Research progress on machine algorithm prediction of liver cancer prognosis after intervention therapy

Abstract

The treatment for liver cancer has transitioned from traditional surgical resection to interventional therapies, which have become increasingly popular among patients due to their minimally invasive nature and significant local efficacy. However, with advancements in treatment technologies, accurately assessing patient response and predicting long-term survival has become a crucial research topic. Over the past decade, machine algorithms have made remarkable progress in the medical field, particularly in hepatology and prognosis studies of hepatocellular carcinoma (HCC). Machine algorithms, including deep learning and machine learning, can identify prognostic patterns and trends by analyzing vast amounts of clinical data. Despite significant advancements, several issues remain unresolved in the prognosis prediction of liver cancer using machine algorithms. Key challenges and main controversies include effectively integrating multi-source clinical data to improve prediction accuracy, addressing data privacy and ethical concerns, and enhancing the transparency and interpretability of machine algorithm decision-making processes. This paper aims to systematically review and analyze the current applications and potential of machine algorithms in predicting the prognosis of patients undergoing interventional therapy for liver cancer, providing theoretical and empirical support for future research and clinical practice.

Introduction

Liver cancer is one of the most common malignant tumors worldwide and ranks third among the leading causes of cancer-related deaths in humans [1]. According to estimates from the International Agency for Research on Cancer (IARC), there were approximately 870,000 new liver cancer cases and 760,000 deaths globally in 2022 [2]. Hepatocellular carcinoma (HCC) is the most common type of liver cancer, accounting for about 80% of all liver cancer cases, followed by intrahepatic cholangiocarcinoma (ICC), which accounts for approximately 15%, with other rare types of liver cancer making up around 5% [2,3]. This review primarily focuses on HCC due to its predominant role in liver cancer; however, we also address ICC and other less common types to provide a comprehensive overview of the current state of prognostic prediction models for liver cancer patients undergoing interventional treatments.

Interventional therapy for liver cancer, such as transcatheter arterial chemoembolization (TACE), is a crucial treatment method that can effectively control tumor growth and extend patient survival across early, intermediate, and advanced stages [4,5]. Interventional treatment strategies for liver cancer are either vascular or non-vascular (Figure 1). Non-vascular interventional techniques include various ablation methods, such as microwave ablation (MWA), radiofrequency ablation (RFA), irreversible electroporation (IRE), cryoablation (CRA), high-intensity focused ultrasound (HIFU), and laser ablation (LSA) [6]. Additionally, radiotherapy, including stereotactic body radiotherapy (SBRT) and iodine-125 seed implantation brachytherapy, is also part of non-vascular interventional therapy. In the vascular interventional treatment domain, transhepatic arterial interventions, particularly TACE and hepatic arterial infusion chemotherapy (HAIC), are mainstream clinical choices due to their significant efficacy. However, patients’ responses and prognoses to interventional treatments vary significantly, with objective response rates (ORRs) ranging from 40% to 80% and overall survival (OS) time varying from 13 to 48 months [7,8] (Figure 1).

Figure 1.

Main interventional treatment methods.

In recent years, with the development of artificial intelligence (AI) technology, machine learning (ML) has been increasingly applied in the medical field, providing new approaches for predicting the prognosis of liver cancer. ML processes large amounts of clinical and imaging data, allowing for advanced feature generation and quantitative radiomics parameter analysis, thus helping to identify hidden patterns in the data [9,10]. Compared to traditional statistical methods, ML algorithms can detect complex patterns in liver cancer treatment data more clearly. These algorithms simulate human learning process, effectively extracting and analyzing key features from multi-source data, such as tumor growth dynamics, changes in imaging performance, and post-treatment response [11,12]. This review aims to explore the latest research advancements of ML algorithms in predicting the prognosis of patients undergoing interventional therapy for liver cancer by summarizing recent key clinical data on interventional treatment outcomes, and discussing how these algorithms can optimize clinical decision-making processes, so as to improve treatment personalization and precision. Flowchart of the study is shown in Figure 2.

Figure 2.

Flowchart of the study.

Overview of liver cancer interventional therapies

Interventional therapy plays a crucial role in managing intermediate and advanced HCC patients, especially for those who are ineligible for surgery or transplantation. Since early-stage HCC often lacks obvious symptoms, most patients are diagnosed at an advanced stage [13]. In the Chinese Guidelines for the Diagnosis and Treatment of Primary Liver Cancer (2022 Edition) [14], interventional therapy is recommended as a treatment strategy spanning from early to late-stage liver cancer, with TACE recognized as one of the most commonly used non-surgical treatments for liver cancer. Similarly, in the internationally recognized Barcelona Clinic Liver Cancer (BCLC) Staging System and Treatment Strategy [15], interventional therapy, particularly ablation therapy, is recommended as the first-line treatment for very early-stage liver cancer (single tumor ≤ 2 cm) in patients not suitable for liver transplantation.

Ablation therapy has been shown to provide efficacy comparable to surgical resection in early-stage liver cancer patients [16]. For intermediate and advanced liver cancer, downstaging conversion therapy strategies, especially local interventional therapy, have played a key role in converting initially unresectable liver cancer into a resectable state [17]. After years of in-depth research, the combination of local interventional therapy and systemic targeted immunotherapy has been proven to significantly improve tumor response rates and conversion resection possibilities, becoming a crucial strategy for downstaging intermediate and advanced liver cancers [18]. A prospective phase II study in China (DoHAICs study, NCT05166772) evaluated the efficacy of HAIC combined with donafenib and sintilimab as a first-line treatment for unresectable liver cancer [19]. The results showed that the ORR of the combined treatment was 78.3%, with a conversion success rate of 65.2%. Recent studies have reported better tumor response rates at 1 month, 3 months, and 6 months following TACE combined with MWA treatment [7].

Although interventional therapy offers diverse treatment options and demonstrates good efficacy for HCC patients, challenges remain in predicting treatment response due to biological heterogeneity of tumors and complications, such as bleeding, liver and kidney function impairment, and ectopic embolization of embolic agents, which are crucial for ensuring patient safety and treatment efficacy [7,20]. These challenges highlight the importance of accurately predicting patient prognosis.

Overview of machine algorithms

Machine algorithms, such as ML and deep learning, have shown significant achievements and application potential in medical prognosis prediction. ML technology can automatically learn and identify patterns from historical data, while deep learning simulates the human brain’s working mechanisms, using complex algorithms to process and analyze data.

Logistic regression (LR), known for its simple model and high computational efficiency, is widely used in liver cancer diagnosis, prediction, and prognosis assessment [21,22]. Support vector machines (SVM) are powerful classification algorithms that find optimal hyperplanes in the feature space to differentiate between categories. In medical image processing, SVMs are often combined with feature extraction techniques [23]. Random forest (RF) algorithms improve classification accuracy and evaluate the importance of different clinical features by constructing multiple decision tree models [24]. Although the k-nearest neighbors (KNN) algorithm has limitations in computational efficiency, its simplicity and intuitive nature still make it suitable for classification tasks in small patient datasets [25]. Naive Bayes simplifies computation by assuming feature independence, making it useful for handling gene expression data with many features and helping identify relevant genetic markers [26]. Gradient boosting decision trees (GBDT) construct models through iterative optimization, accurately evaluating and guiding the latest model with annotations generated from privileged information [27].

Deep learning algorithms, such as convolutional neural networks (CNN), recurrent neural networks (RNN), and deep neural networks (DNN), automatically extract features in medical image analysis, improving the accuracy of liver cancer diagnosis and metastasis prediction [28,29]. Wu et al. [30] proposed a phase difference network (PDN), utilizing phase difference and multi-head self-attention mechanisms to distinguish HCC and ICC from four-phase CT images, showing better performance than traditional deep learning methods. The Transformer, a neural network model based on self-attention mechanisms, consists of multiple encoders and decoders. Tang et al. [31] developed a hybrid model combining graph neural networks and Transformer, which effectively utilized global context information from whole-slide images, significantly improving the accuracy and clinical value of HCC prognosis assessment.

Generative adversarial networks (GANs) demonstrate significant potential in medical image enhancement and data generation. The 2021 AI data challenge successfully showcased GANs’ ability to generate numerous rare malignant tumor MRI images from a few real MRI samples, with qualitative and quantitative evaluations confirming its effectiveness [32]. Bayesian networks, with their advantages in handling uncertainty and causal relationship modeling, provide deeper insights into liver cancer prediction and treatment. Cai et al. [33] identified portal vein tumor thrombus as the most important predictor of survival time in HCC patients after liver resection using Bayesian networks and importance measurement methods.

The clinical application process of ML includes clinical data acquisition and preprocessing, feature selection, model training and tuning, model diagnosis, multi-model fusion, and deploying validated models into clinical practice to assist in diagnosis and treatment decisions (Figure 3). In clinical applications, the ML process must particularly focus on data representativeness and diversity, ensuring the model generalizability to different patient groups while considering interpretability, allowing doctors to understand and trust its outputs.

Figure 3.

Flow of the machine learning algorithm. LR: Logistic Regression; SVM: Support Vector Machine; RF: Random Forest; KNN: K-Nearest Neighbors; XGBoost: eXtreme Gradient Boosting; GBDT: Gradient Boosting Decision Tree; CNN: Convolutional Neural Network; RNN: Recurrent Neural Network; RFE: Recursive Feature Elimination; PCA: Principal Component Analysis; AUC-ROC: Area Under the Receiver Operating Characteristic Curve; MSE: Mean Squared Error; RMSE: Root Mean Squared Error.

Application of ML algorithms in predicting prognosis for liver cancer patients undergoing interventional therapies

TACE

Many HCC patients are diagnosed at intermediate or advanced stages, missing the optimal treatment window. TACE is the gold-standard treatment for intermediate HCC patients [34,35], but patients’ responses to TACE vary significantly, and not all benefit from it [36]. Therefore, developing models to predict the efficacy of TACE is crucial.

Previous studies have constructed various ML models using clinical and radiological variables to predict outcomes for HCC patients post-TACE. Spleen volume (SV) is an undervalued, automatically retrievable imaging biomarker. Muller et al. [37] used CNN algorithms to automatically assess SV and found that high SV correlated significantly with reduced survival in liver cancer patients. SV is also a strong predictor of hepatic decompensation post-TACE. Bartnik et al. [38] analyzed radiomics features from multiple organs of interest in liver cancer patients using deep learning, showing that their radiomics model outperformed traditional clinical models in predicting progression-free survival (PFS) after TACE. These studies highlight the importance of non-tumor regional features in clinical prediction. Bernatz et al. [39] analyzed CT images after three consecutive TACE procedures, combining radiomics features with clinical mHAP-II scores. Their RF model achieved an AUC of 0.70 and an accuracy of 0.72 at the lesion level, an AUC of 0.62 at the patient level, and a C-index of 0.67 for OS prediction, demonstrating potential in improving TACE response prediction. Dong et al. [40] selected clinical data from patients receiving their first TACE for unresectable liver cancer, identified three features (portal vein tumor thrombus type, albumin level, and intrahepatic tumor distribution) using the LASSO algorithm, and built various prognostic models (XGBoost, decision tree, SVM, RF, KNN, and ANN). Among them, the RF model performed best, with an AUC of 0.802, accuracy of 0.784, sensitivity of 0.904, and specificity of 0.480. Ma et al. [41] compared different machine algorithms for predicting responses in unresectable HCC patients receiving lenvatinib combined with TACE, finding that SVM and RF models performed best in accuracy and AUC. The RF model reached an AUC of 0.91, indicating high predictive accuracy. Peng et al. [42] emphasized the high accuracy of radiomics conventional ML and deep learning models in preoperative TACE response prediction (deep learning model AUC = 0.972, integrated model AUC = 0.994). Combining these models with clinical variables offers a novel and accurate method for predicting treatment responses in intermediate-stage liver cancer patients. Zhang et al. [43] developed and validated a fully automated deep learning framework to predict TACE response in real time for HCC patients. Overall, these studies demonstrate the potential of ML models in predicting TACE response and the importance of non-tumor regional features and automated imaging analysis. Relevant studies are summarized in Table 1.

Table 1.

ML-based prognostic model characteristics of HCC patients after TACE

Author	ML Algorithms	Prediction Targets	Key Predictors	Main Results/Performance Indicators	Model Validation and Interpretability	Other Important Findings
Muller 2022 [32]	CNN	OS, PFS, TTUP	SV	Significant correlation between SV and survival rates	Internal validation, S⊘rensen Dice score, Bland-Altman plot	Spleen volume significantly correlates with risk of liver dysfunction after TACE
Bartnik 2024 [33]	DL, RSF, COX	OS, PFS	Tumor VOI and non-tumor VOI	OS: C-index range 0.616 to 0.640. PFS: C-index 0.713	Cross-validation, XAI	Multiple VOI features extracted from CT images, overcoming manual segmentation limitations
Bernatz 2023 [34]	RF	TACE response, OS	Radiomic features and clinical mHAP-II score	Lesion-level AUC 0.70, Accuracy 0.72; Patient-level AUC 0.62; C-index 0.67	Reliability and redundancy analysis	Supports the potential of lipid deposition as an imaging biomarker
Dong 2021 [35]	XGBoost, Decision Tree, SVM, RF, KNN, ANN	Early treatment response post first cTACE	Portal vein tumor thrombus type, Albumin level, Tumor distribution in liver	RF model performed best, AUC 0.802, Accuracy 0.784, Sensitivity 0.904, Specificity 0.480	5-fold cross-validation	Portal vein tumor thrombus type is the most important factor for predicting response to first cTACE treatment
Ma 2023 [36]	CART, AdaBoost, XGBoost, RF, SVM	Response to combination therapy (lenvatinib + TACE)	K, LDL, D-D, Red blood cells, ALT, ALB, Mono, Tumor size, TG, and Age	RF model AUC 0.91, SVM and RF performed best	SHAP algorithm enhanced model interpretability	Lower serum K, older age, higher BMI, and larger tumor size correlate with better efficacy of combination therapy
Peng 2021 [37]	Linear model, LR, SVM, GBM, RF, DL	TACE treatment response	Tumor size	DL model AUC 0.972, Integrated model AUC 0.994	Multicenter data validated model robustness	Tumor size significantly correlates with initial treatment response, while AFP levels do not
Zhang 2022 [38]	ResNet18 and Multilayer Perceptron	TACE treatment response	DSA video information, Demographics, and liver function parameters	Accuracy rates on internal and external validation sets were 78.2% and 75.1% respectively	Internal and external validation	Predictive model performance using segmentation results as input is slightly lower than using true segmentation results, but not significantly

ML: Machine Learning; CNN: Convolutional Neural Network; OS: Overall Survival; PFS: Progression-Free Survival; TTUP: Time to Tumor Progression; SV: Segmentation Volume; TACE: Transarterial Chemoembolization; VOI: Volume of Interest; DL: Deep Learning; RSF: Random Survival Forest; COX: Cox Proportional Hazards Model; RF: Random Forest; AUC: Area Under the Curve; mHAP-II: Modified Hepatoma Arterial Embolization Prognostic Score; SVM: Support Vector Machine; KNN: k-Nearest Neighbors; GBM: Gradient Boosting Machine; LR: Logistic Regression; DL: Deep Learning (used in the context of the algorithm name); AFP: Alpha-Fetoprotein; ALT: Alanine Aminotransferase; ALB: Albumin; Mono: Monocytes; TG: Triglyceride; BMI: Body Mass Index; DSA: Digital Subtraction Angiography; AUC: Area Under the Receiver Operating Characteristic Curve; XAI: Explainable Artificial Intelligence; SHAP: SHapley Additive exPlanations.

TACE models typically combine clinical, radiological, and radiomics features, with non-tumor regional features (such as spleen volume) serving as important predictors. These models employ deep learning and conventional ML algorithms, with RF models showing excellent performance in several studies. AUC values range from 0.62 to 0.994, with most studies reporting AUCs above 0.8. Key predictive features include portal vein tumor thrombus, albumin levels, and intrahepatic tumor distribution. Automated imaging analysis and deep learning frameworks demonstrate significant potential, offering new perspectives for predicting TACE efficacy.

HAIC

HAIC involves catheterizing the hepatic artery supplying the tumor and continuously infusing chemotherapeutic agents. It is advantageous for advanced liver cancer with portal vein tumor thrombus, arterio-venous fistula, and poorly vascularized liver metastases [35].

Recent advances show that combining deep learning with radiomics features significantly improves the predicted accuracy of HAIC treatment response. Xu et al. [44] developed the DLRN model, integrating deep learning, radiomics features, and key clinical variables, achieving high accuracy in training, internal, and external validation cohorts (AUCs of 0.988, 0.915, and 0.896, respectively). The model also predicted survival based on treatment response, with the median OS in the response group significantly higher than in the non-response group. Quan et al. [45] used the InceptionV4 CNN model with preoperative MRI data and clinical factors (HAIC cycle count, tumor thrombus, neutrophil-lymphocyte ratio, and gamma-glutamyltransferase), achieving an AUC of 0.871 in the training cohort and 0.826 in the internal validation cohort. Another retrospective study used a combination model of MRI radiomics and ALBI score to predict HAIC treatment response, providing a nomogram to assess PFS [46]. Patients with high scores had a median PFS of 6.0 months, significantly shorter than 9.0 months in low-score patients. He et al. [47] further explored radiomics features extracted from dual-phase contrast-enhanced CT (CECT), combined with clinical variables and MTM subtypes, and established a multi-task deep learning radiomics (MDLR) model to provide accurate HAIC prognostic risk stratification for HCC patients. Relevant studies are summarized in Table 2.

Table 2.

ML-based prognostic model characteristics of HCC patients after HAIC

Author	ML Algorithms	Prediction Target	Key Predictive Factors	Main Results/Performance Metrics	Model Validation & Interpretability	Other Important Findings
Xu 2022 [39]	DL, XGBoost	OR	APE, RVI, R score, DL score	AUC in training set = 0.988, internal validation set AUC = 0.915, external validation set AUC = 0.896	Internal and external validation	Radiological parameters (APE and RVI) may predict the efficacy of HAIC better than clinical characteristics
Quan 2024 [40]	InceptionV4-CNN	HAIC response	MRI data, HAIC cycles, cancer thrombus, NLR	AUC in training cohort = 0.871, internal validation cohort AUC = 0.826	Cross-validation and independent validation, CAM used for visualization	Age, HAIC cycle number, tumor thrombus, extrahepatic spread, and AST level are independent predictors
Zhao 2023 [41]	LR	PFS	Radiomic score (Radscore) and ALBI score	Combined model AUC in training and validation sets are 0.79 and 0.75, respectively	Internal validation	NA
He 2023 [42]	MDLR	Post-HAIC patient prognosis	CECT radiomic features, portal vein cancer thrombus, HAIC response, HAIC cycles	AUC for survival prediction model in internal and external validation sets are 0.87 and 0.83	Internal and external validation	Tumor burden and distribution as well as tumor microenvironment features are associated with prognosis

XGBoost: Extreme Gradient Boosting; OR: Objective Response; APE: Asymmetry of Parenchymal Enhancement; RVI: Reduction in Viable Tumor on Initial; R score: Radiographic Response Score; DL score: Deep Learning Score; HAIC: Hepatic Arterial Infusion Chemotherapy; MRI: Magnetic Resonance Imaging; NLR: Neutrophil-to-Lymphocyte Ratio; CAM: Class Activation Mapping; PFS: Progression-Free Survival; Radscore: Radiomic Score; ALBI: Albumin-Bilirubin Grade; MDLR: Multitask Deep Learning Radiomics; CECT: Contrast-Enhanced Computed Tomography; AST: Aspartate Aminotransferase; NA: Not Available.

HAIC models often integrate deep learning, radiomics features, and clinical variables, commonly using pre-treatment MRI and CT images for feature extraction. These models achieve AUC values ranging from 0.826 to 0.988, with key predictive features including HAIC cycle count, tumor thrombus, neutrophil-lymphocyte ratio, and gamma-glutamyltransferase. MDLR models exhibit high accuracy, capable of predicting both treatment response and survival outcomes, providing personalized prognostic assessment tools for HAIC treatment.

TARE

Transarterial radioembolization (TARE), also known as selective internal radiation therapy (SIRT), involves injecting the radioactive isotope yttrium-90 (90Y) to treat liver cancer.

Roll et al. [48] extracted and analyzed radiomics features from pre-TARE CT images of patients with colorectal cancer liver metastases. Two independent radiomics features (energy and maximum correlation coefficient) reflected tumor heterogeneity. Their multivariate LR model successfully distinguished high-risk from low-risk patients, with an AUC of 0.75, providing a new prognostic assessment tool. Ince et al. [49] found that ML models (SVM, LR, RF, LightGBM) combining radiomics features from pre-treatment contrast-enhanced MRI and clinical data significantly improved TARE response prediction. Kobe et al. [50] used features from pre-TARE CBCT images, achieving high sensitivity (94.2%) and moderate specificity (67.7%) with a multi-layer perceptron ANN in an external test set. Marinelli et al. [51] collected baseline and early post-TARE MRI of HCC patients, using semi-automatic segmentation to extract radiomics features. Their XGBoost model showed high accuracy (AUC = 0.89) in an independent validation cohort, particularly outperforming models using only clinical parameters and conventional imaging features in predicting complete response. Balli et al. [52] combined dynamic MRI radiomics scores with clinical features using LASSO feature selection and LR to build a radiomics model. Triple cross-validation optimized parameters, with the model predicting response to 90Y TARE in intrahepatic cholangiocarcinoma patients, showing that responders had significantly lower radiomics scores. Axial T2W with fat suppression sequence achieved an AUC of 0.839, indicating high predictive accuracy. Aujay et al. [53] compared the European Association for the Study of the Liver (EASL) criteria, using radiomics combined with MRI data to assess treatment response in patients with locally advanced HCC undergoing TARE. They found that long emphasis, short axis length, surface area, and grayscale non-uniformity in arterial phase images could accurately predict early treatment response, demonstrating the potential of radiomics combined with LR in predicting TARE efficacy. Relevant studies are summarized in Table 3.

Table 3.

ML-based prognostic model characteristics of HCC patients after TARE

Author	ML Algorithms	Prediction Target	Key Predictive Factors	Main Results/Performance Metrics	Model Validation & Interpretability	Other Important Findings
Roll 2024 [43]	Multivariate LR	Treatment Response and Survival Outcome	Energy, Maximal Correlation Coefficient	AUC: 0.75	Feature selection with Boruta algorithm	Radiomic analysis can quantify tumor heterogeneity, including blood supply, cellular vitality, density, and fibrosis
Ince 2023 [44]	SVM, LR, RF, LightGBM	Treatment Response	8 Radiomic features, 4 Clinical features	AUC: 0.88-0.94	5-fold cross-validation	Age and preoperative total bilirubin level significantly correlate with TARE treatment response
Kobe 2021 [45]	Multilayer Perceptron, ANN	Disease Control (PR/SD) and PD	104 Texture Analysis Features from CBCT, 15 features after selection	AUC: 0.85, Sensitivity 94.2%, Specificity 67.7%	10-fold cross-validation	NA
Marinelli 2023 [46]	XGBoost	Treatment Response at 4-6 Months Post-Treatment	Radiomic features from baseline and early post-treatment (1-2 months) MRI images	AUC: 0.89	NA	Combined baseline and early follow-up MRI radiomic data better predict patient treatment response
Balli 2024 [47]	LASSO, LR	Radiological Response at 6 Months Post-Treatment	Rad-score, Bifurcation Lesions	AUC: 0.696-0.880	DeLong test, NRI, IDI	First study to use MRI radiomics to predict TARE treatment response in ICC patients
Aujay 2022 [48]	LR	Treatment Response	Longitudinal Emphasis, Minor Axis Length, Surface Area, and Gray Level Non-uniformity	AUC: 1	Cross-validation	Heterogeneity parameters in arterial and portal venous phase images before and after treatment not significantly different between responders and non-responders

LightGBM: A gradient boosting framework that uses tree-based learning; ANN: Artificial Neural Network; PR: Partial Response; SD: Stable Disease; PD: Progressive Disease; CBCT: Cone Beam Computed Tomography; ICC: Intrahepatic Cholangiocarcinoma; TARE: Transarterial Radioembolization; NRI: Net Reclassification Improvement; IDI: Integrated Discrimination Improvement.

TARE models utilize pre-treatment imaging (CT, MRI, CBCT) for radiomics feature extraction, with the combination of radiomics features and clinical data significantly improving prediction accuracy. These models employ various ML algorithms, including SVM, LR, RF, LightGBM, and XGBoost, with AUC values ranging from 0.75 to 0.89. The models can predict both overall response and complete response, with key radiomics features reflecting tumor heterogeneity. These models offer new perspectives for predicting TARE treatment efficacy, aiding in personalized treatment decisions.

RFA

Treatment guidelines from BCLC, CNLC, and NCCN emphasize RFA as the preferred method for early HCC with small tumors. Studies indicate that the five-year survival rate for RFA-treated patients ranges from 26% to 56.7%, and the five-year disease-free survival rate is between 15% and 28.7% [54-56]. However, the high recurrence rate, rapid tumor growth, and invasiveness post-recurrence remain to be clinical concerns [57].

Recent research explores ML in predicting outcomes for HCC patients undergoing RFA. One study used the XGBoost algorithm on multidimensional data from patients receiving localized RFA between 2018 and 2022 [58]. Their model achieved an accuracy of 78.9% and an AUC of 0.80 in an independent validation set. Tong et al. [59] compared five algorithms (LR, decision tree, GBDT, RF, GBM) in predicting overall mortality post-RFA, and identified platelet count, Alpha-Fetoprotein (AFP), age, tumor size, and total bilirubin as key prognostic factors. GBM was found to have the highest accuracy (0.681), indicating the potential and differences of various algorithms in predicting HCC patient prognosis. Sato et al. [60] developed a transformer-based ML model analyzing data from 1778 treatment-naïve HCC patients undergoing RFA, aiming to improve the prediction of OS. This model used clinical and pathological features, evaluated by Harrel’s c-index, showing superior discrimination compared to traditional deep learning models. Results indicated the transformer’s high discriminative ability in external validation cohorts and its capacity to provide personalized cumulative recurrence prediction curves. Another study analyzed 898 early-stage HCC patients using Lasso and Cox regression analysis to identify independent risk factors like age, gender, BCLC stage, tumor size, globulin, and γ-glutamyl transpeptidase [61]. The nomogram, validated by C-index, ROC, calibration, and decision curve analysis (DCA), showed excellent discrimination, consistency, and clinical utility. RFA treatment showed potential in improving long-term survival for solitary HCC patients with tumor diameters ≤ 5 cm. He et al. [62] revealed the effectiveness of RFA in improving 5-year OS and cancer-specific survival (CSS) rates compared to radiotherapy, chemotherapy, and blank control groups by analyzing data from the SEER database. Further Cox regression analysis and the development of the XGBoost model identified key prognostic factors such as age, race, marital status, grade, cirrhosis, tumor size, and AFP level, and constructed a valid predictive model. The XGBoost model demonstrated good predictive performance in the validation cohort through ROC curve, calibration plot, and DCA, providing a personalized CSS predictive tool for patients with isolated HCC with a diameter of less than 5 cm. The relevant studies are listed in Table 4.

Table 4.

ML-based prognostic model characteristics of HCC patients after RFA

Author	ML Algorithms	Prediction Target	Key Predictive Factors	Main Results/Performance Metrics	Model Validation & Interpretability	Other Important Findings
Hamed 2024 [53]	XGBoost	Disease control at 12 months	Child-Pugh score, WBC count, Heparan concentration, Diabetes, Hypertension, Tumor size, AFP	Accuracy and AUC are 78.9% and 0.80, respectively	Internal validation	1-year survival and local control rates are 94.6% and 61.3%, respectively
Tong 2021 [54]	RF, LR, LightGBM, GBDT, Decision Tree	Total mortality	PLT, AFP, Age, Tumor size, Total bilirubin	GBDT has the highest accuracy (0.681), precision (0.721), AUC: 0.714	Internal validation	NA
Sato 2024 [55]	Transformer-based ML model (SurvTRACE)	OS	Age, Gender, Number and size of tumors, Liver function indicators (Albumin, Total bilirubin, AST, ALT), Tumor markers (AFP, DCP), Hepatitis virus infection status, Platelet count, Prothrombin time	C-index of 0.69	Internal and external validation	5-year and 10-year survival rates are 63.7% and 30.4%, respectively. 1-year, 3-year, and 5-year local tumor recurrence rates are 1.7%, 5.3%, and 6.5%, respectively
Zhang 2024 [56]	Lasso regression and Cox regression analysis	RFS	Age, Gender, BCLC stage, Tumor size, Glob, γ-GT	AUC for 1-year, 3-year, and 5-year RFS are 0.721, 0.756, and 0.779, respectively	Internal validation	NA
He 2023 [57]	LR, SVM, RF, KNN, XGBoost	CSS	Liver cirrhosis, Tumor size, AFP level, Age, Marital status	XGBoost model’s AUC for predicting 1-year, 3-year, and 5-year CSS are 0.88, 0.81, and 0.79, respectively	Internal validation	Compared to other treatment modalities, RFA shows better performance in improving OS and CSS for patients with single HCC ≤ 5 cm, but still lower than hepatectomy

RFA: Radiofrequency Ablation; RF: Random Forest; PLT: Platelet Count; DCP: Des gamma Carboxyprothrombin; RFS: Recurrence-Free Survival; BCLC: Barcelona Clinic Liver Cancer; Glob: Globulin; γ-GT: γ-Glutamyl Transpeptidase; CSS: Cancer-Specific Survival; c-index: Concordance Index; AFP: Alpha-Fetoprotein.

RFA models utilize multidimensional clinical and pathological data, comparing various algorithms including XGBoost, LR, decision tree, GBDT, RF, and GBM. AUC values range from 0.68 to 0.80, with key prognostic factors including platelet count, AFP, age, tumor size, and total bilirubin. Transformer-based models show promise in predicting OS. These models can predict both overall mortality and cancer-specific survival, providing powerful tools for prognostic assessment following RFA treatment.

MWA

MWA is a commonly used ablation method, particularly for tumors with diameters ranging from 3 to 5 cm. It has been shown to have high efficacy and ablation efficiency. Compared to radiofrequency ablation (RFA), MWA significantly shortens the procedure time and is less sensitive to the heat-sink effect of blood flow, thereby reducing the risk of incomplete ablation in the treatment of larger tumors [55]. However, despite its advantages in ablation efficiency, MWA does not show significant differences in local efficacy, complication rates, or long-term survival outcomes compared to RFA [63].

In a study predicting local tumor progression (LTP) in early-stage HCC patients post-MWA, Ren et al. [64] analyzed the clinicopathological data and ablation parameters of 607 untreated early HCC patients. They developed predictive models using four ML algorithms, including CatBoost, RF, XGBoost, and LR. Among these models, the CatBoost algorithm, which combined nine key variables - tumor number, albumin and alpha-fetoprotein levels, tumor size, age, and international normalized ratio - exhibited the highest predictive accuracy (AUC: 0.898). The study suggested that precise ablation planning and personalized treatment based on these predictive factors could significantly reduce the risk of LTP, thereby improving the success rate of MWA in treating early-stage HCC.

Another study [65] employed ML techniques to predict early recurrence (ER) risk by analyzing the clinical data of 1,574 early HCC patients who underwent MWA. This study constructed ML models, including RF, support vector machine (SVM), and XGBoost, and enhanced their interpretability using SHAP and LIME algorithms. The XGBoost model performed best in predicting ER, accurately identifying key risk factors such as tumor number, platelet count, and alpha-fetoprotein level. Their XGBoost-based prediction system is available online, providing clinicians with a practical tool (http://114.251.235.51:8001/).

In the study on the predictability of LTP after MWA in colorectal cancer liver metastases, clinical data and MRI radiomic features were analyzed to develop two combined models [66]. Model 2, which incorporated T2 fat-suppressed and early arterial phase T1 fat-suppressed features, showed better performance in predicting LTP, with an AUC of 0.981. These highly accurate models offer new perspectives for clinical practice, but the studies also emphasize the need for further large-scale research to validate the generalizability and reliability of these models. Related research is summarized in Table 5.

Table 5.

ML-based prognostic model characteristics of HCC patients after MWA

Author	ML Algorithms	Prediction Target	Key Predictive Factors	Main Results/Performance Metrics	Model Validation & Interpretability	Other Important Findings
Ren 2023 [59]	CatBoost, SVM, RF, LR	LTP	Number of tumors, Albumin, AFP, Tumor size, Age, INR	Best performance by CatBoost model, AUC of 0.898	Internal and external validation	NA
An 2022 [60]	LR, RF, SVM, XGBoost	ER	Tumor number, Platelets, AFP, Comorbidity score, WBC, ChE, PT, Neutrophils, Etiology	Best performance by XGBoost model, AUC 0.74 (internal) and 0.76 (external)	Internal and external validation, SHAP and LIME algorithms for model explanation	NA
Shahveranova 2023 [61]	LR	LTP	Preoperative extrahepatic metastasis, Tumor size, CA 19-9	Combined Model 2 (clinical data and Phase 2 radiomic features) has the highest discriminative performance for LTP prediction (AUC 0.981)	NA	LTP group patients have significantly higher radiomic scores in both MRI phases (Phase 1 and Phase 2)

CatBoost: A machine learning algorithm based on decision trees; LTP: Local Tumor Progression; INR: International Normalized Ratio; ER: Early Recurrence; ChE: Cholinesterase; PT: Prothrombin Time; WBC: White Blood Cell Count; CRLM: Colorectal Liver Metastasis; CA 19-9: A tumor marker for gastrointestinal malignancies; LIME: Local Interpretable Model-agnostic Explanations; AFP: Alpha-Fetoprotein.

MWA models focus primarily on predicting LTP and ER, utilizing various ML algorithms including CatBoost, RF, XGBoost, and LR. AUC values range from 0.898 to 0.981, with key predictive features including tumor number, albumin levels, AFP, tumor size, and platelet count. These models combine clinical data with MRI radiomics features, with some studies providing online prediction tools for clinical use, offering important references for prognostic assessment and personalized treatment following MWA.

Overall, these ML-based prognostic models for HCC patients undergoing interventional therapies share common characteristics. They typically combine clinical, radiological, and radiomics features to improve prediction accuracy. Various ML algorithms are applied, with RF and XGBoost frequently performing well. Pre-treatment imaging (CT, MRI) is commonly used for feature extraction. These models can predict various outcomes, including treatment response, survival, and recurrence. AUC values generally range from 0.7 to 0.9, indicating good to excellent predictive performance. Furthermore, there is an increasing trend towards using deep learning and multi-task learning approaches, providing more precise and personalized tools for prognostic assessment following HCC interventional treatments.

In summary, DL and ML models have demonstrated high accuracy in predicting outcomes for various liver cancer treatments. For instance, in PFS prediction, models by Bartnik et al. [38] and Quan et al. [45] outperformed traditional clinical models. In predicting treatment response, studies by Muller et al. [37] and Peng et al. [42] highlighted the advantages of radiomics combined with deep learning. For OS prediction, models by Ma et al. [41] and Sato et al. [60] showed strong predictive capabilities. Additionally, in LTP prediction, the models by Ren et al. [64] and those incorporating MRI radiomics features [66] demonstrated exceptionally high accuracy (Table 6).

Table 6.

Predictive accuracy of machine algorithm models for liver cancer outcomes

Prediction Target	Author	Machine algorithm	AUC/C-index
PFS	Bartnik [38]	Deep Learning	-
PFS	Quan [45]	Deep Learning	0.826
Treatment Response	Muller [37]	CNN	-
Treatment Response	Peng [42]	Deep Learning	0.994
Treatment Response	Xu [39]	DLRN	0.988
CR Prediction	Marinelli [46]	XGBoost	0.89
OS	Ma [41]	RF	0.91
OS	Sato [60]	Transformer Model	-
LTP	Ren [64]	CatBoost	0.898
LTP	Shahveranova [66]	-	0.981

PFS: Progression-Free Survival; LTP: Local Tumor Progression; CNN: Convolutional Neural Network.

Conclusion

Currently, the application of ML algorithms in predicting the prognosis of interventional therapy for liver cancer focuses on two main areas. The first is the prognostic evaluation based on imaging features, such as analyzing CT and MRI imaging data to identify tumor size, morphology, and vascular characteristics, thereby predicting treatment efficacy and recurrence risk [67]. For instance, deep learning algorithms have demonstrated remarkable ability in analyzing liver cancer CT images, accurately identifying tumor boundaries and vascular invasion [68-70]. The second area involves integrating multidimensional clinical information of patients, including age, gender, tumor stage, and liver function, to construct complex prognostic models that help determine the optimal treatment plan.

We found that models integrating clinical, imaging, and radiomics features exhibit superior predictive accuracy. Although RF and XGBoost algorithms perform well in many cases, researchers are exploring a range of algorithms from traditional ML to advanced deep learning methods. Imaging, particularly pre-treatment CT and MRI, plays a crucial role in feature extraction across most models. These models predict not only treatment response but also survival and recurrence risk, demonstrating good to excellent predictive capabilities, with AUC values typically ranging from 0.7 to 0.9. As deep learning and multi-task learning approaches become more prevalent, prediction accuracy and personalization are continually improving.

However, despite the significant advances in applying ML to HCC and ICC interventional treatments, there are areas needing further investigation. For instance, many studies rely on single-center data, so multi-center, large-scale studies are necessary to validate models across different patient populations. Additionally, while certain features (e.g., tumor characteristics, liver function markers) consistently emerge as important predictors across various interventions, a deeper understanding of their biological basis is needed.

With increasing model complexity, ensuring interpretability is crucial for clinical adoption. Techniques such as SHAP and LIME, used in some studies, should be more widely implemented. Moreover, despite promising results, integrating these models into clinical decision-making remains a challenge. Future research should focus on developing user-friendly interfaces and decision support tools.

Most current models emphasize short-term outcomes, so developing models that can predict long-term outcomes and account for changes in patient status over time is a crucial next step. With the growing use of combination therapies, models capable of predicting outcomes for these complex treatment regimens will be increasingly valuable.

Furthermore, the high accuracy of personalized treatment plans highlights the potential for highly individualized therapy. Future research should concentrate on developing dynamic models that can adapt recommendations as patient conditions evolve. For example, Zhang et al.’s [43] automated deep learning framework for TACE response prediction points to the possibility of real-time treatment outcome prediction, potentially allowing immediate adjustments during the procedure.

In the realm of liver cancer prognosis prediction, ML algorithms face core challenges including algorithm bias, data quality, and ethical considerations (Figure 4). Algorithm bias primarily manifests as models potentially overfitting specific datasets, impacting their applicability across diverse patient populations [71]. Addressing this issue requires rigorous cross-validation and generalization capability assessments to ensure model robustness. On the data front, diversity and consistency of data are crucial for enhancing model performance [72]. Promoting multicenter studies and establishing unified data standards can help expand dataset scope and improve model generalizability. Additionally, algorithm transparency and fairness are critical ethical considerations, necessitating that model decision processes be interpretable and unbiased, with effective oversight mechanisms in place [71,73-75].

Figure 4.

Summary of current evidence and future challenges for prognostic indicators in interventional therapy for liver cancer. AFP: Alpha-Fetoprotein; ALB: Albumin; INR: International Normalized Ratio; BCLC: Barcelona Clinic Liver Cancer.

Looking ahead, liver cancer treatment will become more personalized and precise. The integration of high-resolution imaging technologies and advanced ML algorithms will foster the development of intelligent decision support systems based on imaging features. Furthermore, emerging therapies such as targeted therapies and immunotherapies, along with in-depth research into biomarkers and molecular mechanisms, will drive the optimization of liver cancer treatment strategies, improving treatment outcomes and patient quality of life. Interdisciplinary collaboration and in-depth research are essential for continued progress in this field.

Disclosure of conflict of interest

None.