Digital twins in gastroenterology: From computational modeling to precision medicine and clinical translation

Abstract

Digital twin technology, an emerging paradigm in medicine, has the potential for creating dynamic virtual models that integrate multi-modal data to simulate disease trajectories and therapeutic responses of patients. A gastrointestinal (GI) digital twin combines data from electronic health records, imaging, microbiomes and real-time physiological inputs into computational frameworks that include mechanistic modelling, machine learning and hybrid approaches. These systems ensure bi-directional data flow, which allows continuous recalibration, enabling in-silico testing of therapeutic interventions before their implementation in the real-world. In this review, we summarize the current evidence for the architecture, methodologies, and clinical aspects of digital twins in various GI diseases. These advances are promising, but there are still several limitations, such as the integration of data, poor validation across diverse populations, computational demands and ethical concerns around privacy and bias. Most existing models have been limited to research environments. Future directions include the integration of wearable data, the development of multi-organ digital twins, and the incorporation of large language models for data harmonization and the implementation of federated learning frameworks. Digital twins hold immense potential to revolutionize gastroenterology through predictive, personalized, and simulation-guided clinical decision making. However, rigorous validation, standardization and clinical integration need to be ensured before their widespread adoption.

Key Words: Digital twin; Virtual patient; Personalized medicine; Computational modeling; Systems biology; In silico modeling; Artificial intelligence; Machine learning; Predictive modeling

Core Tip: Digital twins are a transformative shift in the field of gastroenterology as they allow dynamic, patient-specific virtual models that combine multimodal clinical, imaging and multi-omics data to mimic disease evolution and response to therapy. Digital twins provide a pathway for truly personalized and predictive care, as they can be updated in real-time and used for in silico testing of intervention strategies, unlike traditional predictive models. Emerging applications in hepatology, gastrointestinal oncology, inflammatory bowel disease, motility disorders and microbiome science underscore their potential. However, challenges related to data integration, validation, scalability and ethical governance need to be addressed before clinical implementation.

INTRODUCTION

Digital twins refer to a virtual model of a real-world entity, such as a patient or an organ, characterized by its ability to integrate continuously updated data and use this information to predict future outcomes or behavior[1]. The concept of a digital twin was first described by Grieves[2] in 2002 in the context of product lifecycle management in aerospace engineering. A digital twin originally consisted of three inseparable components: A physical entity in the real world, a virtual replica in digital space, and a two-way data link that maintains the two in sync all the time. Data flows from the physical object to update its virtual twin, and insights and predictions made by the virtual model flow back to inform decisions on the physical system[2]. Unlike the “one-size-fits-all” approach in medicine, a digital twin can incorporate a biologically realistic disease model that enables care to be tailored to an individual patient’s data. It can estimate the risk of disease and predict how a patient might respond to different interventions. Each follow-up visit refines the digital twin with newly available patient information[1].

This distinction from conventional predictive models is important and deserves emphasis because it is frequently elided in the literature. A traditional prognostic model, even a sophisticated machine-learning classifier, is trained on population-level data, generates a risk score at a single time point, and does not update itself as the patient’s condition evolves. In contrast, digital twins are inherently personalized and longitudinal. It is continuously recalibrated as new patient data arrive; its predictions are therefore not population averages but estimates that reflect the individual’s evolving biological state. Crucially, the twin also enables in silico intervention testing: The clinician can interrogate the virtual patient with a proposed treatment and observe predicted outcomes before committing to any real-world action[3].

The gastrointestinal (GI) tract is complex, integrating mechanical, chemical, immunological, and microbial processes across multiple anatomically and functionally distinct segments, from the enteric nervous system-governed peristalsis of the esophagus to the dense microbial ecology of the colon. The gut microbiome alone, comprising trillions of organisms whose collective genome dwarfs that of the human host, exerts system-wide effects on immune tone, metabolic homeostasis, and drug pharmacokinetics. Motility, mucosal immunity, the gut-brain axis, and host genomics all intersect, creating a system whose behavior cannot be adequately captured by any single biomarker or model[4,5].

The artificial pancreas represents one of the earliest digital twin-like clinical systems, pairing continuous glucose monitoring with automated insulin infusion. Early interstitial glucose sensors, less invasive alternatives to finger-prick blood sampling, required calibration against true blood glucose readings and were prone to noise and sensitivity loss. To address this, signal processing algorithms were integrated into “smart continuous glucose monitoring” systems, combining noise reduction, accuracy improvement, and short-term glucose prediction modules, enabling preventive alerts and real-time therapy adjustment before hypo- or hyperglycemic events occur[6]. The digital twin methodology is more mature in the fields of cardiology and oncology. In cardiology, personalized heart digital twins derived from late gadolinium enhancement magnetic resonance imaging (MRI) have been used to guide ablation procedures for ventricular tachycardia and atrial fibrillation, with in silico ablation targets co-localizing with clinical lesions in prospective studies[7]. In oncology, MRI-based digital twins calibrated to patient-specific imaging data predicted pathological complete response to neoadjuvant chemotherapy in triple-negative breast cancer with an area under the curve (AUC) of 0.82, and patient-specifically optimized treatment schedules yielded a 20%-25% improvement in complete response rates[8]. Gastroenterology, with its extraordinary biological complexity and chronic disease burden, constitutes the logical next frontier.

Literature search and study classification

A comprehensive non-systematic literature search was performed on PubMed, MEDLINE, and Google Scholar from inception to May 2026. Search terms included “digital twin”, “virtual patient”, “in silico model”, “computational gastroenterology”, and disease-specific terms [inflammatory bowel disease (IBD), hepatology, motility, pancreatic, microbiome]. To account for the heterogeneity of the field, the cited systems are explicitly categorized according to the extent to which they satisfy the three criteria defining a true digital twin: (1) Individualized parameterization; (2) Dynamic bidirectional updating; and (3) In silico intervention. Systems that satisfy all these three criteria are called digital twins; those that satisfy one or two are called computational models or digital twin-like frameworks. This classification scheme is applied uniformly to all disease sections.

Architecture of a GI digital twin

A GI digital twin is created by integrating heterogeneous, continuously updated patient data streams into a single computational model that operates across three interacting layers: A data layer, a model layer, and a simulation layer. Each layer draws on distinct input types and modeling approaches, and together they enable the system’s defining capability by creating a bidirectional link between the virtual model and the real patient[9].

The data layer aggregates multimodal inputs that collectively capture the full biological state of the patient. Electronic health records (EHR) provide longitudinal clinical context, including diagnoses, laboratory results, pathology reports, and treatment history. Medical imaging from ultrasound, computed tomography (CT), MRI, and endoscopy contributes structural and morphological information[10]. Multi-omics data, including host genomics, transcriptomics, proteomics, and gut microbiome sequencing, enrich the model with mechanistic molecular detail and enable representation of individual disease drivers that aggregate biomarkers cannot resolve. Physiological sensors and wearables such as wireless pH and pressure capsules provide real-time dynamic inputs, thus transforming them from a static snapshot into a living representation of the patient’s evolving state[10,11].

The predictive quality of a GI digital twin is critically dependent on the quality of molecular data used to parameterize its computational engines. Structural biology approaches that determine the biophysical behavior of therapeutic agents and nucleic acids in living cells under near-physiological conditions, such as in-cell nuclear magnetic resonance spectroscopy, may provide the type of high-resolution mechanistic data inputs that would set a digital twin apart from a coarser statistical model[12]. Similarly, the biophysical and transcriptomic characterization of molecular interactions within the gut microenvironment, including characterization of antimicrobial peptides targeting drug-resistant pathogens, further illustrates that data resolution is required to simulate these specific therapeutic responses against the background of biological noise in a compositionally complex GI ecosystem[13].

The model layer houses the computational engines that translate these inputs into biological understanding through multiple complementary approaches. Physics-based modeling expresses physiological processes through mathematical equations grounded in fundamental physical principles[14]. When applied to patient-specific GI geometries reconstructed from cine-MRI or CT imaging, computational fluid dynamics (CFD) enables simulation of luminal flow, mixing patterns, and transit behavior. These simulations offer mechanistic insight into how various structural variations, such as strictures or surgically altered anatomy, can influence nutrient absorption and drug bioavailability, providing an understanding that cannot be obtained easily through in vivo or in vitro methods alone. Mechanistic models address biological dynamics at the molecular and cellular scale through systems of ordinary differential equations (ODEs). A widely employed ODE-based model in IBD utilized 31 equations to simulate immune cell and cytokine interactions across gut and systemic compartments, incorporating downstream biomarkers such as C-reactive protein and fecal calprotectin[10]. Similarly, another ODE-based framework modeling interactions among Th1, Th17, and regulatory T-cell populations, along with key cytokines including tumor necrosis factor (TNF)-alpha, interleukin (IL)-6, and IL-10, demonstrated that therapeutic responses to the same medication can vary substantially depending on an individual patient’s underlying immune network configuration[11].

A significant roadblock in the creation of GI digital twins is multi-scale coupling, which involves the mathematical translation of molecular-level events into physiological shifts at the organ level. This requires a hybrid architecture for the integration of ODE governing cytokine or receptor dynamics to partial differential equations and CFD models that govern luminal flow and biomechanics using common state variables and consistent parameterization across scales[15]. A specifically promising approach is physics-informed neural networks, where physical constraints are incorporated directly into the network loss function to ensure mechanistic continuity between the molecular and organ layers. In the absence of an explicit coupling strategy, it is impossible to know if a presented model is a real multi-scale digital twin or a set of validated sub-models validated separately. Future publications should detail the coupling methodology sufficiently to allow computational reproducibility[16]. Despite their explanatory strength, these mechanistic models remain limited mainly due to their reliance on accurately defined biological parameters, many of which are still incompletely characterized for GI systems and may differ significantly between individuals[16].

The simulation layer is where the architecture’s clinical value is realized, and it is the bidirectional data loop comprising an afferent and efferent pathway[2,9]. In the afferent pathway, newly acquired inputs, including clinical assessments, laboratory findings, and data from wearable technologies, are incorporated to recalibrate the model’s internal state. Data assimilation strategies such as ensemble Kalman filtering, a class of Bayesian updating methods commonly applied in engineering systems, have been proposed to translate observable clinical variables into updates of underlying model parameters within digital twin architectures[17]. In the efferent pathway, model-generated simulations are returned to the clinical environment to support medical decision-making. These outputs may identify emerging disease trajectories such as progressive fibrosis, predicted non-response to biologic therapies, or risk of post-operative motility complications, and are revised iteratively as additional patient data become available rather than relying solely on historical information. Thus, the hallmark characteristic that distinguishes a digital twin from conventional static models is the presence of a continuously functioning bidirectional interaction between the patient and the virtual representation (Table 1).

Table 1 Traditional predictive model vs digital twin.

Feature	Traditional predictive model	Gastriontestinal digital twin
Data basis	Population cohort (group-level coefficients)	Individual patient (multimodal, longitudinal)
Temporal score	Single time point; static after deployment	Continuous; updated with each new data input
Update mechism	None (model frozen post-training)	Data assimilation (e.g., ensemble Kalman filter, Bayesian methods)
Intervention testing	Not possible	In silico simulation of drug, dose, or procedure before clinical use
Model architecture	Statistical/data-driven (regression, ML)	Mechanistic, data-driven, or hybrid (physics-informed neural networks)
Personalization	Risk score adjusted by a few covariates	Full virtual replica of individual patient physiology
GI example	Biologic response score (single clinic visit)	IBD immune twin; gastric motility CFD model; microbiome MCMM
Key limitation	Cannot adapt to evolving disease; population averages mask individual variation	Data requirements high; computational infrastructure; validation gap

The bidirectional data loop (afferent/efferent) is the defining architectural feature that distinguishes a digital twin from any static predictive model. CFD: Computational fluid dynamics; MCMM: Microbial community-scale metabolic model; IBD: Inflammatory bowel disease; ML: Machine learning; GI: Gastrointestinal.

Perhaps the most transformative capability of digital twins lies in silico intervention testing, which allows clinicians to explore hypothetical treatment strategies within the virtual model rather than on the patient. By enabling prospective simulation of therapeutic options and generating probabilistic predictions of outcomes without real-world exposure to risk, this approach represents a fundamental transition from empirical trial-based care toward hypothesis-driven and individualized clinical decision-making (Figure 1 and Table 2)[18].

Figure 1 Gastrointestinal digital twin architecture. Architecture of a gastrointestinal digital twin. The gastrointestinal (GI) digital twin operates across three interoperating layers. The data layer aggregates and harmonizes multi-modal patient streams, electronic health record, imaging, multi-omics, microbiome sequencing, and physiological sensors. The model layer houses mechanistic, data-driven, and hybrid physics-informed neural network models of GI physiology; hybrid architectures are considered the current gold standard. The simulation layer executes forward modeling to generate disease trajectory projections, therapy response probabilities, in silico intervention scenarios, and clinical risk alerts. A bidirectional link; afferent and efferent defines the system as a true digital twin rather than a static predictive model. EHR: Electronic health record; GI: Gastrointestinal; CT: Computed tomography; MRI: Magnetic resonance imaging; ODE: Ordinary differential equation; CFD: Computational fluid dynamics; PDE: Partial differential equation; CNN: Convolutional neural network; CGM: Continuous glucose monitoring; GNN: Graph neural network; NLP: Natural language processing; IBD: Inflammatory bowel disease.

Table 2 Summary of digital twin applications across gastrointestinal regions.

GI region	Representative models/tools	Key simulation parameters	Target conditions	Potential clinical applications	Current restrictions
Stomach	StomachSim; CFD-based gastric models; dynamic MRI-coupled simulations	Antral contractions. Gastric mixing. Emptying dynamics. Intraluminal pressure. Retropulsive jets. Drug dissolution kinetics. Body posture effects	Gastroparesis; functional dyspepsia	Pre-procedural planning for pyloroplasty and sleeve gastrectomy; pharmacologic efficacy testing; dietary optimization; identification of subtle flow abnormalities	Simplified motility assumptions. Small experimental datasets. Limited patient-specific validation
Small intestine	Peristaltic flow models; CFD luminal transit simulations	Luminal flow velocities. Contractile activity. Transit dynamics. Absorption modeling	Functional bowel disorders; malabsorption states	Nutrient absorption assessment; motility characterization; drug delivery simulation	Early feasibility stage. Complex motility patterns difficult to replicate. Changing luminal content variability
Colon	Digital replicas of experimental colon models; peristaltic fluid dynamics models	Peristaltic fluid movement. Mixing patterns. Flow velocities. Luminal shear stress	Constipation; diarrheal states; colonic dysmotility	Colonic motility disorder characterization; therapeutic intervention planning; microbiome-motility interaction modeling	Validated primarily against experimental models. Minimal patient-specific clinical data. Simplified boundary conditions
Esophagus	CFD bolus transit models; sphincter biomechanical simulations	Bolus transit mechanics. Sphincter function. Intraluminal pressure gradients. Peristaltic wave propagation	Achalasia; spastic motility syndromes; gastroesophageal reflux	Diagnostic support for dysmotility; pre-interventional planning for myotomy or dilation; assessment of sphincter competence	Largely conceptual/early development. Very limited clinical validation data. Complex tissue mechanics not fully modeled

Overview of current computational modeling approaches, simulated parameters, targeted conditions, potential clinical uses, and key limitations for digital twin research across major gastrointestinal organs. CFD: Computational fluid dynamics; GI: Gastrointestinal; MRI: Magnetic resonance imaging.

DIGITAL TWINS IN LIVER DISEASE

Metabolic dysfunction-associated steatotic liver disease/non-alcoholic fatty liver disease: Predicting progression and personalizing therapy

The increasing global prevalence of metabolic dysfunction-associated steatotic liver disease (MASLD) makes the development of precision medicine strategies to predict disease progression and personalize therapy a clinical priority[19]. It is important to distinguish between the two types of evidence discussed in this section from the start: AI-based precision medicine tools informing MASLD care and true digital twin frameworks that meet the bidirectional updating criterion. The former are enabling technologies that constitute essential precursors to digital twins but are not themselves digital twins; the latter are still in an early stage of development. The course of MASLD/non-alcoholic fatty liver disease differs greatly between patients, which complicates the choice of treatment; thus, only certain subgroups benefit, potentially leading to high costs and risks with limited efficacy[20,21].

Artificial intelligence (AI)-based models have continued to improve fibrosis evaluation through multimodal approaches. Deep learning, digital pathology, and imaging-based models can provide accurate, reproducible, and non-invasive staging[22]. The concept of digital twins leverages these advances and represents an obvious next step toward actionable precision medicine by integrating these data streams into dynamic, patient-specific simulations. Based on this, the Whole Body Digital Twin platform (Twin Health) used AI-predicted glycemic responses from continuous glucose monitors and wearable sensors to give personalized nutrition, activity, and sleep recommendations and showed significant improvements in HbA1c, diabetes remission, and hepatic steatosis and fibrosis scores in a multicenter randomized controlled trial[23]. This platform is best described as a digital twin-like precision nutrition platform, as it personalizes recommendations based on real-time physiological inputs but lacks a mechanistic organ-level model of hepatic pathophysiology that can perform in silico drug testing.

The only currently published framework that approaches a true MASLD digital twin is the integrated patient digital twin (PDT)-patient biomimetic twin (PBT) platform described by Miedel et al[21] The framework utilizes longitudinal, multi-modal patient data (clinomics, metabolomics, lipidomics, and genomics) analyzed using causal machine learning models (XGBoost, Random Forest, Supervised Varimax) to generate individualized PDTs and integrates these with patient-specific iPSC-derived liver organoids assembled in the liver acinus micro physiological system platform with hepatocytes, liver sinusoidal endothelial cells, stellate cells, and Kupffer cells. The PDT generates predictions that are then iteratively tested in the matched PBT, and the PBT results feed back to improve the PDT, thus creating a real bi-directional loop. Proof-of-concept experiments with PBTs derived from patient iPSCs recapitulated genotype-dependent MASLD phenotypes and a dose-dependent response to resmetirom consistent with Phase III trial data, validating the platform’s translational potential. However, this framework remains pre-clinical and has not yet been deployed in a clinical decision-making context, underscoring that true MASLD digital twins, while biologically promising, have not yet reached the bedside[22,23].

Liver regeneration and drug-induced injury

Liver regeneration is a complex process with the involvement of different cell types and factors. Despite much research, the understanding is qualitative. Digital twin and multiscale biology models can simulate hepatocyte immune interactions, drug mechanisms, injury propagation, and regeneration to produce testable hypotheses and guide experimental strategies. A realistic digital twin can capture the detailed structure of liver tissue, the interaction of cells and blood flow to simulate how a drug behaves, and to evaluate safety[24,25].

One of the most advanced digital twin models for liver regeneration was developed by Zhao et al[24] They developed a three-dimensional computational model of acetaminophen-induced liver injury and regeneration in mice. The model accounted for intercellular communication of six types of liver cells (hepatocytes, Kupffer cells, and hepatic stellate cells) during injury and healing via physical forces and chemical signaling molecules. What makes their work rigorous is that the authors distinguish between a validated digital twin (one where outputs do match experimental data) and a digital twin candidate (a model that implements a set of hypothesized mechanisms). This framework allows scientists to perform virtual experiments to see which types of interactions between cells are required for efficient liver regeneration and generate testable hypotheses to guide further laboratory research.

Hepatotoxicity is one of the leading causes of acute liver failure in Western populations. It also presents high mortality and a frequent need for liver transplantation. In this context, Virtual Hepatic Lobule was developed as a liver digital twin. It works on multiple scales, incorporates blood flow dynamics and an acetaminophen-induced injury model to predict hepatocyte injury in specific patients. Metabolic zonation incorporation leads to predictions consistent with clinical observations of zonal hepatotoxicity[26]. This is a positive sign that strengthens the belief that further research will enhance the predictive accuracy and translational utility of liver digital twins[21,27]. Both models represent important advances but carry significant limitations, such as the Zhao et al[24] model, which was validated exclusively in mice, while the Virtual Hepatic Lobule was tested on only four patients, thus, still not ready for a clinically deployed model.

Hepatocellular carcinoma: Treatment planning

Hepatocellular carcinoma (HCC) is one of the most common and lethal malignancies worldwide. Intra-arterial therapies such as trans arterial chemoembolization, selective internal radiation therapy (SIRT), and Yttrium-90 radioembolization are established locoregional treatments that exploit the preferential arterial blood supply of the liver tumors to deliver targeted therapy. This is done while sparing the surrounding parenchyma[28].

Current models that are the closest to digital twins are best classified as patient-specific simulation platforms. One such model was created by Cutrì et al[29], which was the most methodologically rigorous example. They constructed a patient-specific CFD model of the hepatic arterial vasculature from cone-beam CT imaging and real-time intraoperative blood pressure measurements in a patient with multifocal HCC who was being treated with Thera sphere SIRT. Using a systematic optimization strategy, the model demonstrated that the longitudinal position of the catheter is the dominant determinant of tumor targeting, with optimal configurations achieving approximately 80% microsphere delivery to the target lesion while preserving the surrounding non-tumor tissue[29]. Similarly, Bomberna et al[28] demonstrated that computational simulations, validated by experimental data, can play an important role in predicting the distribution of particles in the liver during trans arterial delivery, highlighting the importance of catheter position and injection parameters for treatment outcomes.

These advancements have the potential to reduce operator-dependent variability and enable more consistent, optimized treatment delivery across different operators and institutions. While the study represents significant methodological advances, clinical validation is described as limited, highlighting the need for prospective clinical studies to validate the model’s predictions against actual treatment outcomes and patient responses[7].

Liver transplantation: Donor assessment to graft surveillance

Digital twin technology is evolving constantly, with emerging applications already demonstrating clinical utility and a lot of promising uses. In transplant medicine, this now spans the entire care continuum from donor assessment to organ allocation and transport, with additional postoperative optimization and surveillance. Given the time-sensitive nature of transplantation, even marginal improvements in predictive accuracy or dosing precision can significantly improve patient outcomes, graft longevity, and healthcare resource utilization[30].

Among the most technically sophisticated applications is the donor-specific digital twin developed for living donor liver transplantation (LDLT). Halder et al[31] introduced the Personalized Progressive Mechanistic Digital Twin (PePMDT), a hybrid framework that integrates whole-transcriptomic RNA sequencing data from blood samples of LDLT donors with a mechanistic mathematical model of hepatocyte state transitions, spanning quiescent, primed, and replicative phases, to generate donor-specific predictions of liver regeneration trajectories. It was trained on transcriptomic data from twelve LDLT donors over one year, identifying 5876 liver regeneration-specific genes organized into three functional phases-early inflammation, hepatocyte proliferation, and extra-cellular matrix remodeling with strong predictive accuracy even from pre-surgical data (Pearson r ≥ 0.87, mean squared error < 0.15). Unlike conventional black-box models, it embeds a physiologically grounded latent space that enables biologically interpretable predictions and in silico simulation of how varying resection volumes or molecular profiles influence individual recovery trajectories, directly supporting preoperative planning and donor safety decisions[32]. Since this was trained and validated on data from only twelve living donors, there is limited data on generalizability. PePMDT is therefore best classified as a high-fidelity proof-of-concept digital twin at an early clinical development stage[31,33].

In addition to regeneration modelling, early data suggest that machine learning models such as gradient-boosted tree algorithms (XGBoost) can accurately predict hepatic decompensation using routine clinical data from outpatient visits and outperform traditional prognostic scores such as the model for end-stage liver disease, supporting early risk stratification[32]. These tools are key enabling technologies that build the high-quality longitudinal data infrastructure on which future bidirectional liver transplantation digital twins will be built, even though they are not digital twins themselves.

DIGITAL TWIN IN GI ONCOLOGY

To enable a clear evaluation of digital twin applications in GI oncology, it is important to differentiate between the various types of models that are often considered together in the literature: True mechanistic digital twins that continuously update with patient data and allow in silico treatment testing, quantitative systems pharmacology virtual patient models, population-based mathematical simulations calibrated to epidemiological data, and statistical or imaging-based predictive models[28,29,34]. Failure to differentiate between these approaches can lead to an overestimation of their clinical preparedness and obscure where real translational progress has been made.

In colorectal cancer, Ruchlin et al[34] proposed an encoder-decoder neural network framework that used multiomics data of 2100 patients to create latent patient representations that were predictive of overall survival and could simulate counterfactual treatment scenarios. Asghar et al[35] also applied an in-silico trial framework for solid tumors, including colorectal cancer, to predict therapeutic response. Both are methodologically innovative early-stage digital twin candidates, but both are conference or supplementary abstracts without full peer-reviewed methodology, and neither has been independently prospectively validated.

In pancreatic adenocarcinoma, Osipov et al[33] developed a Molecular Twin platform that integrates multi-omic genomic and transcriptomic signatures with clinical outcomes, demonstrating the feasibility of predicting therapeutic responses and survival trajectories in a retrospective proof-of-concept analysis using trial datasets. Similarly, in advanced gastric cancer, Prunella et al[36] used a pharmacometric digital twin framework to simulate adaptive scheduling of combination chemotherapy to optimize timing of treatment while minimizing toxicity, one of the few mechanistically grounded, patient-informed digital twin models in upper GI oncology. However, promising developments notwithstanding, prospective clinical validation remains limited.

Most of the current methods for rectal cancer are imaging-based machine learning models and not true digital twins. Although MRI-based deep learning systems that predict pathological complete response and longitudinal imaging models that assess treatment response are clinically useful, they are mainly statistical predictive tools that lack continuous bidirectional updating or mechanistic recalibration[31,32]. Similarly, 3D surgical planning reconstructions are static digital models, not dynamic patient-specific twins[37,38]. At the population level, Lin et al[39] applied a digital twin simulation approach using faecal immunochemical test data to estimate overdiagnosis rates in colorectal cancer screening, supporting the safety of population-based screening programs.

In HCC, quantitative systems pharmacology (QSP) models represent an important but distinct category from true digital twins. Sové et al[40] used a QSP model of 5000 virtual patients based on CheckMate 040 data to evaluate anti-PD-1 and anti-CTLA-4 therapy and identify immune biomarkers associated with response. Similarly, Huang et al[41] applied a QSP framework to camrelizumab plus apatinib, identifying predictive immune cell ratio biomarkers and suggesting comparable efficacy with lower dose apatinib. While these population-level simulations are crucial for trial design and generating hypotheses, they do not function as individual PDT. Other computational tumor growth models have also explored antiangiogenic dose optimization and metronomic chemotherapy scheduling, although these remain largely preclinical frameworks[42].

Despite these advancements, there are significant challenges that still remain. The present digital twin models are dependent on limited or synthetic datasets. This raises concerns regarding their generalizability, bias, and reliability, while validation frameworks are poorly standardized[43]. Furthermore, many models lack interpretability and may produce inconsistent predictions when applied to heterogeneous real-world populations. Additionally, most applications are still in preclinical or early translational stages, underscoring the need for rigorous prospective validation and regulatory standardization before widespread clinical adoption[44].

DIGITAL TWINS IN MOTILITY AND FUNCTIONAL GI DISORDERS

Functional GI and motility disorders, including gastroparesis, functional dyspepsia, and irritable bowel syndrome, are becoming highly prevalent yet mechanistically complex conditions, for which conventional diagnostics provide only limited, static insights into dynamic GI physiology[45]. Traditional diagnostic tools such as gastric emptying scintigraphy, manometry, and endoscopy provide only limited snapshots of GI physiology.

However, computational modeling has emerged as a promising strategy to overcome these limitations. Most computational models of GI motility currently qualify as digital twin candidates or mechanistic precursors rather than purely clinically deployed digital twins. In recent years, in-silico simulations based on CFD and biomechanical modeling have been developed to reproduce the physiological mechanisms that control GI motility and provide a foundation for constructing digital twins of GI organs capable of simulating physiological and pathological motility patterns in a patient-specific manner[9,46].

One of the most advanced applications of digital twin technology in gastroenterology revolves around computational models of the stomach. StomachSim is a computational model that integrates anatomical geometry derived from imaging with biomechanical representations of gastric motility. Using CFD techniques, the model simulates the transport and mixing of gastric contents under physiological conditions[46,47]. These models aim to replicate gastric biomechanics, including antral contractions, gastric mixing, and emptying dynamics. Simulation outputs have demonstrated strong agreement with experimentally measured gastric emptying and pressure patterns, supporting the validity of the modeling approach. Such models allow investigators to examine how alterations in gastric motility influence digestive processes[46,48]. Simulations involving reduced antral contractions can reproduce features seen in gastroparesis, including impaired grinding of food particles and delayed gastric emptying. Computational studies have shown that reductions in contractile amplitude significantly alter intragastric flow patterns and mixing efficiency. Beyond disease modeling, gastric digital twins can also simulate the effects of therapeutic interventions. Computational studies have explored how procedures such as pyloroplasty or sleeve gastrectomy alter gastric flow dynamics and emptying behavior[47]. These simulations suggest that patient-specific digital twins could eventually support pre-procedural planning by predicting the physiological consequences of surgical or endoscopic interventions[49,50]. Advances have been made to incorporate physiological factors such as body posture, gastric contents, and drug dissolution kinetics. These studies emphasize the complexity of gastric biomechanics and demonstrate how computational approaches can reveal mechanisms that are difficult to study experimentally[51].

Stomach modelling is still the most advanced application, but digital twin concepts are being applied more and more to other parts of the GI tract. The intestine is a particularly challenging environment due to its complex motility patterns and continuously changing luminal contents[9]. Recent investigations have shown the possibility of developing computational models of intestinal flow and contractile activity. In particular, in vitro models of the colon have been reproduced digitally in order to simulate fluid movement and mixing dynamics during peristaltic contractions. These simulations have shown good agreement with experimental measurements of flow velocities and shear stresses within the colon[52].

Advances in imaging techniques are also enabling the creation of subject-specific models of GI motion. Dynamic MRI has been used to quantify contraction speed, luminal occlusion, and gastric geometry in individual subjects. When combined with CFD simulations, these datasets allow the generation of personalized models capable of reproducing intragastric flow patterns and reverse propelling jets associated with gastric peristalsis[51,53]. Gomez et al[54] advanced this concept into clinical practice and designed a framework that was independent of the modality and generated realistic four-dimensional gastric and bowel motion sequences from standard three-dimensional imaging of eleven patients. The model was validated against real patient motion data with submillimeter accuracy and was used to evaluate the deformable image registration for MR-guided adaptive radiotherapy planning. This is one of the first real-world clinical applications of patient-specific GI digital twins in the daily workflow. In the esophagus, computational models provide insights into the mechanics of bolus transit and sphincter function, potentially improving the diagnosis and management of disorders such as achalasia or spastic motility syndromes[55].

The emergence of these GI digital twin frameworks could significantly alter the clinical approach to motility disorders by revealing disturbances in gastric mixing, pyloric resistance, or intraluminal flow that are not detectable using conventional diagnostics[46,50]. But an honest appraisal of the current evidence base is critical. A systematic review of 90 AI studies across the whole spectrum of GI motility diagnostics confirmed that body surface gastric mapping with neural network-based signal processing is the only AI application to date that has received Food and Drug Administration clearance and achieved routine clinical implementation in this field-as a diagnostic aid, not as a bidirectional digital twin[56].

Although most GI digital twins remain in early research stages, continued progress in imaging, computational modeling, and machine learning is likely to accelerate their clinical translation. As these technologies mature, digital twins may offer a powerful framework for understanding and managing complex motility disorders that currently lack precise diagnostic or therapeutic solutions. Large-scale validation using patient-specific clinical data will be necessary before these systems can be integrated into routine clinical decision-making.

DIGITAL TWINS OF THE GUT MICROBIOME

Quinn-Bohmann et al[57] developed a microbial community-scale metabolic modeling approach using cooperative tradeoff flux balance analysis (ctFBA) to predict taxon-specific growth and short-chain fatty acid production (butyrate, acetate, propionate) in response to dietary and microbial interventions. Models are constructed by combining manually curated genome-scale metabolic reconstructions from the AGORA database using MICOM, with taxon abundances constrained from 16S amplicon or shotgun metagenomic sequencing data and solved using ctFBA to yield empirically validated estimates of steady-state growth rates and metabolic uptake and secretion fluxes for each taxon (Figure 2)[57]. Validation was achieved through an elegant in vitro and ex vivo benchmarking strategy: Predicted and measured short-chain fatty acid fluxes showed strong quantitative agreement across four independent ex vivo fermentation studies, with models explaining 25%-35% of inter-individual variance in butyrate and propionate production; these metabolites are of direct relevance to intestinal epithelial integrity, mucosal immune regulation, and colorectal cancer risk[57,58]. Moreover, this framework was also used to simulate personalized interventions, where no single combinatorial strategy was universally effective for all individuals. Some patients benefited most from the addition of prebiotic fiber, while others responded better to a probiotic that produces butyrate. These results indicate that microbiome-based interventions cannot be generalized, but need to be personalized according to the individual’s specific microbial composition and functional profile[57]. This proof-of-concept strongly endorses the precision nutrition application of microbiome digital twins in GI clinical decision-making, especially for diseases such as IBD and colorectal cancer prevention, where diet-microbiome interactions can directly impact clinical outcomes[59].

Figure 2 Microbiome digital twin workflow using microbial community-scale metabolic modelling. Stool samples undergo microbial sequencing (16S rRNA or shotgun metagenomics) to determine taxonomic abundances. These data are integrated with dietary constraints and genome-scale metabolic models from the AGORA database to build patient-specific community models using microbial community-scale metabolic modelling. Cooperative tradeoff flux balance analysis is applied to predict steady-state microbial growth and metabolic outputs, including short-chain fatty acid production. ctFBA: Cooperative tradeoff flux balance analysis; MICOM: Microbial community-scale metabolic modelling.

An alternative modelling strategy aims to predict the temporal evolution of microbial ecosystems and the clinical consequences associated with those changes, rather than only on metabolic outputs. In this context, Sizemore et al[60] developed a digital twin framework of the preterm infant gut microbiome using a generative AI architecture known as Q-net, trained on longitudinal 16S rRNA sequencing datasets. The model incorporated microbiome profiles from 88 preterm infants, encompassing 398 fecal samples and more than 32000 microbial abundance measurements across 91 microbial classes. Using these data, Q-net achieved strong predictive performance for microbiome abundance dynamics, demonstrating a coefficient of determination (R²) of 0.69. Beyond forecasting microbial composition, such models show clinical predictive capability by identifying infants at elevated risk of neurodevelopmental impairment with an accuracy of approximately 76%[60]. These frameworks have broader translational relevance in adult GI disorders, enabling prediction of symptom variability in IBS and recurrence risk in Clostridioides difficile infection under different treatments, including fecal microbiota transplantation (FMT)[61,62].

Machine learning models further show that FMT success depends on donor-recipient microbiome interactions, with specific donor taxa and recipient microbial profiles predicting outcomes and enabling selection of optimal donors to maximize therapeutic efficacy[63]. The central methodological challenge common to both approaches, as discussed above, is the validation of these models. For example, the generalized Lotka-Volterra model does not describe interaction mechanisms and thus does not account for metabolic flexibility, while constraint-based approaches depend on the quality of genome-scale reconstructions and the accuracy of environmental constraints[64]. Prospective human trials testing model-guided dietary or probiotic interventions against the standard of care are the necessary next step before microbiome digital twins can transition from research tools to clinical instruments.

DIGITAL TWINS IN IBD

Crohn’s disease and ulcerative colitis represent ideal candidates for digital twin technology due to their chronic relapsing nature, considerable heterogeneity in disease course, and variable treatment response. Despite the introduction of biologic agents targeting TNF, integrins, and interleukin signaling pathways, treatment selection remains largely empirical, with response rates to first-line biologic therapy of approximately 40%-60% in major randomized trials and a substantial proportion of patients experiencing secondary loss of response[65]. As summarized in Table 3, digital twin approaches in IBD can be organized into five distinct modeling categories, each operating on different data inputs, targeting different clinical questions, and carrying distinct limitations[66,67].

Table 3 Five inflammatory bowel disease modeling approaches as rows.

Modeling method	Representative works	Data inputs	Key targets	Potential clinical applications	Current limitation
Cytokine signaling models	Wendelsdorf et al[67]; systems-level colonic inflammation models	Cytokine profiles. Immune cell populations. Regulatory pathway data	TNF-α, IL-6, IL-12, IL-23, regulatory T-cell networks	In silico drug target validation; identification of dominant inflammatory pathways per patient	Simplified pathway representations. Limited clinical validation
Hybrid immune-clinical model	Shim et al[66]; mechanistic + data-driven framework	Biomarker profiles. Disease activity scores. Longitudinal clinical data	Patient-specific immune parameters; clinical disease activity indices	Individualized disease activity prediction; treatment response stratification prior to biologic initiation	Small derivation cohorts. Requires longitudinal data inputs
Single-cell transcriptomic twins	Karolinska “disease mechanism” twin; scRNA-seq immune network models	scRNA-seq mucosal biopsies. Gene expression networks. Immune cell subset profiles	Regulatory molecular nodes; immune cell subset activity; anti-TNF response predictors	Patient-specific therapy selection; identification of non-responders priorto biologic initiation; precision immune phenotyping	Early development stage. High data acquisition burden. Significant computational complexity
Predictive clinical model	Longitudinal ML models for Crohn’s disease trajectory prediction	CRP, fecal calprotectin. Endoscopic findings. Medication history. Patient-reported symptoms	Disease flare probability; progression to complications; secondary loss of response	Dynamic disease monitoring; early intervention triggering; virtual clinical trial design and patient enrichment	Not full digital twins. Population-level derivation. Limited prospective validation
PK/PD digital twins	Emerging pharmacokinetic-immune signaling coupling models	Drug concentration levels. Antidrug antibody titres. Immune signaling readouts	Biologic drug concentrations; dosing interval optimization; immunogenicity prediction	Precision biologic dosing; prevention of secondary loss of response; individualized therapeutic drug monitoring	Largely conceptual. Requires prospective. PK/PD data. Not yet clinically validated

Cytokine signaling models, hybrid immune-clinical models, single-cell transcriptomic twins, predictive clinical models, and PK/PD digital twins with representative work, data inputs, targets, applications, and limitations. TNF: Tumor necrosis factor; IL: Interleukin; CRP: C-reactive protein; ML: Machine learning.

Wendelsdorf et al[67] laid the earliest mechanistic foundations by building a system-level model of colonic inflammation, which accounted for interactions between cytokines, immune cells, and regulatory pathways. This allowed in silico perturbation of TNF, IL-6, IL-12, and IL-23 signaling pathways to evaluate probable drug response before real-world drug exposure[63]. Later, Shim et al[66] developed a hybrid framework that integrated mechanistic modelling with machine learning to leverage patient-specific biomarker profiles and longitudinal clinical data for informing model parameters and predicting individualized disease activity scores[67]. Building on this, a hybrid mechanistic-statistical platform applied to the VERSIFY study (n = 69) assigned each Crohn’s disease patient a digital twin from an archetypical response cohort and forecast vedolizumab response over 26 weeks, achieving sensitivities of 80% and 75% and specificities of 69% and 70% for endoscopic remission and mucosal healing respectively, with tissue damage predictions rated good or fair in 94% of validation patients[68].

Biomarker classifiers have further refined patient stratification by operating at the molecular rather than the mechanistic level. The GIMATS module, comprising IgG plasma cells, inflammatory mononuclear phagocytes, activated T cells, and stromal cells, identified in ileal Crohn’s disease lesions across four independent cohorts (n = 441), predicted failure to achieve durable corticosteroid-free remission on anti-TNF therapy, and its simplified 6-gene MIN score now enables biologic class selection by classifying patients into metabolic or immune subtypes[69,70]. Machine learning models have improved treatment prediction in Crohn’s disease, with an XGBoost model achieving an AUC of 0.91 for infliximab response and DNA methylation panels showing AUCs of 0.86-0.89 for multiple biologics, outperforming traditional clinical tools and enabling treatment stratification before biologic initiation[71].

The third and separate application is represented by precision dosing frameworks. These frameworks do not deal with treatment selection but rather with optimizing drug exposure after a therapy has been selected. Pharmacokinetic-immune signaling coupling models and model-informed precision dosing strategies have demonstrated significantly higher remission rates than the standard reactive approaches, with proactive therapeutic drug monitoring reducing secondary loss of response by individualization of biologic drug concentrations and dosing intervals[72,73].

At the clinical implementation level, the MED2ICIN digital twin, a CDSS evaluated by 31 gastroenterologists across Germany, demonstrated significantly positive impact on time efficiency, quality of care, and cost reduction, with clinicians estimating a mean time saving of approximately 5 minutes per consultation (nearly 20% of total consultation time), though improvement in therapy adherence was not significant, reflecting the multifactorial nature of non-adherence that no single system is likely to resolve alone[74]. Real-world data integration further supports these systems: The Finnish IBD digital twin initiative, for example, integrates data from over 16000 patients from six university hospitals using the OMOP Common Data Model, demonstrating how large, federated datasets can power scalable, privacy-preserving decision-support tools[75]. Similarly, Colwill et al[76] described how multimodal digital twins that integrate clinical, molecular, imaging, and real-world data could aid synthetic control arms, adaptive trial design, and prediction of relapse or treatment response in IBD.

Despite these promising possibilities, several challenges remain before digital twins can be implemented in routine clinical practice. Many current models are based on relatively small research datasets and require validation in large, diverse patient populations. Furthermore, integrating complex molecular data into clinical decision-support systems presents technical and logistical challenges. As these technologies mature, digital twins may become a central component of precision medicine strategies for IBD, enabling clinicians to simulate disease mechanisms and therapeutic responses in silico before making treatment decisions in the clinical setting[27,68].

DIGITAL TWINS IN PANCREATIC DISEASE

One of the first clinical applications equivalent to a digital twin system was the artificial pancreas, which integrated continuous glucose monitoring and automated insulin delivery to dynamically manage glycemic control[77]. The conceptual basis for pancreatic digital twins is built on decades of mathematical modelling of glucose-insulin dynamics. The UVA/padova type 1 diabetes simulator was approved by the United States Food and Drug Administration (FDA) in 2008 to replace animal testing in preclinical insulin therapy studies and offered the proof of concept that patient-specific computational models could assist in clinical decision-making[78]. This simulator has been continuously refined over the years and now serves as the core computational engine underlying multiple automated insulin delivery systems, enabling in silico clinical trials to evaluate novel control algorithms, glucose sensors, and insulin formulations before human testing[79]. Based on this, ReplayBG was developed as a digital twin-based platform that utilizes patient-specific type 1 diabetes data to simulate glucose responses and evaluate alternative insulin or dietary strategies, showing high predictive accuracy and outperforming existing approaches[80]. Moreover, a recent randomized trial demonstrated that human-machine adaptation based on digital twin improved both time-in-range and HbA1c via personalized insulin optimization and simulation of “what-if” treatment scenarios compared with standard automated insulin delivery alone[81]. These advances are mostly in diabetes care, but they introduced key digital twin concepts such as personalized simulation, in silico therapy testing, and adaptive closed-loop modeling, which are now influencing wider pancreatic and GI applications[79-81].

The more direct application of digital twin thinking to pancreatic disease within gastroenterology encompasses pancreatitis and pancreatic cancer. Machine learning models are already capable of simulating digital twin-like predictive functions for acute pancreatitis: The PANCREATIA study developed early clinical variable-based models with AUCs of 0.849 for mortality, 0.786 for ICU admission, and 0.783 for persistent organ failure, which were superior to existing scoring systems such as Acute Physiology and Chronic Health Evaluation-II and Bedside Index of Severity in Acute Pancreatitis[82]. In pancreatic ductal adenocarcinoma, the Molecular Twin platform, an AI-driven multi-omic framework integrating 6363 clinical, genomic, proteomic, lipidomic, and computational pathology features from resected pancreatic ductal adenocarcinoma patients, predicted disease survival with an accuracy of 0.85 and positive predictive value of 0.87, validating across four independent cohorts and demonstrating that plasma protein is the top single-omic predictor of survival, outperforming CA 19-9[33]. In addition, digital twin frameworks have enabled precise in silico simulation of clinical trials in pancreatic cancer and accurately predicted the relative efficacy of regimens, such as gemcitabine vs 5-F-U, and combination therapies[83]. Multimodal machine learning integrating clinical, radiologic, and genomic data improves outcome prediction, while patient-derived 3D tumor avatars provide physiologically relevant platforms that mirror treatment response and support personalized therapeutic decision-making[83-85]. A multi-scale digital twin for adiposity-driven insulin resistance, modelling interactions across whole-body, organ, and cellular levels, extends these principles to metabolic-pancreatic disease progression and shows how mechanistic models can predict responses to dietary intervention and pharmacotherapy across timescales relevant to type 2 diabetes and its pancreatic sequelae[83]. Recently, Sanchez-Castro et al[86] developed a digital twin of pancreatic islet differentiation based on temporal multi-omic data of 400603 cells, enabling the causal inference of transcription factor perturbations that control beta cell vs exocrine lineage specification, directly relevant to regenerative approaches to pancreatic disease[84].

Despite this progress, pancreatic digital twins remain less mature than those in cardiology or oncology. Continued progress will depend on improved integration of metabolic, inflammatory, and oncologic modeling within a unified patient-specific framework.

CHALLENGES AND LIMITATIONS

Data integration, quality, and computational infrastructure

Digital twins are relatively new to gastroenterology, and they pose challenges both theoretical and practical. A meta-review of 25 literature reviews identified three primary categories of barriers: Data quality and integration challenges, ethical concerns, and socioeconomic barriers, with significant gaps in scalability, interoperability, and clinical validation. Currently, there are only 12% of the total claimed digital twins in healthcare that fully meet the National Academies of Sciences, Engineering, and Medicine criteria requiring personalization, dynamic updating, and predictive capabilities, while only two studies mentioned verification, validation, and uncertainty quantification-a critical standard for model reliability[87]. The integration of multimodal patient data from heterogeneous sources (EHRs, imaging, omics) remains a major challenge for GI digital twins, particularly due to a lack of real-time, bidirectional data exchange and issues such as data latency[88]. Data quality is equally critical, as noisy, incomplete, or inconsistent inputs can compromise model accuracy, highlighting the need for robust data governance frameworks addressing both intrinsic quality and contextual relevance[89].

Explainability, algorithmic bias, and the complexity of molecular inputs

For clinicians to be on board with digital twins, their output needs to be interpretable. Explainable AI architectures are an effective way of generating trust by ensuring that the biochemical reasoning behind each result is made clear and not hidden. Transparency of digital twin models is critical for clinicians to interrogate the biological rationale driving each recommendation. This is beyond superficial interpretability to remove spurious correlations and include under-recognized molecular mechanisms, such as the context-dependent regulatory roles of long non-coding RNAs in intestinal epithelial reorganization and inflammatory gene expression. Black-box models can model high-dimensional biological complexity, but they can’t do it alone. In the short term, a balanced architecture that integrates the predictive power of deep learning with the interpretability of mechanistic or rule-based layers will be required[87].

AI ethics is becoming increasingly important, especially in this new era of rapidly evolving AI technology[90]. Biases in training data or algorithms can lead to discriminatory outcomes, potentially exacerbating existing health disparities when deployed at scale[91]. In gastroenterology specifically, there is a risk that AI/machine learning technologies could generate or worsen systematic racial, ethnic, and sex disparities across applications, including diagnosis of esophageal cancer, management of IBD, liver transplantation, and colorectal cancer screening. Digital twins also pose significant data privacy challenges due to continuous use of sensitive patient data, compounded by security concerns and limitations of traditional informed consent in opaque, continuously learning systems[92].

Rare GI diseases and data sparsity

A particular and under-addressed challenge is the application of the digital twin framework to rare GI pathologies, such as rare hereditary polyposis syndromes, eosinophilic GI disorders, and uncommon motility disorders, where the currently available training datasets are small. Transferable learning from larger related datasets and synthetic data generation using generative AI models such as generative adversarial networks, variational autoencoders, and diffusion models may help to address limited data availability by creating realistic and biologically plausible patient trajectories for model training[93]. Despite this promise, the systematic application of these methods to rare GI diseases has not yet been evaluated.

The ASME Verification and Validation 40 standard provides a credibility framework for computational models used in medical device evaluation and represents the most mature existing regulatory pathway for digital twin evidence. The FDA and EMA are currently developing guidance on the use of in-silico trial evidence in device and drug submissions. Participation in these frameworks is a fundamental prerequisite before GI digital twins can be used in regular clinical practice[94,95]. As these technologies mature through rigorous validation and thoughtful implementation, digital twins may fundamentally reshape the practice of precision gastroenterology, enabling clinicians to simulate disease mechanisms and therapeutic responses in silico before making treatment decisions in the clinical setting.

FUTURE DIRECTIONS

The limitation in most of the existing digital twins for healthcare is that they are static, which means that these models are only able to get data on a single point, without being able to update in real time over time. This limits their practical clinical value, especially in conditions with relapsing and remitting courses such as IBD. Wearable devices and ingestible capsules offer a way to overcome this limitation through continuous physiological monitoring. Hirten et al[96] demonstrated that physiological signals from consumer wearables such as Apple Watch could be used to predict IBD flares several weeks before symptom onset, showing the early-warning capability that could eventually be incorporated into a bidirectional clinical update loop for GI digital twins.

Another limitation today is that most current digital twin models focus on a single organ. Though this is a starting point for research, it can misrepresent conditions that gastroenterologists often manage, which could be more systematic. One such example is MASLD, which cannot be adequately modelled without accounting for the gut microbiota composition, adipose tissue signaling, and pancreatic insulin secretion. They interact continuously with hepatic metabolism.

Computational methods for twin development are advancing rapidly, thereby directly addressing current limitations. Large language models (LLMs) available, for example, are getting quite good at extracting meaningful structured information, enriching digital twin data layers without adding burden to clinical workflows[97,98]. The current models present a partial answer to the persistent problem of training data for rare GI conditions[99]. With LLM expansion, the computational constraints on GI twin development are likely to ease over the coming decade.

Implementing twins into the clinic is a key challenge. A digital twin that is only in a research paper and has no impact on clinical outcomes is of no use. Every model requires validation before implementing it in clinical settings. Therefore, this field requires well-designed pilot programs conducted in real clinical settings, because they expose these models to underlying challenges that are not currently visible[99]. Additionally, a twin-guided closed-loop delivery system in gastroenterology could be transformative. application, but it remains far from realization. Such technology for the gut, primarily due to microbiome composition, is far more complex than glucose and insulin levels[100].

A digital twin requires real-time data acquisition and multimodal longitudinal data, which cannot be provided by a single institution. Therefore, multi-institutional collaboration is necessary. Under federated learning frameworks, model training proceeds across distributed datasets without patient data leaving its host institution. This offers a technically mature solution to the privacy constraints that have historically impeded this kind of collaboration[101].

While important for the progress of medicine, it is equally urgent that international organizations develop a standardization of digital twin fidelity. Without agreed criteria, institutions and vendors are free to label any patient-specific model as a “digital twin”, eroding the clinical and scientific meaning of the term and making comparative evaluation impossible. An internationally recognized Digital Twin Fidelity Framework, similar to technology readiness levels in aerospace engineering, that grades GI digital twins against defined benchmarks on personalization, update cadence, validation scope, and clinical integration would be helpful for the field. Professional societies are well placed to convene this standardization effort. Until such a framework is established, authors should at least adopt and consistently apply the three-criterion definition, as highlighted here, for something to qualify as a digital twin.

CONCLUSION

The advancement of digital twin technology represents an evolution in gastroenterology. Digital twin models are moving the field from reactive, population-based care to predictive, individualized, and simulation-driven medicine. Digital twins are dynamic, evolving virtual models that integrate multimodal clinical, imaging, and molecular data, offering unprecedented insight into disease mechanisms, therapeutic response, and patient-specific trajectories. Initial applications in hepatology, oncology, IBD, motility disorders, pancreatic disease, and microbiomes have demonstrated the potential for optimizing treatment selection, improving procedural planning, and enabling in-silico clinical experimentation. However, challenges such as data integration, limited validation, scalability, interpretability, and ethical governance remain to be addressed. We need multidisciplinary collaboration, stringent frameworks, and future clinical validation studies to overcome these barriers. As the healthcare data ecosystems continue to grow and mature, digital twins will eventually go on to become a cornerstone of precision gastroenterology, enabling clinicians to predict disease progression, personalize interventions, and improve patient-care related outcomes through data-driven decision-making.