Artificial Intelligence in Cirrhosis: Diagnosis, Screening, and Evaluation for Liver Transplantation

Medical imaging analysis

Deep learning models: Deep learning algorithms have demonstrated tremendous potential for improving diagnostic accuracy for cirrhosis by analyzing medical images and histologic data with minimal human intervention. By offering reproducible and objective assessments critical to early diagnosis and disease monitoring, these systems are being increasingly applied in clinical research. For instance, Ahn et al. developed a deep learning model to generate an ACE (AI-Cirrhosis-ECG) score, which identified patients with cirrhosis based on ECG findings (AUC = 0.91).⁸ Building on this, a follow-up study evaluated the ACE score’s ability to detect hepatic decompensation, revealing high discriminative performance with an AUC of 0.933 (95% CI: 0.923–0.942).⁹ In another area of histologic examination, the HEPASS (HEPatic Adaptive Steatosis Segmentation) algorithm was developed as a fully automated deep learning program to quantify liver steatosis in H&E-stained biopsy specimens. It utilizes object detection and semantic segmentation to identify both microsteatosis and macrosteatosis regardless of size, shape, or staining intensity. This model enables accurate, reproducible, and user-independent assessment and is the first to automatically measure both types of liver fat with high precision⁷ ( Table 1). In addition, liver fibrosis—an essential predictor of disease progression and mortality—is best assessed by histopathologic biopsy but remains limited by observer variability in manual methods. To address this, a model named MUSA-UNet (Multiple Up-sampling and Spatial Attention guided UNet) was developed to automatically and accurately segment portal tract regions in whole-slide liver biopsy images using deep learning. This model outperforms traditional methods by enabling standardized and efficient assessment of fibrosis, demonstrating strong potential to support AI-aided diagnosis of cirrhosis¹⁰ ( Table 1).

Radiomics: The combination of Computed Tomography (CT) imaging with laboratory and demographic data has been shown to improve the prediction of undiagnosed cirrhosis, outperforming the Fibrosis-4 (Fib-4) score or morphomics when used alone, with an AUC of 0.85.¹¹ Building on imaging-based approaches, deep learning-based spectral CT—a novel form of rapid kilovolt-switching CT (FKSCT)—enables higher energy separation and neural network-based computerized image reconstruction. This technique allows accurate quantification of iodine density and extracellular volume fraction (ECV) in liver parenchyma, which correlates strongly with the extent of fibrosis. As a noninvasive radiomic technique, it has the potential to be valuable in AI-assisted staging and diagnosis of liver fibrosis¹² ( Table 1). Complementing these imaging advancements, a point-of-care near-infrared spectroscopy (NIRS) device integrated with machine learning can quickly and accurately diagnose liver fibrosis from freshly excised tissue specimens. It offers a biopsy-free solution that holds promise for real-time monitoring of fibrosis in assessing liver transplantation¹³ ( Table 1). Similarly, a radiomics-based machine learning model utilizing diffusion-weighted MRI (DWI) was developed to non-invasively identify liver fibrosis and early-stage cirrhosis. Built on a support vector machine (SVM), the model achieved high accuracy (AUC 0.948 for healthy vs. abnormal liver and AUC 0.968 for fibrosis vs. cirrhosis). The technique surpasses current Magnetic resonance imaging (MRI) and deep learning techniques and has high potential for early and accurate diagnosis of cirrhosis.¹⁴ Further expanding AI applications, a recent innovation explored the use of ChatGPT-4 in liver ultrasound radiomics, demonstrating its ability to outline significant texture features and classify fibrosis, steatosis, and normal liver tissue with an accuracy of up to 76%. Although it was slightly less sensitive than standard software, it significantly reduced the analysis time by 40%, allowing for efficient high-throughput evaluation. This suggests that ChatGPT-4 has the potential to be a fast and scalable tool in AI-assisted liver disease diagnosis.¹⁵

Pattern recognition

AI-based pattern recognition and omics analysis are emerging as powerful tools in improving the diagnosis and outcome prediction in liver cirrhosis. In one study, Guo et al. used various ML models to evaluate laboratory data and related diagnoses to predict mortality in liver cirrhosis. All ML models outperformed the traditional Model for End-Stage Liver Disease–Sodium (MELD-NA) scores. A deep neural network (DNN) demonstrated AUC values of 0.88 for 90-day mortality, 0.86 for 180-day mortality, and 0.85 for 1-year mortality.¹⁶ In parallel, omics technologies hold tremendous potential in detecting and monitoring early liver fibrosis, with the possibility of delaying disease progression, lowering the risk of developing end-stage liver diseases such as hepatocellular carcinoma, and, eventually, mortality. While conventional statistical methods have traditionally been the foundation, artificial intelligence (AI) methods—especially data-driven AI—have become increasingly applied to omics data analysis. Deep learning (DL), a subset of AI that entails multilayered neural networks, is demonstrating itself to be particularly well-suited for recognizing complex patterns in high-dimensional omics data and is seeing growing applications in this field.¹⁷

Natural Language Processing (NLP)

Electronic Health Records (EHRs) analysis: NLP tools can analyze electronic health records to identify patients with undiagnosed cirrhosis by extracting relevant information from clinical notes, lab results, and imaging reports. Vleck et al. utilized natural language processing (NLP) to extract vast amounts of clinical information from electronic health records (EHRs) by mapping patient data to Systematized Nomenclature of Medicine – Clinical Terms (SNOMED CT) expressions. This method outperformed traditional methods, such as ICD codes and free-text search, in identifying and tracking the progression of NAFLD into nonalcoholic steatohepatitis (NASH) or cirrhosis. It showed superior precision, recall, and F1 scores, indicating high diagnostic accuracy. Their combination with ICD codes improved sensitivity without losing specificity. Findings highlight the strength of NLP in improving cirrhosis detection through EHR mining.¹⁸

Screening

Predictive modelling

Risk stratification: To support early identification and individualized management in chronic liver diseases, machine learning models have been increasingly applied for risk stratification by predicting disease progression and clinical outcomes. An ensemble machine learning model has been developed to predict outcomes in Alpha-1 antitrypsin deficiency-associated liver disease (AATD-LD), demonstrating Area Under the Receiver Operating Characteristic Curve (AUROC) values of 68.1% for all-cause mortality, 75.9% for liver-related death, and 91.2% for liver transplantation, indicating strong predictive capability.¹⁹ In patients with Primary Sclerosing Cholangitis (PSC), artificial intelligence has been utilized to analyze digitized histological slides, with a particular emphasis on cytokeratin 7 (K7) expression. It predicted a composite end-liver transplant, liver-related mortality, or cholangiocarcinoma, with an AUC of 0.81.²⁰ Another AI-based model combined Magnetic Resonance Cholangiopancreatography+ (MRCP+) imaging with clinical markers such as bilirubin and aspartate aminotransferase (AST) levels to predict the risk of transplant or death in PSC, achieving superior performance compared to the Mayo risk score, with an AUC of 0.86.²¹ Additionally, gradient boosting algorithms have been utilized to classify disease severity in NAFLD, with AUROC values of 0.90 for NAFLD detection, 0.82 for NASH, and 0.84 for advanced fibrosis, supporting the role of AI in noninvasive disease staging.²²

Biomarker analysis

Advanced algorithms: AI can analyze EHR data to identify early signs of liver disease, enabling timely screening and monitoring. A machine learning model was developed using serum, stool samples, and portal venous blood to investigate changes in the gut microbiome and enterohepatic bile acid circulation, identifying a combinatorial biomarker panel (AUC = 0.92) for submassive hepatic necrosis (SMHN). This panel integrated selected bacterial species and bile acid markers from multiple sources, offering substantial predictive value through a multi-omics machine learning approach.²³

Evaluation for liver transplantation

Patient selection

Predictive analytics: Artificial intelligence models can predict outcomes of liver transplantation based on a range of patient-specific factors, aiding in the selection of suitable candidates. One study compared deep learning models with the MELD-sodium score and demonstrated that deep learning was superior in predicting pretransplant mortality, achieving an AUC of 0.936.²⁴ Another machine learning model, the Optimized Prediction of Mortality (OPOM), outperformed MELD in predicting mortality, particularly among patients with higher disease severity. OPOM showed greater accuracy in identifying the sickest candidates, significantly reduced waitlist and post-transplant deaths, and maintained overall transplant volume, supporting a more effective “sickest-first” allocation approach.²⁵ In pediatric transplant centers, a Random Forest (RF) model was employed to identify the important predictors of mortality, and changes in creatinine and listing status were identified as the most significant factors. These high-risk feature identifications would improve survival by prioritizing the most vulnerable children for transplantation.²⁶ Additionally, an artificial intelligence model was developed to predict harmful alcohol use after liver transplantation for alcoholic hepatitis using psychosocial variables such as lack of a primary support person, caregiving responsibilities, and prior opioid use. These factors were key to delineating personalized risk profiles, and the model demonstrated outstanding predictive capability, with an AUC of 0.93, and outstanding performance in both training and external validation, supporting its utility in guiding post-transplant care.²⁷

Survival prediction: Several studies have used ML to predict post-transplant survival. Kazemi et al. applied a data mining approach to 902 adult liver transplant recipients and identified key pre-, peri-, and post-transplant factors—such as graft failure, infections, and vascular complications—as major determinants of survival. Their model highlighted how machine learning can uncover clinically relevant predictors that align closely with real-world outcomes, supporting its role in refining prognostic assessment²⁸ ( Table 1). One study applied an RF model to identify recipient weight, BMI, age, and INR as relevant predictors of survival, factors that were not highlighted by traditional models²⁹ ( Table 1). In another large-scale analysis using a U.S. transplant database, machine learning models, including random survival forests and neural networks, were employed for post-transplant survival prediction. Among these, partial logistic artificial neural networks (PLANN) demonstrated the most accurate long-term calibration and outperformed Cox models in predicting long-term survival. Key predictors identified included re-transplantation, donor age, life support, and Intensive Care Unit (ICU) admission before transplant, supporting their utility for personalized graft survival assessment.³⁰

Additionally, a scoring system was developed using ML to predict 90-day, 1-year, and 5-year post-transplant survival based solely on preoperative characteristics. This model identified age, MELD score, BMI, diabetes, and dialysis as strong predictors of 3-month mortality, achieving an AUC of 0.95.³¹ ML approaches have also been employed to predict graft dysfunction³² ( Table 1). For example, Lin et al. employed AI to graft proteomics and perfusate metabolomics to predict early allograft dysfunction (EAD), reaching an AUC of 0.83.³³ In the pediatric setting, analysis of the Studies in Pediatric Liver Transplantation (SPLIT) registry using Random Forest Analyses (RFA) evaluated whether demographic and biochemical factors one year after pediatric liver transplant could influence outcomes at three years. Predictors of adverse outcomes included racial factors, longer surgery duration, early biliary and vascular complications, and location within UNOS regions 2 and 5, with an accuracy of 0.71.³⁴ Similarly, a systematic review of studies using artificial neural networks (ANN) to predict graft survival found that ANN models outperformed traditional tools like the balance of risk (BAR), donor risk index (DRI), and Survival Outcome Following Liver Transplantation (SOFT), with ANN achieving AUCs between 0.82 and 0.84.³⁵

Complications of liver transplant: Various studies have employed AI to predict the probability of sepsis, pneumonia, Acute Kidney Injury (AKI), biliary, and cardiovascular complications post-LT^{36–38,40–42} ( Table 1).³⁹ These complications are potential causes of mortality post-transplant, and accurate prediction of their risk may help stratify patients appropriately to improve outcomes. One study identified raised preoperative indirect bilirubin, lower intraoperative urine output, longer anesthesia duration, lower preoperative platelet count, and graft steatosis classified as nonalcoholic steatohepatitis, Clinical Research Network grade 1 (NASH CRN1), as predictors of acute kidney injury (AKI)⁴⁰ ( Table 1). Similarly, infectious complications can pose a risk after LT. Colonization with Carbapenem-resistant Enterobacterales (CRE) at the time of liver transplant can increase the risk of post-transplant infections. Use of certain antibiotics, hepato-renal syndrome, and poor chronic liver failure sequential organ failure assessment (CLIF-SOFA) scores are associated with CRE.⁴³ An ML model was designed to predict CRE colonization with a sensitivity of 66%, a specificity of 83%, and a negative predictive value of 97%.⁴³ This could help tailor antibiotic prophylaxis at the time of liver transplant. Another study proposed an ultrasound-spectrogram fusion network that achieves greater accuracy, sensitivity, and specificity for the early and rapid detection of allograft dysfunction (EAD), surpassing radiologists and providing clinically helpful results. The four-ultrasound-image and clinical-data-fused model achieved an AUC of 0.968 (95% CI: 0.940–0.991), with approximately 30% greater performance on all metrics compared to radiologists. Furthermore, AI assistance significantly improved diagnostic performance (P < 0.050) for both less experienced and experienced clinicians.⁴⁴

In addition, Wehrle et al. used ML to study the outcomes of combined LT and cardiac surgery, which is regarded as a hazardous procedure. The model identified risk factors such as preoperative kidney dysfunction and Coronary Artery Bypass Grafting (CABG) surgery as predictors of mortality.⁴⁵ This could help in risk stratification and appropriately assess the safety of this procedure for each patient.

Liver transplant in patients with comorbidities: LT remains the curative option for acute on chronic liver failure (ACLF). Ge et al. and Yang et al. used ML to predict outcomes in the same way. Yang et al. found RF (AUC = 0.94) to be the most accurate among ML methods, while MELD (AUC = 0.70) proved to be an accurate traditional method for predicting the short-term survival of ACLF patients after LT⁴⁶ ( Table 1).⁴⁷ Several studies have been conducted to evaluate the prognosis after LT among patients with pre-existing comorbidities. ML was used in one study to predict post-LT survival among patients with diabetes mellitus (DM) and concluded that patients with pre-existing DM had poorer outcomes compared to non-diabetics.⁴⁸

Furthermore, higher serum creatinine, blood pressure, and sirolimus therapy were linked with higher mortality.⁴⁸ In another study, a deep learning model was trained to predict outcomes in Hepatitis C patients undergoing LT. The model demonstrated 100% accuracy in predicting postoperative complications, supporting its use in risk stratification and tailored monitoring for this population.⁴⁹ Ischaemic-type biliary lesions (ITBL), a potential complication post-LT, were also evaluated using AI. A convolutional neural network (CNN) was applied to detect and denoise CT angiography (CTA) images, thereby improving their quality and facilitating the accurate identification of hepatic artery lesions and thrombosis. Early CTA abnormalities were associated with the later development of ITBL.⁵⁰

Thus, it can be inferred that AI has been used extensively and shown promising capability to predict patient mortality and post-LT survival.

Preoperative assessment

Comprehensive evaluation: Preoperative assessment is one of the most critical parts of liver transplantation, and AI-based models have increasingly assisted in making it efficient and precise. An RF model was developed to predict 30-day survival after liver transplantation using preoperative blood investigations, with an AUC of 0.71. An imputation method proposed for handling missing data outperformed existing alternatives.⁵¹ Machine learning was also applied to evaluate the validity of the Global Leadership Initiative on Malnutrition (GLIM) criteria in cirrhotic patients, identifying midarm muscle circumference as a strong predictor of malnutrition and 1-year mortality. Another application of ML involved identifying patients at high risk of graft-versus-host disease post-transplant who may require closer monitoring.^52,53 Given the cardiovascular demands of liver transplantation, a machine learning algorithm was used to non-invasively compute CT-derived fractional flow reserve (CT-FFR), offering 85% accuracy in ruling out coronary artery disease and potentially reducing reliance on invasive coronary angiography.⁵⁴ In pediatric liver transplant recipients, transcriptomic signatures were integrated with protein interactome data through ML to identify targets for personalized antirejection therapies.⁵⁵ Deep learning has also been used to predict long-term survival post-transplant, with AUCs ranging from 0.69 to 0.85 for infection-related and graft rejection-related deaths, respectively.⁵⁶ As sarcopenia is known to increase post-transplant mortality risk, deep learning models were developed to assess muscle mass from CT scans. These tools showed equal predictive accuracy compared to manual assessments and helped associate sarcopenia with poor survival and increased risk of post-transplant diabetes.^57–59 Transfer learning, combined with the SHapley Additive exPlanation (SHAP) algorithm, was utilized to enhance the prediction of post-LT complications, yielding high precision, recall, and F1 scores.³⁹ Finally, the GraftIQ model—a fusion of a multiclass neural network and clinical expertise—achieved an AUC of 0.902, outperforming other ML models and showing strong potential as a decision support tool by accurately discriminating between graft etiologies and identifying high-risk patients for targeted interventions.⁶⁰

Resource allocation

Optimizing organ allocation: ML models have been used for the accurate matching of donors with recipients^61,62 ( Table 1). Briceño et al. found both ANN and RF superior to traditional methods, such as the Model for End-Stage Liver Disease (MELD) score. While RF models were able to assess the significance of different variables to build a decision tree, ANN models were better for analyzing large groups (Success rate > 80%).⁶¹

Several possible donor livers carry a risk of transmitting diseases, and it is essential to weigh the risks of receiving a potentially unhealthy liver against the dangers of waiting longer for a standard transplant^63,64 ( Table 1). A novel study utilized a deep learning model (Dl4jMLP) to analyze over 350,000 UNOS entries and predict short- and long-term survival after liver transplantation, incorporating 23 significant clinical factors. The model showed exceptional performance in both living donor (99.91%) and deceased donor (99.86%) recipients. The long-term survival prediction was above 98% within one year. Interestingly, there was no significant difference by donor type, dispelling assumptions in favor of living donors and facilitating broader use of deceased donor grafts. Deep learning outperformed traditional methods in terms of prediction accuracy.⁶⁵ ML has also been used to assess hepatic steatosis^66–68 ( Table 1). In most studies, ML is superior to pathologists’ assessments^68,69 ( Table 1).^70,71 Narayan et al. developed a computer vision artificial intelligence (CVAI) program. They found an association between donor liver steatosis and EAD that remained significant after adjusting for age, diabetes, MELD score, and other factors (2.9% vs 1.9%, p = 0.02). Additionally, CVAI showed slightly better calibration than pathologists in detecting steatosis.⁷⁰ The Banff Liver Working Group released recommendations to assess steatosis in biopsies of donor livers. However, they have not been implemented widely so far. Gambella et al. developed a CNN to assess steatosis and compared it with pathologists’ scores. They found that CNN and the scores assigned by pathologists showed greater overlap when the Banff approach was employed compared to pre-Banff recommendations.⁷¹ Contrary to the majority of literature, a study in the Republic of Korea used various ML models as well as logistic regression to assess for macrovesicular steatosis and found logistic regression to be the most accurate, with an accuracy of 80%.⁷² Another study applied AI by integrating transcriptomics and histopathology to evaluate biopsies from donor livers that were initially rejected for transplantation. Findings revealed that many of these rejected livers shared transcriptomic profiles with accepted organs, suggesting they might be viable for transplant.⁶ These insights highlight AI’s potential to expand the donor pool and enhance organ allocation decisions.