Realizing personalized prediction for the reversal of chronic hepatitis B-associated liver fibrosis through machine learning

Abstract

Although long-term antiviral therapy can reverse liver fibrosis in a considerable proportion of patients with chronic hepatitis B, treatment response differs markedly among individuals, and reliable methods for predicting who will benefit remain lacking. Conventional clinical tools such as liver stiffness measurement and the aspartate aminotransferase-to-platelet ratio index were originally dewsigned for cross-sectional staging rather than longitudinal prediction, and their ability to capture the dynamic process of fibrosis reversal is limited. In this opinion review, we discuss how machine learning (ML) may help address this unmet need. We focus on three areas where ML has shown early but encouraging results: Quantitative analysis of images using ML, integration of pathological and clinical data through multimodal models, and dynamic prediction of treatment response based on serial measurements during antiviral therapy. In our view, the greatest value of ML lies not in replacing existing assessment tools but in complementing them, particularly through the early identification of patients who are unlikely to achieve fibrosis reversal with antiviral therapy alone. However, we also caution that most current models have been developed from small, single-center retrospective cohorts, and their clinical adoption will require prospective multi-center validation, improved model interpretability, and adherence to standardized reporting frameworks.

Key Words: Machine learning; Chronic hepatitis B; Liver fibrosis reversal; Personalized prediction; Clinical translation

Core Tip: Predicting liver fibrosis reversal in chronic hepatitis B remains a major clinical challenge due to the limitations of conventional assessment tools. Leveraging machine learning-driven histopathological image analysis and multimodal data fusion, enhances the accuracy and individualization of predicting treatment response during antiviral therapy. Integrating quantitative pathological features, clinical parameters, and proteomic biomarkers into unified machine learning frameworks optimizes early identification of non-responders, supports personalized therapeutic decision-making for chronic hepatitis B patients.

INTRODUCTION

Han et al[1] developed a deep learning-based multimodal model. By integrating histopathological images and clinical features, the model achieves personalized prediction of liver fibrosis reversal in patients with chronic hepatitis B (CHB). CHB is a highly prevalent chronic viral liver disease globally[2]. As a moderately endemic area for CHB, China ranks among the top countries globally in terms of the number of existing infected individuals[3]. Persistent chronic infection can trigger sustained liver inflammation[4,5], gradually initiating the process of liver fibrosis[6,7]. CHB-associated liver fibrosis, a critical precursor to liver cirrhosis and hepatocellular carcinoma (HCC), remains a clinical challenge due to significant heterogeneity in treatment response[8-10].

Liver fibrosis is not a unidirectional or irreversible pathological process, but rather a reversible intermediate stage in the progression of CHB to cirrhosis and HCC[11,12]. Numerous clinical studies have confirmed that early standardized intervention, accurate identification of the potential for fibrosis reversal, and dynamic monitoring of treatment response are crucial for halting disease progression, reducing the long-term risk of HCC, and improving patients’ long-term prognosis[13,14]. Although long-term antiviral therapy induces histological fibrosis reversal in approximately 51% of patients with CHB[11], residual risk remains a primary concern in clinical practice[15-17]. Some patients may still progress to cirrhosis or even develop HCC. This residual risk is closely related to persistent abnormalities in the liver microenvironment, latent low-level inflammatory damage, incomplete fibrous tissue remodeling, and an imbalance in the intrahepatic immune microenvironment[18-20]. Addressing this has become a core clinical challenge in the long-term management of CHB and the reduction of adverse long-term outcomes.

At present, non-invasive fibrosis assessment tools commonly used in clinical practice, such as liver stiffness measurement (LSM) and the aspartate aminotransferase-to-platelet ratio index (APRI), possess advantages such as repeatability and ease of implementation, enabling preliminary screening and broad stratification of liver fibrosis[21-24]. However, there are significant limitations that are difficult to avoid in clinical practice: These single-dimensional indicators struggle to accurately reflect the dynamic characteristics of fibrosis evolution, fail to capture subtle early changes in inflammatory fluctuations, collagen deposition, and degradation in the liver microenvironment, and are susceptible to interference from confounding factors such as acute liver inflammation and fatty liver[25-27]. They lack early predictive efficacy for fibrosis reversal and fall short of meeting the core needs of current personalized precision medicine. Liver biopsy, as the gold standard for liver fibrosis assessment, has inherent limitations, such as its invasive nature, sampling error, poor patient compliance, the inability to perform repeated dynamic monitoring, and the difficulty of reflecting the global state of liver fibrosis[28-30]. These limitations severely restrict its widespread use in routine clinical practice.

Recently, machine learning (ML), with its capacity to analyze high-dimensional data, has made breakthroughs in digital pathology[31,32], multimodal fusion[33], and dynamic prediction[34], offering novel solutions for the precise assessment of liver fibrosis reversal in CHB. In this opinion review, we take the study by Han et al[1] as a starting point and draw on related recent work to discuss where ML stands in the assessment of CHB-associated fibrosis reversal, what it can realistically offer at this stage, and what obstacles must be overcome before these methods can enter routine clinical practice.

ML-BASED LIVER FIBROSIS REVERSAL PREDICTION

Quantitative image analysis

Conventional pathological grading of liver fibrosis depends on visual assessment by trained observers, a process that is inevitably affected by inter-reader variability and the coarse resolution of semi-quantitative scoring systems[35,36]. ML-based image analysis, specifically convolutional neural networks, offers a different approach that allows the algorithm to learn microscopic tissue patterns directly from pixel-level data, independent of predefined scoring criteria[37,38].

The study by Han et al[1] illustrates this strategy well. Using hematoxylin and eosin (HE) and Masson’s trichrome-stained biopsy sections, their deep learning model was trained to detect pathological features relevant to fibrosis outcome, including hepatocyte degeneration, disorganized hepatic cords, and thick-bridging fibrous septa. In the validation sets, the model yielded area under the curve (AUC) values of 0.657 (HE-based) and 0.727 (Masson-based) for predicting fibrosis reversal, which were modest but consistently higher than the performance of manual scoring. What we find particularly noteworthy is the model’s feature visualization capability: It could highlight the tissue regions that contributed most to a prediction of non-reversal, such as areas of ductular reaction, inflammatory cell infiltration, and disrupted hepatic architecture. This kind of spatial mapping, in our view, is essential if pathologists and clinicians are to trust and meaningfully interact with artificial intelligence (AI)-generated predictions.

Similarly, Wang et al[39] developed a deep learning radiomics model based on shear wave elastography images, achieving AUCs of 0.97 and 0.98 for identifying cirrhosis (F4) and advanced fibrosis (≥ F3), respectively. While these values are substantially higher than those reported for conventional LSM, it should be noted that the clinical question addressed (cross-sectional staging) is different from, and arguably less challenging than, predicting future fibrosis reversal. A tool that accurately stages fibrosis at a single time point does not necessarily help clinicians decide whether a given patient’s fibrosis will regress with treatment.

Multimodal data fusion

In practice, the likelihood of fibrosis reversal depends not on pathology alone but on a constellation of virological, biochemical, and host factors, including hepatitis B virus DNA levels, hepatitis B e antigen (HBeAg) status, platelet count, and serum albumin[40]. Han et al[1] addressed this by combining histopathological features with clinical parameters in a unified ML framework. The resulting multimodal model achieved an AUC of 0.694 in an external validation set, with stronger performance observed in the advanced fibrosis subgroup (Ishak score 3-4; AUC = 0.779) and the HBeAg-positive subgroup (AUC = 0.755). These subgroup findings are clinically relevant because they suggest that ML-based stratification may be most useful precisely in the patient populations where clinical uncertainty is greatest.

Liu et al[41] took a complementary approach, combining acoustic radiation force impulse imaging (ARFI) and transient elastography (TE) with the APRI to construct a linear combination algorithm for non-invasive fibrosis assessment. Their model achieved AUCs of 0.92 and 0.98 for diagnosing significant fibrosis (≥ F2) and cirrhosis (F4), respectively, with 10-fold cross-validation accuracies of 83.86% and 91.88%. These results exceeded the performance of any single modality used alone. However, the study cohort was small and most participants had a normal body mass index; patients with atypical body habitus were largely absent, leaving open the question of whether the complementary value of APRI relative to ARFI and TE holds across a broader population.

Dynamic prediction of treatment response

From a clinical standpoint, perhaps the most immediately valuable application of ML is in predicting treatment response over time. Under the current standard of care, clinicians typically wait 72 weeks or longer after starting antiviral therapy before reassessing fibrosis status on biopsy. During this interval, patients who will ultimately prove to be non-responders continue on a regimen that may need to be supplemented or modified. Identifying non-responders earlier would provide clinicians with a wider window to adjust therapeutic strategies, such as adding anti-fibrotic agents or intensifying surveillance for disease progression.

The multimodal model reported by Han et al[1] moves in this direction. Using baseline pathological images and clinical data, it was able to flag individuals at elevated risk of fibrosis progression before treatment was initiated, with a sensitivity of 0.647 for detecting non-reversal. While this sensitivity is far from perfect, it nonetheless represents a quantitative starting point that did not previously exist for pre-treatment risk stratification.

Kong et al[42] took a dynamic, longitudinal approach. They prospectively followed a cohort of CHB patients treated with entecavir (ETV) and used ML to model the trajectory of LSM values during the first six months of therapy. A rapid early decline in LSM proved to be a strong predictor of histological fibrosis reversal at 18 months. Their prediction model, which incorporated both baseline indicators and early on-treatment changes and was evaluated against the Ishak scoring system[43] and PIR classification (predominantly progressive, indeterminate, and predominantly regressive)[25,44,45], reached an AUC of 0.81. This approach is appealing because it mirrors actual clinical workflow. Since physicians routinely obtain serial measurements during treatment, an ML model capable of extracting predictive information from these temporal trends would fit naturally into existing practice. It should be noted that this model was only validated at the 18-month point. Its applicability in other antiviral therapies still requires further testing.

ML-guided drug selection and treatment optimization

Beyond predicting fibrosis outcomes, ML-based analyses have begun to inform the choice of antiviral agent itself. Hur et al[46] developed the prediction of liver cancer using AI-driven model for network-antiviral selection for hepatitis B model, which stratified CHB patients into a tenofovir disoproxil fumarate (TDF)-superior group and a TDF-non-superior group. In the TDF-superior group, TDF was associated with a significantly lower HCC risk compared with ETV (hazard ratios 0.60-0.73; P < 0.05), whereas no significant difference was observed in the TDF-non-superior group. This type of ML-guided drug selection, if validated in independent cohorts, could provide direct evidence for tailoring antiviral therapy to individual patient profiles rather than relying on a uniform treatment algorithm.

Dong et al[47] addressed a somewhat different clinical scenario: Inactive hepatitis B surface antigen (HBsAg) carriers receiving 48 weeks of pegylated interferon therapy. Using least absolute shrinkage and selection operator regression and the Boruta algorithm for feature selection, they identified three predictors of HBsAg seroclearance (baseline HBsAg level, HBsAg decline > 1 log IU/mL at week 12, and the ratio of alanine aminotransferase to HBsAg at week 12). A random forest model built on these variables achieved AUCs of 0.829 in the training set and 0.838 in external validation. The authors further translated the model into an online calculator and a simplified scoring system, making it accessible for use at the point of care. We consider this effort to build user-friendly clinical tools from ML outputs to be an important and often neglected step in the translational pipeline.

Proteomic biomarker discovery

At the biomarker level, Zhang et al[48] combined ML with serum proteomics (four-dimensional data-independent acquisition and parallel reaction monitoring mass spectrometry) to identify candidate protein markers for fibrosis reversal. Two non-invasive panels were developed: A 7-protein panel for short-term assessment and a 3-protein panel optimized for patients on long-term antiviral therapy. Among the identified markers, complement factor H related 4 (CFHR4) emerged as a particularly noteworthy candidate. Both panels outperformed conventional non-invasive tools (LSM and APRI) in distinguishing patients who achieved fibrosis reversal from those who did not, and the 3-protein long-term panel showed stable performance across different treatment durations. These findings suggest that integrating proteomic data into ML frameworks could add a biological dimension that imaging and routine blood tests cannot provide.

Cross-disease applicability

ML technology has also demonstrated its potential for application in other chronic liver diseases[49-52]. In the assessment of fibrosis related to metabolic-associated fatty liver disease, the AI digital pathology model developed by Abdurrachim et al[53] can assist pathologists in scoring fibrosis, improving diagnostic consistency. In the field of autoimmune liver diseases, AI models analyze the immune cell infiltration patterns in liver tissues and the characteristics of bile duct damage to distinguish disease subtypes and predict prognosis[54]. These cross-disease application experiences provide a reference for the optimization of CHB-associated liver fibrosis models and suggest that a multi-etiology compatible liver fibrosis assessment platform can be constructed in the future.

We should note, however, that the diversity of methodologies across these studies (including differences in sample size, patient selection criteria, reference standards, and validation strategies) makes direct comparison difficult. This heterogeneity underscores the need for standardized benchmarking if the field is to move beyond proof-of-concept toward clinically deployable tools. Supplementary Table 1 provides a comparison of the ML models discussed in this review.

CHALLENGES AND FUTURE PERSPECTIVES

Figure 1 illustrates the process of predicting the reversal of CHB-associated liver fibrosis using ML. Although ML demonstrates immense potential, its clinical translation faces significant hurdles.

Figure 1 The process of predicting the reversal of chronic hepatitis B-associated liver fibrosis using machine learning. CHB: Chronic hepatitis B.

Data standardization and external generalizability. The most pressing concern, in our assessment, is that nearly all existing models have been built on single-center or geographically restricted datasets. The external validation cohort used by Han et al[1], for example, comprised only 74 patients. At this scale, confidence in the model's stability across different clinical settings is inevitably limited. The problem is compounded by technical variability: Differences in tissue staining protocols, slide scanning hardware, and image preprocessing pipelines across institutions can introduce systematic biases that degrade model performance when applied outside the training environment. Before any of these models can be recommended for clinical adoption, they will need to be tested prospectively in larger, multi-ethnic, and multi-center cohorts. We believe that federated learning, which allows models to be trained collaboratively across institutions without sharing raw patient data, offers a practical path toward expanding sample diversity while respecting data privacy constraints. At the same time, the field would benefit from broader adoption of standardized reporting guidelines, such as the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis-AI[55] and the consolidated standards of reporting trials-AI[56], to ensure that model development, validation procedures, and performance metrics are described in a transparent and reproducible manner.

Model interpretability and clinical trust. Even when an ML model performs well statistically, clinicians are understandably reluctant to act on predictions they cannot understand. Feature visualization techniques, such as those employed by Han et al[1] to highlight pathological regions associated with non-reversal, represent a step in the right direction, but they address only part of the problem. Many of the texture-level and associative features extracted by deep learning algorithms lack a clear biological correlate, and their pathophysiological meaning remains obscure. Zhang et al[48], for instance, identified CFHR4 as a key proteomic marker but acknowledged that its regulatory role in the complement-coagulation pathway during long-term antiviral treatment is not yet well understood; elucidation may require single-cell sequencing and spatial transcriptomic studies. In our view, efforts to bridge this interpretability gap should proceed along two parallel tracks. On the computational side, post-hoc explanation tools such as SHapley Additive exPlanations and local interpretable model-agnostic explanations are increasingly being applied to link model outputs to candidate biological mechanisms[57-60]. On the clinical side, structured collaboration between data scientists and hepatologists, which should ideally commence during study design rather than following model development, is essential to ensure that the learned features are clinically meaningful and not merely statistical artifacts.

Implementation costs and accessibility. A practical obstacle that receives less attention in the literature but matters greatly in real-world settings is cost. High-resolution digital pathology scanners and the computational infrastructure needed to run deep learning models require substantial capital investment, placing these technologies out of reach for many primary care facilities and resource-limited regions[61]. If ML-based fibrosis assessment is to benefit more than a small number of tertiary referral centers, the development of lightweight, computationally efficient algorithms that can run on portable ultrasound devices or standard clinical workstations should be a priority. Regulatory and governance issues also warrant attention: Clear frameworks for data sharing, patient consent, and algorithmic accountability are prerequisites for responsible clinical deployment but remain underdeveloped in most jurisdictions.

Toward dynamic, multidimensional prediction. It is important to build a dynamic prediction model, which integrates the longitudinal data collected during the treatment process. The proteomic panels identified by Zhang et al[48] (a 7-protein panel for early assessment and a 3-protein panel for long-term monitoring) illustrate how serial biomarker measurements can be combined with ML algorithms to provide evolving risk estimates rather than a single baseline prediction. Pairing such molecular data with serial imaging follow-ups could yield genuinely real-time decision support tools. In addition, factors such as medication adherence, lifestyle, psychological well-being, and socioeconomic context all influence treatment outcomes in chronic liver disease, yet they are rarely incorporated into current models. Building a broader, biopsychosocial predictive framework, which captures both the biological trajectory of fibrosis and the human context in which treatment occurs, may ultimately be necessary to achieve truly personalized CHB management.

CONCLUSION

ML technology is advancing the prediction of liver fibrosis reversal in CHB. By integrating digital pathology, clinical indicators, and multi-omics data, ML not only achieves precise stratification of treatment responses but also provides interpretable and actionable quantitative evidence for clinical decision-making. However, the path of technology transformation still requires collaborative efforts from the academic and clinical communities. Through standardized data construction, in-depth mechanism research, and technology democratization, this innovative tool can truly benefit CHB patients.