Artificial intelligence to predict hepatocellular carcinoma risk in cirrhosis

Abstract

Hepatocellular carcinoma (HCC) is the second leading cause of cancer-related deaths worldwide. The primary risk factor for HCC is cirrhosis. Identifying individuals who are at high risk of developing HCC will have numerous benefits for patient outcomes, patient quality of life, and the global healthcare system. Artificial intelligence (AI) has the capability to develop systems that emulate human intelligence. Recent studies have highlighted the potential of AI in the management of HCC, and the application of AI appears promising for identifying high-risk groups among patients with cirrhosis who require closer monitoring. Ultimately, the aim of AI in the field of HCC clinical care is to enable earlier diagnosis and consequently improve prognosis.

Key Words: Hepatocellular carcinoma; Cirrhosis; Prediction; Artificial intelligence; Prognosis; Machine learning

Core Tip: Hepatocellular carcinoma (HCC) is a global health problem and cirrhosis is the principal risk factor for its development. Early diagnosis of HCC is associated with a better prognosis because curative treatment is feasible. The recommended screening strategy for patients with cirrhosis is ultrasound evaluation twice per year. However, the risk estimation of HCC is not the same in patients with cirrhosis, and as such risk stratification scores have been a focus of study in this population. The application of artificial intelligence has shown promise in screened patients with cirrhosis who are at the highest risk of developing HCC.

INTRODUCTION

Liver cancer is the fifth most common cancer globally, and hepatocellular carcinoma (HCC) represents 90% of primary liver cancers. The incidence of HCC has increased worldwide[1,2], with a 75% increase in new HCC cases reported between 1990 and 2015[3]. There are several risk factors for HCC, but cirrhosis is the primary one. Cirrhosis represents the final stage in the progression of chronic liver diseases, and it is estimated that one-third of patients with cirrhosis will develop HCC[4]. Unfortunately, comorbidity of HCC and cirrhosis is associated with a poor prognosis, unless it is diagnosed early. Curative treatment is recommended only for very early stage (0) and early stage (A) patients (by the modified Barcelona Clinic Liver Cancer staging system)[1].

Systematic screening of at-risk populations, particularly individuals with cirrhosis, is of critical importance. Since the risk of developing HCC is not uniform among patients with cirrhosis, several prognostic scoring systems have been developed to identify those patients with cirrhosis at the highest risk. Among these, the Toronto Hepatocellular Carcinoma Risk Index (THRI)[5-7] is the most extensively studied and validated.

The goal of artificial intelligence (AI) is to create machines able to simulate human intelligence. AI has been increasingly applied in the health care field. As a multidisciplinary field, incorporating computer science and mathematics to develop and implement computer algorithms facilitates improved predictive accuracy from static or dynamic data sources via analytical or probabilistic models. Recent studies have focused on the contribution of AI in the diagnosis and prediction of treatment results in patients with HCC. AI has shown promising results in predicting HCC in patients with cirrhosis. By stratifying patients at high risk for HCC, clinicians can recommend special monitoring to diagnose HCC at an early stage, leading to curative intervention and improving prognosis[8,9].

This article reviews the current application of AI in the prediction of HCC in patients with cirrhosis and discusses the progress of accurate stratification enabling better screening and prognosis.

SCREENING FOR HCC

HCC develops in 90% of patients with cirrhosis[10]. Therefore, the European Association for the Study of the Liver guidelines recommend HCC screening with abdominal ultrasound every 6 months for patients with cirrhosis[1]. Despite regular surveillance, however, less than 20% of patients are diagnosed early, making them ineligible for curative treatment options[11].

Several scoring models have been developed to better predict the risk of HCC and include the Risk Evaluation of Viral Load Elevation and Associated Liver Disease/Cancer-Hepatitis B Virus nomogram[12], the Guide with Age, Gender, hepatitis B virus (HBV) DNA, Core Promoter Mutations and Cirrhosis risk score[13], the Platelet, Age, Gender, hepatitis B (PAGE-B) score[14], the Risk Estimation for Hepatocellular Carcinoma in Chronic Hepatitis B[15], and the Hepatitis C Antiviral Long-Term Treatment Against Cirrhosis model[16]. The PAGE-B score was recently modified to include albumin as a biological parameter, resulting in a better performance score than the original PAGE-B score[17].

These models were developed in patients with viral infections and varying stages of liver fibrosis; few scores have been developed for patients with cirrhosis. The ADRESS-HCC model was developed to predict the 1-year risk of HCC in patients with cirrhosis in the United States[18]. The THRI score was also developed to stratify HCC risk in patients with cirrhosis, and distinctly includes age, sex, platelet count, and etiology[5]. It has been validated to identify patients at high risk of developing HCC and to offer personalized surveillance recommendations[5-7]. According to this score, patients with cirrhosis are classified into three groups: Low-risk patients (< 120); intermediate risk patients (120-240); and high risk patients (> 240)[6].

Traditional statistical methods are quite limited compared with AI-based predictive models. Conventional approaches depend on assumptions of linearity, normality, and restricted variable interactions. However, AI models can detect linear and nonlinear relationships between predictors and outcomes as well as subtle patterns and emergent neo trends that are often undetected by the traditional statistical analyses. The capacity of AI to rapidly process[19] and learn from extensive, multifaceted datasets makes it particularly valuable for contemporary clinical applications, especially in complex conditions like cirrhosis. Because of the intricate nature of cirrhosis, traditional statistical analysis has fallen short. AI, however, has been shown to have superior predictive performance through its ability to simultaneously integrate and analyze multiple factors.

AI

Definitions

AI encompasses a diverse range of technologies designed to enable machines to perform tasks that typically require human reasoning and cognitive abilities.

Machine learning (ML) allows the system to automatically learn and improve from experience rather than requiring explicit programming. ML is performed as supervised, unsupervised, semi-supervised, or reinforcement learning. AI methodologies have been used to predict HCC development in individuals with cirrhosis, and the predictive models were constructed using supervised ML approaches, including logistic regression (LR), decision trees (DT), support vector machines, K-nearest neighbors, and artificial neural networks. These models were developed through the analysis of relationships between predictor variables and annotated outcome data[8,20].

Deep learning (DL) is a specialized approach to ML and is characterized by neural networks containing multiple nonlinear processing layers. Unlike ML methods that may require manual feature extraction, DL automatically discovers intricate patterns through its hierarchical structure. This structure has numerous hidden layers and rich interconnections, having superior performance to conventional artificial neural networks, particularly for complex tasks involving unstructured data like images, speech, and text.

There are several types of ML algorithms that we will briefly mention. LR is a supervised learning algorithm used to predict the probability of a categorical outcome (binary or multinomial) based on one or more independent variables[21]. LR is commonly employed to model the association between risk factors (predictors) and a categorical outcome. DTs are a supervised learning model that utilizes a hierarchical tree-like structure to predict the value of a target variable by recursively partitioning the data based on decision rules derived from feature attributes[22]. Random forests (RFs), are constructed of multiple DTs[23] during training and then aggregate their predictions to improve accuracy and reduce over fitting[23].

Extreme gradient boosting (XGBoost) sequentially builds DTs to correct errors from the previous one while incorporating regularization to prevent over fitting[24]. Adaptive boosting is a supervised ensemble method that sequentially constructs weak classifiers (e.g., decision stumps). The algorithm begins with uniform weights and performs successive reweighting rounds where incorrectly predicted samples receive higher importance. Final predictions combine all weak learners through weighted aggregation, enhancing classification performance and diagnostic robustness[24]. Finally, multilayer perceptron is a class of feed forward artificial neural networks[24] composed of an input layer, one or more hidden layers of nonlinear processing units, and an output layer. The network is then able to learn complex hierarchical representations for classification or regression tasks.

AI models for predicting HCC in patients with cirrhosis

As summarized in Table 1, recent studies have demonstrated the superior performance of AI over conventional models. For example, an XGBoost model achieved areas under the curve (AUCs) of 0.829-0.832 in stratifying patients with HBV-related cirrhosis[24]. While RF models excelled in hepatitis C virus (HCV) cohorts (AUC up to 0.9507 for 1-year prediction)[25].

Table 1 Application of artificial intelligence for the prediction of hepatocellular carcinoma in patients with cirrhosis.

Ref.	Cohort	Data source	AI-based machine learning algorithms	Input	Results
Xu et al[24]	6980 patients; training cohort: 20%; validation cohort: 80%	Hospital of Nanchang University (patients with HBV-related cirrhosis)	XGBoost; LR; RF; AdaBoost; MLP	Clinical and biological data	XGBoost was the most efficient model (AUC: 0.829, 95%CI: 0.804-0.852) in the test set; AUC: 0.832, 95%CI: 0.807-0.857 in the validation set)
Zou et al[25]	400 patients with HCV-related cirrhosis who achieved SVR with direct-acting antivirals	Chronic Hepatitis C Research Program of Jiangsu (China)	RF	Clinical and biological data	AUC for the longitudinal models were 0.9507 (0.8838-0.9997), 0.8767 (0.6972-0.9918), and 0.8307 (0.6941-0.9993) for 1-year, 2-year, and 3-year risk prediction, respectively
Nam et al[26]	424 patients with HBV-related cirrhosis	Two tertiary hospitals (Republic of Korea)	DL-based model	Clinical and biological data	Better performance of DL-based model compared with previously reported risk models
Ioannou et al[27]	48151 patients with HCV-related cirrhosis; no external validation	National Veterans Health Administration	LR with cross-sectional inputs (cross-sectional LR); LR with longitudinal inputs (longitudinal LR); DL model with longitudinal inputs	Baseline and longitudinal predictors	DL models outperformed conventional LR models: [mean (SD) AUC: 0.806 (0.025); mean (SD) Brier score: 0.117 (0.007)]
Audureau et al[28]	836 patients with compensated biopsy-proven hepatitis C virus-cirrhosis	French ANRS CO12 CirVir Cohort	Three prognostic models for HCC occurrence: (1) Fine-Gray regression as a benchmark; (2) DT; and (3) RF	Parameters before and after SVR	Externally validated C-indexes before/after SVR were 0.64/0.64 (Fine-Gray), 0.60/62 (DT), and 0.71/0.70 (RF)
Singal et al[29]	442 patients with Child-Pugh A or B cirrhosis	University of Michigan cohort and HALT-C cohort (for independent validation)	RF	Clinical and biological data	Machine learning algorithm had significantly better diagnostic accuracy, a net reclassification improvement (P < 0.001), and an integrated discrimination improvement (P = 0.04)

AdaBoost: Adaptive boosting; AI: Artificial intelligence; AUC: Area under the curve; DL: Deep learning; DT: Decision tree; HALT-C: Hepatitis C Antiviral Long-Term Treatment Against Cirrhosis; HBV: Hepatitis B virus; HCV: Hepatitis C virus; LR: Logistic regression; MLP: Multilayer perceptron; RF: Random forest; SD: Standard deviation; SVR: Sustained virological response; XGBoost: Extreme gradient boosting.

DL outperformed previously reported risk models in HBV related cirrhosis[26] and conventional LR in large-scale datasets (AUC: 0.806) in HCV related cirrhosis[27]. External validations, such as the French CO12 cohort, confirmed the robustness (C-indexes: 0.60-0.71)[28]. AI consistently improved risk stratification, with significant net reclassification improvements (P < 0.001) in Child A/B cirrhosis[29]. These findings highlight AI's potential to enhance HCC surveillance in diverse etiologies and clinical settings.

Notably, Lee et al[30] validated these findings in a large sample of patients with cirrhosis (training: n = 425, 44.3%; validation: n = 485, 25.0%), with ML models reaching exceptional AUCs (0.930-0.946). Predictive models established by ML are more efficient than traditional statistical models, especially when comparing with DL models[19]. These results highlight the efficacy of ML, particularly DL, in stratifying HCC risk.

Despite these advantages including dynamic risk assessment and integration of complex variables[19,31,32], limitations persist. Key challenges include the lack of a standardized model design, limited generalizability across diverse populations, and insufficient clinical implementation framework[12,19]. Addressing these barriers is essential to translate the potential of AI into routine practice for patients with cirrhosis.

CONCLUSION

Cirrhosis is the primary risk factor for developing HCC. Therefore, ultrasound screening for HCC every 6 months is recommended for patients with cirrhosis, to diagnose HCC at an early stage and improve the patient’s prognosis. However, the risk level is not the same for all patients, and individualized screening recommendations are needed. ML seems to be a powerful tool for HCC risk stratification in patients with cirrhosis; however, several challenges need to be overcome to integrate this promising and powerful tool into the clinic.