Artificial intelligence in transarterial chemoembolization for hepatocellular carcinoma: from treatment response prediction to refractoriness assessment.

Background: Hepatocellular carcinoma (HCC) is the most prevalent primary hepatic malignancy worldwide and a leading cause of cancer-related mortality. Transarterial chemoembolization (TACE) is the standard of care for Barcelona Clinic Liver Cancer (BCLC) intermediate-stage HCC; however, treatment response varies considerably among individuals, and a substantial proportion of patients develop TACE refractoriness. Objective: This review systematically examines the current applications of artificial intelligence (AI) in TACE for HCC, encompassing treatment response prediction, survival prognostication, refractoriness prediction, and the underlying molecular mechanisms, while critically appraising methodological quality of the existing literature and delineating TACE-specific future directions. Methods: We conducted a comprehensive literature search of PubMed, Embase, Web of Science, Scopus, and Google Scholar for studies published through 2025 that investigated AI, machine learning (ML), or deep learning (DL) in TACE response prediction, prognostic stratification, and refractoriness assessment. Included studies were appraised against PROBAST, IBSI, TRIPOD+AI, and CLAIM frameworks. Results: AI models based on radiomics and DL demonstrated high discriminative performance for predicting TACE outcomes, with meta-analytic area under the receiver operating characteristic curve (AUROC) values ranging from 0.81 to 0.92. Combined clinico-radiological models-incorporating albumin-bilirubin (ALBI) grade, BCLC stage, alpha-fetoprotein (AFP) level, tumor diameter, distribution, and peritumoral arterial-phase enhancement-consistently outperformed single-source models. Convolutional neural networks (CNNs) and gradient-boosting, support-vector-machine, and random-forest models were the consistently top-performing algorithms. However, methodological quality was uneven: most studies showed high or unclear PROBAST risk of bias in the analysis domain, IBSI-compliant feature reporting and public code release were rare, and the generalization gap between internal and external validation cohorts averaged 0.08-0.15 AUROC. Conclusions: AI offers a powerful toolkit for individualized TACE decision-making, with the potential to shift clinical practice from experience-driven to data-driven precision therapy. Nevertheless, critical challenges remain-including insufficient external validation, limited sample sizes, geographic bias, low standardization, and inadequate interpretability-which together necessitate large-scale, multicenter, prospective studies adhering to harmonized reporting standards.