: 2.4 vs. 14.6 months, p < 0.001; AP-HP: 2.1 vs. 6.1 months, p = 0.046). Furthermore, low-risk patients had significantly higher immune checkpoint inhibitor response rates than high-risk patients (35.6% vs. 21.4%; p = 0.038).
Multivariable Cox regression analysis identified the radiomic risk group as the strongest independent predictor of OS (hazard ratio [HR] 3.22, 95% CI 1.99–5.20, p < 0.001) and PFS (HR 1.82, 95% CI 1.18–2.80, p = 0.010), underscoring its prognostic value beyond traditional clinical metrics.
Expert Commentary
This study represents a significant advancement in precision oncology for HCC by integrating advanced imaging analytics with machine learning to address a critical unmet need: accurately predicting patient outcomes following immunotherapy. The use of deep learning for automated hepatic segmentation standardizes radiomic extraction, minimizing user-dependent variability and enhancing model reproducibility. The multi-institutional validation strengthens the generalizability of the findings.
Nevertheless, certain limitations warrant consideration. The retrospective design introduces potential biases, and although the sample size is reasonable, larger prospective studies are needed to definitively establish clinical utility. Additionally, future investigations could explore integration of genomic and blood-based biomarkers alongside radiomics for further refined prediction. Despite these caveats, this approach exemplifies how leveraging routine clinical imaging data can improve risk stratification and guide treatment personalization in advanced HCC.
Current guidelines do not yet incorporate radiomic biomarkers, partly due to the novelty and technical demands of these methodologies. However, this study’s findings may catalyze future guideline updates as evidence accumulates. On a mechanistic level, radiomic features likely capture tumor heterogeneity, vascular characteristics, and immune microenvironment correlates affecting immunotherapy responsiveness.
Conclusion
Integrated machine learning-based radiomic models combining pre-treatment CT imaging and clinical variables outperform traditional clinical biomarkers to predict survival and immunotherapy response in unresectable HCC patients treated with atezolizumab plus bevacizumab. These predictive tools enable robust risk stratification into distinct prognostic groups, facilitating precision treatment approaches and potentially improving patient outcomes. Prospective validation and incorporation of multimodal biomarkers remain important next steps to advance personalized medicine in advanced hepatocellular carcinoma.
References
Vithayathil M, Koku D, Campani C, Nault JC, Sutter O, Ganne-Carrié N, Aboagye EO, Sharma R. Machine learning based radiomic models outperform clinical biomarkers in predicting outcomes after immunotherapy for hepatocellular carcinoma. J Hepatol. 2025 Oct;83(4):959-970. doi: 10.1016/j.jhep.2025.04.017. Epub 2025 Apr 17. PMID: 40246150.
