Artificial intelligence in autoimmune liver diseases: Diagnosis, risk stratification, and personalized care

Abstract

Autoimmune hepatitis (AIH), primary biliary cholangitis (PBC), and primary sclerosing cholangitis (PSC) require clinicians to combine heterogeneous clinical, biochemical, histological, and imaging data. This minireview examined disease-specific artificial intelligence studies that focused on validation, comparators, and clinical use. Multicenter models distinguished drug-induced liver injury from AIH and AIH from PBC, although performance falls in external cohorts and treatment response prediction in AIH was largely absent. PBC models identified phenotypes and predicted the ursodeoxycholic acid response; however, none has shown benefit beyond the GLOBE and UK-PBC pathways. In PSC, primary sclerosing cholangitis risk estimate tool and quantitative magnetic resonance cholangiopancreatography had the strongest comparative evidence; computed tomography, proteomics, and cholangiocarcinoma-risk models were still retrospective. Most studies used small samples, single-center data, imbalanced classes, incomplete calibration, and few prospective validations. Current tools should remain adjunct tools. Before using a model, clinicians should check its intended use, comparators, external validation, calibration, decision curve benefit, and subgroup performance. Multinational prospective validation, transparent reporting, workflow testing, and post-deployment monitoring are required.

Key Words: Artificial intelligence; Machine learning; Deep learning; Autoimmune hepatitis; Primary biliary cholangitis; Primary sclerosing cholangitis; Autoimmune liver disease; Precision medicine; Digital pathology; Large language models

Core Tip: Artificial intelligence can be used to quantify images and combine multimodal data in autoimmune liver disease; however, most models remain investigational. The best-supported evidence currently comes from externally validated differential diagnosis models for autoimmune hepatitis, established machine learning prognostic tools for primary sclerosing cholangitis, and treatment response models for primary biliary cholangitis. However, these tools require local validation and monitoring during clinical testing. Clinicians should seek calibration, comparison with current care, decision curve benefit, and subgroup performance; however, the area under the curve alone is insufficient. No current model supports autonomous diagnosis or changes in therapy or surveillance without guideline-based clinical judgment.

INTRODUCTION

Autoimmune hepatitis (AIH), primary biliary cholangitis (PBC), and primary sclerosing cholangitis (PSC) are immune-mediated hepatobiliary diseases[1]. Although uncommon, they account for a growing proportion of chronic liver diseases, cirrhosis, and liver transplantation in several populations[1]. Although their clinical features differ, hepatologists face the same underlying difficulty: Diagnosis often depends on several partly overlapping findings, biological heterogeneity is high, and the disease course and treatment response vary between patients who initially look similar[2]. AIH is diagnosed using composite scores that combine autoantibodies, immunoglobulin G, exclusion of viral hepatitis, and liver histology because no single test is decisive[3,4]. PBC risk stratification uses biochemical responses to ursodeoxycholic acid and continuous prognostic scores such as GLOBE and UK-PBC[5,6]. No medical treatment has been proven to alter the natural history of PSC; therefore, clinicians rely on imperfect prognostic scores and mainly qualitative imaging[7,8].

Current scores reduce complex information to few variables and can be calibrated in cohorts that do not resemble the patient being assessed. They have also compressed whole-slide histology, cross-sectional and magnetic resonance cholangiographic imaging, serological patterns, and genomic or other omics data to a few thresholds. Artificial intelligence (AI), particularly machine learning (ML) and deep learning (DL), can directly learn relationships from such inputs rather than requiring every predictor and interaction to be specified in advance. Image-based studies in other specialties achieved specialist-level performance under defined test conditions[9-12]. Hepatology studies now apply methods related to steatotic liver disease, viral hepatitis, fibrosis, and hepatocellular and biliary malignancies[13-16].

Autoimmune liver disease exposes a clear limitation of this approach. Its diagnostic and prognostic questions suit multimodal pattern recognition, but the diseases are too uncommon to generate large and varied datasets on which reliable models usually depend[17]. Overlapping syndromes further blur the boundaries. AIH may coexist with features of PBC or PSC, leaving clinicians to decide based on incomplete evidence in patients who do not fit a single category[2,18]. Expertise is concentrated in referral centers, and many patients are first assessed by clinicians who encounter these diseases infrequently. AI can extend the specialist interpretation of histology, autoantibody patterns, and cholangiographic images to other centers; however, this benefit remains to be tested[17].

This minireview explains the main methods and appraises disease-specific evidence in AIH, PBC, and PSC (Figure 1; representative studies are summarized in Table 1). The clinical questions include computational histopathology, noninvasive fibrosis and outcome prediction, treatment-response modeling, omics-based biomarker discovery, AI-assisted imaging, and large language models (LLMs). We also examined automated serological interpretation, digital pathology, and prospects of combining these inputs into a multimodal care model. Digital pathology and radiology, structured electronic health records, increased computing capacity, and modern neural networks have made large-scale multimodal analyses technically possible[11,14]. The difficult question is whether small fragmented datasets can support models that remain accurate across centers and improve decisions in practice[13,17].

Figure 1 Artificial intelligence applications across the three autoimmune liver diseases. Representative applications are shown for autoimmune hepatitis (computational histopathology, differential diagnosis, non-invasive fibrosis staging, and biomarker- and language-model-based tools), primary biliary cholangitis (risk stratification and sub-phenotyping, treatment-response prediction, genetic-risk modeling, and prognostic and dynamic monitoring), and primary sclerosing cholangitis (quantitative and artificial intelligence-enhanced imaging, radiomic and topological prognostication, cholangiocarcinoma-risk modeling, and proteomic biomarkers). Cross-cutting applications (automated serological interpretation, digital pathology, large language models, and multi-omics integration) apply across all three diseases. AIH: Autoimmune hepatitis; PBC: Primary biliary cholangitis; PSC: Primary sclerosing cholangitis; DILI: Drug-induced liver injury; MRCP: Magnetic resonance cholangiopancreatography; UDCA: Ursodeoxycholic acid.

Table 1 Representative artificial intelligence studies across autoimmune liver diseases.

Ref.	Clinical task	Model and input	Cohort and validation	Performance/conventional comparator	Clinical interpretation
Eaton et al[7]	PSC decompensation	PREsTo; gradient boosting, routine clinical data	509 derivations; 278 external validations	C-statistic 0.90 vs MELD 0.72, Mayo 0.85, ALP < 1.5 × ULN 0.65	Externally validated against clinical scores; no implementation trial
Zhang et al[19]	AIH significant fibrosis	Explainable XGBoost; clinical/Laboratory variables	261; multicenter retrospective validation	AUROC 0.791 vs APRI 0.557 and FIB-4 0.625	Moderate discrimination; prospective calibration and biopsy-impact testing needed
Gerussi et al[20]	PBC phenotypes	Unsupervised/supervised machine learning	11819 development; 1069 validation	Four reproducible phenotypes; not an AUC-comparator study	Phenotypes reproduced in a large validation cohort; treatment benefit remains untested
Gerussi et al[27]	PBC genetic risk	Logic learning machine; genetic variants	1345 training; 834 validation	Validation AUROC 0.73; sensitivity 0.55, specificity 0.74	Transparent rules; screening utility unproven
Hu et al[31]	PSC cholangiocarcinoma risk	Cox regression, random survival forest, and gradient boosting	1459 single center; bile-acid subset 300	Best C-indices across assessment times 0.68-0.71; combined subset 0.67; better than common PSC scores	Modest discrimination; no prospective surveillance benefit
Wang et al[33]	DILI vs AIH	BJ-AID gradient boosting; five laboratory variables	2554 across 10 hospitals; external and prospective cohorts	AUROC 0.94 development, 0.91 external, 0.93 prospective; atypical 0.85	Adjunct only; geographic concentration and no management-impact trial
Gerussi et al[37]	AIH vs PBC histology	ALNE deep learning; whole-slide histology	354 patients; 92 external validation	External AUROC 0.81; general-pathologist κ 0.09	Externally tested adjunct; binary task and small external set
Allaume et al[38]	AIH histology	Weakly supervised deep learning; whole slides	402 training; 185 external	AUROC 0.92 internal, 0.74 external; HBV 0.42, DILI 0.76	Marked external degradation; transportability remains limited
Ercan et al[39]	AIH histological features	AI(H); manually annotated biopsy features	123 biopsies; internal assessment	Accuracy 72.4%-99.4% across features	Annotation-intensive feasibility study, not externally validated diagnosis
Wei et al[42]	AIH advanced fibrosis	Random forest; routine variables	233 discovery; 33 validation	AUROC 0.951 training, 0.863 test, 0.843 validation	Small validation set and probable optimism
Kimura et al[54]	PBC UDCA response	XGBoost; baseline/early clinical variables	245 from 18 hospitals; 51 tests	AUROC 0.811 development, 0.856 test; Paris II outcome	Small test set; model-guided treatment not evaluated
Yang et al[55]	PBC UDCA response	Metabolomic machine learning	347 non-cirrhotic patients	AUROC 0.91 response, 0.87 deep response; better than clinical-only	Selected population and platform dependence; no treatment trial
Fu et al[61]	PBC survival	Random survival forest	90; single center, internal validation	C-index 0.7805; 1-/3-/5-year AUROC 0.9595/0.8898/0.9088	High estimates in a small cohort require external replication
Vuppalanchi et al[73]	PSC outcomes	M + BA quantitative MRCP metric	102 derivations; 50 validations	AUROC 0.86; hazard ratio 5.8; better than Mayo score	Quantitative biomarker; limited validation size
Ragab et al[66]	PSC diagnosis	DeePSC; multi-view convolutional neural network on MRCP	342 PSC + 264 controls; small internal tests; 37 external	Accuracy 80.5% (3 T), 82.6% (1.5 T), 92.4% external; radiologist differences not significant	Case-control task, scanner effects, and small test cohorts
Singh et al[74]	PSC decompensation	3-dimensional dense convolutional network; computed tomography	277; 128 events; five-fold cross-validation	AUROC 0.89 ± 0.04; no external comparator cohort	Internal-only estimate; center-specific overfit possible
Snir et al[77]	PSC proteomics	Elastic net; 2870 serum proteins	45 PSC + 30 controls; discovery study	Top markers AUROC > 0.90 for disease and > 0.80 for ELF-defined disease severity	High-dimensional small-sample discovery; requires independent replication
Colapietro et al[92]	AIH counseling	GPT-4; 12 questions from four patients	11 expert raters; no patient outcomes	Median accuracy 5/6, completeness 2/3, safety 3/3; poor/slight agreement	Interface study; version dependence and subjective evaluation

AIH: Autoimmune hepatitis; PBC: Primary biliary cholangitis; PSC: Primary sclerosing cholangitis; DILI: Drug-induced liver injury; UDCA: Ursodeoxycholic acid; PREsTo: Primary sclerosing cholangitis risk estimate tool; XGBoost: EXtreme Gradient Boosting; BJ-AID: Beijing-AID model for discriminating drug-induced liver injury from autoimmune hepatitis; ALNE: Autoimmune liver neural estimator; M + BA: Composite risk score combining quantitative magnetic resonance cholangiopancreatography metrics with total bilirubin and aspartate aminotransferase; MRCP: Magnetic resonance cholangiopancreatography; DeePSC: Deep learning model for primary sclerosing cholangitis; GPT-4: Generative pre-trained transformer 4; ALP: Alkaline phosphatase; ULN: Upper limit of normal; MELD: Model for end-stage liver disease; APRI: Aspartate aminotransferase-to-platelet ratio index; AUROC: Area under the receiver operating characteristic curve; FIB-4: Fibrosis-4 index; HBV: Hepatitis B virus; ELF: Enhanced liver fibrosis; AUC: Area under the curve.

METHODOLOGY

This narrative minireview searched PubMed, Scopus, and Web of Science from inception to June 2026, using combinations of “autoimmune hepatitis”, “primary biliary cholangitis”, “primary biliary cirrhosis”, “primary sclerosing cholangitis”, “autoimmune liver disease”, “artificial intelligence”, “machine learning”, “deep learning”, “neural network”, “radiomics”, “large language model”, “natural language processing”, “digital pathology” and “prognostic model”. We screened reference lists and relevant societal guidelines for additional sources, and prioritized primary studies, validation studies, systematic reviews, and authoritative guidelines. For each main clinical model, we recorded the task, model, inputs, sample size, validation design, principal performance metric, conventional comparator, and main limitations. Study selection reflected the clinical relevance and methodological strength. We did not undertake a formal systematic review protocol or quantitative synthesis because the evidence was heterogeneous and the literature was fluctuating.

A PRIMER ON AI FOR THE HEPATOLOGIST

AI refers to computational methods that perform tasks typically associated with human cognition[12]. In clinical hepatology, the most relevant subset is ML, in which algorithms learn statistical relationships from data rather than following only prespecified rules. Supervised learning uses labeled examples. A model might learn from biopsies assigned to a fibrosis stage or from patients classified by later development of cirrhosis and then predict the label for a new case. Common models include logistic regression, support vector machines, random forests, and gradient-boosted decision trees. eXtreme Gradient Boosting (XGBoost) is widely used for tabular clinical data[12,19]. Unsupervised learning instead searches unlabeled data for structure and can identify previously unrecognized patient groups or disease subphenotypes[20].

DL uses multilayer artificial neural networks[9]. Convolutional neural networks learn image features directly from pixels and have achieved specialist-level classification in defined medical imaging tasks[10,11]. Relevant inputs in autoimmune liver disease include whole-slide histology, magnetic resonance cholangiopancreatography (MRCP), and indirect immunofluorescence imaging for autoantibody testing. Radiomics follows a different route; it extracts quantitative measures of texture, shape, and intensity from standard scans and then supplies those measures to a machine-learning model[21].

Natural language processing handles free text. Current applications include LLMs and transformer-based neural networks trained on large text corpora that can answer questions, summarize documents, and draft clinical texts[22]. Fluency does not ensure accuracy, and LLMs may produce confident statements that are false. Retrieval-augmented generation attempts to limit this problem by requiring the model to draw from a selected knowledge base[23]. Multimodal foundation models extend the same idea across images, text, laboratory values, and omics and are often described as generalist medical AI[24].

Clinicians need more than a high area under the receiver operating characteristic curve (AUROC) to judge these models. Discrimination separates patients who will and will not experience an outcome, but the area under the precision-recall curve may be more informative when the outcomes are uncommon. A useful model must also be calibrated so that the predicted probabilities match the observed frequencies and show a net benefit in the decision curve analysis. The training and test data must be separated, and external validation must be distinguished from internal resampling[25,26]. The data type also matters: Tabular variables suit classical ML, whole-slide histology and radiological images suit convolutional neural networks or radiomics, free text suits natural language processing and LLMs, and high-dimensional omics require dimensionality reduction and methods that combine data sources. Transfer learning, data augmentation, and weak or self-supervised training can reduce the amount of labeled data required for rare diseases[17,24]. Explainability methods, including attention maps, rule extraction, and Shapley value-based feature measures, may help clinicians examine a model output, although they do not replace validation[27,28]. Predictive models return a classification or risk estimate, whereas generative models produce new text or images[22]. Both methods require representative data, independent testing, and clinical oversight[25].

WHY AUTOIMMUNE LIVER DISEASES ARE A NATURAL AND DIFFICULT TARGET FOR AI

Autoimmune liver disease is suitable for ML because its diagnosis rarely relies on a single discriminating test. AIH, for example, is diagnosed by a combination of autoantibodies, immunoglobulin G, histology, and the exclusion of other causes; distinguishing it from drug-induced liver injury (DILI) or cholestatic autoimmune disease remains difficult[3,29]. PBC and PSC also cover a broad range of severities and progression that categorical classifications describe poorly[2]. These are pattern recognition problems in which several data types must be considered together[17].

A single patient may contribute to longitudinal biochemistry, autoantibody imaging, digitized liver biopsies, cross-sectional and magnetic resonance cholangiographic studies, elastography, and, in research settings, genomic, transcriptomic, proteomic, metabolomic, or microbiome data[17,30]. Reducing these inputs to one fibrosis stage or binary treatment response category discards this information. ML can use combined data to identify sub-phenotypes and estimate individual rather than group-average risk[20].

Rarity was the main constraint. Even large referral centers encounter a limited number of patients with AIH, PBC, and especially PSC. Therefore, the available datasets are often small, imbalanced, and drawn from one center with few examples of cirrhosis, transplantation, or cholangiocarcinoma[2,31]. Such conditions favor overfitting; a model learns features peculiar to its training cohort and performs poorly elsewhere. The data were split across institutions and stored in different formats. Transfer learning, data augmentation, federated learning, and multinational data sharing may be helpful; however, none removes the need for independent testing[17]. The most useful models may identify latent subphenotypes and update risk over time, rather than simply automating existing categories[20,32]. This depends on cohorts that are sufficiently large and varied to represent the intended population[28]. Therefore, evidence maturity needs to be judged separately for each disease and clinical task.

AI IN AIH

Diagnosis and differential diagnosis

The strongest evidence for AIH differential diagnosis was obtained from two multicenter studies. The Beijing-AID (BJ-AID) gradient-boosting model used five routine laboratory variables in 2554 patients (1750 DILI and 804 AIH) recruited across 10 Chinese hospitals; the AUROC was 0.94 in development, 0.91 in external validation, and 0.93 in a prospective cohort, although discrimination fell to 0.85 in atypical cases[33]. The multicenter design and prospective tests support transportability within the study setting. Geographic concentration and the absence of evidence that model use changes treatment decisions limit its adoption. Pharmacovigilance analyses can identify drug-induced autoimmune-like hepatitis signals, but spontaneous-reporting data cannot establish causality or replace expert DILI assessment[34].

AIH scoring systems demonstrate why a model trained on patient data may be helpful. The original International AIH Group research score and the later simplified clinical score combine autoantibodies, immunoglobulin G, histology, and the exclusion of viral hepatitis, because no single feature is decisive[3,4]. Both are additive and each variable contributes a fixed weight. ML can model nonlinear interactions among serology, biochemistry, histology, omics, and clinical features. Any future model requires preserving the integrative, judgment-dependent diagnosis described in the current guidelines rather than replacing it with an unexamined probability[35,36].

Computational histopathology

The autoimmune liver neural estimator (ALNE) was trained on whole-slide images from 354 patients and tested in an external set of 92 cases to distinguish AIH from PBC. External AUROC was 0.81, whereas agreement among general pathologists was poor (κ = 0.09)[37]. These findings support ALNE as a reproducible adjunct, but the small external sample and binary AIH-PBC task do not cover overlapping syndromes or other common mimics. A newer weakly supervised model trained on 402 slides and externally tested on 185 slides achieved an internal AUROC of 0.92 but an external AUROC of 0.74; performance was particularly poor for hepatitis B (0.42) and only moderate for DILI (0.76)[38]. This decline demonstrates how internal estimates overestimate transportability.

The AI(H) system was used to evaluate 123 biopsies after manual annotation by experts. Reported accuracies were 99.4% for tissue detection, 88.0% for microanatomy, 83.9% for necroinflammation, 81.7% for bile-duct damage, 79.2% for portal inflammation, and 87.6% for fibrosis; immune-cell classification was lower at 72.4%[39]. These task-level results establish feasibility, but do not produce an independently validated diagnostic or management tool. A single-cohort annotation-intensive design may also overstate reproducibility. Methods developed for steatohepatitis digital pathology may be transferred to AIH, but AIH systems require validation across scanners, laboratories, and populations[40,41].

Non-invasive assessment of fibrosis and severity

Two non-invasive AIH fibrosis studies showed which current models are supported. In a multicenter retrospective cohort of 261 patients, explainable XGBoost predicted significant fibrosis with an AUROC of 0.791, outperforming aspartate aminotransferase-to-platelet ratio index (0.557) and fibrosis-4 index (0.625; both P < 0.001)[19]. A random forest study involving 233 discovery cases and 33 additional validation cases reported AUROC of 0.951 for training, 0.863 for testing, and 0.843 for validation[42]. The gap between the training estimate and the small validation result was compatible with overfitting. No studies have tested whether this model improves biopsy selection, monitoring, or immunosuppression decisions. Calibration, decision curve analysis, and wider external validation are required before clinical use[43].

Predicting treatment response and relapse

Compared with PBC, AIH has few published studies on predicting the immunosuppressive response or relapse after drug withdrawal. Relapse is common and difficult to predict; prolonged treatment induces toxicity, whereas unsuccessful withdrawal risks flares. Incomplete response and difficult-to-treat disease add a second management problem[2,18]. Methods previously used to predict the biochemical response in PBC or the fibrosis stage in AIH could be tested for sustained remission and relapse[19,42]. Before such predictions alter the duration or intensity of immunosuppression, they need independent, prospective validation and evidence that their use improves decisions.

Omics and biomarker discovery

AIH studies have also used ML to analyze molecular data. Clinical proteomics has identified pathways and protein signatures that distinguish AIH from other chronic liver diseases, including metabolic dysfunction-associated steatotic liver disease, which is a common diagnostic confounder[29]. Metabolomic studies have examined the biochemical patterns associated with AIH and disease activity[44]. Salivary proteomics have also been combined with ML to classify autoimmune liver disorders[45,46]. However, these studies remain at the discovery stage and remain unvalidated diagnostic tests. Combining molecular signals with clinical data and histology may be useful; however, the resulting multimodal models require the same independent evaluation as imaging or laboratory models.

AI IN PBC

Risk stratification and sub-phenotyping

PBC already has well-validated conventional prognostic tools; therefore, an AI model must demonstrate incremental value over GLOBE and UK-PBC rather than merely a high AUROC[5,6,47,48]. Gerussi et al[20] applied unsupervised and supervised learning to 11819 patients and tested the approach on 1069 additional patients and identified four clinically distinct phenotypes. The scale and validation support reproducible phenotyping, but clustering is exploratory; it does not prove that phenotype-guided treatment improves transplant-free survival or outperforms existing pathways at the point-of-care.

PBC treatment options have changed, making early response prediction more relevant but also more dependent on current jurisdiction-specific guidance. Obeticholic acid is no longer Food and Drug Administration-approved for PBC, and its European Union marketing authorization was revoked; newer second-line therapies and fibrates have different regulatory status and safety considerations[49,50]. Therefore, AI prediction should not select or trigger a therapy independently. A more defensible method would be to flag patients for earlier specialist reviews within a guideline-based pathway, with the treatment chosen after conventional biochemical assessment, contraindication review, and shared decision-making[47,51,52]. Risk models may also inform surveillance intensity; however, their effects on management must be tested rather than inferred from discrimination alone[53].

Predicting treatment response

Kimura et al[54] developed an XGBoost model in 245 patients from 18 hospitals and tested it on 51 additional patients; the AUROC for the Paris II biochemical response was 0.811 for development and 0.856 for testing. A metabolomic model of 347 non-cirrhotic patients reported an AUROC of 0.91 for biochemical response and 0.87 for deep response, outperforming clinical variables alone[55]. Both models may identify patients who are unlikely to respond, but neither study randomized patients to model-guided treatment nor showed better clinical outcomes. Small test sets, selected non-cirrhotic populations, and limited access to metabolomic platforms have reduced generalizability. Microbiome findings are still early and require standardized sampling and independent validation before they can be used to guide therapy[56-60].

Genetic, multi-omic, and prognostic modeling

Two PBC studies illustrated an explainable classification and survival modeling, but neither supported its clinical use. The logic learning machine was trained in 1345 individuals (444 PBC and 901 controls) and validated in 834 (255 PBC and 579 controls); the validation AUROC was 0.73, with a sensitivity of 0.55 and specificity of 0.74[27]. Its rules can be inspected, but its screening utility remains unproven. A random survival forest study included only 90 patients at one center; its final C-index was 0.7805 and reported 1-, 3-, and 5-year AUROCs of 0.9595, 0.8898, and 0.9088, respectively[61]. Without an external validation, these estimates are vulnerable to overfitting. Multi-omic studies and longitudinal elastography may support dynamic prediction in future studies; however, GLOBE and UK-PBC remain the clinical benchmarks[30,32,62-64].

AI IN PSC

PSC combines high clinical requirements with difficult data. It is rare and heterogeneous, and the events that drive outcomes (dominant strictures, recurrent cholangitis, decompensation, and cholangiocarcinoma) are uncommon and unpredictable[8,65]. Imaging is central to diagnosis and follow-up, but remains largely qualitative. Therefore, quantitative image analysis is a logical target for AI provided that the performance holds across scanners, centers, and clinically representative patients.

Quantitative and AI-enhanced imaging

Deep learning model for primary sclerosing cholangitis (DeePSC) used MRCP images from 342 patients with PSC and 264 controls, with internal test sets of 39 samples each from the 3-T and 1.5-T datasets and an external set of 37 cases. The accuracy was 80.5% for 3-T imaging, 82.6% for 1.5-T imaging, and 92.4% externally[66]. Performance exceeded four radiologists by 5.5 and 10.1 percentage points on the internal tests, but neither difference was statistically significant (P = 0.34 and P = 0.13). Quantitative MRCP measures are reproducible[67,68]; however, small test cohorts, scanner dependence, and case-control tasks indicate that DeePSC cannot replace expert MRCP interpretations. Pediatric quantitative magnetic resonance imaging also requires independent validation before it can inform care[69-71].

Quantitative imaging is important for PSC, because adverse events can be difficult to predict and may be detected late. Many patients have inflammatory bowel disease and face a lifelong risk of hepatobiliary and colorectal malignancies. Benign dominant strictures must be distinguished from malignant strictures; however, brush cytology and cross-sectional imaging have limited sensitivity for cholangiocarcinoma[8,72]. Automated analyses of the biliary tree, imaging, and laboratory data may improve risk estimates and patient selection for endoscopy or closer surveillance[31,73]. Whether such use leads to an earlier diagnosis or better outcomes has not yet been tested.

Radiomics, prognostic modeling, and prediction of complications

PSC prognostic studies differ greatly in terms of validation quality. Primary sclerosing cholangitis risk estimate tool (PREsTo) was derived in 509 patients and externally validated in 278; its C-statistic was 0.90 (95% confidence interval: 0.84-0.95), compared with 0.72 for model for end-stage liver disease, 0.85 for the Mayo risk score, and 0.65 for alkaline phosphatase below 1.5 times the upper limit of normal[7]. A composite quantitative MRCP metric was derived in 102 patients and validated in 50 patients, achieving an AUROC of 0.86 and a hazard ratio of 5.8, and outperforming the Mayo score[73]. A 3-dimensional dense convolutional network computed-tomography model used 277 patients with 128 decompensation events and reported AUROC 0.89 ± 0.04 using five-fold cross-validation only[74]. Therefore, the PREsTo had the strongest external comparative evidence. Quantitative MRCP requires implementation testing, whereas cross-sectional DL remains vulnerable to center-specific overfitting. Radiomics and topological models require the same comparison and external validation[21,75,76].

Biomarkers and cholangiocarcinoma risk

The largest reported PSC cholangiocarcinoma model used 1459 patients from one center. Its clinical/laboratory models achieved best cross-validated C-indices of 0.68-0.71 across assessment times; in a 300-patient subset with bile-acid data, the best combined model reached 0.67[31]. The model exceeded the common PSC scores; however, its modest discrimination, single-center design, and lack of prospective evidence did not support its use in escalating invasive testing. A serum-proteomic discovery study included 45 patients with PSC and 30 controls, tested 2870 proteins with elastic-net models, and reported an AUROC above 0.90 for disease presence and above 0.80 for classifying ELF-defined disease severity[77]. Small-cohort, high-dimensional feature searches require independent replicates. These tools remain hypothesis-generating and complement guideline-based imaging, endoscopy, cytology, and established biomarkers[72,78-84].

OVERLAP SYNDROMES, SPECIAL POPULATIONS AND UNMET NEEDS

Several of the most difficult clinical problems in autoimmune hepatology have received little attention in AI research. Overlapping syndromes, in which AIH features coexist with PBC or PSC, lack a firm diagnostic standard and are managed on a case-by-case basis. In principle, a histological model that distinguishes AIH from PBC could be extended to intermediate phenotypes; however, dedicated models are largely absent[2,37]. Acute severe or fulminant AIH also requires rapid decisions regarding intensified immunosuppression and transplantation; however, outcome prediction models for this setting are scarce[36].

Pediatric diseases may benefit from noninvasive assessment, and quantitative magnetic resonance studies have already provided a basis for further testing[69,70]. Other unanswered questions included recurrence after liver transplantation, immunosuppression to minimize toxicity, and pregnancy-related flares, all of which are managed with limited evidence and individual judgment[2,18]. Published models have concentrated on data-rich diagnoses and prognoses of established diseases. Progress in rarer, higher-risk settings depends on shared data and clinically meaningful outcome definitions.

CROSS-CUTTING APPLICATIONS ACROSS AUTOIMMUNE LIVER DISEASES

Several methods have been applied across all three diseases, particularly automated autoantibody imaging, digital pathology, language models, and the analysis of multiple omics layers.

Automated serological interpretation

Autoantibodies detected by indirect immunofluorescence in human epithelial (HEp-2) cells contribute to the diagnosis of AIH and PBC; however, reading the images is labor-intensive, subjective, and dependent on local expertise[85]. DL is suitable for this image-classification task. Computer-aided systems quantify HEp-2 fluorescence intensity, whereas newer models segment cells and classify antinuclear antibody patterns. Some have also reported reliability measures or analyzed whole specimens end-to-end[86-90]. Automated reporting could standardize recognition of antimitochondrial reactivity in PBC and disease-relevant antinuclear patterns, shorten turnaround time, and support laboratories with few trained immunofluorescence readers. Clinical validation must still show that this improves reporting accuracy and patient care.

Digital pathology as a shared substrate

Digitized liver histology allows for the computational pathology of autoimmune liver diseases. Most studies have focused on steatohepatitis, in which ML models quantify histological features and fibrosis more reproducibly than categorical assessments alone[40,41]. The same whole-slide methods can measure interface hepatitis in AIH and bile duct lesions in PBC or PSC; fibrosis remains a shared target for quantitative analysis[91]. Quantitative ultrasound and elastography provide noninvasive measures of fibrosis and inflammation that can be combined with ML[43]. Transfer from another liver disease has not yet been validated; however, each autoimmune application requires disease-specific training, external testing, and comparison with expert assessment.

LLMs for decision support, documentation, and patient communication

LLMs can summarize records, retrieve guidance, draft documentation, and support patient education; however, fluent answers are not necessarily reliable[22,23]. In the focused AIH evaluation, generative pre-trained transformer 4 answered 12 questions derived from four patients and was rated by 11 experts. The median accuracy was 5/6, completeness was 2/3, and comprehensiveness and safety were 3/3; however, the inter-rater agreement was slight or poor[92]. The small vignette set, subjective ratings, dependence on the model version, and absence of patient-outcome testing make this an interface study rather than a validation of a clinical tool. Near-term use should remain clinically supervised, source-grounded, privacy-aware, and limited to tasks where each output can be checked[93-97].

Multi-omics integration and biomarker discovery

ML can be used to combine genomic, transcriptomic, proteomic, metabolomic, and microbiome data with the clinical data. These models identified candidate immune biomarkers and possible disease mechanisms in PBC. Immune repertoire sequencing and spatial profiling have also made the hepatic immune environment available for computational studies and have linked specific cell populations to cholangiopathy and fibrosis progression[30,98,99]. Methods that use repeated biomarker measurements may be useful in treating diseases with risk changes over time[32]. However, these studies remain exploratory. They have not yet shown that combining molecular data with histology and imaging improves diagnosis, treatment selection, or prognosis.

TOWARD A PRECISION, MULTIMODAL FRAMEWORK FOR AUTOIMMUNE LIVER DISEASE

Figure 2 illustrates the intended clinical use of the multimodal model for prediction, personalization, and monitoring. At presentation, ML can combine clinical and biochemical variables, autoantibody patterns, quantitative histology, and measurements from cross-sectional or magnetic resonance cholangiographic imaging to estimate the diagnosis, disease sub-phenotype, and outcome risk[20,37,73].

Figure 2 A multimodal, predict-personalize-monitor framework for artificial intelligence in autoimmune liver disease. Multimodal patient data (clinical and biochemical variables, autoantibody patterns, digitized whole-slide histology, cross-sectional and magnetic resonance cholangiographic imaging, and multi-omics) are integrated by machine learning and deep learning methods, including gradient boosting, convolutional neural networks, radiomics, natural language processing and large language models, and emerging multimodal foundation models. The resulting clinical applications are grouped as prediction (diagnosis and differential diagnosis, disease sub-phenotype, fibrosis stage, and outcome and cancer risk), personalization (choice and intensity of therapy, immunosuppression duration in autoimmune hepatitis, second-line therapy timing in primary biliary cholangitis, and surveillance intensity in primary sclerosing cholangitis), and monitoring (dynamic risk updating, treatment-response tracking and deviation alerts). The feedback loop denotes continuous monitoring and model updating. AI: Artificial intelligence; CT: Computed tomography; MRCP: Magnetic resonance cholangiopancreatography; MRI: Magnetic resonance imaging.

Personalization and monitoring are proposed uses rather than established care. The predicted risk may inform treatment intensity, duration of immunosuppression in AIH, timing of second-line treatment in PBC, or surveillance in PSC[31,54,55]. Repeated elastography, quantitative MRCP, biochemistry, and molecular biomarkers can then update the risk and detect deviations from the expected course[32,64,68]. None of these components was ready for routine use. Foundation models can technically combine heterogeneous inputs; however, clinical adoption depends on representative data, external validation, interpretability, workflow testing, and governance[24].

PRACTICAL GUIDANCE FOR CLINICIANS

How to judge an AI model

Before acting on a model, clinicians should ask the following six questions: (1) What decision and patient population was the model designed for? (2) Was the effective sample, including outcome events, adequate? (3) Were the training and test data separated by patient, time, and site without leakage? (4) Was the performance reproduced in a representative external cohort? (5) Did the study compare the model with current care, and report calibration and decision curve benefits rather than discrimination alone? And (6) Were the subgroup performance, missing data handling, workflow burden, and monitoring plans reported? The lack of external validation or calibration should stop clinical deployment, even when the AUROC is high[25,26].

Which tools are closest to practice?

In AIH, BJ-AID, ALNE, and artificial intelligence on liver immunity (AIOLI) may serve as diagnostic adjuncts; however, each requires local and prospective validation before it influences immunosuppression[33,37,38]. In PBC, GLOBE and UK-PBC remain the clinical standards; ML response and phenotype models are currently in research settings[5,6,20,54,55]. PREsTo is the closest to the practical risk estimation in PSC because it has been externally validated against conventional models. Quantitative MRCP requires implementation testing, and imaging or cholangiocarcinoma models remain investigational[7,31,66,73,74]. None of them were autonomous decision-makers.

A staged implementation pathway

Implementation should pass through four gates: Retrospective external validation in the intended local population; silent prospective testing without changing care; Monitored clinician-facing use with predefined override and accountability rules; and post-deployment surveillance for calibration drift, subgroup harm, and workflow effects. A model should move between gates only after meeting the prespecified safety and utility thresholds. If the performance deteriorates, the model must be recalibrated or retired.

CHALLENGES AND LIMITATIONS

Most of the evidence comes from retrospective development studies, and small samples, internal resampling, class imbalance, and information leakage can overstate performance (Table 2). AIOLI declined from internal AUROC 0.92 to external AUROC 0.74[38]; the AIH fibrosis random forest fell from 0.951 in training to 0.843 in a 33-patient validation set[42]; and the PSC computed-tomography model reported 0.89 using cross-validation without an external cohort[74]. The PREsTo and BJ-AID have strong multicenter and external evidence, although neither has shown that deployment improves patient outcomes[7,33]. To support clinical use, the decision, effective sample and event count, patient- and site-level data split, external validation, comparison with the accepted clinical standard, calibration, decision curve benefit, confidence interval, and subgroup performance are required. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) + AI provides a reporting framework[25,26]. A statistically sound model also requires workflow testing, privacy safeguards, accountability and override rules, drift monitoring, and recalibration or retirement plans. No published autoimmune-liver AI studies have shown better outcomes in prospective implementation trials. Claims of routine readiness were premature.

Table 2 Key challenges and minimum requirements before clinical use of artificial intelligence.

Challenge	Why performance may be misleading	Minimum requirement
Small effective sample and few events	Unstable estimates, wide uncertainty, and overfitting especially when n < 100	Justified sample/event count; confidence intervals; multicenter collaboration
Internal validation only	Random splits retain site, scanner, and practice signatures	Temporal and site-level external validation in the intended population
Information leakage	Patient, slide, or future information appears in both training and test data	Patient-level separation; locked preprocessing; independent test handling
Weak comparator	A high AUROC may not improve on an accepted score or specialist assessment	Head-to-head comparison with current care and incremental-value analysis
Discrimination without calibration	Rank ordering can be good while absolute risk is wrong	Calibration plot/slope, Brier score, and decision-curve benefit
Unrepresentative cohort	Performance and harm may differ by ancestry, sex, age, disease stage, or site	Prespecified subgroup and fairness analyses; representative recruitment
No workflow or outcome study	Offline accuracy does not establish better decisions or outcomes	Silent prospective testing followed by monitored implementation trial
Model and data drift	Assays, scanners, populations, and practice change after deployment	Performance monitoring, human override, recalibration, and retirement plan

AUROC: Area under the receiver operating characteristic curve.

FUTURE DIRECTIONS

Progress depends on the sharing of multinational data. No single center can recruit a sufficient number of patients with these rare diseases to represent their full clinical range. The International AIH Group, Global PBC Study Group, and International PSC Study Group have aggregated data that could support wider validation. Federated learning allows institutions or countries to train models without transferring identifiable patient data[17]. These cohorts still require a prospective design, independent testing, and reporting that follows TRIPOD + AI[26].

Future studies should test whether combining histology, imaging, longitudinal biochemistry, serology, and omics can improve decision-making beyond simple models. Foundation models provide one possible architecture, whereas self-supervised pre-training may reduce the need for large labeled datasets[24]. Synthetic data may help with some analyses; however, they cannot be substituted for representative patients or for external validation. Quantitative pathological and imaging measures may also be studied as trial endpoints as they have been in steatohepatitis[40].

Dynamic prediction is a nearer clinical aim than speculative “digital twins”. Repeated measurements could update a patient’s risk as treatment and disease course change[32,55]. Grounded, supervised LLMs may provide an interface for retrieving and explaining such estimates[23], but every clinical output must remain traceable to the underlying data and evidence.

Implementation requires prospective trials, interpretable outputs, regulations, clear accountability, and monitoring after deployment[11,28]. Equity must be designed for recruitment and validation because the disease burden and access to specialist care differ between regions. The deliberate inclusion of populations from low- and middle-income settings will determine whether these tools narrow or widen disparities[28,100]. Experience with noninvasive scores in viral and metabolic liver diseases shows that ML tools can enter practice when cohorts and validation are adequate[13]. Clinicians also require sufficient statistical and computational literacy to challenge a model’s output[12,28]. The intended endpoint is supervised decision support; the model detects patterns while the clinician supplies the context, judgment, and responsibility.

CONCLUSION

AI has produced credible signals in autoimmune liver disease; however, evidence of maturity varies sharply depending on the task. PREsTo has the strongest comparative external prognostic evidence in PSC, BJ-AID and ALNE may serve as differential diagnosis adjuncts in AIH, and PBC treatment-response models remain research tools in addition to the established GLOBE and UK-PBC pathways. High internal AUROCs, especially in samples below 100 or without site-level external testing, should not be interpreted as clinical readiness. Before adoption, clinicians should require a defined intended use, representative external validation, calibration, comparison with current care, net clinical benefit, subgroup analysis, and monitored workflow with a human override. No reviewed models support autonomous diagnosis, treatment selection, or surveillance escalation. The evidence now needed is prospective proof that clinician-AI use improves decisions and patient outcomes and not another isolated performance estimate.