BPG is committed to discovery and dissemination of knowledge
Minireviews Open Access
Copyright: ©Author(s) 2026. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution-NonCommercial (CC BY-NC 4.0) license. No commercial re-use. See permissions. Published by Baishideng Publishing Group Inc.
World J Hematol. Nov 20, 2026; 12(3): 126171
Published online Nov 20, 2026. doi: 10.5315/wjh.126171
Artificial intelligence and the rebalanced hemostasis of chronic liver disease: From viscoelastic testing to digital twins
Ahmed Salman, Department of Internal Medicine, Kasr Alainy School of Medicine, Cairo 11559, Al Qāhirah, Egypt
ORCID number: Ahmed Salman (0000-0003-0026-0841).
Author contributions: Salman A conceived the review, designed and performed the literature search, appraised and selected the included studies, prepared the figures and tables, drafted and critically revised the manuscript, and approved the final version submitted for publication.
AI contribution statement: Anthropic Claude (Opus 5, as recorded in the earlier manuscript disclosure) and OpenAI Codex (GPT-6) assisted with language editing, manuscript organization, reference and reported-data checking, and preparation or revision of conceptual figures and tables. The tools helped compare published study details and correct wording. No new research data were generated. The literature-selection process and study conclusions remain the author’s responsibility. AI-assisted statements, citations, numerical summaries, and figures require author verification against the cited sources. The author accepts responsibility for the accuracy, integrity, and originality of the submitted manuscript.
Conflict-of-interest statement: The author declares no conflict of interest.
Corresponding author: Ahmed Salman, FRACP, FRCP, Department of Internal Medicine, Kasr Alainy School of Medicine, 1 Al-Saray Street, Al-Manial, Cairo 11559, Al Qāhirah, Egypt. awea844@gmail.com
Received: July 28, 2026
Revised: September 4, 2026
Accepted: October 16, 2026
Published online: November 20, 2026
Processing time: 114 Days and 11.8 Hours

Abstract

Chronic liver disease, particularly cirrhosis, produces a fragile, rebalanced hemostatic state. Patients may develop bleeding or thrombosis. The international normalized ratio and platelet count incompletely reflect concurrent procoagulant and anticoagulant changes and poorly predict procedural bleeding. Viscoelastic testing and thrombin-generation assays assess hemostasis more broadly. Randomized trials support viscoelastic testing-guided reductions in blood product use, but these trials do not establish prediction of later bleeding or thrombosis. Machine-learning models address gastrointestinal bleeding outcomes, variceal hemorrhage, portal vein thrombosis (PVT), and transfusion requirements using clinical records, imaging, or elastography. Most cirrhosis models address one outcome, omit direct hemostatic measurements, and have limited external validation. Within the English-language eligibility criteria and focused queries of a PubMed update on July 29, 2026, no adult model jointly estimating future bleeding and PVT in cirrhosis was identified. This structured narrative minireview evaluates the evidence for both outcomes and specifies research requirements for paired-risk models and a possible digital twin of the coagulation-portal axis. Treatment-specific predictions would require additional causal validation.

Key Words: Artificial intelligence; Chronic liver disease; Cirrhosis; Hemostasis; Portal vein thrombosis; Viscoelastic testing

Core Tip: Bleeding and thrombosis in cirrhosis are competing failures of a fragile hemostatic system; however, published prediction models usually quantify only one failure. Viscoelastic testing describes the current clot behavior more directly than routine tests, but it is rarely incorporated into future-risk models. Before developing a digital twin, a paired-risk model should combine clinical, hemodynamic, and hemostatic data and be tested in external cohorts.



INTRODUCTION

Chronic liver disease denotes the broader clinical problem in this review. Cirrhosis is used for studies that enrolled patients with cirrhosis and for discussion of portal-hypertensive complications. The terms are related but differ in clinical scope.

Hemostasis in cirrhosis is not a unidirectional problem. Factors II, V, VII, IX, X, and XI decrease along with protein C, protein S, antithrombin, and plasminogen levels. Thrombocytopenia coexists with increased von Willebrand factor (VWF) and reduced levels of its cleaving protease, a disintegrin and metalloproteinase with a thrombospondin type 1 motif, member 13 (ADAMTS13). Profibrinolytic and antifibrinolytic changes can also occur in the same patient. This results in a rebalanced system rather than a bleeding diathesis[1]. Its reserves are narrow, and infection, renal failure, portal hypertension, and acute decompensation can cause bleeding or thrombosis.

Clinical guidelines reflect this physiology. The 2020 practice guidance of the American Association for the Study of Liver Diseases advises against routine correction of the international normalized ratio (INR) solely to normalize the result before procedures and recommends individualized assessment of severe thrombocytopenia[2]. Similarly, the Baveno VII consensus states that the INR does not reflect hemostatic status in cirrhosis[3]. Procedural decisions therefore require clinical assessment of the intervention, disease severity, and modifiable risk factors, rather than reliance on a single laboratory threshold.

Two developments warrant another look at this problem. Global hemostatic assays are now more mature, with viscoelastic testing (VET) evaluated in randomized trials as well as case series[4]. Machine learning (ML) has progressed in both specialties. Formal reviews now assess artificial intelligence (AI) applications in hematology[5,6] and hepatology[7]. Using these measurements in care depends on hematology expertise in assay standardization, thrombosis consultation, perioperative blood management, and interpretation for each testing platform.

Reported portal vein thrombosis (PVT) prevalence varies widely across cirrhosis cohorts, reflecting differences in populations, risk factors, and classification[2]. Variceal hemorrhage is one of the most dangerous consequences of portal hypertension. Differences in imaging surveillance can also influence outcome labels used to train a model. Procedures, such as paracentesis, endoscopy, biopsy, line insertion, and surgery, require decisions regarding prophylactic transfusion, whereas newly detected thrombi require assessment for anticoagulation. An estimate integrating both bleeding and thrombotic risk could inform these decisions, but would require validation for the intended setting.

These lines of evidence have not been fully integrated. Studies summarized here address variceal hemorrhage, PVT, and transfusion as separate prediction targets. Both bleeding and thrombosis can arise from the same perturbed system, and an intervention may affect both outcomes. Anticoagulation in a patient with large varices illustrates why paired patient-level estimates could be useful, provided that prediction is distinguished from estimation of treatment effects.

This review addresses the following three questions: First, what do routine tests, global assays, and current clinical variables reveal about present and future portal-hypertensive bleeding, procedure-related bleeding, transfusion needs, and PVT? Second, what have ML models contributed to each outcome, and which design or validation limitations prevent their clinical use? Third, what data, harmonization, validation, and governance would be required for a paired bleeding-thrombosis model, and ultimately, a digital twin of the coagulation-portal axis?

METHODOLOGY
Search strategy

This structured narrative minireview used a non-systematic search. The primary evidence base was defined by searching PubMed, MEDLINE, EMBASE, and the Cochrane Library from inception through June 30, 2026, using Scopus and Web of Science for forward citation tracing. The core free-text strategy was: (cirrhosis OR “chronic liver disease” OR “portal hypertension” OR “liver failure”) AND (hemostasis OR coagulopathy OR “rebalanced hemostasis” OR thromboelastography OR thromboelastometry OR “viscoelastic test” OR “thrombin generation” OR transfusion OR “variceal bleeding” OR “portal vein thrombosis” OR anticoagulation) AND (“artificial intelligence” OR “machine learning” OR “deep learning” OR “neural network” OR “prediction model” OR “digital twin” OR “large language model”). The controlled vocabulary and syntax were adapted for each source. ClinicalTrials.gov, reference lists, guidance documents, and consensus statements were also examined. June 30, 2026 denotes the primary search coverage endpoint; a separate execution date was not retained in the available search record. Iterative narrative searches did not retain database-level hit counts, and no retrospective counts are presented. No Preferred Reporting Items for Systematic Reviews and Meta-Analyses flow diagram, duplicate independent screening, or quantitative risk-of-bias instruments were used.

A separate, narrower PubMed-only update was completed on July 29, 2026, solely to test the central absence claim; it did not extend the date range of the primary multi-database evidence search. The query (cirrhosis[Title/Abstract] OR “chronic liver disease”[Title/Abstract]) AND (thromboelastography[Title/Abstract] OR thromboelastometry[Title/Abstract] OR “viscoelastic test”[Title/Abstract] OR “thrombin generation”[Title/Abstract]) AND (“machine learning”[Title/Abstract] OR “artificial intelligence”[Title/Abstract]) returned one conference-supplement record and no eligible adult clinical model. The query (cirrhosis[Title/Abstract] OR “chronic liver disease”[Title/Abstract]) AND (“bleeding risk”[Title/Abstract]) AND (“thrombosis risk”[Title/Abstract]) AND (“machine learning”[Title/Abstract] OR “prediction model”[Title/Abstract]) returned no records. The absence finding is restricted to these PubMed queries, the July 29, 2026 update, and the English-language eligibility criteria. It does not establish absence across other databases, terminology, or languages.

Selection criteria

Adult human studies were prioritized in this review. Eligible sources included randomized trials, cohorts, model development or validation studies, systematic reviews, meta-analyses, societal guidance, and reporting standards for clinical AI. English-language articles with retrievable full texts were included. Model studies had to report a clinical endpoint related to bleeding, thrombosis, or transfusion. Technical-only studies, pediatric cohorts, and non-cirrhotic splanchnic vein thrombosis were excluded. Liver transplantation was considered separately, because intraoperative models predict events over hours in continuously monitored patients, whereas the outpatient models discussed here predict events over weeks to years. Transplant studies were cited only when they addressed relevant methodological questions. Conference abstracts and preprints were not treated as eligible evidence of clinical model validation. Non-English full-text reports were excluded from the main synthesis; accessible English abstracts of related non-English studies were considered as contextual evidence, including the concurrent-outcome nomogram discussed below. The retained records do not establish the screening stage at which every language exclusion occurred.

One author screened the retrieved records and extracted the population, design, target outcome, prediction horizon, data modality, model class, sample size, event count, discrimination with confidence intervals, calibration, and validation methods into a standardized evidence table. The validation was classified as random split or resampling internal validation, temporal validation, geographic external validation, or multicenter external testing. Table entries were checked against the available primary reports or their indexed abstracts. The revision response records which sources were accessible and which details could not be resolved. The DOI, PMID, title, and publication metadata of all 45 cited records were checked against PubMed and publisher records. These covered hemostasis and assay studies, guidance, randomized trials, reviews and meta-analyses, prediction studies, reporting standards, and contextual literature.

Synthesis

The evidence was grouped into six themes: Measurement of rebalanced hemostasis; ML for bleeding; ML for thrombosis; transfusion and procedural stewardship; mechanistic simulation; and translation into practice. Prospective studies, multicenter studies, and studies tested in independent external cohorts carried more weight in the interpretation. Random splits, temporal splits from the same hospital, and validation at another geographic site were treated as different levels of testing. The studies differed in their outcomes, prediction horizons, case mix, available predictors, and validation methods. Discrimination estimates were therefore described without treating them as directly comparable or suitable for pooling.

REBALANCED HEMOSTATIC SYSTEM AND ITS MEASUREMENT
Changes underlying rebalanced hemostasis

Rebalanced hemostasis reflects several changes, each with a different margin of safety. Vitamin K-dependent factor synthesis falls together with protein C, protein S, and antithrombin levels. Thrombin generation is therefore less impaired than the prothrombin time would suggest. Factor VIII, which is produced by endothelial cells, rises as clearance becomes less efficient. VWF also rises, while ADAMTS13 falls, partly supporting primary hemostasis despite thrombocytopenia. Fibrinogen concentration may remain adequate despite qualitative dysfunction. Plasminogen and antiplasmin levels both decline, and the balance between hyperfibrinolysis and hypofibrinolysis changes with the disease stage and infection[1,2].

These compensations preserve the balance, not reserve. Sepsis may increase factor VIII and VWF levels, whereas renal failure can impair platelet function without producing a parallel change in the routine tests most often ordered. Therefore, the system can move quickly toward bleeding or thrombosis, even when the INR and platelet count are stable.

What conventional tests measure, and what they miss

The INR standardizes prothrombin time for monitoring vitamin K antagonists. These tests register reduced procoagulant activity but do not capture the accompanying reduction in anticoagulant activity, including protein C and antithrombin. Conventional tests perform poorly at predicting perioperative bleeding[8]. In a prospective study of acute decompensation and acute-on-chronic liver failure, laboratory coagulopathy did not distinguish patients who developed in-hospital procedure-related or spontaneous non-portal-hypertensive bleeding from those who did not[9]. The platelet count has a similar limitation because increased VWF partly compensates for thrombocytopenia, and a low platelet count reflects portal hypertension. Figure 1 shows how a rebalanced system can fail in either direction and where computational tools may add information.

Figure 1
Figure 1 Conceptual framework of rebalanced hemostasis in chronic liver disease and its bleeding and thrombotic complications. Pro-bleeding and prothrombotic drivers can coexist in a fragile equilibrium with limited reserve. Arrows indicate hypothesized clinical relationships rather than quantified causal effects. Current assays and target-specific artificial intelligence models assess different aspects of this system; the diagram is not a validated clinical algorithm. AI: Artificial intelligence; INR: International normalized ratio; VET: Viscoelastic testing; VWF: Von Willebrand factor.

Operational constraints also matter. The INR is widely available and embedded in prognostic scores, transplant allocation, procedural checklists, and local transfusion protocols. Any additional assay or prediction model should provide timely, interpretable information that improves decisions beyond these established workflows.

Global assays and their prognostic reach

VET, performed using thromboelastography (TEG) or rotational thromboelastometry (ROTEM), examines clot initiation, propagation, strength, and lysis in whole blood; therefore, platelets, fibrinogen, and fibrinolysis all contribute to a single trace. Thrombin-generation assays assess related aspects of coagulation, particularly when thrombomodulin is added to expose the protein C pathway, and near-patient whole-blood formats have been described for patients with cirrhosis[10].

A small longitudinal study suggests that global assays can provide prognostic information. ROTEM parameters were associated with later bleeding in patients with cirrhosis[11]. A prospective study included 231 patients: 80 with compensated cirrhosis, 70 with stable decompensation, and 81 with acute decompensation. All 11 procedure-related bleeds occurred in the acute-decompensation group; within that subgroup, whole-blood endogenous thrombin potential (ETP) had an area under the receiver operating characteristic curve (AUROC) of 0.854 [95% confidence interval (CI) 0.732-0.976]. A preprocedure thrombomodulin-modified whole-blood ETP < 350 nmol/L·minute identified major bleeding in this small cohort, but requires independent validation[12]. Neither whole-blood nor platelet-poor-plasma thrombin generation predicted the 15 thrombotic events, most of which were PVT. These findings do not establish a bidirectional assay or model.

The main constraint is comparability across time and laboratories. A trace obtained today may change after infection, renal deterioration, diuretic adjustment, or decompensation. TEG, ROTEM, and thrombin-generation results also depend on the analyzer, cup or channel, activator, reagent lot, specimen handling, and the interval from collection to testing. Cutoffs cannot be assumed to transfer between platforms. Multicenter modeling, therefore, requires prespecified operating procedures, complete platform and reagent metadata, external quality control, and either platform-specific normalization or separate platform models before the pooled thresholds are considered.

What the randomized trials answered

Table 1 summarizes five randomized controlled trials of VET-guided transfusion in cirrhosis and a meta-analysis of earlier trials. In acute variceal bleeding, TEG guidance sharply reduced blood product use with no statistically significant difference in early rebleeding or mortality and was associated with less rebleeding by 42 days[13]. Similar reductions in transfusion occurred in nonvariceal upper gastrointestinal bleeding[14], before high-risk invasive liver-related procedures[15], in patients with severe thrombocytopenia undergoing procedures[16], and with ROTEM-guided avoidance of preemptive transfusion before elective procedures[17]. The meta-analysis found reduced product exposure across its included trials; it preceded the more recent trials in Table 1 and did not pool all five listed studies[4]. Four of the five listed trials were conducted in India, which limits transportability to other transfusion practices. More importantly, the trials randomized the transfusion strategy; they did not develop individualized predictions of portal-hypertensive bleeding, procedure-related bleeding, transfusion needs, or PVT.

Table 1 Randomized controlled trials and pooled analysis of viscoelastic testing-guided transfusion in cirrhosis.
Ref.
Design and sample size
Clinical setting
Assay
Transfusion outcome
Bleeding and mortality outcome
Rout et al[13]RCT, 30 vs 30, IndiaAcute variceal bleedingTEG4 patients (13.3%) vs all patients (100%) transfused, P < 0.001Rebleeding at 5 days 3.3% vs 13.3%, P = 0.167; rebleeding at 42 days 10.0% vs 36.7%, P = 0.012; 6-week mortality 13.3% vs 26.7%, P = 0.176
Kumar M et al[14]RCT, 49 vs 47, IndiaNonvariceal upper GI bleeding in advanced cirrhosisTEGAll three components (plasma, platelets, and cryoprecipitate) given to 26.5% vs 87.2%, P < 0.001Failure to control bleeding, rebleeding, and mortality similar in both arms
Vuyyuru et al[15]RCT, 29 vs 29, IndiaHigh-risk invasive liver-related proceduresTEG9 patients (31%) vs all patients (100%) transfusedNo procedure-related bleeding through 5 days in either arm
Biswas et al[16]RCT, three arms of 29, IndiaSevere thrombo­cytopenia before invasive proceduresTEGPlatelets: 4/29 (13.8%) with TEG, 29/29 (100%) with standard care, and 4/29 (13.8%) on demand; P < 0.001 for TEG vs standard careMinor bleeding in 3 (10%) with TEG vs 1 (3.4%) in each comparator arm, P = 0.43; no bleeding-related death
Kumar R et al[17]RCT, 20 vs 20, SingaporeElective procedures, predominantly low to intermediate riskROTEM40% vs 100% transfused, P < 0.001; less plasma, P < 0.001, and fewer platelets, P = 0.046No clinically significant bleeding; 30-day and 90-day mortality did not differ significantly
Hartmann et al[4]Meta-analysis of 5 earlier RCTs (302 patients)Cirrhosis, including liver transplantationTEGPlatelet use RR 0.17, 95%CI 0.03-0.90, P = 0.047-day mortality RR 0.52, 95%CI 0.30-0.91, P = 0.02

These trials support transfusion stewardship in the studied settings. Their small samples and event counts limit conclusions about uncommon bleeding or thrombotic outcomes. They evaluated transfusion strategies rather than developing models to predict both complications in an individual patient.

MACHINE LEARNING AND THE BLEEDING PHENOTYPE
Triage in acute gastrointestinal bleeding

The most mature application is pre-endoscopic risk stratification. A gradient-boosting model trained on presenting variables from 1958 patients in four countries predicted hospital intervention or death within 30 days with an AUROC of 0.91, compared with 0.88 for the Glasgow-Blatchford score, and achieved 0.90 in an external Asia-Pacific cohort[18]. The same group later validated a model based entirely on structured electronic health record (EHR) data[19]. Both studies included patients with cirrhosis but were not specifically designed for cirrhosis or portal hypertension.

Varices: Detection, first hemorrhage, and rebleeding

Noninvasive identification of varices has shifted from tabular predictors to imaging. A vision transformer combined with reinforcement learning applied to computed tomography (CT) angiography classified first variceal hemorrhage in a retrospective cohort, with an AUROC of 0.985 in a randomly split validation subset[20]; despite multicenter recruitment, this was not an external cohort. Deep learning has also been used to estimate portal pressure gradients in patients with decompensated cirrhosis without catheterization[21]. A recent review summarized the endoscopic and imaging literature in detail[22] but did not focus on thrombosis prediction or direct global hemostatic measurements.

Rebleeding after endoscopic therapy has produced some of the more credible cirrhosis-specific models. A prospective multicenter study developed a rebleeding model in 153 patients with hepatitis B virus (HBV)-related cirrhosis and spontaneous portosystemic shunts (SPSS), with 103 SPSS patients in external validation. AUROCs were 0.814 internally and 0.776 externally[23]. Another deep-learning model estimated 6-week treatment failure and 1-year mortality in 1227 patients treated with endoscopic variceal ligation plus drugs. A separate group of 1863 patients receiving preemptive transjugular intrahepatic portosystemic shunts (TIPS) was used for retrospective treatment-stratification analyses, which do not establish individualized causal benefit[24]. Across 21 studies and 7011 patients, pooled C-indices were 0.85 for esophageal and 0.89 for esophagogastric variceal bleeding[25]. These values should not be ranked as though they describe the same target. Credibility depends on the outcome definition, event count, 95%CI, calibration, and testing in a genuinely independent population. A lower externally validated AUROC is not inherently better or worse than a higher internally validated estimate.

Which model classes have earned their complexity

Model complexity should be justified against the available data and an appropriate benchmark. For structured predictors, penalized regression provides a useful comparator for tree-based methods. Images, videos, and waveforms may support deep-learning approaches, but their incremental value should be tested against simpler features and models. Whether direct hemostatic measurements improve performance more than changing the classifier remains an empirical question; the studies summarized here do not establish that comparison.

What critical appraisal reveals

Methodological appraisal identifies important limitations. A systematic review with formal methodological analysis of machine-learning models for variceal bleeding risk found inadequate participant numbers, particularly in validation cohorts, unclear handling of missing data, and incomplete reporting of endoscopic assessment, with most included studies judged to be at a high or unclear risk of bias[26]. Several cirrhosis-specific cohorts in Table 2 are HBV-predominant Asian populations. These models’ performance in alcohol-related or metabolic dysfunction-associated cirrhosis requires independent evaluation.

Table 2 Representative machine-learning models across the hemostatic spectrum in chronic liver disease.
Ref.
Prediction target
Data modality
Model
Cohort
Reported performance
Validation design
Shung et al[18]Hospital-based intervention or death within 30 days in upper GI bleedingPresenting clinical and laboratory variablesGradient boosting1958 derivation across 4 countries; 399 geographic externalAUROC 0.91 internal and 0.90 external vs 0.88 and 0.87 for the Glasgow-Blatchford scoreGeographic external validation
Shung et al[19]Red cell transfusion, hemostatic intervention, or 30-day all-cause mortality in GI bleedingStructured EHR dataGradient boosting (XGBoost)2546 training; 850 internal; 926 geographic externalExternal AUROC 0.92, vs 0.89 for both Glasgow-Blatchford and Oakland scoresGeographic external validation
Wan et al[20]First variceal hemorrhage classification in a retrospective cohortCT angiography imagesVision trans­former plus reinforce­ment learning184 patients (107 with and 77 without first hemorrhage); random 7:3 splitAUROC 0.956 (95%CI 0.919-0.988) training; 0.985 (95%CI 0.955-1.000) validationRandom-split internal validation; no external cohort
Liu et al[21]Portal pressure gradient and high-risk portal hypertensionLaboratory and imaging variablesArtificial neural network428 patients with decompensated cirrhosis; 2 external datasetsReported accuracy 87.5%; external accuracy 85.4% and 90.8%; external AUROC 0.842 for high-risk portal hypertensionExternal validation in 2 datasets
Zheng et al[23]Rebleeding after endoscopic therapy in HBV cirrhosis with SPSSClinical and imaging featuresExtreme gradient boostingSPSS model cohorts: 153 development and 103 external; 77/153 and 68/103 rebleeds. Overall study populations: 295 and 190AUROC 0.814 (95%CI 0.785-0.843) internal; 0.776 (95%CI 0.748-0.805) externalProspective geographic external validation
Xiang et al[24]6-week treatment failure and 1-year mortality in acute variceal bleedingClinical data within 24 hours of admissionDeep learning1227 EVL-plus-drug patients: 668 training, 167 internal, 392 external. Separate p-TIPS treatment-stratification cohort: 1863AUROC 0.842 and 0.954 internal; 0.814 and 0.889 externalMulticenter external validation
Lian et al[25]Esophageal and esophagogastric variceal bleedingClinical, radiomic, and endoscopic inputsSystematic review and meta-analysis21 studies; 7011 patients; 1412 esophageal and 733 esophagogastric bleeding eventsPooled validation C-index 0.85 (95%CI 0.77-0.92) and 0.89 (95%CI 0.85-0.94), respectivelyPooled across heterogeneous studies
Li et al[29]Diagnosis of existing PVT in cirrhosisClinical, blood, and ultrasound variablesStacking ensemble468 development at the First Hospital of Lanzhou University; 348 external validation at Jilin Hepato-Biliary Diseases Hospital, Changchun, ChinaAUROC 0.73 development and 0.86 validationGeographic external validation
Meng et al[30]Portal vein thrombosis at 3 years and 5 years in HBV-related cirrhosisClinical and laboratory variablesArtificial neural networkDevelop­ment (2011-2014) and temporal validation (2015-2016) at the same hospital; cohort sizes are inconsistently reported in the sourceTemporal-validation AUROC 0.958 (95%CI 0.944-0.973) at 3 years and 0.973 (95%CI 0.958-0.987) at 5 yearsTemporal internal validation; no external site
Lou et al[37]Intraoperative red cell transfusion across surgical specialtiesPreoperative EHR dataGradient boosting (S-PATH)3275956 surgical cases at 45 United States hospitalsMedian AUROC 0.929 (IQR 0.915-0.946) vs 0.857 (IQR 0.822-0.884) for the maximum surgical blood order scheduleMulticenter external testing without local retraining

The summarized models mainly use structured EHR variables, endoscopic images and videos, contrast-enhanced CT and radiomic features, or liver and spleen stiffness. Within the focused PubMed queries, July 29, 2026 search date, and English-language eligibility criteria, no eligible adult non-transplant cirrhosis model incorporating VET or thrombin-generation measurements was identified. This finding is limited by the narrow search terminology and scope.

Related applications demonstrate technical feasibility in other clinical settings. Admission ROTEM supported prediction of 30-day mortality after trauma[27]. Separately, machine-learning models predicted laboratory changes after transfusion during liver transplantation, but lacked TEG or ROTEM inputs and showed limited external generalization[28]. These studies differ in population, predictor timing, and outcome. Neither validates future paired bleeding and PVT prediction in non-transplant cirrhosis.

A useful next study would test whether archived VET traces add predictive information to routine clinical and portal-hemodynamic variables. Such a study would need to account for changing physiology and test performance in independent cohorts. The focused search supports this as a research question, without establishing that all potentially related models have been excluded.

MACHINE LEARNING AND THE THROMBOTIC PHENOTYPE
Why the thrombotic endpoint matters

PVT is associated with immediate clinical consequences. Extension into the superior mesenteric vein may complicate or preclude transplantation, and mesenteric ischemia is associated with substantial mortality. PVT is associated with more severe portal hypertension, but its independent contribution to decompensation remains uncertain. Prediction is relevant because prophylaxis and anticoagulation are available, and the risk may be detectable months before established occlusion[2].

Predicting PVT

PVT has a smaller prediction-model literature than variceal bleeding. A stacking ensemble trained on clinical, blood, and ultrasound variables in 468 patients from Lanzhou reached an AUROC of 0.86 in an external cohort of 348 patients from Jilin[29]. All external-validation patients had undergone splenectomy, whereas none in the development cohort had, limiting generalizability to other cirrhosis populations. An artificial neural network developed in patients with HBV-related cirrhosis showed an AUROC of 0.967 at 3 years in training and 0.958 in temporal validation at the same hospital[30]. The first study addressed PVT diagnosis; the second estimated PVT occurrence over follow-up. They therefore evaluate different prediction tasks.

These studies raise two concerns. PVT is often found incidentally on imaging performed for another reason; therefore, the outcome label in retrospective datasets partly reflects imaging intensity rather than biology. A patient scanned more often is more likely to be labeled as positive. Beyond that, none of these models includes the variables that a hematologist would consider central, such as thrombin generation with thrombomodulin, VWF, or the factor VIII to protein C ratio. These associations do not establish a hemostatic mechanism, and high discrimination in a single-etiology cohort requires evaluation in other populations.

Anticoagulation decision

Once PVT is diagnosed, the two failure modes collide. The evidence on the safety of anticoagulants and antiplatelet agents in cirrhosis has been reviewed in detail[31], and a systematic review and meta-analysis of direct oral anticoagulants (DOACs) for PVT in cirrhosis reported recanalization and bleeding outcomes that support their use in selected patients[32]. A tool that reports recanalization without bleeding risk provides an incomplete account of the clinical trade-off. The reviewed evidence does not establish whether a paired model improves decisions compared with specialist assessment.

Individualized anticoagulation also requires attention to renal function and genetic variation. Renal function can change anticoagulant exposure and susceptibility to bleeding; genetic variants affect vitamin K antagonist dose requirements. In a Japanese outpatient cohort, estimated glomerular filtration rate had the strongest association with warfarin dose. Variants in VKORC1, CYP4F2, GGCX, and CYP2C9 remained associated with dose after adjustment for renal function[33]. These findings concern neither cirrhosis specifically nor DOACs. They support retaining renal function, drug choice, concomitant therapy, and relevant pharmacogenetic data in models of individualized anticoagulation.

Within the specified PubMed queries, July 29, 2026 search date, and English-language eligibility criteria, no adult cirrhosis model jointly estimating future bleeding and PVT was identified. Related studies nevertheless require consideration. A retrospective study of 385 patients with established cirrhotic PVT reported a nomogram for concurrent esophagogastric variceal bleeding (AUROC 0.820, 95%CI 0.707-0.843), rather than paired future outcomes[34]. Its English-language abstract was considered as contextual evidence; the Chinese full text was outside the main synthesis. Outside liver disease, paired models for venous thromboembolism and major bleeding were derived from the same arthroplasty cohort[35]. Therefore, multi-outcome and competing-risk methods are available; however, their use with shared hemostatic and portal variables in cirrhosis was not established by the eligible literature.

Inflammation as the shared driver

A shared model also has biological plausibility. Inflammation, bacterial translocation, renal dysfunction, and infection can promote thrombosis through increased VWF and factor VIII levels, platelet activation, and impaired protein C signaling, as well as worsening portal pressure and decompensation. Observational data identify overlapping factors for PVT and esophagogastric variceal bleeding, including portal vein diameter and hemoglobin level, but do not provide paired future-risk estimates[36]. In acute decompensation, both risks may increase together, helping to explain why static coagulation tests fail to distinguish patients who bleed from those who do not[9]. Therefore, a joint model should test inflammatory markers, renal function, infection status, and hemodynamic variables as shared predictors, rather than assuming that the two outcomes are simple opposites.

TRANSFUSION AND PERIPROCEDURAL HEMOSTATIC STEWARDSHIP

Transfusion is a direct meeting point for the fields of hematology, hepatology, and resource stewardship. A surgical model tested across 3275956 cases at 45 hospitals in the United States achieved a median AUROC of 0.929, compared with 0.857 for the maximum surgical blood order schedule[37]. Although not developed for cirrhosis, this study shows that routinely collected preoperative data can support transportable, patient-specific blood-ordering decisions.

Transfusion has additional consequences in chronic liver disease. Plasma can increase intravascular volume and portal pressure, so unnecessary transfusion may cause harm as well as consume resources. Baveno VII recommends a conservative red blood cell transfusion strategy in acute variceal bleeding and advises against fresh frozen plasma because it does not correct the underlying hemostatic disturbance and may worsen volume overload and portal hypertension[3].

A VET-informed model could be tested for two uses: Estimating blood product needs before a procedure and identifying patients who might benefit from targeted hemostatic management. For either use, it would have to improve on routine clinical assessment and existing transfusion protocols. Some tertiary centers already collect the required data. Prospective studies would still need linked records, harmonized assays, adjudicated outcomes, and resources to put the model into practice.

FROM PREDICTION TO SIMULATION: A DIGITAL TWIN OF THE COAGULATION-PORTAL AXIS

A prognostic model’s estimates reflect the treatment patterns in the data used to develop it. Such a model does not, by itself, estimate what would happen to the same patient under alternative treatments. Paired prediction estimates two outcomes from shared predictors; multi-outcome models may account for their dependence, while competing-risk methods account for events such as death or transplantation that alter observation of the target outcome. Bleeding and PVT can both occur in one patient and should not automatically be treated as mutually exclusive competing events. Estimating the effects of alternative treatments requires a separate causal framework.

Mechanistic modeling offers a complementary approach. Mechanistic models of coagulation have been used to explore thrombin generation across varying factor and inhibitor concentrations, providing a framework for testing mechanistic hypotheses[38]. More recently, AI methods have been combined with mechanistic thrombosis models to connect simulations to clinical translation[39]. In cardiology, digital twin technology combined with generative AI is being developed for individualized cardiovascular care[40].

No clinically validated equivalent for hepatic hemostasis was identified within the literature reviewed. Figure 2 proposes four components of a conceptual research framework. The first is a mechanistic coagulation core parameterized by measured factor and inhibitor concentrations rather than population averages to reproduce the individual thrombin-generation curve. The second is a hemodynamic layer for portal pressure, splanchnic flow, and shear. Because PVT is as much a flow problem as a coagulation problem, and because portal pressure responds to volume, beta-blockade, and TIPS, noninvasive estimates of portal pressure could be investigated as candidate inputs, with their measurement uncertainty retained[21]. The third assimilates new laboratory, endoscopic, elastographic, and imaging data as they arrive, with multimodal foundation models as possible substrates[41]. The fourth would explore hypothetical interventions, such as starting a DOAC, banding varices, transfusion, inserting a TIPS, or continuing observation, and generate candidate paired bleeding and thrombosis estimates for research evaluation. These outputs would not establish individualized treatment effects.

Figure 2
Figure 2 Proposed digital twin of the coagulation-portal axis in chronic liver disease. This proposed research design brings together longitudinal data assimilation, a patient-specific coagulation core, portal-hemodynamic modeling, and simulations of hypothetical interventions. Arrows show the proposed flow of information. Producing paired estimates of bleeding and thrombosis remains a research aim; these estimates would not establish treatment effects for an individual patient. Clinical use would require causal validation and prospective evaluation. TIPS: Transjugular intrahepatic portosystemic shunt; VET: Viscoelastic testing.

However, evidence for multimodal fusion remains indirect. In clear cell renal cell carcinoma, a DeepInsight model combining CT radiomics with gene-expression data predicted 2-year disease-free survival more accurately than conventional ML or single-modality deep learning[42]. This proof-of-concept shows that imaging and molecular features can share one representation; however, it does not establish its clinical utility in cirrhosis.

The likely role of a large language model in such a system is orchestration rather than prediction: Reading the unstructured record, assembling the parameters required by the mechanistic layers, and rendering the output as something that a clinician can interrogate.

Intervention simulation would require explicit control of treatment confounding, temporal ordering of predictors and treatment, and assessment of whether the necessary causal assumptions are credible. Mechanistic plausibility alone does not remove these requirements. Reproducing effects from completed randomized trials could test selected components. Prospective silent evaluation could assess feasibility, calibration, and disagreements with clinical decisions, but agreement with clinicians would not establish treatment benefit. Synthetic simulations could help design studies; clinical recommendations would require prospective evaluation of patient outcomes.

The proposed system is not clinically validated. Its components can be specified and evaluated separately, with implementation contingent on evidence of accuracy and clinical benefit.

An initial study could evaluate prognostic paired-risk prediction without a full digital twin. Data from consecutive adults with cirrhosis and archived VET traces could be linked to prespecified 90-day outcomes: Adjudicated major or clinically relevant non-major bleeding and incident or progressive PVT on imaging reviewed without knowledge of the VET results. Candidate TEG features would include R time, K time, alpha angle, maximum amplitude, and lysis at 30 minutes (LY30), while ROTEM features would include clotting time, clot-formation time, alpha angle, maximum clot firmness, and lysis measures. Death and transplantation should be treated as competing events, with censoring at last verified follow-up. Missing predictors should be handled using multiple imputation within each training resample to avoid information leakage. Penalized regression should be the benchmark, with gradient boosting tested only when there are enough events to support it. Nested cross-validation would estimate optimism. Performance reporting should include the AUROC, area under the precision-recall curve, calibration intercept and slope, Brier score, and decision-curve analysis. Records must identify the analyzer, activator, reagent, and processing details. Raw cutoffs should not be pooled across platforms; a common threshold would first require platform-specific normalization and validation at external sites.

VALIDATION, CALIBRATION, AND GOVERNANCE

Before a model is used in care, it needs independent validation and reliable estimates of absolute risk. Its methods and results should be reported transparently, and its clinical usefulness tested prospectively. Discrimination in the development cohort alone is insufficient.

Further external validation is needed. Discrimination often decreases when a model is tested in a new population: In one prospective rebleeding study, the estimate fell from 0.814 to 0.776[23]. A lower estimate alone does not make the evidence more credible. Appraisal also depends on cohort independence, event count, uncertainty, calibration, handling of missing data, and prespecified analyses. Calibration is reported less often than discrimination, yet a model may rank patients correctly and still misstate their absolute risks. Changes in the cirrhosis population make this a concern: Untreated hepatitis C has declined, while metabolic and alcohol-related diseases have increased. The studies differed in what they reported beyond discrimination. The later Shung EHR model reported a calibrated Brier score of 0.106; both Shung studies used complete-case analyses[18,19]. Meng et al[30] reported calibration and decision curves, with temporal validation limited to one hospital. Xiang et al[24] removed variables with more than 20% missingness and used median or mode imputation for the remaining values. No calibration or decision-curve results were found in the main article inspected. AUROCs should be read with these reporting and design differences in mind.

TRIPOD+AI provides reporting guidance for prediction models using regression or ML[43]. DECIDE-AI applies to early-stage live clinical evaluation of AI-based decision-support systems[44]. These standards should guide reporting. The published roadmap for diagnostic translation in thrombosis and hemostasis also describes the steps between publication of a model and its use as a clinical test[45].

Uncommon events make it difficult to obtain sufficient counts for analysis, especially within clinically relevant subgroups. Working across centers could increase sample size and include a broader mix of patients. When patient-level records cannot be pooled, federated learning is an option, provided data definitions are compatible and governance and external validation are in place. Where resources are limited, routine clinical and laboratory data could form the basis of a benchmark model. Adding VET, thrombin-generation testing, or longitudinal imaging would be justified only if the extra benefit supports the cost and the tests can be made available.

Clinicians may give too much weight to a model estimate, particularly when the evidence is uncertain. This risk of automation bias should be assessed before deployment. Silent evaluation should come before live use, with disagreements between clinicians and the model reviewed without presuming that either is right.

Study objectives should specify in advance who will use the model, the clinical decision it will inform, and the thresholds for action. A tool intended to help weigh treatment choices should show both bleeding and thrombotic outcomes, the uncertainty in each, and the limits of any treatment-effect inference. Reporting only one could hide a clinically relevant competing harm.

Equity also needs assessment. A model trained mainly in HBV-related cirrhosis may not be reliable in women or patients with fragmented records, or in patients with alcohol-related or metabolic dysfunction-associated steatotic liver disease. Missing tests are often a marker of access rather than normal physiology; treating an absent elastography result as normal can systematically underestimate the risk. Performance by sex, etiology, geography, and data completeness should, therefore, appear in the main report.

LIMITATIONS

This narrative review has limitations. The main synthesis was restricted to English-language full-text reports, with related non-English abstracts used only as context, and the original iterative searches did not retain database-level hit counts. Most prediction studies were retrospective, used heterogeneous outcomes, and reported calibration inconsistently. Therefore, Table 2 describes performance in the studied population rather than a property that can be assumed to transfer. HBV-related Asian cohorts are overrepresented, and unsuccessful models are likely to be underpublished. Finally, the digital twin is a proposed architecture, not an implemented or clinically validated system, and Figure 2 presents a conceptual framework whose integrated clinical performance has not been established.

NOVELTY OF THE STUDY

This review makes three linked contributions to the existing literature. It frames bleeding and PVT as potentially coexisting complications of one rebalanced system; it shows that the reviewed liver-disease models rely mainly on portal pressure, imaging, and routine clinical variables rather than direct hemostatic measurements or paired outcomes; and it translates the long-term idea of a hepatic digital twin into four testable layers with an explicit validation problem. The immediate proposal is deliberately smaller: Link archived, platform-annotated VET traces to adjudicated 90-day bleeding and PVT outcomes, test whether they add to routine clinical data, and validate the results at another site.

CONCLUSION

Cirrhosis produces a fragile hemostatic balance. Routine tests incompletely describe this state, and VET-guided protocols can reduce transfusion exposure in selected clinical settings. Thrombin-generation and VET studies also provide preliminary prognostic evidence, but do not establish joint prediction of bleeding and thrombosis. Existing models address heterogeneous outcomes and often have limited external validation or calibration reporting. The focused PubMed update on July 29, 2026 identified no eligible adult cirrhosis model jointly estimating future bleeding and PVT within its specified queries and English-language criteria.

The immediate research priority is to test whether direct hemostatic measurements improve paired prognostic estimates beyond routine clinical data, using prespecified outcomes and independent validation. A digital twin would additionally require validated mechanistic components and a credible causal framework before treatment-specific outputs could inform care. Current evidence supports further investigation of this architecture, rather than clinical deployment.

References
1.  Tripodi A, Mannucci PM. The coagulopathy of chronic liver disease. N Engl J Med. 2011;365:147-156.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 1172]  [Cited by in RCA: 1006]  [Article Influence: 67.1]  [Reference Citation Analysis (10)]
2.  Northup PG, Garcia-Pagan JC, Garcia-Tsao G, Intagliata NM, Superina RA, Roberts LN, Lisman T, Valla DC. Vascular Liver Disorders, Portal Vein Thrombosis, and Procedural Bleeding in Patients With Liver Disease: 2020 Practice Guidance by the American Association for the Study of Liver Diseases. Hepatology. 2021;73:366-413.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 540]  [Cited by in RCA: 511]  [Article Influence: 102.2]  [Reference Citation Analysis (8)]
3.  de Franchis R, Bosch J, Garcia-Tsao G, Reiberger T, Ripoll C; Baveno VII Faculty. Baveno VII - Renewing consensus in portal hypertension. J Hepatol. 2022;76:959-974.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 2244]  [Cited by in RCA: 2353]  [Article Influence: 588.3]  [Reference Citation Analysis (56)]
4.  Hartmann J, Dias JD, Pivalizza EG, Garcia-Tsao G. Thromboelastography-Guided Therapy Enhances Patient Blood Management in Cirrhotic Patients: A Meta-analysis Based on Randomized Controlled Trials. Semin Thromb Hemost. 2023;49:162-172.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 26]  [Cited by in RCA: 24]  [Article Influence: 8.0]  [Reference Citation Analysis (1)]
5.  Nazha A, Elemento O, Ahuja S, Lam B, Miles M, Shouval R, McWeeney S, Sirhan S, Srisuwananukorn A, Haferlach T. Artificial intelligence in hematology. Blood. 2025;146:2283-2292.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 1]  [Cited by in RCA: 22]  [Article Influence: 22.0]  [Reference Citation Analysis (1)]
6.  Kuan YKI, Kok YJ, Liu NSH, Ong BJA, Chee YJ, Xu C, Chow M, Ramanathan K, Dalan R, Ho P, Fan BE. Artificial intelligence in clinical thrombosis and hemostasis: A review. Res Pract Thromb Haemost. 2025;9:102984.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 14]  [Cited by in RCA: 12]  [Article Influence: 12.0]  [Reference Citation Analysis (0)]
7.  Balsano C, Alisi A, Burra P, Calvaruso V, Cammà C, Campagner A, Donatelli P, Germani G, Gerussi A, Giuffrè M, Lleo A, Panebianco V, Persico M, Pugliese N, Rossi S; Artificial Intelligence Committee. Artificial Intelligence in hepatology: A position paper by the Italian Association for the Study of the Liver. Dig Liver Dis. 2026;58:747-758.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in RCA: 1]  [Reference Citation Analysis (2)]
8.  Tripodi A, Primignani M, D'Ambrosio R, Tosetti G, La Mura V, Lampertico P, Peyvandi F. Reappraisal of the conventional hemostasis tests as predictors of perioperative bleeding in the era of rebalanced hemostasis in cirrhosis. Hepatology. 2025;82:755-765.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 8]  [Cited by in RCA: 10]  [Article Influence: 10.0]  [Reference Citation Analysis (0)]
9.  Campello E, Zanetto A, Bulato C, Maggiolo S, Spiezia L, Russo FP, Gavasso S, Mazzeo P, Tormene D, Burra P, Angeli P, Senzolo M, Simioni P. Coagulopathy is not predictive of bleeding in patients with acute decompensation of cirrhosis and acute-on-chronic liver failure. Liver Int. 2021;41:2455-2466.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 53]  [Cited by in RCA: 54]  [Article Influence: 10.8]  [Reference Citation Analysis (1)]
10.  Wan J, Roberts LN, Hendrix W, Konings J, Ow TW, Rabinowich L, Barbouti O, de Laat B, Arya R, Patel VC, Roest M, Lisman T, Bernal W. Whole blood thrombin generation profiles of patients with cirrhosis explored with a near patient assay. J Thromb Haemost. 2020;18:834-843.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 20]  [Cited by in RCA: 32]  [Article Influence: 5.3]  [Reference Citation Analysis (2)]
11.  Janko N, Majeed A, Commins I, Gow P, Kemp W, Roberts SK. Rotational thromboelastometry predicts future bleeding events in patients with cirrhosis. Scand J Gastroenterol. 2024;59:1062-1068.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in RCA: 5]  [Reference Citation Analysis (2)]
12.  Zanetto A, Campello E, Bulato C, Willems R, Konings J, Roest M, Gavasso S, Nuozzi G, Toffanin S, Burra P, Russo FP, Senzolo M, de Laat B, Simioni P. Impaired whole blood thrombin generation is associated with procedure-related bleeding in acutely decompensated cirrhosis. J Hepatol. 2025;82:1023-1035.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 19]  [Cited by in RCA: 22]  [Article Influence: 22.0]  [Reference Citation Analysis (0)]
13.  Rout G, Shalimar, Gunjan D, Mahapatra SJ, Kedia S, Garg PK, Nayak B. Thromboelastography-guided Blood Product Transfusion in Cirrhosis Patients With Variceal Bleeding: A Randomized Controlled Trial. J Clin Gastroenterol. 2020;54:255-262.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 107]  [Cited by in RCA: 106]  [Article Influence: 17.7]  [Reference Citation Analysis (6)]
14.  Kumar M, Ahmad J, Maiwall R, Choudhury A, Bajpai M, Mitra LG, Saluja V, Mohan Agarwal P, Bihari C, Shasthry SM, Jindal A, Bhardwaj A, Kumar G, Sarin SK. Thromboelastography-Guided Blood Component Use in Patients With Cirrhosis With Nonvariceal Bleeding: A Randomized Controlled Trial. Hepatology. 2020;71:235-246.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 157]  [Cited by in RCA: 144]  [Article Influence: 24.0]  [Reference Citation Analysis (3)]
15.  Vuyyuru SK, Singh AD, Gamanagatti SR, Rout G, Gunjan D, Shalimar. A Randomized Control Trial of Thromboelastography-Guided Transfusion in Cirrhosis for High-Risk Invasive Liver-Related Procedures. Dig Dis Sci. 2020;65:2104-2111.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 72]  [Cited by in RCA: 78]  [Article Influence: 13.0]  [Reference Citation Analysis (3)]
16.  Biswas S, Anand A, Vaishnav M, Mehta S, Swaroop S, Aggarwal A, Arora U, Agarwal A, Elhence A, Mahapatra SJ, Agarwal S, Gunjan D, Sehgal T, Aggarwal M, Dhawan R, Gamanagatti S, Shalimar. Thromboelastography-Guided versus Standard-of-Care or On-Demand Platelet Transfusion in Patients with Cirrhosis and Thrombocytopenia Undergoing Procedures: A Randomized Controlled Trial. J Vasc Interv Radiol. 2024;35:1508-1518.e2.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in RCA: 6]  [Reference Citation Analysis (0)]
17.  Kumar R, Ng LXL, Wong YJ, Tan CK, Wang LZ, Qiu TY, Wong B, Lin KW, Li JW, Kwek ABE, Ang TL, Gokhle RS, Sivanath TP. Rotational Thromboelastometry Reduces the Need for Preemptive Transfusion in Cirrhosis: A Randomized Controlled Trial (NCT:05698134). J Clin Exp Hepatol. 2025;15:102409.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 6]  [Cited by in RCA: 11]  [Article Influence: 11.0]  [Reference Citation Analysis (1)]
18.  Shung DL, Au B, Taylor RA, Tay JK, Laursen SB, Stanley AJ, Dalton HR, Ngu J, Schultz M, Laine L. Validation of a Machine Learning Model That Outperforms Clinical Risk Scoring Systems for Upper Gastrointestinal Bleeding. Gastroenterology. 2020;158:160-167.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 188]  [Cited by in RCA: 169]  [Article Influence: 28.2]  [Reference Citation Analysis (7)]
19.  Shung DL, Chan CE, You K, Nakamura S, Saarinen T, Zheng NS, Simonov M, Li DK, Tsay C, Kawamura Y, Shen M, Hsiao A, Sekhon JS, Laine L. Validation of an Electronic Health Record-Based Machine Learning Model Compared With Clinical Risk Scores for Gastrointestinal Bleeding. Gastroenterology. 2024;167:1198-1212.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 24]  [Cited by in RCA: 23]  [Article Influence: 11.5]  [Reference Citation Analysis (0)]
20.  Wan S, Gao Y, Li T, Tian X, Liu M, Wei Y, FengChe, Ren W, Liu M, Hu F, Song B. Deep Learning-Enhanced CTA for Noninvasive Prediction of First Variceal Haemorrhage in Cirrhosis: A Multi-Centre Study. Liver Int. 2025;45:e70274.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 3]  [Cited by in RCA: 3]  [Article Influence: 3.0]  [Reference Citation Analysis (0)]
21.  Liu ZW, Song T, Wang ZH, Sun LL, Zhang S, Yu YZ, Wang WW, Li K, Li T, Hu JH. Predicting Portal Pressure Gradient in Patients with Decompensated Cirrhosis: A Non-invasive Deep Learning Model. Dig Dis Sci. 2024;69:4392-4404.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in RCA: 2]  [Reference Citation Analysis (1)]
22.  Furuichi Y, Nishiguchi R, Furuichi Y, Kobayashi S, Fujiwara T, Sato K. The Role of Artificial Intelligence, Including Endoscopic Diagnosis, in the Prediction of Presence, Bleeding, and Mortality of Esophageal Varices. Dig Endosc. 2025;37:1261-1272.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in RCA: 1]  [Reference Citation Analysis (0)]
23.  Zheng L, Shi N, Li P, Ge H, Tu C, Qu Y, Wang Y, Lin Y, Chen S, Sun D, Weng C, Wu S, Jiang W. Development and validation of machine learning models to predict esophagogastric variceal rebleeding risk in HBV-related cirrhosis after endoscopic treatment: a prospective multicenter study. EClinicalMedicine. 2025;87:103436.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in RCA: 4]  [Reference Citation Analysis (2)]
24.  Xiang Y, Yang N, Zheng TL, Huang YF, Liu TY, Ma DQ, Hu SJ, Zhang WH, Xiang HL, Zhang LY, Yuan LL, Wang X, Dang T, Zhang G, Wu B, Peng LJ, Gao M, Xia DL, Liu ZB, Li J, Song Y, Zhou XQ, Qi XS, Zeng J, Tan XY, Deng MM, Fang HM, Qi SL, He S, He YF, Ye B, Wu W, Shao JB, Wei W, Hu JP, Yong X, He CH, Bao JL, Zhang YN, Ji R, Bo Y, Yan W, Li HJ, Li SL, Geng S, Zhao L, Liu B, Qi XL. Development of a deep learning model for guiding treatment decisions of acute variceal bleeding in patients with cirrhosis. World J Gastroenterol. 2025;31:111361.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in RCA: 3]  [Reference Citation Analysis (4)]
25.  Lian Y, Qiu X, Wang G, Liu N, Zhao J, Song S, Wang S, Sun M. Machine Learning in Predicting the Risk of Esophagogastric Variceal Bleeding Among Patients With Liver Cirrhosis: Systematic Review and Meta-Analysis. J Med Internet Res. 2026;28:e78203.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in RCA: 1]  [Reference Citation Analysis (0)]
26.  Malik S, Tenorio BG, Moond V, Dahiya DS, Vora R, Dbouk N. Systematic review of machine learning models in predicting the risk of bleed/grade of esophageal varices in patients with liver cirrhosis: A comprehensive methodological analysis. J Gastroenterol Hepatol. 2024;39:2043-2059.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in RCA: 14]  [Reference Citation Analysis (4)]
27.  Kildal VV, Dahlberg M, Ek CH, Oldner A, Wikman A, Wahlgren CM, Günther M. Prognostic value of admission ROTEM in trauma: enhancing 30-day all-cause mortality prediction using machine learning. Eur J Trauma Emerg Surg. 2025;51:301.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 3]  [Cited by in RCA: 3]  [Article Influence: 3.0]  [Reference Citation Analysis (0)]
28.  Duranteau O, Popoff B, Abels A, Lucidi V, Savier E, Blanchard F, Martinez T, Loi P, Germanova D, Demulder A, Creteur J, Tuna T. Prediction of biological evolution following blood product transfusion during liver transplantation: the contribution of machine learning to decision-making. BMJ Health Care Inform. 2025;32:e101466.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 2]  [Cited by in RCA: 2]  [Article Influence: 2.0]  [Reference Citation Analysis (1)]
29.  Li Y, Gao J, Zheng X, Nie G, Qin J, Wang H, He T, Wheelock Å, Li CX, Cheng L, Li X. Diagnostic Prediction of portal vein thrombosis in chronic cirrhosis patients using data-driven precision medicine model. Brief Bioinform. 2023;25:bbad478.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 12]  [Cited by in RCA: 11]  [Article Influence: 3.7]  [Reference Citation Analysis (2)]
30.  Meng PP, Xiong FX, Chen JL, Zhou Y, Liu XL, Ji XM, Jiang YY, Hou YX. Establish and validate an artificial neural networks model used for predicting portal vein thrombosis risk in hepatitis B-related cirrhosis patients. World J Hepatol. 2025;17:97767.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Reference Citation Analysis (3)]
31.  Ma J, Chalasani NP, Schwantes-An L, Björnsson ES. Review article: the safety of anticoagulants and antiplatelet agents in patients with cirrhosis. Aliment Pharmacol Ther. 2023;57:52-71.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 1]  [Cited by in RCA: 9]  [Article Influence: 3.0]  [Reference Citation Analysis (0)]
32.  Boccatonda A, Brighenti A, Simion C, Napolitano A, Serra C, Simioni P, Campello E. Direct oral anticoagulants in portal vein thrombosis in patients with cirrhosis: a systematic review and meta-analysis. Blood Adv. 2026;10:4399-4411.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 1]  [Cited by in RCA: 1]  [Article Influence: 1.0]  [Reference Citation Analysis (0)]
33.  Nishiba H, Nagamine A, Yashima H, Takahashi Y, Higuchi Y, Sekizaki N, Nakamura H, Araki T, Takama N, Koitabashi N, Nakajima T, Kaneko Y, Ohyama Y, Yokoyama T, Imai K, Kurabayashi M, Yamamoto K, Obayashi K. Effect of genetic factors on the interindividual variability of warfarin dosage requirements in Japanese patients after adjusting for renal function. Pharmazie. 2024;79:173-177.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in RCA: 1]  [Reference Citation Analysis (0)]
34.  Shen Y, Zhao ZB, Li X, Chen L, Yuan H. [Risk factors and construction of a nomogram model for cirrhotic portal vein thrombosis combined with esophagogastric variceal bleeding]. Zhonghua Gan Zang Bing Za Zhi. 2023;31:1035-1042.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Reference Citation Analysis (0)]
35.  Shohat N, Ludwick L, Sherman MB, Fillingham Y, Parvizi J. Using machine learning to predict venous thromboembolism and major bleeding events following total joint arthroplasty. Sci Rep. 2023;13:2197.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in RCA: 38]  [Reference Citation Analysis (0)]
36.  Lin C, Huang Y. Portal vein thrombosis and Esophageal-Gastric variceal bleeding in cirrhosis: shared risk factors and causal relationship. BMC Gastroenterol. 2025;25:500.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 4]  [Cited by in RCA: 4]  [Article Influence: 4.0]  [Reference Citation Analysis (0)]
37.  Lou SS, Kumar S, Goss CW, Avidan MS, Kheterpal S, Kannampallil T; Multicenter Perioperative Outcomes Group. Multicenter Validation of a Machine Learning Model for Surgical Transfusion Risk at 45 US Hospitals. JAMA Netw Open. 2025;8:e2517760.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 4]  [Cited by in RCA: 4]  [Article Influence: 4.0]  [Reference Citation Analysis (0)]
38.  Link KG, Stobb MT, Monroe DM, Fogelson AL, Neeves KB, Sindi SS, Leiderman K. Computationally Driven Discovery in Coagulation. Arterioscler Thromb Vasc Biol. 2021;41:79-86.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 1]  [Cited by in RCA: 11]  [Article Influence: 1.8]  [Reference Citation Analysis (0)]
39.  Bannoud MA, Martins TD, Montalvão SAL, Annichino-Bizzacchi JM, Filho RM, Maciel MRW. Artificial intelligence in computational modeling of thrombosis: Bridging mechanistic insights and clinical translation. J Thromb Thrombolysis. 2026;59:727-745.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 2]  [Cited by in RCA: 1]  [Article Influence: 1.0]  [Reference Citation Analysis (0)]
40.  Thangaraj PM, Benson SH, Oikonomou EK, Asselbergs FW, Khera R. Cardiovascular care with digital twin technology in the era of generative artificial intelligence. Eur Heart J. 2024;45:4808-4821.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 38]  [Cited by in RCA: 84]  [Article Influence: 42.0]  [Reference Citation Analysis (0)]
41.  Moor M, Banerjee O, Abad ZSH, Krumholz HM, Leskovec J, Topol EJ, Rajpurkar P. Foundation models for generalist medical artificial intelligence. Nature. 2023;616:259-265.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 1358]  [Cited by in RCA: 977]  [Article Influence: 325.7]  [Reference Citation Analysis (4)]
42.  Kawahara D, Kishi M, Kadooka Y, Hirose K, Murakami Y. Integrating radiomics and gene expression by mapping on the image with improved DeepInsight for clear cell renal cell carcinoma. Cancer Genet. 2025;292-293:100-105.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 14]  [Cited by in RCA: 12]  [Article Influence: 12.0]  [Reference Citation Analysis (0)]
43.  Collins GS, Moons KGM, Dhiman P, Riley RD, Beam AL, Van Calster B, Ghassemi M, Liu X, Reitsma JB, van Smeden M, Boulesteix AL, Camaradou JC, Celi LA, Denaxas S, Denniston AK, Glocker B, Golub RM, Harvey H, Heinze G, Hoffman MM, Kengne AP, Lam E, Lee N, Loder EW, Maier-Hein L, Mateen BA, McCradden MD, Oakden-Rayner L, Ordish J, Parnell R, Rose S, Singh K, Wynants L, Logullo P. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 1587]  [Cited by in RCA: 2654]  [Article Influence: 1327.0]  [Reference Citation Analysis (17)]
44.  Vasey B, Nagendran M, Campbell B, Clifton DA, Collins GS, Denaxas S, Denniston AK, Faes L, Geerts B, Ibrahim M, Liu X, Mateen BA, Mathur P, McCradden MD, Morgan L, Ordish J, Rogers C, Saria S, Ting DSW, Watkinson P, Weber W, Wheatstone P, McCulloch P; DECIDE-AI expert group. Reporting guideline for the early-stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. Nat Med. 2022;28:924-933.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 372]  [Cited by in RCA: 500]  [Article Influence: 125.0]  [Reference Citation Analysis (8)]
45.  Nagler M, Nilius H, Hastings J, Nakas C. Making Machine Learning Clinically Useful in Thrombosis and Hemostasis: A Roadmap for Diagnostic Translation. Semin Thromb Hemost.  2026.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in RCA: 1]  [Reference Citation Analysis (0)]
Footnotes

Peer review: Externally peer reviewed.

Peer-review model: Single blind

Specialty type: Gastroenterology and hepatology

Country of origin: Egypt

Peer-review report’s classification

Scientific quality: Grade A, Grade A, Grade B, Grade B

Novelty: Grade A, Grade A, Grade B, Grade B

Creativity or innovation: Grade A, Grade A, Grade B, Grade B

Scientific significance: Grade A, Grade A, Grade B, Grade B

P-Reviewer: Fan XC, MD, PharmD, PhD, Post Doctoral Researcher, Postdoc, Postdoctoral Fellow, Research Assistant Professor, China; Gugulothu D, Academic Fellow, Assistant Professor, PhD, India S-Editor: Liu JH L-Editor: A P-Editor: Zhao YQ

Write to the Help Desk