Role of artificial intelligence in the management of chronic hepatitis B infection

Download video file ^{(24.7MB, mp4)}

INTRODUCTION

According to the World Health Organization (WHO), 296 million people were living with chronic hepatitis B (CHB) infection in 2019 and 1.5 million new infections each year. CHB resulted in an estimated 820,000 deaths, mostly from cirrhosis and HCC.¹ Currently, antiviral therapy can effectively suppress the virus; however, it still takes a long time to cure HBV. Most patients with CHB need monitoring of hepatitis activity and disease progression, maintaining good compliance with antiviral therapy, and receiving HCC surveillance.

Integrating artificial intelligence (AI) into the health care landscape has emerged as a promising tool for revolutionizing the precision management of CHB. The multifaceted AI extends from early detection through advanced diagnostic tools to personalized treatment plans and continuous patient monitoring. This transformative potential not only enhances clinical decision-making but also addresses the evolving needs of patients, paving the way for a more effective and patient-centric approach to managing CHB.

AI prediction models often outperform traditional models like Cox regression in accuracy by handling complex, nonlinear relationships, processing high-dimensional and unstructured data, and automatically detecting feature interactions. They also adapt well to new data input. However, AI models do not uniformly surpass all other models; their effectiveness relies on the data quality and quantity and the specific task they are applied to.

Up to now, several studies using AI have been published to facilitate the management of CHB in terms of diagnosis, disease stratification, and outcome prediction for HBsAg clearance or the development of HCC (Table 1).

TABLE 1.

The selected publications using AI in the management of chronic hepatitis B

Tasks	Data, model, and finding	Reference
Diagnosis of chronic hepatitis B	Data: Use demographic and clinical records from the UCI Machine Learning Repository Model: Machine learning Finding: 92% accuracy to predict the diagnosis of chronic hepatitis B	2
Prediction of HBsAg levels	Data: 90 HBV inactive carriers Model: Deep neural network model Finding: A mean absolute percentage error of 15% compared with the multiple regression model (58%)	3
Prediction of liver fibrosis	Data: 398 patients with 1990 images Model: Deep learning radiomics of elastography Finding: AUC of DLRE were 0.97 for F4, 0.98 for ≥F3, and 0.85 for ≥F2 compared with 2D-SWE and biomarkers	4
Prediction of hepatitis B relapse	Data: 43 patients who were HBeAg-negative from the ABX-203 vaccine trial, validated in another 49 patients Model: Machine learning Finding: The combination of IL-2, monokine induced by interferon γ/CCL9, RANTES/CCL5, stem cell factor, and TRAIL was reliable in predicting virological relapse (AUC: 0.89; 95% CI: 0.5–1.0)	5
Prediction of HBsAg seroclearance	Data: 2235 patients with CHB Model: Machine learning model, XGBoost, random forest, decision tree, and logistic regression Finding: The XGBoost model discriminates the HBsAg seroclearance best with an AUC of 0.891.	6
Antiviral therapy selection	Data: 13,970 patients with CHB on TDF or ETV Model: Machine learning model (PLAN-S model) Finding: In the TDF-superior group, TDF was associated with a significantly lower risk of HCC than ETV. In the TDF-nonsuperior group, there was no significant difference between TDF and ETV.	7
Prediction of HBV-related HCC	Data: 6051 patients with CHB received ETV or tenofovir Model: Gradient-boosting machine algorithm Finding: C-index 0.79, significantly better than other models, validated by Korean and Caucasian cohorts	8
Prediction of HBV-related HCC	Data: 960 patients with CHB who received ETV or tenofovir Model: A soft voting ensemble model by combining the random forest, XGBoost, and logistic regression models Finding: An AUC of 0.872, better than other models	9

Abbreviations: 2D-SWE, two-dimensional shear wave elastography; CHB, chronic hepatitis B; DLRE, deep learning radiomics of elastography; ETV, entecavir; RANTES, regulated on activation, normal T cell expressed and secreted; TDF, tenofovir disoproxil fumarate; XGBoost, extreme gradient boosting.

Using AI to help the diagnosis of hepatitis B and disease stratification

Obaido et al used demographic and clinical records of patients obtained from the UC Irvine Machine Learning Repository and applied the Shapley Additive exPlanations interpretable machine learning approach to predict the diagnosis of hepatitis B. The AdaBoost model achieved an accuracy score of 92%, among other models,² and they identified bilirubin level as the most significant feature contributing to a higher mortality rate, followed by ascites.²

The HBsAg levels correlated with disease outcomes, such as HBsAg seroclearance or the development of HCC. Deep learning algorithms have been shown to predict the HBsAg levels in inactive carriers. Kamimura et al conducted a retrospective study of 72 inactive patients with CHB whose HBsAg levels were evaluated over 10 years.³ They used SONY Neural Network Console’s algorithms to build their deep neural network models. Another 18 patients were used for the validation. Multiple regression analysis revealed a mean absolute percentage error of 58%, and deep learning showed a mean absolute percentage error of 15%; thus, deep learning is an accurate predictive discriminant tool.

AI for the prediction of liver fibrosis

In a prospective multicenter study, Wang et al developed a deep learning radiomics of elastography (DLRE) model, which is a radiomic strategy for quantitative analysis of the heterogeneity in two-dimensional shear wave elastography images. They included 398 patients with 1990 images and found the AUC of DLRE was 0.97–0.98 for cirrhosis (F4) and advanced fibrosis (≥ F3) prediction, which was significantly better than liver stiffness measurement or serum biomarkers in patients with CHB.⁴ Notably, the diagnostic accuracy of DLRE improves as more images are acquired from each individual for training, and the DLRE performance did not change significantly regardless of which hospital supplied the training images.⁴

AI for the prediction of hepatitis relapse and HBsAg seroclearance

Hepatitis relapse after discontinuation of antiviral therapy is common, and the risk predictors for early virological relapse after stopping nucleos(t)ide analogs have been explored.¹⁰ In a post hoc analysis of a prospective, multicenter therapeutic vaccination trial (ABX-203, NCT02249988), HBsAg, hepatitis B core-related antigen, and 47 soluble immune markers were repeatedly determined before nucleos(t)ide analogs was stopped in 43 patients who were HBeAg-negative. A supervised machine learning approach was used to identify predictive markers for early virological relapse and was validated in another cohort of 49 patients treated with entecavir. The authors found that the best predictor was IL-2, IL-17, and regulated on activation, normal T cell expressed and secreted/CCL5 with a maximum AUC of 0.65. The combination of IL-2, monokine induced by interferon γ/chemokine (C-C motif) ligand 9 (CCL9), regulated on activation, normal T cell expressed and secreted/CCL5, stem cell factor, and TRAIL was reliable in predicting VR (0.89) and showed viable results in the validation cohort (0.63). Machine learning can help find predictive soluble immune marker patterns that precisely identify patients particularly suitable for nucleos(t)ide analogs cessation.⁵

Another study used machine learning algorithms to predict HBsAg seroclearance. Tian et al obtained the laboratory and demographic information for 2235 patients with CHB from the South China Hepatitis Monitoring and Administration cohort. Among them, HBsAg seroclearance occurred in 106 patients. A total of 30 variables, including radiological indicators, were included. Four prediction models, extreme gradient boosting (XGBoost), random forest, decision tree, and logistic regression, were developed. The AUCs reflecting the total discriminative abilities of the XGBoost, random forest, decision tree, and logistic regression were 0.891, 0.829, 0.619, and 0.680, respectively. They found that machine learning algorithms, especially XGBoost, perform appropriately in predicting HBsAg seroclearance.⁶ The level of HBsAg was the most crucial predictor of HBsAg seroclearance followed by age and DNA.

Personalized antiviral drug selection for CHB

There have been some debates regarding the secondary prevention for HCC development between the usage of tenofovir disoproxil fumarate (TDF) and entecavir (ETV). A multinational study included 13,970 patients with CHB to establish the PLAN-S model, derived from the random survival forest algorithm, to predict HCC risks.⁷ Patients were categorized into the TDF-superior and TDF-nonsuperior groups. In the TDF-superior group, TDF was associated with a significantly lower risk of HCC than ETV (HR = 0.60–0.73, all p < 0.05). In the TDF-nonsuperior group, however, there was no significant difference between TDF and ETV. Therefore, TDF and ETV treatment may be recommended for the TDF-superior and TDF-nonsuperior groups.

Using AI for the prediction of HBV-related HCC

HCC is the most dreadful outcome after HBV infection, and many risk predictors have been created for clinical usage. Kim et al conducted a retrospective study to develop and validate an AI-assisted prediction model of HCC risk in patients with CHB. They included 6051 patients with CHB who received ETV or tenofovir therapy from 4 hospitals in Korea and built a gradient-boosting machine algorithm.⁸ A model using 10 parameters at baseline was derived and showed good predictive performance (c-index 0.79), which was significantly better discrimination than previous models (PAGE-B, modified PAGE-B, REACH-B, and CU-HCC) in both the Korean (n = 5817, c-index 0.79 vs. 0.64–0.74; all p < 0.001) and Caucasian validation cohorts (n = 1640, c-index 0.81 vs. 0.57–0.79; all p < 0.05 except modified PAGE-B, p = 0.42). Patients in the minimal-risk group had less than 0.5% risk of HCC during 8 years of follow-up.⁸

Lee et al recently included a training cohort of 960 patients with CHB on EV or tenofovir therapy from South Korea and developed a novel machine learning model (a soft voting ensemble model) for predicting the 5-year HCC risk. Three popular machine learning algorithms, namely random forest, XGBoost, and logistic regression, were applied to predict HCC using standard clinical and laboratory information. The discriminatory performance of the novel machine learning model remained excellent in the large external validation cohort of 1937 Korean patients. The novel model achieved an AUC of 0.872, which was again significantly superior to that of CAMD (0.788), REAL-B (0.801), HCC-RESCUE (0.798), modified PAGE-B score (0.775), and PAGE-B scores (0.765).⁹

Limitations

Despite the substantial potential of AI applications, challenges exist in AI research, such as data preparation, collection, labeling, biases, and privacy. Furthermore, the selection, evaluation, validation of algorithms, and the real-world implementation of AI models, especially in resource-limited regions where HBV infection is endemic, face considerable hurdles.¹¹ The black box issue concerns the lack of transparency and interpretability in AI models, while overfitting is a common issue where a model fails to generalize well to new data due to capturing noise or irrelevant patterns in the training data. Both are important considerations in the development and deployment of AI models.

Perspectives

AI holds great promise in addressing unmet needs by analyzing vast datasets that include patient demographics, genetic information, treatment history, and laboratory results. By integrating these data, AI has the great potential to predict individual responses to specific anti-HBV treatments, allowing for more personalized and precision therapies, thereby enhancing efficacy and minimizing adverse effects. AI can also play a crucial role in early detection and risk stratification by continuously analyzing patient data, including liver function tests, viral load, and imaging results. AI has the potential to enhance patient engagement and adherence to antiviral therapy. AI-driven applications, including mobile health platforms, can provide personalized reminders, educational resources, and support mechanisms to encourage patients to adhere to management plans. This holistic approach may contribute to better treatment outcomes and improved overall quality of life for individuals living with CHB. In summary, the unmet needs in managing CHB provide opportunities for AI to make meaningful contributions. By leveraging advanced data analytics and machine learning, AI has the potential to revolutionize CHB management by facilitating personalized treatment strategies, early risk identification, and patient engagement.