Artificial intelligence in prognostication and treatment response modeling in advanced hepatocellular carcinoma

Abstract

Advanced hepatocellular carcinoma (HCC) presents a dual challenge of tumor progression and underlying hepatic dysfunction, resulting in substantial heterogeneity in survival and treatment response. Artificial intelligence (AI) has been increasingly applied to address limitations of conventional stage-based systems; however, its impact on clinical decision-making remains uncertain. We conducted a narrative critical review of primary validation studies and high-quality systematic reviews (2020-2026) evaluating AI-based survival prediction and treatment response modeling in advanced HCC. We focused on methodological rigor, external and temporal validation, calibration, causal inference, and translational readiness. Several retrospective studies report improved discrimination for radiomic, deep learning, and multimodal machine learning models compared with conventional frameworks such as Barcelona Clinic Liver Cancer staging and albumin-bilirubin grade. Absolute area under the receiver operating characteristic curve improvements of approximately 0.14-0.40 have been reported in selected cohorts. These models improve intra-stage risk stratification and have the potential to predict response to systemic and locoregional therapies. However, most studies remain retrospective and focus primarily on prognostic rather than predictive modeling. External validation, calibration assessment, and competing-risk adaptation are inconsistently reported, and causal approaches to estimate treatment-effect heterogeneity are rarely used. Evidence demonstrating improvement in therapeutic allocation or patient-centered outcomes remains limited. AI-based modeling in advanced HCC offers improved risk stratification, yet improved discrimination metrics alone do not establish clinical utility. Prospective validation, causal inference frameworks, standardized reporting, and demonstration of decision impact are essential prerequisites for clinical integration. Without such rigor, AI may remain technologically sophisticated yet clinically unproven.

Key Words: Artificial intelligence; Hepatocellular carcinoma; Radiomics; Prognostic modeling; Predictive modeling; Survival analysis; Causal inference; Precision oncology

Core Tip: Artificial intelligence models in advanced hepatocellular carcinoma frequently demonstrate improved discrimination compared with conventional staging systems; however, most remain retrospective, prognostic rather than predictive, and inconsistently validated across institutions and therapeutic eras. This review critically distinguishes statistical performance gains from clinically meaningful advancement, emphasizing the need for causal modeling, competing-risk adaptation, calibration assessment, and prospective decision-impact evaluation before routine integration into multidisciplinary hepatology practice.

INTRODUCTION

Hepatocellular carcinoma (HCC) is the sixth most common malignancy worldwide and the third leading cause of cancer-related mortality. Its incidence and related death rates continue to rise globally[1-3]. Unlike many solid tumors, HCC typically develops in the background of chronic liver disease or cirrhosis. Hence, it is a dual disease in which both tumor burden and hepatic dysfunction independently influence prognosis and therapeutic decision-making[3]. Despite advances in systemic therapy, including immune checkpoint inhibitors and combination regimens, treatment allocation in advanced HCC remains largely stage-based and empirical[3,4]. Current prognostic systems were not developed in the immunotherapy era and do not adequately capture tumor biology, treatment-response heterogeneity, and competing risks from hepatic decompensation. Consequently, clinicians lack validated tools to individualize therapy beyond conventional staging frameworks[5].

Recent therapeutic advances, including multikinase inhibitors and immune checkpoint inhibitors, have modestly improved survival. However, clinical outcomes remain highly variable even within the same staging category. This intra-stage heterogeneity highlights the limitations of conventional prognostic systems. Conventional prognostic systems often rely on a restricted set of clinical and radiologic variables and may inadequately capture the biological complexity of advanced disease. Furthermore, multidisciplinary decision-making, considered the standard of care in HCC, relies heavily on clinicians' interpretation of diverse, often fragmented datasets. This highlights the need for more integrative and scalable analytic approaches[6].

Artificial intelligence (AI) has emerged as a promising tool capable of addressing these limitations. AI encompasses computational algorithms that learn patterns from large-scale datasets, including machine learning (ML) and its subset, deep learning, which leverages multilayered neural networks to model complex, nonlinear relationships[2]. In the context of HCC, AI models can integrate heterogeneous data sources, including imaging, radiomics, histopathology, electronic health records, and molecular biomarkers. This integration enhances risk stratification, prognostication, and prediction of therapeutic response[2,3].

However, whether AI improves clinical decision-making, beyond statistical discrimination, remains uncertain[7,8]. Key unanswered questions include: (1) Does AI improve survival prediction when accounting for competing risks? (2) Can AI reliably predict treatment benefit rather than prognosis alone? (3) Are performance gains robust across institutions and imaging platforms? And (4) Does improved model discrimination translate into changes in therapeutic allocation or patient outcomes?

Recent studies demonstrate that AI-driven models often outperform traditional statistical approaches in predicting survival outcomes and treatment responses in advanced HCC. These models achieve superior discrimination metrics, such as the concordance index (C-index) and the area under the receiver operating characteristic curve (AUROC)[2,3]. Beyond predictive accuracy, AI offers efficiency and scalability in processing high-dimensional data. This enables real-time clinical insights that can support personalized treatment selection.

Despite these promising developments, challenges remain regarding model generalizability, interpretability, data standardization, and external validation. Nonetheless, the integration of AI into advanced HCC care could represent a shift toward precision oncology. It has the potential to refine prognostic assessment and optimize therapeutic strategies in a disease marked by biological complexity and clinical uncertainty[7,9].

This review is a narrative critical review synthesizing recent primary validation studies and high-quality systematic reviews published from 2020 to 2026. We emphasized methodological rigor, external validation, and translational readiness. Accordingly, this review critically evaluates AI-driven survival prediction and treatment-response modeling in advanced HCC and compares their performance with that of conventional prognostic systems. It also examines methodological limitations, including overfitting and gaps in external validation, and delineates the translational barriers that must be addressed before clinical implementation.

Advanced HCC is characterized by substantial biological heterogeneity arising from diverse etiologies, including viral hepatitis, alcohol-related liver disease, and metabolic dysfunction. Complex molecular and microenvironmental alterations also contribute to this heterogeneity. Tumor progression and clinical outcomes reflect the interplay between tumor biology and underlying hepatic reserve.

AI models aim to capture this complexity by integrating multiscale data. These include imaging-derived phenotypes (radiomics), histopathological architecture (pathomics), and clinical variables reflecting liver function and systemic status. In this context, AI does not operate independently of disease biology but rather serves as a computational framework to model the multidimensional interactions underlying prognosis and treatment response[2-4]. Rather than advocating indiscriminate adoption, we aim to distinguish improvements in statistical performance from clinically meaningful advancement, thereby defining the realistic trajectory of AI integration in precision hepatology.

This is a narrative critical review synthesizing evidence from primary validation studies and high-quality systematic reviews published from 2020 to 2026. We identified relevant literature through searches of PubMed, Scopus, and Web of Science. We used keywords related to AI, HCC, prognostic modeling, and treatment response.

Studies were selected based on relevance to survival prediction and treatment response modeling in advanced HCC, with emphasis on methodological rigor, external validation, calibration, and translational applicability. This review does not follow a formal systematic review methodology (e.g., PRISMA)[10], and no quantitative meta-analysis was performed.

AI METHODOLOGIES IN ADVANCED HCC

AI has evolved beyond descriptive modeling to support biologically informed risk stratification and prediction of treatment response[11]. AI applications in advanced HCC can be categorized along four methodological axes (Table 1): (1) Learning paradigm (supervised vs unsupervised); (2) Clinical objective (prognostic vs predictive); (3) Outcome structure (binary classification vs time-to-event modeling); and (4) Data modality (radiomics, deep learning imaging, pathomics, clinical ML, and multimodal fusion). This framework provides a structured approach to evaluating methodological robustness and translational readiness.

Table 1 Methodological taxonomy of artificial intelligence applications in advanced hepatocellular carcinoma.

Axis	Categories	Typical application in advanced HCC	Key methodological risk
Learning paradigm	Supervised/unsupervised	Survival prediction, response modeling vs clustering phenotypes	Overfitting in supervised models
Clinical objective	Prognostic/predictive	OS estimation vs treatment benefit estimation	Confounding by indication
Outcome structure	Binary/time-to-event/competing risk	12-month mortality vs OS vs liver failure-specific death	Improper censoring handling
Data modality	Radiomics/deep learning imaging/pathomics/clinical ML/multimodal	CNN imaging models; LASSO radiomics; radiopathomics	Feature instability, dimensionality inflation

HCC: Hepatocellular carcinoma; OS: Overall survival; ML: Machine learning; CNN: Convolutional neural network; LASSO: Least absolute shrinkage and selection operator.

Supervised prognostic modeling (binary and survival)

The majority of AI applications in advanced HCC are supervised prognostic models trained to predict overall survival or fixed-time mortality. These models typically fall into two categories: Binary classification frameworks and time-to-event survival modeling approaches[12-15]. AI-based survival modeling in advanced HCC generally follows two distinct methodological approaches: (1) Binary classification models, which predict fixed-time endpoints such as 6- or 12-month mortality using metrics such as AUROC; and (2) Time-to-event survival models, which account for censoring and variable follow-up using Cox-based ML models, random survival forests, or deep survival networks.

Binary classifiers are simpler to implement but may ignore censoring and temporal dynamics. In contrast, survival-specific ML models use the C-index to evaluate discrimination across time-to-event outcomes. They are more appropriate for advanced HCC, where follow-up duration and event timing vary considerably.

Importantly, patients with advanced HCC frequently experience competing risks, such as liver failure-related mortality, independent of tumor progression. Models that fail to account for competing events may overestimate prognostic accuracy. Competing-risk-adapted ML frameworks remain underrepresented in current literature.

Within these frameworks, radiomics-based models have been particularly prominent in advanced HCC. Various models incorporating Barcelona Clinic Liver Cancer (BCLC), albumin-bilirubin (ALBI), alpha-fetoprotein level, tumor diameter, and peritumoral enhancement have been evaluated in the setting of transarterial chemoembolization (TACE). These models significantly outperformed clinical models[7]. Integrated radiomic-clinical models demonstrated improved discrimination compared with clinical-only approaches in TACE-treated cohorts. These findings suggest incremental prognostic value from integrating imaging-derived texture and vascular features with clinical variables.

A central methodological concern in radiomic survival modeling is the imbalance between the high-dimensional feature space and a limited number of events. Many radiomic studies extract hundreds of imaging features while operating within relatively small cohorts. This raises concerns about inadequate event-per-variable ratios and model overfitting[13,16].

To mitigate this risk, dimensionality reduction strategies such as least absolute shrinkage and selection operator regularization, principal component analysis, or feature stability filtering are commonly employed. However, the robustness of these approaches depends heavily on the validation strategy. Cross-validation alone may overestimate performance if feature selection occurs before data partitioning. Nested cross-validation frameworks are more appropriate to prevent information leakage during hyperparameter tuning and feature selection[16]. Data leakage, particularly when preprocessing, normalization, or feature filtering is conducted on the full dataset before validation, remains an underreported but significant threat to inflated performance metrics in HCC radiomics studies[17].

Radiomic feature stability and reproducibility remain critical methodological concerns. Radiomic signatures are sensitive to variations in image acquisition parameters, reconstruction algorithms, and scanner type, which may compromise inter-institutional transportability. Test-retest reproducibility studies have demonstrated that a substantial proportion of handcrafted radiomic features exhibit variability under repeated imaging conditions[16,18,19].

Furthermore, segmentation-dependent variability, particularly interobserver differences in region-of-interest delineation, can significantly influence extracted feature values. Stability assessment using intraclass correlation coefficients and robustness filtering before model development is therefore essential[20].

Statistical harmonization techniques, such as ComBat, have been increasingly employed to mitigate interscanner variability. However, their adoption in advanced HCC survival modeling remains inconsistent. Most radiomic survival models in advanced HCC are validated primarily for short- to intermediate-term endpoints (e.g., 6-12 months). Robust long-term overall survival prediction remains less developed.

More recently, survival modeling in the immunotherapy era has further illustrated AI’s capacity to refine risk discrimination. In immunotherapy-treated cohorts, integrated radiomic-clinical models demonstrated improved discrimination compared with conventional staging systems. This was observed in both derivation and external validation settings[21]. Notably, these models were developed for fixed-time mortality classification rather than dynamic survival prediction. Whether similar performance gains would persist under time-dependent discrimination or competing-risk survival modeling remains uncertain.

Survival modeling in advanced HCC is inherently complex due to censoring, variable follow-up duration, and competing risks such as liver failure-related mortality. Current radiomic and ML models demonstrate methodological promise. However, broader adoption of survival-specific and competing-risk-adapted frameworks, along with calibration assessment and external validation, is required to align AI prognostic tools with real-world cirrhotic populations[22].

Predictive modeling for treatment benefit

It is critical to distinguish between prognostic and predictive modeling in advanced HCC (Table 2). A prognostic model estimates clinical outcomes (e.g., overall survival or progression-free survival) irrespective of treatment exposure. In contrast, a predictive model estimates the differential treatment benefit, identifying which patients derive greater benefit from one therapy than from alternative strategies[23,24].

Table 2 Prognostic vs predictive modeling in advanced hepatocellular carcinoma.

Feature	Prognostic model	Predictive model
Core question	What is the patient’s outcome risk?	Does the patient benefit more from treatment A vs B?
Treatment consideration	Ignored or uniform	Central to model structure
Typical dataset	Single-treatment cohort	Multi-treatment or counterfactual framework
Key bias risk	Overfitting	Confounding by indication
Statistical requirement	Discrimination and calibration	Causal inference methods (propensity modeling, counterfactual ML, uplift models)
Clinical utility	Risk stratification	Therapy selection guidance

ML: Machine learning.

In the current HCC literature, most AI models described as “treatment response predictors” are in fact prognostic models developed in single-treatment cohorts. Such models may identify patients with favorable tumor biology, but do not establish treatment-specific causal benefit.

True predictive modeling requires explicit handling of confounding by indication. This is because treatment allocation in advanced HCC is influenced by tumor burden, liver function, performance status, and clinical judgment. Without causal adjustment, apparent predictive performance may reflect underlying prognostic imbalance rather than differential therapeutic effect[23,24].

Within this conceptual framework, existing AI applications in advanced HCC can be broadly categorized into prognostic stratifiers within treated cohorts and exploratory predictive models that estimate therapy-specific benefit. Beyond prognostication, AI models are increasingly used to predict therapeutic response, particularly in locoregional and systemic therapy settings.

A systematic review evaluating the role of AI in predicting TACE response showed that convolutional neural networks, gradient boosting, support vector machines, and random forests consistently outperformed conventional statistical models[7]. The addition of clinicoradiologic features markedly enhanced predictive discrimination, reinforcing the importance of structured multimodal feature integration. AI-based pathology-derived stratification (e.g., ABRS-P scoring) has identified biologically distinct subgroups among patients treated with atezolizumab-bevacizumab. These groups demonstrated differences in progression-free survival.

Radiomics-based ML approaches have shown potential to model treatment response in advanced HCC; however, robust causal validation remains limited[25]. Emerging studies in other solid tumors have applied causal ML frameworks, including counterfactual and uplift modeling, to estimate differential treatment benefit across patient subgroups[26]. These approaches aim to distinguish baseline prognosis from true treatment effect heterogeneity. They may provide a methodological framework for future predictive modeling in advanced HCC.

Deep learning in imaging and pathomics

Multimodal AI frameworks integrate imaging, radiomics, clinical variables, histopathology, and molecular biomarkers within unified modeling architectures. These systems can be operationalized using distinct integration strategies that influence interpretability, robustness, and transportability[11,14,27-29].

Early fusion refers to feature-level concatenation. Here, variables from multiple modalities (e.g., radiomic features, clinical parameters, molecular markers) are combined into a single feature vector before model training. This approach allows unified optimization but may amplify dimensionality and increase the risk of overfitting[30-32].

Late fusion involves model-level integration. In this approach, independent modality-specific models are trained separately, and their outputs are subsequently combined (e.g., via weighted averaging or meta-learning). This strategy may enhance modularity and robustness across heterogeneous datasets[33,34].

More advanced approaches incorporate attention-based integration mechanisms, which dynamically weight modality-specific contributions during model training. While potentially improving predictive performance, these architectures increase complexity and reduce interpretability, and pose additional challenges for clinical translation[11,35]. Most multimodal HCC models employ early fusion strategies, and few explicitly compare architectural approaches or evaluate transportability across institutions[36].

Despite these architectural distinctions, most studies report performance metrics without stratifying results by fusion strategy. Consequently, it remains unclear whether observed performance gains reflect true multimodal synergy or increased feature dimensionality. Direct comparative evaluation of early vs late fusion architectures within the same dataset remains limited in advanced HCC[7,32,33].

Such integrative architectures allow simultaneous modeling of tumor phenotype, liver functional reserve, and systemic disease state. However, empirical validation of incremental clinical utility remains limited. Despite theoretical advantages, multimodal architectures introduce additional challenges. These include data harmonization, modality-specific missingness, infrastructure requirements, and an increased risk of overfitting due to dimensionality. Comparative evaluation of early vs late fusion strategies, particularly with external validation and calibration assessment, remains limited in advanced HCC cohorts.

Although multimodal AI frameworks demonstrate methodological promise in retrospective cohorts, most remain supervised prognostic models trained on relatively limited datasets. Predictive modeling, survival-specific ML frameworks, and competing-risk approaches remain underdeveloped. In addition, multimodal systems introduce complexity, data harmonization challenges, and reproducibility concerns that must be addressed before clinical implementation.

CLINICAL PERFORMANCE COMPARED WITH CONVENTIONAL MODELS

Retrospective studies of advanced HCC have reported improved discrimination with AI-based models compared with conventional staging systems and clinicopathologic scores (Table 3). However, prospective validation remains limited. While these studies demonstrate measurable improvements in discrimination, several considerations should be noted. Differences in performance between derivation and external cohorts suggest potential overfitting and heterogeneity across datasets. Moreover, most models were developed on retrospective, single-center, or limited-center datasets, which limit generalizability. Calibration and competing-risk modeling were inconsistently reported, and decision-analytic evaluation was largely absent, limiting assessment of real-world clinical utility. Collectively, these findings indicate improved discrimination; however, the extent to which these gains translate into clinically meaningful decision support remains uncertain.

Table 3 Discrimination gains of artificial intelligence vs conventional models.

Clinical context	AI model type	AI performance	Comparator	Comparator performance	Absolute gain	Validation type
Atezo-Bev (advanced HCC)	Radiomic + clinical ML	AUC 0.89 (derivation), 0.75 (external)	BCLC/ALBI	0.61/0.48	+0.14-0.41	Multicenter external
TACE	Radiomic + clinical	AUC approximately 0.79-0.81	Clinical-only	Approximately 0.60	+0.19-0.21	Internal validation

Values represent reported discrimination metrics from individual retrospective studies and are not pooled estimates. Variability reflects differences in study design, patient populations, and validation strategies across cohorts. AI: Artificial intelligence; HCC: Hepatocellular carcinoma; Atezo-Bev: Atezolizumab plus bevacizumab; ML: Machine learning; AUC: Area under the receiver operating characteristic curve; ALBI: Albumin-bilirubin; BCLC: Barcelona Clinic Liver Cancer; TACE: Transarterial chemoembolization.

C-index comparisons

The C-index, a survival discrimination metric that accounts for censored data, is commonly used to evaluate time-to-event performance in AI-based prognostic models. In immunotherapy-treated advanced HCC, integrated radiomic-clinical models achieved a C-index of 0.75 in the external AP-HP (Assistance Publique-Hôpitaux de Paris) validation cohort. They outperformed conventional staging systems such as BCLC and ALBI grade, which showed substantially lower discrimination[21]. These findings indicate improved time-to-event discrimination in selected cohorts; however, evidence remains limited to retrospective studies with variable external validation.

AUROC comparisons

AUROC remains the most commonly reported performance metric in AI-HCC studies. In TACE-treated HCC, radiomic-clinical models have reported AUROCs in the range of 0.79-0.81, compared with approximately 0.60 for clinical-only models[7]. Similarly, in advanced HCC treated with atezolizumab-bevacizumab, integrated radiomic-clinical models achieved AUROCs of 0.89 in the derivation cohort and 0.75 in the external validation cohort. They outperformed conventional staging systems such as BCLC and ALBI[21]. In contrast, BCLC stage and ALBI grade demonstrated area under the receiver operating characteristic curves of 0.61 and 0.48, respectively (P < 0.001), highlighting the limited discriminatory capacity of conventional staging in systemic therapy. Although no universally accepted performance thresholds exist, AUROC or C-index values below approximately 0.70 are generally considered to reflect limited discrimination for individual-level clinical decision-making, and discrimination alone does not establish clinical utility[37].

Across retrospective cohorts, AI-based multimodal models have reported absolute AUROC improvements of approximately 0.14-0.40 relative to conventional clinical or stage-based models. Representative discrimination comparisons across systemic therapy and TACE cohorts are shown in Figure 1. Reported performance metrics vary across studies and are not directly comparable because of heterogeneity in study design, population, and validation strategy.

Figure 1 Values represent reported discrimination metrics from individual retrospective studies and are not pooled estimates. Error bars were not included due to inconsistent reporting of variance measures across the included studies. ML: Machine learning; ALBI: Albumin-bilirubin; BCLC: Barcelona Clinic Liver Cancer.

However, discrimination metrics alone do not establish clinical utility. AUROC and the C-index quantify rank-order discrimination but do not assess the accuracy of predicted absolute risk. Calibration assessment is inconsistently reported in advanced HCC AI studies. Brier scores and formal calibration measures (e.g., calibration plots) are also rarely used. Improved statistical discrimination alone may not translate into meaningful clinical decision support. This requires evaluation of calibration and clinically relevant risk thresholds. For example, two models with similar discrimination may differ clinically if one accurately predicts a 30% vs 10% 1-year mortality risk, thereby influencing treatment intensity and clinical decision-making[37,38]. Decision curve analysis and net benefit estimation are also essential for evaluating clinical utility.

Intra-stage heterogeneity

A recognized limitation of conventional staging systems is their inability to fully capture biological heterogeneity within the same BCLC stage (Table 4). AI-based models attempt to address this limitation by leveraging high-dimensional radiomic, pathomic, and clinical features. These approaches enable more granular risk stratification.

Table 4 What artificial intelligence adds beyond conventional staging.

Capability	Conventional staging	AI-based modeling
Feature handling	Limited predefined variables	High-dimensional nonlinear integration
Interaction modeling	Limited	Complex nonlinear interactions
Spatial imaging patterns	Qualitative	Quantitative radiomic features
Multimodal fusion	Rare	Integrated imaging + pathology + clinical
Intra-stage stratification	Limited	Fine-grained risk clustering

AI: Artificial intelligence.

AI-based pathology models (e.g., ABRS-P scoring) have identified biologically distinct subgroups within immunotherapy-treated cohorts. These groups demonstrate significant differences in progression-free survival[39]. Similarly, structured prognostic modeling approaches such as the Modified Albumin-Bilirubin plus Neutrophil-Platelet nomogram have been proposed to improve survival stratification after stereotactic body radiotherapy in advanced HCC[40]. In parallel, integrated radiomic models have stratified patients into high- and low-risk groups. These groups demonstrate divergent survival outcomes despite similar baseline staging[21].

Multimodal frameworks may further refine stratification by integrating radiomics, pathomics, genomics, and clinical variables. They may also address the coarse categorization inherent in stage-based systems. Collectively, these findings suggest that AI approaches may complement rather than replace conventional staging by addressing intra-stage prognostic variability, although prospective validation remains limited[41].

Real-world limitations

Despite reported improvements in discrimination metrics, several barriers to clinical translation remain. First, performance attenuation in external validation cohorts (e.g., a decline in area under the receiver operating characteristic curves from 0.89 to 0.75) likely reflects dataset heterogeneity and potential overfitting[7]. Second, many studies rely on retrospective, single-center datasets with limited diversity, raising concerns regarding generalizability. Third, variability in imaging acquisition, radiomic feature extraction pipelines, and algorithm selection further limits reproducibility.

Jia et al[41] describe an “iceberg effect”, wherein strong internal validation metrics may obscure reduced external performance due to data drift and publication bias. Additionally, black-box model architectures limit interpretability and clinical trust. This underscores the need for standardized reporting frameworks such as Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD)-AI and for greater emphasis on explainable and causal AI approaches.

In summary, AI-based models in advanced HCC have shown improved discrimination in selected retrospective cohorts and may improve characterization of intra-stage heterogeneity. However, several limitations constrain immediate clinical translation. These include performance attenuation in external validation, methodological heterogeneity, limited calibration assessment, and a lack of prospective evaluation. Broader multicenter validation, methodological standardization, and improved interpretability are required before routine implementation. From a clinical perspective, integration of AI models into routine practice also depends noninterpretability, workflow compatibility, and clinician trust.

CURRENT LIMITATIONS AND CHALLENGES

While retrospective studies report improved discrimination in selected cohorts, significant structural, statistical, and translational limitations reduce the reliability, reproducibility, and clinical applicability of current AI models in advanced HCC (Tables 5 and 6).

Table 5 Major methodological limitations in current artificial intelligence-hepatocellular carcinoma studies.

Domain	Specific limitation	Consequence
Statistical design	High-dimensional feature space with low event count	Model instability, optimism bias
Validation	Limited external and temporal validation	Poor transportability
Causal modeling	Lack of treatment-effect estimation frameworks	Prognostic misinterpreted as predictive
Radiomics	Scanner variability, segmentation inconsistency	Reduced reproducibility
Calibration	Rarely reported	Poor absolute risk estimation
Clinical utility	No decision curve or impact analysis	Uncertain real-world benefit

Table 6 Barriers and potential solutions for clinical translation of artificial intelligence in advanced hepatocellular carcinoma.

Limitation	Impact	Potential solution
Overfitting/low EPV	Model instability	Regularization, larger datasets
Limited external validation	Poor generalizability	Multicenter validation
Lack of causal modeling	Misinterpretation of treatment benefit	Counterfactual ML, uplift modeling
Poor calibration	Inaccurate risk estimation	Calibration plots, Brier score
Radiomic variability	Reduced reproducibility	Standardization, ComBat harmonization
Black-box models	Low clinical trust	Explainable AI
Lack of prospective validation	Limited clinical adoption	Prospective trials

EPV: Events per variable; ML: Machine learning; AI: Artificial intelligence; ComBat: Empirical Bayes-based batch effect harmonization method.

Statistical fragility and model instability

A central concern in advanced HCC AI research is statistical fragility arising from the use of high-dimensional feature spaces in small cohorts. Radiomic and multimodal models frequently incorporate hundreds of candidate features despite limited event numbers. As a result, the risk of unstable coefficient estimates and optimism bias increases[7,42].

Although regularization techniques such as least absolute shrinkage and selection operator and cross-validation are commonly employed, inconsistent use of nested validation frameworks and safeguards for hyperparameter tuning raises concerns about performance inflation. In advanced HCC, survival outcomes are influenced by tumor burden, liver function, and competing mortality. Consequently, overfitting may yield risk estimates that do not perform well under temporal or geographic validation.

Limited external and temporal validation

Most AI models in advanced HCC are derived from retrospective single- or dual-center datasets. External validation, when performed, is often limited to geographically similar institutions or to contemporaneous cohorts[43]. Few studies evaluate temporal validation across evolving therapeutic eras, such as pre- vs post-immunotherapy periods. As systemic therapies and imaging technologies evolve, distributional shifts may compromise model performance. Without robust multicenter and temporal validation, transportability across diverse healthcare settings remains uncertain.

Causal inference gaps in treatment modeling

Many AI models are described as predictors of treatment response. However, most are trained within single-treatment cohorts and therefore function primarily as prognostic stratifiers rather than true estimators of differential treatment benefit. Treatment allocation in advanced HCC is influenced by multiple factors, whereas current models often lack causal frameworks. Incorporating causal inference approaches may enable the development of more accurate tools for therapy selection[5,44].

Emerging directions in AI for advanced HCC include competing-risk-adapted survival models that explicitly account for liver failure-related mortality alongside tumor progression. Their application, however, remains limited. In parallel, explainable AI approaches, such as feature attribution methods and interpretable modeling frameworks, are increasingly explored to enhance clinical transparency. These approaches, however, remain underutilized in HCC. Model performance may also vary across etiologic subgroups (e.g., hepatitis B virus, hepatitis C virus, alcohol-related liver disease, and metabolic dysfunction). These differences reflect underlying biological heterogeneity that is rarely accounted for in existing datasets[2,22,41,45]. Furthermore, real-world evidence evaluating AI model performance outside controlled retrospective cohorts remains limited. This underscores the need for prospective and real-world validation.

Radiomic reproducibility and feature instability

Radiomic signatures remain sensitive to variability in image acquisition parameters, scanner platforms, reconstruction algorithms, and segmentation practices. Interobserver differences in region-of-interest delineation can substantially alter feature extraction. Although statistical harmonization techniques such as ComBat adjustment have been proposed, adoption remains inconsistent. Limited reporting of test-retest reproducibility metrics and intraclass correlation coefficients further constrains the assessment of feature stability. These factors collectively limit reproducibility across institutions[7,11].

Interpretability and regulatory readiness

Deep learning architectures, particularly convolutional neural networks and multimodal attention-based systems, often operate as opaque “black-box” models. In high-risk conditions such as advanced HCC, clinicians require transparent reasoning to guide therapeutic decisions.

Regulatory frameworks increasingly demand traceability, reproducibility, and post-deployment monitoring of AI systems. Limited use of explainable AI methods and inconsistent adherence to reporting standards such as TRIPOD-AI and CLAIM limit regulatory readiness and clinical trust[45,46].

Lack of prospective clinical utility evaluation

To date, few AI models in advanced HCC have been evaluated in prospective decision-impact studies or interventional trials. Demonstration of statistical discrimination alone does not establish clinical utility. Decision curve analysis, net benefit estimation, cost-effectiveness analysis, and workflow integration studies remain limited. Without evidence that AI-guided stratification improves patient-centered outcomes or alters therapeutic allocation, clinical translation remains limited[7].

FUTURE DIRECTIONS AND CLINICAL TRANSLATION

Although AI models in advanced HCC demonstrate methodological promise, meaningful clinical integration requires a transition from retrospective performance optimization to prospectively validated, causally informed, and impact-oriented research frameworks (Table 7). Future progress must prioritize clinical utility, reproducibility, and regulatory readiness rather than incremental improvements in discrimination metrics alone.

Table 7 Roadmap for clinical translation of artificial intelligence in advanced hepatocellular carcinoma.

Priority area	Required action	Goal
Validation	Multicenter, temporal, prospective evaluation	Transportability
Causal modeling	Incorporate counterfactual ML/uplift modeling	True treatment guidance
Reporting	Adherence to TRIPOD-AI/CONSORT-AI	Transparency
Calibration	Brier score, calibration plots	Reliable absolute risk
Clinical impact	Decision curve analysis, outcome trials	Demonstrate net benefit
Infrastructure	Federated learning, harmonization protocols	Data diversity and robustness

ML: Machine learning; TRIPOD-AI: Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis-Artificial Intelligence; CONSORT-AI: Consolidated Standards of Reporting Trials-Artificial Intelligence.

Transition from retrospective modeling to prospective clinical evaluation

Most current AI models in advanced HCC are derived from retrospective datasets and evaluated primarily through internal or limited external validation. To establish clinical utility, future research must move toward prospective, protocol-driven evaluation frameworks.

Prospective cohort validation, temporal validation across evolving therapeutic eras, and AI-assisted decision-impact trials are essential. Demonstrating that AI-guided stratification can alter therapeutic allocation, improve multidisciplinary decision-making, and patient-centered outcomes is critical before routine adoption[47,48]. Without prospective evidence, high AUROC or C-index values remain indicators of statistical discrimination rather than clinical effectiveness.

Incorporation of causal ML for treatment effect estimation

A major limitation of current AI applications is the predominance of prognostic modeling over true predictive modeling. Future systems must incorporate causal inference frameworks capable of estimating treatment effect heterogeneity.

Counterfactual modeling approaches, uplift modeling, and causal ML methodologies may enable differentiation between baseline prognosis and true treatment benefit[25,41,49]. Such methods are particularly relevant in advanced HCC, where treatment allocation is influenced by tumor burden, liver function, and performance status. The integration of causal modeling represents a necessary step for AI to guide therapy selection rather than merely stratify risk within treated cohorts.

Multicenter collaboration, data diversity, and federated learning

Advanced HCC exhibits substantial geographic and etiologic heterogeneity, including variations in hepatitis B virus, hepatitis C virus, alcohol-related liver disease, and metabolic-associated steatotic liver disease. AI systems trained on limited regional datasets risk reduced transportability.

Future development should emphasize multicenter training datasets, diverse etiologic representation, and heterogeneous imaging platforms. Federated learning frameworks, which enable decentralized model training without direct data sharing, offer a promising mechanism to improve generalizability while preserving patient privacy[25,50]. In radiomics research, the systematic implementation of harmonization techniques such as ComBat adjustment may further improve cross-institutional robustness[41,51]. However, standardized harmonization protocols must accompany broader collaborative efforts.

Regulatory readiness and reporting standardization

Clinical deployment of AI systems requires alignment with evolving regulatory standards. Beyond statistical validation, models must demonstrate reproducibility, traceability, and stability across software updates and clinical workflows[47]. Adherence to reporting frameworks such as TRIPOD-AI, CONSORT-AI, SPIRIT-AI, and CLAIM should become routine to enhance transparency and methodological rigor[41,51]. Additionally, post-deployment monitoring strategies, including performance auditing, drift detection, and structured model updating, will be necessary to maintain safety over time. Regulatory readiness will increasingly determine whether AI tools transition from investigational prototypes to clinically endorsed decision-support systems.

Evaluation of clinical utility and health-system impact

Future AI studies must extend beyond discrimination metrics to include calibration, clinical decision thresholds, and net benefit. Decision curve analysis, cost-effectiveness analysis, and workflow integration studies remain underrepresented in advanced HCC research. In addition, the real-world implementation of AI systems entails substantial costs, including storage of imaging data, high-performance computing infrastructure (e.g., graphic processing units), software development and maintenance, and trained personnel[47]. These economic considerations are critical for evaluating the feasibility and scalability of AI integration into routine clinical practice.

Demonstrating that AI-guided strategies improve survival, reduce unnecessary interventions, optimize resource allocation, or enhance quality-of-life outcomes will be essential to justify implementation. Without such evaluation, AI risks remaining a technically impressive but clinically marginal tool.

Long-term vision: Precision hepatology through multimodal integration

The long-term trajectory of AI in advanced HCC may involve large-scale multimodal foundation models integrating imaging, pathology, genomics, and longitudinal clinical data[2,50,52]. Emerging concepts such as digital twin architectures may simulate individualized disease trajectories and treatment responses.

However, technological expansion must remain grounded in methodological rigor, equitable data governance, and interdisciplinary collaboration among hepatologists, oncologists, radiologists, data scientists, and regulatory authorities. If future research aligns statistical innovation with prospective validation and causal reasoning, AI may evolve from retrospective pattern recognition toward a reliable component of precision hepatology.

CONCLUSION

AI has introduced advanced analytical approaches to prognostication and treatment response modeling in advanced HCC. These approaches have demonstrated improved statistical discrimination compared with conventional stage-based systems in selected retrospective cohorts. However, improved AUROC or C-index values alone do not establish clinical utility.

Current AI applications remain predominantly retrospective, are primarily prognostic rather than predictive, and are inconsistently validated across institutions and therapeutic eras. Limitations related to causal inference, calibration, competing-risk modeling, reproducibility, and prospective evaluation limit immediate clinical implementation. The conceptual structure underlying multimodal AI-based prognostic and predictive modeling in advanced HCC is summarized in Figure 2.

Figure 2  The framework illustrates the integration of radiomic, pathomic, and clinical data, along with methodological considerations, including validation, calibration, and causal inference, for clinical translation.

The future of AI in advanced HCC depends not on algorithmic complexity but on methodological rigor: Integration of survival-specific and causal modeling frameworks, multicenter and temporal validation, transparent reporting, and prospective assessment of decision impact. Only through such structural refinement can AI move beyond statistical enhancement toward clinically meaningful decision support.

Future research should prioritize prospective validation, causal modeling, and multicenter collaborative frameworks to improve generalizability and clinical applicability. Integration of explainable AI methods and real-world evaluation will be critical to bridge the gap between statistical performance and clinical utility[41].

If these standards are met, AI may evolve into a complementary tool within multidisciplinary hepatology practice, augmenting rather than replacing clinical expertise. Absent such rigor, it risks remaining a technologically sophisticated yet clinically peripheral innovation.