Multimodal data-driven prediction of postoperative recurrence and survival in hepatocellular carcinoma: a narrative review

Introduction

Primary liver cancer is a malignant tumor that poses a serious threat to global health, with persistently high incidence and mortality rates. According to 2022 Chinese cancer statistics, liver cancer ranks 5th in incidence among malignant tumors in China, with 367,700 new cases annually, and 2nd in mortality, with 316,500 deaths (1,2). Primary liver cancer includes hepatocellular carcinoma (HCC, 75–85%), intrahepatic cholangiocarcinoma (ICC, 10–15%), and combined hepatocellular-cholangiocarcinoma (cHCC-CCA) (3,4). Due to its high invasiveness and complex heterogeneity, HCC is the primary focus of clinical diagnosis and treatment; this review primarily discusses research progress in predicting postoperative recurrence and survival risk for HCC.

Although the 5-year survival rate for early-stage HCC patients has improved with advances in treatment options such as surgical resection, radiofrequency ablation, liver transplantation, targeted drugs, and immunotherapy, the postoperative recurrence rate remains above 50%, severely impacting long-term survival. Predicting postoperative recurrence and survival risk is crucial for implementing individualized treatment and precision management. Postoperative prediction involves building models to infer the timing of recurrence and metastasis after HCC resection. Accurate prediction can not only facilitate timely assessment of disease progression and improve patient survival but also avoid the waste of medical resources due to overtreatment, holding significant importance for treatment planning and efficacy evaluation (5-8).

Traditional methods for predicting HCC recurrence and survival risk mainly depend on clinical experience, tumor staging systems [e.g., China liver cancer (CNLC), Barcelona Clinic Liver Cancer (BCLC), and tumor-node-metastasis (TNM)], alpha-fetoprotein (AFP) levels, pathological scores, and other indicators. While widely used, these methods suffer from limited information dimensions and suboptimal predictive accuracy, as single data sources are inadequate for comprehensively capturing the complex biological characteristics of HCC recurrence and progression. In recent years, with the deep integration of bioinformatics, molecular pathology, medical imaging, and artificial intelligence (AI), the medical field has accelerated its digital and intelligent transformation, shifting towards data-driven research paradigms (9). Previously, medical research often followed a hypothesis-driven approach comprising clinical observation, hypothesis formulation, and experimental validation. Currently, in the context of big data, actively mining information from medical data to discover new insights has become a viable alternative. In computational data analysis, utilizing multimodal data to solve problems has become a consensus due to the rich information contained within (10-12). Multimodal data provide different perspectives on the same subject, with various types of information complementing each other to form a more comprehensive description of the research object. Compared to traditional fragmented data analysis methods, multimodal data analysis avoids the limitations of single indicators and the impact of physician subjective bias in HCC diagnosis and treatment analysis.

This narrative review synthesizes the latest literature on HCC, multimodal data, and AI-based predictions related to key terms. We summarize the commonly used multimodal data types and their characteristics in current research on predicting postoperative recurrence and survival risk in HCC, discuss research trends and modeling strategies for multimodal fusion models, analyze existing technical and clinical challenges, and outline future directions. Table 1 presents the conceptual framework of this article. We present this article in accordance with the Narrative Review reporting checklist (available at https://jgo.amegroups.com/article/view/10.21037/jgo-2025-aw-848/rc).

Methods

This article is designed as a narrative review summarizing recent research progress in multimodal data-based prediction of postoperative recurrence and survival risk in HCC. Relevant literature was identified through searches of PubMed, Google Scholar, and Web of Science, with a primary focus on studies published from 2020 onward.

The search terms included combinations of “hepatocellular carcinoma”, “liver cancer”, “recurrence”, “survival”, “prognosis”, “multimodal”, “machine learning”, and “deep learning”. The retrieved publications were screened based on their relevance to the topic, with particular attention to studies that explored clinical data, imaging data, omics data, or their combinations for prognostic modeling.

Rather than applying a formal systematic selection framework, this review emphasizes representative and influential studies that illustrate key methodological trends, data modalities, and multimodal fusion strategies in the field. The aim is to provide an integrated narrative overview of current modeling approaches, emerging techniques, and existing challenges, rather than an exhaustive or quantitative synthesis of all available evidence.

Definition and content of multimodal data

Multimodal data refers to information integrated from multiple sources. Common multimodal data in HCC diagnosis and treatment include clinical data [e.g., tumor differentiation degree, alanine aminotransferase (ALT), AFP levels], imaging data [e.g., computed tomography (CT), magnetic resonance imaging (MRI), whole slide images (WSI)], and omics data (e.g., genomics, transcriptomics, proteomics). By reflecting multi-level characteristics of HCC from macroscopic imaging manifestations to microscopic molecular mechanisms, multimodal data contribute to improving the accuracy and stability of risk prediction.

Clinical data

Clinical data represent the earliest and most widely used modality in postoperative risk prediction, serving as a foundational data source due to their accessibility, low cost, and strong clinical interpretability. Recent studies have further explored the potential value of clinical data to provide a more precise basis for prediction. In 2022, Wang et al. (13) used least absolute shrinkage and selection operator and Cox proportional hazards model (LASSO-COX) regression and random survival forest (RSF) to build models for early-stage HCC patients receiving minimally invasive treatment, finding that the LASSO-COX model had a slightly higher concordance index (C-Index) than the RSF model. In 2024, Liu et al. (14) identified eight independent risk factors based on clinical features from a large cohort of postoperative patients and constructed the VERM-pre model for predicting early recurrence, which demonstrated a high C-index (>0.7) in an independent validation cohort. Similarly, Moazzam et al. (15) proposed the SARScore model to predict long-term survival risk in postoperative HCC patients. However, clinical data often contain noise and exhibit strong dynamism, making it challenging to achieve higher accuracy and stability in predictive models.

Imaging data

Imaging data reveal potential recurrence risk factors primarily through feature extraction and quantitative analysis. Traditional methods relied heavily on manual selection and quantitative measurement of radiological features. With the development of machine learning and deep learning technologies, AI-based automated feature extraction and prediction models have gradually become a research hotspot. In 2023, Kucukkaya et al. (16) used a deep learning approach for automated feature extraction from MRI data, employing the Visual Geometry Group Network 16 (VGG16) convolutional neural network (CNN) to extract image feature vectors, which were then input into an extreme gradient boosting (XGBoost) model to predict HCC postoperative recurrence risk. This study suggested that radiomics models combining machine learning and deep learning outperform traditional manual feature selection methods in automation and prediction accuracy, particularly suitable for high-throughput analysis of large-scale data.

Omics data

Omics models based on metabolism, non-coding RNA, and immune-related proteins provide new biomarkers and targets for prognosis prediction by revealing the molecular mechanisms of HCC. In 2022, Tian et al. (17) integrated metabolic pathway analysis to identify six key metabolic genes (ADPGK, GOT2, MTHFS, etc.), constructing a metabolic score model. They found that high-score patients were significantly associated with TP53 mutation and advanced tumor stage. In 2024, Wang et al. (18) discovered that high expression of TIGIT and NKG2A proteins in HCC tissue was an independent risk factor for postoperative recurrence (P<0.05). The nomogram model based on these factors showed C-indices greater than 0.7 for predicting 1–5 years recurrence-free survival (RFS), with well-fitted calibration curves, suggesting its potential as a biomarker for predicting immunotherapy efficacy. Furthermore, omics markers such as metabolic genes, microRNAs (miRNAs), and immune proteins can not only independently predict HCC prognosis but also reveal characteristics of the tumor microenvironment, promoting the shift of HCC precision treatment from clinicopathological stratification to molecular subtyping. However, omics data are often limited by small sample sizes, high costs, and significant batch effects, restricting their widespread application as independent data models.

Existing models for predicting postoperative recurrence and survival risk

Constructing prediction models for postoperative recurrence and survival risk is a crucial step towards precision medicine and individualized therapy for HCC. Current mainstream prediction models can be divided into traditional statistical models and AI-based methods. Traditional models are widely adopted due to their strong interpretability and ease of clinical application, while AI models demonstrate superior performance in mining high-dimensional, non-linear relationships.

Conventional models

Traditional prediction models for recurrence and survival risk mainly include nomograms and risk scores, which are widely used in postoperative HCC patient management due to their good visualization and clinical operability.

Nomogram

Nomograms are typically based on statistical methods like Cox proportional hazards regression or logistic regression. They visually represent individual risk levels by assigning weighted scores to independent risk factors. In 2023, Wei et al. (19) constructed pre-operative and post-operative nomograms based on indicators such as pre-operative circulating tumor cell (CTC) count, tumor size, and lymph node metastasis to screen high-risk patients for extrahepatic recurrence for precise decision-making. In 2025, Chun et al. (20) developed a nomogram model focusing on early recurrence risk in patients with cHCC-CCA. In 2025, Su et al. (21) built two nomogram models based on Logistic regression alone and a combination of random forest (RF) with LASSO regression, respectively. Comparative evaluation revealed that the model combining LASSO and RF had significantly higher predictive accuracy than the Logistic regression model.

Risk scoring system

Risk scoring systems primarily involve scoring key prognostic variables and performing risk stratification through statistical modeling, characterized by simple calculation and ease of use. In 2022, Yao et al. (22) utilized five independent risk factors related to post-recurrence survival (PRS) to develop a simple risk stratification model for predicting PRS in recurrent HCC patients. In 2025, Zheng et al. (23) constructed an Early Recurrence Outside Milan (EROM) score based on MRI for predicting early RFS in HCC patients via Cox regression analysis. Compared to the BCLC staging system, this score performed better on an independent test dataset (2-year C-index, 0.69 vs. 0.52, P<0.001).

In summary, as traditional statistical models, nomograms and risk scoring systems hold an important position in predicting postoperative recurrence and survival risk in HCC due to their simplicity and good interpretability. However, these models are mostly based on single-modal variables and struggle to integrate the high-dimensional and non-linear characteristic information from imaging or omics data, limiting their predictive precision.

AI models

With the continuous development of AI technology, AI models have shown great potential in predicting postoperative recurrence and survival risk in HCC. AI models are mainly divided into non-deep learning algorithms and deep learning algorithms, which improve prediction accuracy by mining latent features within multimodal data.

Non-deep learning algorithms

Non-deep learning algorithms include RF, support vector machine (SVM), XGBoost, etc. In 2021, Zhan et al. (24) proposed a two-stage Cox-nnet model that innovatively integrated pathological images and transcriptomic data from HCC patients. The results showed that the median C-index of the two-stage Cox-nnet (0.75) was significantly higher than that of the Cox-nnet model based solely on gene expression data (0.70). In 2024, Xie et al. (25) constructed a multimodal model combining clinical data, CT radiomics scores, and WSI pathomics scores. Using an SVM classifier, they built four feature fusion models: CRP (clinical + radiomics + pathomics), CRp, CrP, and rCP. They found that fusing multi-source heterogeneous features effectively improved prognosis prediction accuracy, with the CRP model performing best, achieving an area under the curve (AUC) value of 0.863.

Deep learning algorithms

Deep learning algorithms, such as deep neural networks (DNN), CNN, and residual networks (ResNet), demonstrate considerable advantages in processing high-dimensional and complex medical data. DNN is capable of learning complex non-linear relationships across multimodal inputs; CNN excels in extracting spatial and hierarchical features from imaging data; and ResNet, with its residual connections, mitigates the vanishing gradient problem and enables training of very deep networks, thereby enhancing feature representation and model performance. Particularly in integrating clinical information, imaging data, and molecular omics features, they provide higher accuracy and robustness for predictive models. In 2020, Nam et al. (26) developed and validated a MoRAL-AI prediction model, which achieved a C-index of 0.75; the largest weighted parameter in the model was tumor diameter, followed by AFP, age, and Protein Induced by Vitamin K Absence or Antagonist II (PIVKA-II). A 2022 study (27) built a neural network model (MobileNetV2_HCC_class) integrating clinical and pathological image data. Its hazard ratio (HR) was significant, and its discriminatory ability surpassed that of traditional clinicopathological factors. It could also identify pathological features in tumor regions (e.g., stroma presence, cellular atypia) with high predictive value for recurrence. In 2023, Liu et al. (28) mined immune-related marker genes from transcriptomic data and constructed a multi-module prediction model integrating DNN and COX regression analysis, which outperformed traditional methods in predicting metastasis (AUC =0.85) and survival time.

In summary, AI models possess multi-dimensional advantages in modeling postoperative recurrence and prognosis risk in HCC. On one hand, their powerful feature extraction capabilities can deeply mine high-order associations hidden within multimodal data. On the other hand, by integrating different model structures and optimization strategies, they can achieve more stable and accurate risk prediction.

Fusion models based on multimodal data

Multimodal fusion models aim to integrate heterogeneous information from different data sources to achieve a more comprehensive and robust assessment of postoperative recurrence and survival risk in HCC. According to the stage at which multimodal information is integrated, current approaches can be broadly categorized into feature-level fusion and decision-level fusion (29). Importantly, the choice between these strategies depends on data characteristics, clinical context, and computational resources (see Table 2 for a comparative summary of recent studies).

Feature-level fusion

Feature-level fusion integrates raw data or extracted features from multiple modalities into a unified feature space prior to model training, enabling deep cross-modal interactions. This strategy is particularly suitable for scenarios in which all modalities are available for most patients, preprocessing pipelines are standardized, and sample size is sufficient to support high-dimensional modeling. In 2021, He et al. (31) proposed an imageomics and multi-network-based deep learning model (i-RAPIT), which independently extracted multimodal features and designed feature interaction modules within the network architecture to achieve feature-level fusion, successfully integrating clinical data, MRI images, and pathological images. The model was developed in a single-center cohort of 109 patients, and internal validation showed superior recognition ability compared to single-modality models (AUC =0.87). However, it did not include an external validation cohort, and the confidence intervals (CIs) were not reported, limiting conclusions regarding generalizability. Similarly, Huang et al. (32) constructed a clinical model and a deep learning radiomics (DLR) model based on pre-operative grayscale ultrasound and contrast-enhanced ultrasound images from 414 HCC patients who underwent radical resection. They integrated these at the feature level to build a clinical + DLR model. The results showed that the multimodal model integrating clinical and DLR features outperformed single modalities in predicting postoperative recurrence and prognosis with AUC values exceeding 0.75 in the internal validation cohort. Notably, although decision curves were reported, external validation was not performed, and the study population was restricted to a single institution.

From a clinical perspective, feature-level fusion is most appropriate when the goal is maximizing predictive accuracy in controlled research environments or tertiary centers with comprehensive data availability. However, its reliance on complete multimodal inputs and high-dimensional feature spaces increases susceptibility to overfitting and reduces robustness in real-world settings where missing data are common.

Decision-level fusion

Decision-level fusion combines predictions from modality-specific models rather than directly merging features. This approach is more flexible and robust to heterogeneous data availability, making it particularly attractive for clinical scenarios in which certain modalities (e.g., omics or advanced imaging) are unavailable for all patients. In 2024, Schmauch et al. (33) employed decision-level fusion, modeling deep features extracted from pathological images by ResNet50 and clinical variables separately, and then fusing the results using Cox proportional hazards regression. The study included 469 patients. Regrettably, it did not utilize an independent external validation cohort, but reported the model’s CIs, thereby enhancing interpretability and clinical credibility. Yan et al. (34) used VGGNet-19 to extract deep learning features from enhanced MRI and combined them with clinical data to construct a nomogram, integrating the risk scores through the nomogram’s weight assignment. The study included 285 patients from two centers, and the fused model achieved an AUC of 0.909 (95% CI: 0.842–0.976), significantly outperforming the clinical model (AUC =0.715, 95% CI: 0.586–0.843). In 2025, 519 patients with HCC were included from three medical centers. Peng et al. (30) utilized CT imaging data covering hepatic artery phase (HAP), portal venous phase (PVP), delayed phase (DP), and plain scan (PS). They built a radiomics model using LASSO regression and SVM algorithms, and a deep learning model using ShuffleNet as the base framework. By combining the radiomics and deep learning models, they developed a multimodal radiomics-deep learning model (MM-RDLM). The results showed that the MM-RDLM model achieved an AUC of 0.930 (95% CI: 0.876–0.984) in the validation cohort, outperforming single models, the radiomics model, and the deep learning model.

Discussion

Although studies employing multimodal data to predict postoperative recurrence and survival risk in HCC are increasing, substantial challenges remain with respect to clinical applicability, methodological robustness, and dynamic modeling capability. This section provides a systematic discussion from three perspectives: the clinical translation of multimodal prediction models, methodological limitations, and future directions.

Clinical translation of multimodal prediction models

The ultimate goal of multimodal prediction models is to support clinical decision-making, with the continuous improvement of predictive performance, multimodal data-driven models for postoperative recurrence and survival risk in HCC are gradually transitioning from methodological exploration toward potential clinical application. Compared with conventional staging systems such as the BCLC, CNLC, and TNM classifications, multimodal prediction models offer individualized, continuous risk estimates by integrating complementary information from clinical variables, imaging features, and molecular profiles. This shift aligns with the clinical demand for precision postoperative management rather than coarse risk stratification (35).

From a clinical decision-making perspective, accurate postoperative risk prediction has several potential implications. First, patients identified as high risk for early recurrence may benefit from intensified surveillance strategies, including shorter imaging follow-up intervals or the use of advanced imaging modalities. Second, multimodal risk stratification may support the selection of candidates for adjuvant or neoadjuvant therapies, such as targeted agents, immunotherapy, or locoregional interventions, particularly in patients who are classified as early-stage by conventional staging but harbor aggressive biological features revealed by imaging or omics data. Third, these models may assist clinicians in postoperative counseling by providing more individualized prognostic information, thereby facilitating shared decision-making.

Despite these prospects, the clinical translation of multimodal prediction models remains limited. Most published studies focus primarily on improving discrimination metrics, such as the C-index or AUC, while providing insufficient guidance on how predicted risk categories should concretely alter clinical pathways. Moreover, few models have been prospectively evaluated or compared head-to-head with established clinical frameworks. Barriers to implementation also include the limited availability of high-quality multimodal data in routine practice, the additional costs associated with omics profiling, and the limited interpretability of complex AI-based models. These challenges underscore the need for clinically oriented validation studies that emphasize usability, interpretability, and incremental value over existing decision-support tools.

Methodological limitations

Although multimodal fusion models consistently outperform single-modality approaches (36), current research is characterized by substantial methodological heterogeneity and several unresolved limitations. One major challenge lies in the lack of consensus regarding optimal fusion strategies (37). Feature-level fusion enables deep interaction among heterogeneous data types and often yields superior predictive performance; however, it requires strict alignment of modalities, large sample sizes, and complex preprocessing pipelines. In contrast, decision-level fusion offers greater flexibility and robustness to missing data but may underutilize cross-modal correlations (38). Existing studies rarely provide systematic justification for choosing one fusion strategy over another, making cross-study comparison difficult.

Another critical limitation is the predominance of single-center, retrospective studies with internal validation only. Deep learning-based multimodal models, while powerful, are particularly susceptible to overfitting in such settings. Differences in imaging protocols, disease etiology [e.g., hepatitis B virus (HBV), portal vein tumor thrombosis (PVTT), or microvascular invasion (MVI)], and treatment strategies across institutions further hinder external validation and cross-center deployment. Consequently, many reported models demonstrate impressive internal performance yet lack evidence of robustness in real-world clinical environments.

Taken together, these limitations indicate that current multimodal fusion research remains largely exploratory. Future studies should prioritize standardized reporting of model architecture, fusion strategy, validation cohorts, and performance metrics, as well as conduct multi-center external validation to establish clinical credibility.

Future perspectives

A major limitation of existing postoperative HCC prediction models is their reliance on static baseline data. In clinical practice, however, patient status evolves continuously, with dynamic changes in tumor burden, laboratory indices, imaging findings, and treatment exposure during follow-up. This temporal nature of disease progression has prompted growing interest in time-dependent and longitudinal prediction approaches. Recent studies have begun to incorporate temporal information using time-updated Cox regression models or by modeling dynamic biomarkers such as AFP trajectories (39-41). For example, time-updated survival models integrating longitudinal tumor burden and biomarker dynamics have demonstrated improved RFS prediction compared with static baseline models (42). In parallel, advances in deep learning have enabled the application of recurrent neural network (RNN), long short-term memory (LSTM) networks, and, more recently, transformer-based architectures to survival analysis, allowing for the modeling of complex temporal dependencies across multimodal inputs.

Despite these advances, longitudinal multimodal prediction remains underexplored in HCC. Challenges include irregular follow-up intervals, missing data, and the difficulty of synchronizing heterogeneous time-series data from different modalities. Moreover, few studies explicitly evaluate whether dynamic models provide clinically actionable advantages over simpler time-updated statistical approaches. Future research should focus on developing standardized frameworks for longitudinal multimodal data integration, emphasizing interpretability and clinical relevance. Prospective studies incorporating repeated measurements and real-world follow-up data will be essential to validate whether dynamic prediction models can meaningfully improve postoperative surveillance, early intervention, and long-term outcomes in HCC patients.

Conclusions

This narrative review summarizes recent advances [2020–2025] in multimodal data-driven models for predicting postoperative recurrence and survival risk in HCC. By integrating clinical, imaging, pathological, and omics data, these models enable more individualized and accurate prognostic assessment than traditional staging systems. Overall, multimodal approaches consistently outperform single-modality models, reflecting the complementary nature of heterogeneous data. Conventional statistical models remain clinically valuable due to their interpretability and simplicity but are limited in capturing complex non-linear relationships. In contrast, AI-based models, particularly deep learning methods, offer superior feature extraction and integration capabilities. Feature-level fusion can achieve higher predictive performance in settings with complete data, whereas decision-level fusion provides greater flexibility for real-world clinical practice.

Despite encouraging results, most studies are retrospective and single-center, with limited external validation and reliance on static baseline data. Future work should prioritize longitudinal multimodal modeling, standardized fusion strategies, and multi-center prospective validation to enhance clinical translation and support personalized postoperative management in HCC.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://jgo.amegroups.com/article/view/10.21037/jgo-2025-aw-848/rc

Peer Review File: Available at https://jgo.amegroups.com/article/view/10.21037/jgo-2025-aw-848/prf

Funding: This work was supported by National Natural Science Foundation of China (grant No. 82473424 for Y.W.), Science and Technology Committee of Xuhui District, Shanghai, China (grant No. 23XHYD-26 for Y.W.), Shanghai Hospital Development Center (grant No. SHDC2024CRI094 for Y.W.) and Shanghai Engineering Technology Research Center (grant No. 18DZ2250900).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jgo.amegroups.com/article/view/10.21037/jgo-2025-aw-848/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Han B, Zheng R, Zeng H, et al. Cancer incidence and mortality in China, 2022. J Natl Cancer Cent 2024;4:47-53. [Crossref] [PubMed]
2. Zheng R, Zhang S, Zeng H, et al. Cancer incidence and mortality in China, 2016. J Natl Cancer Cent 2022;2:1-9. [Crossref] [PubMed]
3. Zhou M, Wang H, Zeng X, et al. Mortality, morbidity, and risk factors in China and its provinces, 1990-2017: a systematic analysis for the Global Burden of Disease Study 2017. Lancet 2019;394:1145-58. [Crossref] [PubMed]
4. Rumgay H, Ferlay J, de Martel C, et al. Global, regional and national burden of primary liver cancer by subtype. Eur J Cancer 2022;161:108-18. [Crossref] [PubMed]
5. Xu XF, Wu H, Gu LH, et al. Development and Validation of an Individualized Prediction Model for Postoperative Late Recurrence After Hepatectomy for Hepatocellular Carcinoma (POLAR-HCC): A Multicenter Study. Ann Surg Oncol 2025;32:9573-83. [Crossref] [PubMed]
6. Xiang Z, Deng J, Liang H, et al. Artificial intelligence for the prediction of posthepatectomy recurrence in hepatocellular carcinoma: a systematic review and meta-analysis. Ann Med 2025;57:2568118. [Crossref] [PubMed]
7. Zhang X, Chen C, Wang Y, et al. Recurrence risk prediction models for hepatocellular carcinoma after liver transplantation. J Gastroenterol Hepatol 2024;39:2272-80. [Crossref] [PubMed]
8. Zhang W, Zhang B, Chen XP. Adjuvant treatment strategy after curative resection for hepatocellular carcinoma. Front Med 2021;15:155-69. [Crossref] [PubMed]
9. Stahlschmidt SR, Ulfenborg B, Synnergren J. Multimodal deep learning for biomedical data fusion: a review. Brief Bioinform 2022;23:bbab569. [Crossref] [PubMed]
10. Tan X, Su AT, Hajiabadi H, et al. Applying Machine Learning for Integration of Multi-Modal Genomics Data and Imaging Data to Quantify Heterogeneity in Tumour Tissues. Methods Mol Biol 2021;2190:209-28. [Crossref] [PubMed]
11. Huang B, Yang F, Yin M, et al. A Review of Multimodal Medical Image Fusion Techniques. Comput Math Methods Med 2020;2020:8279342. [Crossref] [PubMed]
12. Huang SC, Pareek A, Seyyedi S, et al. Fusion of medical imaging and electronic health records using deep learning: a systematic review and implementation guidelines. NPJ Digit Med 2020;3:136. [Crossref] [PubMed]
13. Wang Q, Qiao W, Zhang H, et al. Nomogram established on account of Lasso-Cox regression for predicting recurrence in patients with early-stage hepatocellular carcinoma. Front Immunol 2022;13:1019638. [Crossref] [PubMed]
14. Liu L, Qin S, Lin K, et al. Development and comprehensive validation of a predictive prognosis model for very early HCC recurrence within one year after curative resection: a multicenter cohort study. Int J Surg 2024;110:3401-11. [Crossref] [PubMed]
15. Moazzam Z, Alaimo L, Endo Y, et al. A Prognostic Model To Predict Survival After Recurrence Among Patients With Recurrent Hepatocellular Carcinoma. Ann Surg 2024;279:471-8. [Crossref] [PubMed]
16. Kucukkaya AS, Zeevi T, Chai NX, et al. Predicting tumor recurrence on baseline MR imaging in patients with early-stage hepatocellular carcinoma using deep machine learning. Sci Rep 2023;13:7579. [Crossref] [PubMed]
17. Tian Y, Lu J, Qiao Y. A metabolism-associated gene signature for prognosis prediction of hepatocellular carcinoma. Front Mol Biosci 2022;9:988323. [Crossref] [PubMed]
18. Wang J, Cao Y, Tian Y, et al. A Novel Prognostic Nomogram Based on TIGIT and NKG2A Can Predict Relapse-Free Survival of Hepatocellular Carcinoma After Hepatectomy. Cancer Med 2024;13:e70419. [Crossref] [PubMed]
19. Wei HW, Qin SL, Xu JX, et al. Nomograms for postsurgical extrahepatic recurrence prediction of hepatocellular carcinoma based on presurgical circulating tumor cell status and clinicopathological factors. Cancer Med 2023;12:15065-78. [Crossref] [PubMed]
20. Chun SJ, Jung YJ, Choi Y, et al. Prognostic Evaluation and Survival Prediction for Combined Hepatocellular-Cholangiocarcinoma Following Hepatectomy. Cancer Res Treat 2025;57:229-39. [Crossref] [PubMed]
21. Su BB, Zhu CJ, Cao J, et al. Enhanced prediction of 5-year postoperative recurrence in hepatocellular carcinoma by incorporating LASSO regression and random forest models. Surg Endosc 2025;39:2540-50. [Crossref] [PubMed]
22. Yao LQ, Chen ZL, Feng ZH, et al. Clinical Features of Recurrence After Hepatic Resection for Early-Stage Hepatocellular Carcinoma and Long-Term Survival Outcomes of Patients with Recurrence: A Multi-institutional Analysis. Ann Surg Oncol 2022; Erratum in: Ann Surg Oncol 2022;29:5206. [Crossref] [PubMed]
23. Zheng T, Sheng L, Wu Y, et al. Imaging-based prediction of early recurrence and neoadjuvant therapy outcomes for resectable beyond Milan HCC. Eur J Radiol 2025;184:111945. [Crossref] [PubMed]
24. Zhan Z, Jing Z, He B, et al. Two-stage Cox-nnet: biologically interpretable neural-network model for prognosis prediction and its application in liver cancer survival using histopathology and transcriptomic data. NAR Genom Bioinform 2021;3:lqab015. [Crossref] [PubMed]
25. Xie Q, Zhao Z, Yang Y, et al. A clinical-radiomic-pathomic model for prognosis prediction in patients with hepatocellular carcinoma after radical resection. Cancer Med 2024;13:e7374. [Crossref] [PubMed]
26. Nam JY, Lee JH, Bae J, et al. Novel Model to Predict HCC Recurrence after Liver Transplantation Obtained Using Deep Learning: A Multicenter Study. Cancers (Basel) 2020;12:2791. [Crossref] [PubMed]
27. Liu Z, Liu Y, Zhang W, et al. Deep learning for prediction of hepatocellular carcinoma recurrence after resection or liver transplantation: a discovery and validation study. Hepatol Int 2022;16:577-89. [Crossref] [PubMed]
28. Liu J, Qu J, Xu L, et al. Prediction of liver cancer prognosis based on immune cell marker genes. Front Immunol 2023;14:1147797. [Crossref] [PubMed]
29. Wu C, Chen Q, Wang H, et al. A review of deep learning approaches for multimodal image segmentation of liver cancer. J Appl Clin Med Phys 2024;25:e14540. [Crossref] [PubMed]
30. Peng J, Wang J, Zhu H, et al. Three-dimensional multimodal imaging for predicting early recurrence of hepatocellular carcinoma after surgical resection. J Adv Res 2026;81:865-75. [Crossref] [PubMed]
31. He T, Fong JN, Moore LW, et al. An imageomics and multi-network based deep learning model for risk assessment of liver transplantation for hepatocellular cancer. Comput Med Imaging Graph 2021;89:101894. [Crossref] [PubMed]
32. Huang Z, Shu Z, Zhu RH, et al. Deep learning-based radiomics based on contrast-enhanced ultrasound predicts early recurrence and survival outcome in hepatocellular carcinoma. World J Gastrointest Oncol 2022;14:2380-92. [Crossref] [PubMed]
33. Schmauch B, Elsoukkary SS, Moro A, et al. Combining a deep learning model with clinical data better predicts hepatocellular carcinoma behavior following surgery. J Pathol Inform 2024;15:100360. [Crossref] [PubMed]
34. Yan M, Zhang X, Zhang B, et al. Deep learning nomogram based on Gd-EOB-DTPA MRI for predicting early recurrence in hepatocellular carcinoma after hepatectomy. Eur Radiol 2023;33:4949-61. [Crossref] [PubMed]
35. Xia H, Huang Q, Huang Z, et al. Multimodal deep learning model for predicting prognosis following radiotherapy-based combination therapy in unresectable hepatocellular carcinoma. Cancer Lett 2026;636:218122. [Crossref] [PubMed]
36. Zheng T, Zhu Y, Jiang H, et al. MRI-Based Topology Deep Learning Model for Noninvasive Prediction of Microvascular Invasion and Assisting Prognostic Stratification in HCC. Liver Int 2025;45:e16205. [Crossref] [PubMed]
37. Wang T, Chen H, Chen Z, et al. Prediction model of early recurrence of multimodal hepatocellular carcinoma with tensor fusion. Phys Med Biol 2024; [Crossref] [PubMed]
38. Wang W, Chen Q, Iwamoto Y, et al. Deep Fusion Models of Multi-Phase CT and Selected Clinical Data for Preoperative Prediction of Early Recurrence in Hepatocellular Carcinoma. IEEE Access 2020;8:139212-20.
39. Ma X, Huang L, Yu M, et al. Dynamic Prediction of the Risk of Hepatocellular Carcinoma After DAA Treatment for Hepatitis C Patients. Cancer Control 2025;32:10732748251316609. [Crossref] [PubMed]
40. Shen L, Jiang Y, Zhang T, et al. Machine Learning for Dynamic Prognostication of Patients With Hepatocellular Carcinoma Using Time-Series Data: Survival Path Versus Dynamic-DeepHit HCC Model. Cancer Inform 2024;23:11769351241289719. [Crossref] [PubMed]
41. Yang H, Lu L, Guo W, et al. A Longitudinal Study of AFP Trajectories and Clinical Outcomes in Intermediate-Stage Hepatocellular Carcinoma After Hepatectomy. J Hepatocell Carcinoma 2024;11:219-28. [Crossref] [PubMed]
42. Akabane M, Kawashima J, Altaf A, et al. Enhancing Recurrence-Free Survival Prediction in Hepatocellular Carcinoma: A Time-Updated Model Incorporating Tumor Burden and AFP Dynamics. Ann Surg Oncol 2025;32:5648-56. [Crossref] [PubMed]