AI in Hepatology: Revolutionizing the Diagnosis and Management of Liver Disease

1. Introduction

As artificial intelligence (AI) continues to evolve, its integration into various medical subspecialties has grown significantly, offering innovative approaches to diagnosis, treatment, and disease management. Hepatology is no exception. With the rising global burden of liver conditions such as metabolic-associated steatotic liver disease (MASLD), the need for efficient and accurate tools for early detection and management has never been more critical. AI technologies have emerged as powerful tools, leveraging vast datasets to improve clinical decision-making and patient outcomes. Machine learning (ML), a subfield of AI, leverages diverse algorithms to enhance predictive capabilities and decision-making. ML techniques share a common goal: systematically organizing inputs into features, labels, clusters, or neural networks to improve prediction accuracy and refine outcomes across successive iterations. ML models span a spectrum of human involvement, from supervised learning, where labeled data guides training, to unsupervised learning, which identifies patterns in unlabeled data. Reinforcement learning, another subset, involves algorithms learning optimal actions through trial and error. Deep learning (DL), a specialized branch of ML, mimics human brain processing by employing multi-layered neural networks with minimal human supervision but requires significant computational power [1]. These approaches collectively drive advancements in AI applications, transforming complex data into actionable insights in medicine and beyond.

A review of the current literature highlights AI’s applications in diagnosing and managing a range of hepatic conditions, including autoimmune hepatitis (AIH), primary biliary cholangitis (PBC), primary sclerosing cholangitis (PSC), MASLD, and hepatitis B (HBV). Beyond diagnosis, AI shows promise in predicting disease progression, such as advanced fibrosis and cirrhosis, and plays a growing role in hepatocellular carcinoma (HCC) diagnosis and management. By offering insights into how AI is transforming hepatology, this review aims to provide a comprehensive overview of its applications, limitations, and future potential in advancing liver health.

2. Applications of AI in the Diagnosis and Management of Cirrhosis and Its Complications:

AI has demonstrated significant potential in improving the accuracy of advanced fibrosis and cirrhosis detection through advanced imaging techniques, ML algorithms, and biomarker analysis. Furthermore, AI is being applied to predict complications such as hepatic encephalopathy, variceal bleeding, and ascites, enabling more personalized and timely interventions. This section explores the current applications of AI in diagnosing cirrhosis, managing its complications, and its potential to revolutionize care in this critical area of hepatology.

2.1. Detecting Advanced Fibrosis and Cirrhosis

Chang et al. evaluated the accuracy of various ML models, including logistic regression, random forest, and artificial neural networks (ANN), for predicting histological stages of fibrosis in 1,370 patients with biopsy-proven MASLD [2]. These models, which incorporated seventeen demographic and clinical features, demonstrated superior predictive accuracy compared to traditional non-invasive fibrosis assessments such as Fibroscan, FIB-4, NFS, and FAST. Fan et al. investigated ML algorithms that included liver stiffness measurements (LSM) alongside aMAP scores (a composite index incorporating age, gender, albumin, bilirubin, and platelet count) for diagnosing fibrosis and cirrhosis in 946 biopsy-confirmed MASLD patients from China and the United States. The study reported diagnostic accuracy (96.8% for advanced fibrosis and 91.2% for cirrhosis), with high specificity (92.6% and 98.3%) and sensitivity (82.4% and 88.9%) [3]. In another large, multinational study, Anushiravani et al. utilized the random forest algorithm to develop the FIB-6, a scoring system which combines readily available laboratory parameters to reliably rule out cirrhosis in patients with hepatitis C (HCV). The FIB-6 index was subsequently validated in a large study including 2,472 patients from Egypt, Iran, KSA, Greece, Turkey, and Oman who had biopsy-proven MASLD and was again noted to be superior to other non-invasive indices in ruling out advanced fibrosis and cirrhosis with a sensitivity of 97% and negative predictive value of 94.7% [4]. Similarly, Wang et al. constructed an ML model using the Light Gradient Boosting Machine (Light GBM) algorithm to predict cirrhosis in patients with chronic HBV, leveraging platelet count and bile acid levels. This model demonstrated superior predictive accuracy compared to conventional non-invasive measures, underscoring the significance of bile acids as biomarkers of liver dysfunction [5]. In addition to the study by Wang et al., several other studies have explored the use of the LightGBM algorithm to identify cirrhosis. For instance, Orhanbulucu et al. employed ensemble learning techniques, including LightGBM, to predict chronic liver disease (CLD) [6]. In this study, LightGBM achieved the highest accuracy (98.8%) among various boosting algorithms, including Random Forest and AdaBoost. Kuzhippallil et al., used LightGBM to predict CLD alongside other algorithms like XGBoost and stacking. After feature selection and outlier elimination, LightGBM demonstrated competitive accuracy, reinforcing its utility in disease prediction models [7]. LightGBM’s success in these studies highlights its strengths: memory efficiency, fast processing, and suitability for large-scale data analysis, making it an optimal choice for identifying cirrhosis in patients with CLD. In another study, Pu et al. explored a comprehensive ML algorithm incorporating 55 clinical and laboratory features to predict fibrosis in 1,023 patients with chronic HBV that had biopsy-proven fibrosis. This study’s extensive use of diverse data points offers insight into the benefits of large-scale feature integration, which may enhance the model’s clinical applicability in Hepatology [8]. Chen et al. used EmpowerStats to develop an ML model that incorporated general clinical data, complete blood counts (CBC), 24-hour urinary copper levels, and serum ceruloplasmin, to accurately predict the presence of cirrhosis in a cohort of 346 patients with Wilson disease [9].

Various ML models that incorporate imaging have also been developed to predict cirrhosis. Mazumder et al. used a cohort of 1,590 computed tomography (CT) scans to train an automated liver segmentation model using 3D-U-Net and Google’s DeepLabv3+. Their approach also integrated routine lab values and demographic data, achieving an area under the curve (AUC) of 0.85 in a validation cohort of 352 patients, including 72 liver transplant recipients [10]. Similarly, Nowak et al. utilized deep transfer learning with a ResNet50 convolutional neural network (CNN) pre-trained on the ImageNet archive to accurately detect cirrhosis in a sample of 713 patients, outperforming a board-certified radiologist [11]. Luetkens et al. utilized DL to distinguish alcohol- from non-alcohol- related cirrhosis using magnetic resonance imaging (MRI). In a cohort of 465 patients—221 with alcohol-related cirrhosis and 244 with non-alcohol-related cirrhosis—T2-weighted images were analyzed using ResNet50 and DenseNet121 CNN models. ResNet50 achieved the highest performance, with an AUC of 0.82 and an accuracy of 0.75, demonstrating the potential of DL models to differentiate alcohol-related cirrhosis from other causes [12]. Wei et al. used Pyradiomics software on MRI images from 280 patients with CLD to extract radiomics features, combining MRI sequences to predict fibrosis staging and inflammation grading [13]. Destrempes et al. developed a model using ultrasound and shear wave elastography data from 82 patients with CLD, achieving more accurate identification of steatosis, inflammation, and fibrosis than shear wave elastography alone [14]. Similarly, Meng et al. used ultrasound and shear wave elastography data from 618 patients with diabetes and MASLD, identifying 25 fibrosis-related radiomics features and highlighting five as most predictive using a support vector machine model [15].

Other ML models incorporate a broad range of clinical, imaging, and laboratory data to predict cirrhosis. For example, Ahn et al. developed a DL-based model using electrocardiogram (ECG) signals to identify patients with cirrhosis and generated an AI-Cirrhosis-ECG (ACE) score which correlated with disease severity. The model was validated and tested in a large sample of patients with cirrhosis and had a sensitivity of 84.9% and specificity of 83.2%, highlighting the potential of AI-enabled ECG analysis to assist in cirrhosis prediction [16]. Xiao et al. used DL to analyze images from slit lamps and retinal fundal examinations in 1252 patients with hepatobiliary disorders including cirrhosis and created models using these images that were highly effective in detecting cirrhosis with an AUC of 0.90 for slit lamp and 0.83 for retinal fundal imaging [17]. In another interesting study, Dou et al. developed a support vector machine algorithm using urine fluorescence spectroscopy to detect cirrhosis, showing a sensitivity of 73.85% and specificity of 89.34% [18]. These studies collectively illustrate the effectiveness of ML in combining a broad array of clinical data points with advanced algorithmic frameworks to enhance diagnostic accuracy for cirrhosis and advanced fibrosis.

2.2. Clinically Significant Portal Hypertension

Early detection of clinically significant portal hypertension (CSPH) is essential since early interventions can prevent decompensation. Hepatic venous pressure gradient (HVPG) measurement is considered the gold standard for the diagnosis of portal hypertension. However, this is an invasive procedure that is not readily available outside of tertiary care centers. Identifying non-invasive tests and models that can accurately detect CSPH is therefore paramount. Bosch et al. used trichrome-stained liver biopsy samples obtained from patients with cirrhosis due to MASLD to develop a ML model that accurately identified patients with CSPH with AUCs of 0.85 and 0.76 in the training and test phases [19]. In another study, Yu et al. used an auto-machine-learning CT radiomics HVPG quantitative model (aHVPG) that was effective in assessing portal hypertension severity and outperformed several other non-invasive measures, including elastography [20]. Noureddin et al. obtained liver biopsies from patients with cirrhosis due to MASLD and used ML to create a score based on septa, nodules, and fibrosis (SNOF) which predicted CSPH with high accuracy [21], while Reinis et al. used ML models based on readily available lab parameters (bilirubin, platelet count, and international normalized ratio) to predict CSPH in patients with compensated advanced CLD with an AUC of 0.775 [22].

2.3. Esophageal Varices

ML models have proven valuable in managing esophageal varices. Liu et al. demonstrated the effectiveness of extreme gradient boosting (XGBoost) in predicting variceal rebleeding, while Agarwal et al. used ML to create a model for predicting bleeding risk in compensated advanced CLD, offering high accuracy and a non-invasive alternative to endoscopic evaluations [23,24]. Bayani et al. applied ensemble learning techniques that combined clinical data and lab parameters, outperforming traditional methods like the platelet-to-spleen ratio with an AUC of 0.87 [25]. By integrating elastography and lab data, this approach provides a viable, non-invasive method for identifying patients at risk of variceal bleeding, potentially reducing reliance on invasive diagnostics.

Recent advances in ML for variceal management also include DL-based imaging techniques. Gao et al. developed an imaging-based ML model using endoscopic images and structured data, outperforming traditional clinical risk scores in predicting variceal bleeding in patients with cirrhosis [26]. Similarly, Jin et al. introduced an AI-driven non-invasive method for measuring esophageal varix diameter via endoscopy, enabling precise assessment of variceal size and bleeding risk without direct visualization [27]. Yan et al. developed a novel radiomics model for diagnosing high-risk esophageal varices in patients with cirrhosis using portal venous-phase enhanced CT images. The model demonstrated high accuracy, outperforming Baveno VI criteria [28].

CNNs have also been used to predict variceal bleeding risk. Hou et al. created an ANN model to predict the 1-year risk of esophagogastric variceal bleeding in patients with cirrhosis using data from 12 independent risk factors. This model demonstrated superior predictive accuracy compared to traditional indices [29]. In another large study, Chen et al. evaluated ENDOANGEL, a deep convolutional neural network, for diagnosing gastroesophageal varices (EVs and GVs) and predicting rupture risk. Trained on 14,718 endoscopic images, ENDOANGEL outperformed endoscopists in detecting EVs and GVs, identifying red color signs, and recognizing risk factors for bleeding with high accuracy [30].

These studies collectively demonstrate the potential of AI to enhance the accuracy and efficiency of variceal management by leveraging non-invasive methods that minimize patient burden, optimize clinical workflows, and refine risk-based surveillance strategies.

2.4. Ascites:

ML models can help guide management of ascites. For example, Wurstle et al. developed a model incorporating 34 readily available clinical data points to identify patients with infected ascites. This model, demonstrated high negative predictive values for infected ascites across various pre-test probabilities, potentially providing a reliable non-invasive alternative to paracentesis [31]. Another study by Hatami et al., utilized ML models like k-nearest neighbors (KNN), support vector machines, and neural networks to predict ascites grades based on routine clinical data. The KNN model demonstrated the highest accuracy at 94%, underscoring ML’s effectiveness in non-invasively predicting ascites severity [32]. These ML-based approaches can reduce the need for invasive procedures and thus minimize unnecessary risks and costs, and allow more personalized care for cirrhotic patients with ascites.

2.5. Hepatic Encephalopathy

ML models can also predict outcomes in patients with cirrhosis and hepatic encephalopathy (HE). In a retrospective study utilizing the MIMIC-IV database, Yang et al. developed and validated ML models to predict early mortality in HE patients. This study found that the artificial neural network NNET model provided the highest predictive accuracy, achieving an AUC of 0.837. The NNET model also demonstrated superior calibration and discrimination compared to traditional clinical scores such as the MELD and MELD-Na [33]. These findings suggest that NNET can identify patients with HE who are at high risk for mortality and who may benefit from more aggressive treatment and possibly early referral for a liver transplant; however, external prospective validation is needed to confirm the model’s utility across diverse patient populations. In another study involving patients with cirrhosis complicated by HE, Zhang et al. employed weighted ML algorithms to accurately predict mortality risk. The weighted random forest model achieved high sensitivity and specificity, with an AUC of 0.82, and effectively distinguished between high-risk and uncomplicated HE cases further underscoring the potential of ML to support clinical decision-making in complex liver disease cases [34].

3. Applications of AI in the Diagnosis and Management of HCC

HCC represents a significant global health challenge due to its high mortality rate and complex management requirements. HCC is the most common form of primary liver cancer and the third leading cause of cancer-related death worldwide [35]. The high mortality rate associated with HCC is largely due to late-stage diagnosis and the limited effectiveness of conventional treatments, with 5-year survival of only 15% for unresectable disease [36]. Traditional predictive and risk stratification models for HCC, such as PAGE-B, CAMD, REAL-B, HCC-RESCUE, AASL-HCC, and modified REACH-B scores, rely on demographic data, laboratory findings, and elastography to generate risk profiles ranging from low to high using multivariate regression modeling [37,38,39]. Current guidelines for HCC are informed by its natural history, lesion detection sensitivity, and cost-effectiveness of screening intervals. However, advancements in big data processing and the refinement of modern algorithms offer exciting possibilities for improving risk stratification, diagnostic accuracy, and tailored therapeutic approaches. In the following sections we will outline efforts to implement ML methods to identify risks of developing HCC among patients with CLD; diagnose HCC with imaging, pathology and genomic markers; and predict treatment response.

3.1. HCC Risk Stratification

ML models have shown effectiveness in predicting the risk of HCC in patients with cirrhosis. Singal et al. developed both a random forest ML model and a regression model using a cohort of 442 patients with cirrhosis who were enrolled in an HCC surveillance program and found that the ML model outperformed the regression model in identifying patients at high risk for developing HCC [40]. Similarly, Kim et al. used gradient machine boosting algorithms in a cohort of 6051 Korean patients with HBV to create an ML model that was compared to other predictive models in Korean and Caucasian validation cohorts. The ML model was found to be superior in predicting HCC risk in both validation cohorts outperforming various scoring indices including PAGE-B, modified PAGE-B, REACH-B and CU-HCC [41]. In a large retrospective cohort of 4242 patients with HBV and HCV, Sato et al. evaluated multiple ML methods for the prediction of HCC. The authors found that gradient boosting produced the highest accuracy (87.34%; AUC = 0.940) followed by random forest (86.08%; AUC = 0.923) for the prediction of HCC compared to logistic regression (79.74%; AUC = 0.866). Traditional decision tree-based models such as random forest outperformed neural networks in this study [42]. In a similar study from Hong Kong, Wong et al. compared multiple ML models for predicting HCC risk among 124,006 patients with HBV and HCV. Ridge regression (AUROC) 0.844, sensitivity = 0.90, specificity = 0.90 with dual cutoffs) and random forest (AUROC = 0.837, sensitivity = 0.90, specificity = 0.910 with dual cut-offs) produced the highest model accuracy in the validation cohort [43]. More recently, Lin et al. used a sample of 5155 Korean patients with multiple etiologies of CLD to develop an ML- based model named SMART-HCC incorporating elastography measurements and validated this model in Korean and Caucasian cohorts. SMART-HCC outperformed multiple other traditional predictive models in both cohorts. The authors also successfully used this model to stratify patients into low, intermediate, and high-risk groups with significantly higher incidence of HCC in the high-risk group (p<0.001) [44]. Sarkar el. compared multiple ML models to predict HCC risk in a cohort of 2247 patients with MASLD and demonstrated that tree algorithms such as gradient boost, decision tree, and random forest produced the highest accuracy in predicting HCC compared to a probabilistic neural network [45]. Audureau et al. used several ML models to determine HCC risk in a cohort of 836 patients with chronic HCV based on SVR status. The random survival forest model performed best, suggesting that ML can enhance HCC risk assessment and support more cost-effective, personalized surveillance programs [46]. Phan et al. evaluated various DL models in a large cohort of patients with viral hepatitis and found that the CNN model outperformed other models in predicting HCC among patients with chronic hepatitis with an accuracy of 0.980 (AUC = 0.886) [47]. Nam et al. also developed a CNN model that outperformed six older generation models in predicting HCC risk in a sample of 424 patients with cirrhosis due to HBV [48].

3.2. HCC Diagnosis

CNNs are highly effective for analyzing radiology and pathology images, as they can create feature layers and pool layers to emphasize predictive characteristics. Several studies have investigated CNN applications for diagnosing HCC.

Chaiteeraakij et al. conducted a retrospective analysis of ultrasound data from 5,444 patients using a CNN, achieving an HCC detection rate of 82.3% (95% CI: 77.1–87.5) [49]. Mokrane et al. analyzed data from 178 patients with cirrhosis across 27 institutions from 2010 to 2015, using 1,160 imaging features extracted from multiphase CT to classify indeterminate liver nodules. They identified a radiological signature reflecting changes between arterial and venous phases, yielding an AUC of 0.66 to 0.70 (sensitivity = 0.70, specificity = 0.54) for HCC prediction [50]. In another study, Wang et al. developed two CNN models using CT data from 7,512 patients, achieving an AUROC of 0.883 (95% CI: 0.855–0.911) on the validation set [51]. Zhang et al. applied a three-dimensional CNN to MRI data from 267 patients to predict microvascular invasion, with a combined model yielding an AUC of 0.81, sensitivity of 69%, and specificity of 79% in the training set [52]. CNNs have also been applied in limited studies to assess tumor progression on CT following locoregional therapy [53]. Kim et al. used CNN on contrast-enhanced MRI from 549 patients, achieving 87% sensitivity and 93% specificity for HCC detection [54]. Additionally, a meta-analysis of 44 studies by Salehi et al. found that AI algorithms across various imaging modalities had a pooled sensitivity of 85% (95% CI: 78–89) and specificity of 84% (95% CI: 72–91) [55]. Chen et al. used CNNs to analyze pathology slides from 377 patients in the Genomics Data Commons Database, training a neural network (Inception V3) to classify whole slide images (WSI) cropped to smaller tiles at 20x magnification. The model performed exceptionally well in distinguishing tumor from normal tissue (AUC = 0.961) and showed near-perfect agreement with an experienced pathologist. Additionally, the model achieved good accuracy in differentiating tumor grades [56]. In another study, Wang et al. trained a neural network using WSI from 304 tumors, also developing the model to perform single-cell segmentation based on 246 image features, which characterized individual nuclei and their relation to infiltrating lymphocytes. This enabled the identification of three distinct histological subtypes and, by integrating survival data, the prediction of prognosis for each subtype. This study emphasizes the potential of DL not only in diagnosing HCC but also in distinguishing tumor subtypes based on genomic and tumor microenvironment characteristics, supporting more personalized treatment approaches [57]. Neural networks have also been successfully applied to differentiate HCC from other liver tumors as demonstrated by Jang et al. who evaluated three neural networks on 20x tile images, achieving an AUC of 0.995 in distinguishing HCC from cholangiocarcinoma and metastatic colon cancer [58].

3.3. Predicting HCC Treatment Response

Numerous ML models have been created to forecast patient survival, tumor recurrence, microvascular invasion, and treatment response in HCC. A notable approach is radiomics, which analyzes regions of interest from imaging data retrospectively collected from HCC patient cohorts [59]. Many of these models have undergone validation and demonstrated superior performance compared to traditional prognostic tools [60]. For instance, Sheen et al. developed a radiomics normogram using CT-derived Rad score, T stage, alpha-fetoprotein (AFP) level and bi-lobar distribution that accurately identified tumors refractory to trans-arterial chemoembolization (TACE) with AUC ranging from 0.91 to 0.95 [61]. Ji et al. developed validated predictive models that incorporated ML to analyze contrast-enhanced CT images, along with variables such as the ALBI grade, AFP level, cirrhosis, and tumor margin characteristics. Their ML models demonstrated effective prediction of HCC recurrence, achieving concordance index (C-index) values ranging from 0.633 to 0.699 across various datasets and outperforming traditional models and staging systems [62]. Additionally, Qiao et al. designed a prognostic ANN model to predict early HCC outcomes after partial hepatectomy in a cohort of 543 patients. This model demonstrated superior predictive accuracy, achieving an AUC of 0.855 in the training cohort and was better than established staging systems such as TNM, BCLC, and IHPBA [63].

DL models can identify conventional and novel histopathological patterns associated with HCC recurrence after surgical resection or transplantation. By integrating these patterns, DL models generate reliable predictions that improve prognostic accuracy and aid clinical decision-making [64,65,66,67]. For example, Saillard et al. utilized two DL algorithms based on digitalized WSI to predict patient survival post-surgery. The first model, SCHMOWDER, incorporated an attention mechanism focused on tumoral areas annotated by a pathologist, while the second model, CHOWDER, operated independently of human expertise. In a discovery set of 194 patients, SCHMOWDER achieved a C-index of 0.78, and CHOWDER achieved a C-index of 0.75, both outperforming traditional survival prediction methods. Validation in an independent dataset from The Cancer Genome Atlas (TCGA) confirmed the superior discriminatory power of these models [68]. Chaudhary et al. investigated DL based multi-omics approaches to predict survival outcomes in a cohort of 360 patients with HCC. The model incorporated data encompassing RNA sequencing, microRNA sequencing, and methylation profiles sourced from TCGA and demonstrated effective prognostic capabilities, achieving a C-index of 0.68 and identifying two distinct patient subgroups with significant survival disparities (p = 7.13e-6). Validation across five external datasets confirmed the model’s strong predictive performance, with C-indices ranging from 0.67 to 0.82 [69].

AI also has the potential to tailor treatment plans based on individual patient characteristics [70]. ML can predict waitlist dropout rates among liver transplant candidates with HCC [71], tumor recurrence in post-liver transplant patients [72], as well as graft survival [73]. Such predictive capabilities can allow for the appropriate prioritization of patients with HCC on the transplant waitlist. Furthermore, AI algorithms can analyze genetic and molecular data to identify potential therapeutic targets and predict responses to specific treatments. For example, AI-driven analyses of genomic data can identify mutations associated with HCC, thereby guiding targeted therapies [74]. Large Language Models can also process extensive literature, electronic health records, and clinical guidelines, to provide evidence-based treatment recommendations tailored to individual patient profiles. In a recent study, Ge et al. utilized the Versa generative pretrained transformer (GPT-4) application programming interface to extract data from unstructured text in imaging reports related to HCC. This study compared the performance of Versa GPT-4 versus manual chart reviews for extracting eight specific data elements, including maximum LI-RADS score, size of HCC lesions, and the presence of macrovascular invasion or metastases from 1,101 imaging reports of 753 patients with HCC. The results demonstrated that Versa GPT-4 achieved an overall accuracy of 0.934, with specific accuracies ranging from 0.886 to 0.989, depending on the element being assessed [75]. While the model performed well in classification tasks, it encountered challenges with more complex data extraction.

4. Applications of AI in Autoimmune Liver Disease

AI has also been applied in the diagnosis and management of various liver disorders including AIH, PBC and PSC.

Ercan et al. created a DL tool AI(H) using the Aiforia platform that was able to analyze liver biopsies and accurately identify features of AIH [76]. Guadalupi et al. analyzed salivary samples from patients with AIH, PBC, and healthy controls using a gel-based proteomic approach and machine learning to distinguish AIH from PBC. Mass spectroscopy identified protein variations, with random forest and multidimensional scaling pinpointing a distinct set of proteins. These proteins were mostly linked to immune function and liver fibrosis. This AI-assisted approach offers a promising, non-invasive method to understand the pathophysiology of AIH and PBC [77]. Singh et al. extracted MRI data from 169 patients with large-duct PSC and trained a decision tree model to identify patients that were at high risk for decompensation within one year of the initial MRI [78]. Additionally, two studies by Cristoferi et al. and Vuppalanchi et al., used MRCP+ software to produce three-dimensional models of the biliary tree to calculate biliary tree metrics and predict disease progression of PSC [79,80]. Other studies have employed ML techniques to identify biomarkers associated with more severe disease and progression to cirrhosis amongst patients with PSC. Snir et al. used ML to analyze blood samples from 45 patients with PSC and 30 controls, identifying 2,870 proteins linked to PSC severity and progression to cirrhosis. An elastic net regression model highlighted 16 proteins strongly correlated with the enhanced liver fibrosis (ELF) score, which mapped to 13 enriched cell pathways, including those involving CCL24, a protein associated with PSC pathogenesis and cirrhosis risk [81]. Mousa et al. profiled plasma bile acids in 508 PSC patients and 302 controls using liquid chromatography-tandem mass spectrometry and used gradient boosting to create a multivariate model that identified high taurine-conjugated bile acids as predictive of hepatic decompensation. The resulting PSC-BAP score, derived from six bile acid variables, shows promise for predicting decompensation risk and assessing anti-cholestatic drug efficacy. These findings highlight potential biomarkers and therapeutic targets for PSC [82].

5. Challenges and Future Directions:

Despite notable advancements, the integration of AI into the routine clinical management of chronic liver disease presents substantial challenges, necessitating strategic planning and careful consideration. Several critical issues must be addressed to enable the effective application of AI in hepatology, including the following:

5.1. Data Limitations and Overfitting Risks

One of the primary challenges in developing AI models for hepatology is the risk of overfitting [83]. Overfitting occurs when a model learns specific patterns from the training data without capturing underlying, generalizable features, leading to strong performance during model development but poor external validity in varied clinical contexts. Deep learning models trained on small datasets are particularly susceptible to capturing “noise” or spurious correlations rather than meaningful clinical signals, thereby reducing their predictive reliability [84]. Effective mitigation strategies include employing cross-validation, conducting rigorous external validation on independent datasets, and utilizing interpretable models when data size is limited [85].

5.2. Biases and Generalizability Issues

AI models in hepatology often encounter biases resulting from imbalanced or non-representative training data [86]. For instance, many studies rely on datasets derived from specific geographic regions, ethnic groups, or healthcare environments, which may not reflect the global patient population affected by liver diseases. This issue is particularly evident in models developed using predominantly Western or Asian cohorts, which may struggle to generalize to patient populations with distinct disease etiologies, such as MASLD or ALD in Western countries versus viral hepatitis in South Asia [87]. Addressing these challenges requires incorporating diverse patient demographics and leveraging bias-aware machine-learning techniques to detect and correct skewed data distributions [88].

5.3. Challenges in Model Interpretability and Adoption

A significant obstacle to the clinical adoption of AI is the “black box” nature of complex models, particularly deep learning algorithms [89]. These models often lack transparency, making it difficult for clinicians to trust their diagnostic or prognostic outputs. Explainable AI (XAI) methodologies, such as Shapley Additive Explanations (SHAP) [90] and Local Interpretable Model-Agnostic Explanations (LIME⁾ [91]^, are increasingly being employed to demystify the decision-making processes of AI models. Moving forward, developing models that balance high accuracy with clear, interpretable insights aligned with clinical reasoning frameworks will be essential for facilitating clinician acceptance and integration into practice.

5.4. Regulatory and Ethical Considerations

The evolving regulatory landscape of healthcare poses challenges for the approval and deployment of AI-based tools in hepatology. While regulatory agencies such as the FDA and EMA have issued guidelines for AI-driven medical devices, the lack of standardized protocols for evaluating model safety, efficacy, and equity remains a significant hurdle. Ethical concerns, including data privacy, patient consent, and equitable access, must also be addressed. Transparent and comprehensive policy frameworks, coupled with standardized evaluation metrics and robust validation processes, are necessary to ensure AI applications meet clinical and ethical standards before widespread adoption.

5.5. Data Integration and Interoperability Issues

Effective AI integration in hepatology requires the analysis of diverse data types, including imaging, genomics, clinical notes, and laboratory results [92]. However, inconsistencies in electronic health record (EHR) systems across institutions exacerbate this challenge, hindering the creation of comprehensive datasets necessary for robust AI model training [93]. Advances in multimodal learning frameworks, which enable the simultaneous analysis of multiple data sources, hold promise for addressing these challenges but require further refinement to improve data integration and overall model performance.

5.6. Need for Prospective, Multi-Center Validation

Most AI models in hepatology are developed and validated on retrospective datasets, which limits their applicability to real-world clinical settings. Prospective, multi-center studies are crucial to evaluate model performance across diverse patient populations and healthcare environments, enhancing their generalizability and reliability [94].

5.7. Continuous Learning and Adaptive AI Systems

To maintain clinical relevance, AI models must adapt to new data and evolving clinical practices. Continuous learning frameworks allow models to update as new data becomes available, thereby mitigating the risk of performance degradation over time [95]. However, implementing such systems requires robust monitoring mechanisms to detect shifts in data distribution—referred to as model drift—and ensure sustained accuracy and reliability in predictions.

5.8. Economic and Resource Considerations

The development, implementation, and maintenance of AI systems in healthcare entails substantial costs. Beyond initial investments in technology and infrastructure, ongoing expenses related to data storage, model updates, clinician training, and system integration must be considered. Conducting cost-effectiveness analyses will be vital to determine whether AI tools deliver sufficient clinical benefits to justify their financial impact, particularly in resource-limited settings [96].

5.9. Patient and Public Engagement

Engaging patients and the public in the development and deployment of AI tools can enhance transparency and trust in these technologies. Education initiatives that inform patients about AI’s capabilities and limitations, along with feedback mechanisms for data usage, can foster more ethical and patient-centered applications [97].

6. Conclusions

Addressing these challenges will require a multidisciplinary approach involving collaboration between clinicians, AI researchers, regulatory bodies, and patient advocacy groups. By adopting transparent, inclusive, and evidence-based strategies, the field of AI in hepatology can advance toward the development of effective, equitable, and trustworthy clinical applications.