Using Machine Learning to Predict Advanced-Stage Progression of Intermediate-Stage Hepatocellular Carcinoma after Transarterial Chemoembolization

Purpose To develop and test a machine learning (ML)-based model that integrates preoperative variables for prediction of advanced-stage progression (ASP) after transarterial chemoembolization (TACE). Materials and Methods This multicenter retrospective study (ResearchRegistry.com identifier no. researchregistry9425) included patients with intermediate-stage hepatocellular carcinoma (HCC) who underwent TACE at seven hospitals from June 2008 to December 2022. Thirty-four preoperative clinical and CT imaging variables were input into six ML-based models for prediction of ASP, and model performances were compared. Furthermore, the best-performing ML model was compared with the major staging systems, and its utility in performing post-TACE therapies was assessed. The performances of the models were compared by using area under the receiver operating characteristic curve (AUC) with DeLong test. Kaplan-Meier survival curves were compared using the log-rank test. Results A total of 2333 eligible patients (mean age, 54 years ± 12 [SD]; 2051 male patients) were categorized into the training set (<i>n</i> = 1026), the internal test set (<i>n</i> = 257), and the external test set (<i>n</i> = 1050). ASP was found in 8.4% (86 of 1026), 8.2% (21 of 257), and 6.7% (70 of 1050) of patients in the three datasets, respectively. Among all ML models, the Categorical Gradient Boosting (CatBoost) model yielded the highest AUC: 0.97 (95% CI: 0.95, >0.99) for the training set, 0.94 (95% CI: 0.92, 0.97) for the internal test set, and 0.93 (95% CI: 0.90, 0.95) for the external test set. Furthermore, it yielded better discriminatory ability with higher concordance indexes than the five staging systems (all <i>P</i> < .001). The time-dependent AUC of the CatBoost model was also higher than that of the clinical staging systems at various time points (all <i>P</i> < .001). Moreover, post-TACE systemic therapy improved progression-free survival and overall survival for patients in the high-risk group (both <i>P</i> < .001) but not in the low-risk group. Conclusion The CatBoost model demonstrated higher predictive performance compared with existing staging systems in predicting ASP after TACE in patients with intermediate-stage HCC. This model effectively stratified patients by risk level and identified those who benefited from post-TACE systemic therapy. <b>Keywords:</b> Liver, Oncology, Transarterial Chemoembolization, Hepatocellular Carcinoma, Advanced-stage Progression, Machine Learning, Risk Differentiation ResearchRegistry.com identifier no. researchregistry9425 <i>Supplemental material is available for this article.</i> © RSNA, 2025 See also commentary by Rouzbahani in this issue.