Digital twins in pediatric metabolic dysfunction-associated steatotic liver disease: Developmentally referenced multimodal risk stratification

Abstract

Metabolic dysfunction-associated steatotic liver disease (MASLD) is common in children, but its course is heterogeneous and current risk tools are largely static and population-based. No validated pediatric MASLD digital twin currently exists, and no prior synthesis has defined how digital-twin methodology should be adapted to childhood, when growth and puberty change the interpretation of clinical data. This narrative review evaluates digital-twin models and candidate inputs for pediatric MASLD, including noninvasive diagnostics, imaging, body composition, molecular and genetic markers, lifestyle, and social context. We explicitly distinguish direct pediatric MASLD evidence from adult MASLD or metabolic evidence, evidence from other pediatric diseases, and generic digital-twin and artificial-intelligence methodology. We then propose an author-derived, investigational three-layer architecture: Longitudinal multimodal inputs; an age- and puberty-referenced twin engine; and clinician-facing candidate outputs. Initial research use should focus on association-based prognostic monitoring and validation of modular components. Adaptive follow-up and individualized risk estimates require prospective validation, while causal treatment-effect simulation is a separate and more demanding objective. Clinical deployment would require interoperable electronic health record data, prospective multicenter validation, transparent reporting across subgroups, explainability, family-centered consent and assent, cost-effective implementation, and regulatory oversight of adaptive software. The near-term priority is rigorous evaluation of a minimum viable core model rather than deployment of a complete twin. No evidence-based timeline for clinical implementation can yet be specified.

Key Words: Metabolic dysfunction-associated steatotic liver disease; Digital twin; Pediatrics; Risk stratification; Body composition; Precision medicine; Machine learning; Longitudinal modeling

Core Tip: Pediatric metabolic dysfunction-associated steatotic liver disease is dynamic, but routine monitoring relies heavily on isolated measurements. No validated pediatric metabolic dysfunction-associated steatotic liver disease (MASLD) digital twin currently exists. This review separates direct pediatric MASLD evidence from adult, other-pediatric, and generic digital-twin evidence and proposes an author-derived, investigational architecture. A minimum viable core should begin with routinely available developmental, anthropometric, laboratory, treatment, and clinically obtained imaging data, with advanced imaging, body composition, genetics, multi-omics, microbiome, and digital phenotyping added only as separately validated modules. Near-term research should emphasize association-based prognostic monitoring, explicit uncertainty, missing-not-at-random data, false alerts and false reassurance, and silent prospective validation. Treatment-effect simulation requires separate causal validation and should not be inferred from ordinary risk prediction.

INTRODUCTION

Metabolic dysfunction-associated steatotic liver disease (MASLD), the term that has replaced non-alcoholic fatty liver disease, is the most common chronic liver disease of childhood. Its long monitoring horizon begins during a period of rapid physiological change, so early risk assessment must account for growth and puberty[1,2].

Population estimates place MASLD prevalence near 10% among all children and above one-third among children with obesity[3-5]. In gastroenterology cohorts selected for biopsy, approximately 25%-50% have steatohepatitis and 10%-25% have advanced fibrosis at presentation. These ranges reflect referral and verification bias and should not be generalized to community populations[3].

The consequences extend beyond childhood. Pediatric cohorts show extrahepatic morbidity and long-term mortality, while steatohepatitis has become a rapidly increasing indication for liver transplantation in young adults[6,7].

The central clinical problem is heterogeneity. Two children with similar baseline anthropometrics and aminotransferases may follow different courses, and pediatric cohorts show both progression and improvement over relatively short follow-up[8]. Current tools remain largely static and population-based. Pediatric reviews and guidance also emphasize the limits of alanine aminotransferase (ALT) and other routine measures[9,10]. These tools rarely integrate serial liver tests, imaging, body composition, treatment exposure, molecular susceptibility, lifestyle, and social context into one developmentally referenced trajectory.

A healthcare digital twin is a computational representation of an individual that updates as new data arrive, estimates the current state, and forecasts plausible future trajectories. It may also compare bounded intervention scenarios. Rather than showing only that a child’s ALT is elevated today, a twin could display its two-year trend against age- and sex-referenced values. It could then integrate liver stiffness and body composition to estimate the likely direction of risk during the next year. A prediction model becomes a twin only when it is individualized, repeatedly synchronized with the patient, and used prospectively for monitoring or simulation[11-21].

This review first defines digital-twin methodology and summarizes early applications. It then examines candidate data domains, including non-invasive diagnostics, body composition, genetic and metabolic factors, lifestyle, and social determinants. Finally, it proposes a developmentally referenced architecture and outlines the evidence, governance, and implementation requirements for clinical translation.

DIGITAL TWINS: CONCEPT AND EARLY APPLICATIONS IN MASLD

A digital twin differs from a conventional risk calculator in three ways. It maintains a longitudinal state rather than producing one prediction from one visit[11,12]. It updates bidirectionally as patient data and model outputs inform subsequent monitoring[13,14]. It can combine prediction with explicit uncertainty and bounded scenario testing, although those functions require stronger validation than descriptive risk estimation[15-21].

Pediatric applications remain early. Investigators have described frameworks for asthma, infectious diseases, maternal-fetal health, rare disease, and digital-health infrastructure[22-26]. A prototype federated platform for pediatric obesity and cardiometabolic complications also illustrates how longitudinal data could support developmentally referenced models[27].

Cardiology provides more mature technical examples. Electrocardiography and cardiac magnetic resonance imaging have been integrated to personalize electrophysiological models, simulate drug effects, and reconstruct valvular anatomy[28,29]. Metabolic applications include adult case studies in polycystic ovary syndrome and type 2 diabetes (T2D)[30,31], while broader healthcare-infrastructure reviews emphasize the need for secure synchronization and interoperable data streams[32].

EVIDENCE HIERARCHY AND SCOPE OF INFERENCE

For orientation, Table 1 contrasts traditional monitoring with the author-proposed investigational digital-twin approach. No clinically validated pediatric MASLD digital twin has been reported. Accordingly, the evidence supporting the proposed framework is indirect to different degrees. We classify evidence by directness to pediatric MASLD digital-twin use rather than treating all cited studies as equivalent (Table 2). Tier 1 is direct pediatric MASLD evidence and supports candidate variables, observed associations, disease phenotypes, and candidate outcome definitions. Tier 2 is adult MASLD or metabolic evidence and supports biological or technical plausibility but not pediatric performance. Tier 3 is evidence from other pediatric diseases and supports developmental, ethical, family-centered, and implementation considerations. Tier 4 is generic digital-twin, artificial-intelligence, methodological, regulatory, and interoperability evidence and supports model architecture and evaluation principles.

Table 1 Traditional pediatric metabolic dysfunction-associated steatotic liver disease monitoring compared with an author-proposed investigational digital-twin approach.

Dimension	Traditional monitoring	Author-proposed investigational digital-twin approach
Data domains	Periodic anthropometrics, liver tests, and selected imaging	Minimum viable core of time-stamped developmental, clinical, biochemical, treatment, and clinically obtained imaging data; optional advanced imaging, body composition, genetics/omics, microbiome, and digital context modules
Developmental reference	Age or sex considered at individual visits	Age-, sex-, growth-, and puberty-referenced trajectories when appropriate
Update frequency	Clinic visits at fixed intervals; tests repeated when indicated	Prospective synchronization when validated new data arrive; cadence depends on data type and intended use
Primary output	Thresholds, population risk groups, and clinician synthesis	Candidate association-based trajectory or calibrated risk range with uncertainty and main drivers; not for care until prospectively validated
Therapeutic use	Retrospective assessment of response	Hypothesis-generating scenario comparison only; no individualized treatment-effect claim without a separately validated causal model
Requirements and risks	Established workflows but fragmented longitudinal context	Interoperable records, label-quality controls, explicit missing-not-at-random handling, external validation, explainability, governance, equity monitoring, and safeguards against false alerts, false reassurance, and alert fatigue

The digital-twin column describes an author-proposed research framework, not validated clinical guidance.

Table 2 Evidence hierarchy for claims in this review.

Evidence tier	Evidence source	What it supports in this review	What it does not establish	Ref.
Tier 1: Direct	Pediatric MASLD	Candidate clinical variables, observed associations, epidemiology, ALT/histology limitations, imaging, body composition, genetics, cardiometabolic outcomes, microbiome, and social context	A validated pediatric MASLD digital twin; calibrated individualized twin-based risk; adaptive follow-up benefit; causal treatment-effect simulation	[2-10,33-44,57-61]
Tier 2: Extrapolated	Adult MASLD or metabolic disease	Biological plausibility, adult feasibility of digital-twin-enabled metabolic management, and selected adult prognostic or genetic concepts	Pediatric calibration, safety, developmental validity, or pediatric treatment effects	[21,30,31,46,56]
Tier 3: Pediatric extrapolation	Other pediatric diseases/child-health digital twins	Developmental referencing, pediatric digital-twin feasibility, family perspectives, ethics, and child-centered implementation principles	MASLD-specific predictive performance or clinical utility	[11,22-27]
Tier 4: Methodological	Generic digital twins, AI, prediction-model methods, regulation, and interoperability	Definitions, architecture, explainability, federated learning, validation/reporting methods, governance, interoperability, and regulatory principles	Clinical validity or effectiveness in pediatric MASLD	[12-20,28,29,32,62-82]

Evidence tiers reflect directness to pediatric metabolic dysfunction-associated steatotic liver disease (MASLD) digital-twin clinical use, not an assessment of study quality. Evidence from Tiers 2-4 is extrapolative and should not be interpreted as validation of pediatric MASLD digital-twin outputs. AI: Artificial intelligence; ALT: Alanine aminotransferase; MASLD: Metabolic dysfunction-associated steatotic liver disease.

Evidence from tiers 2-4 is reasonable extrapolation, not validation of pediatric MASLD clinical outputs. The three-layer architecture, adaptive follow-up strategy, individualized risk displays, and translation phases in this review are therefore author-proposed investigational conceptual frameworks. In silico scenario comparison is more speculative: Unless a model is explicitly designed and validated to estimate causal treatment effects, it should be interpreted only as hypothesis generation.

COMPUTATIONAL APPROACHES TO A CLINICAL DIGITAL TWIN

Mechanistic models

These models encode physiological relationships, such as lipid flux, insulin signaling, or tissue biomechanics. Their structure can support causal interpretation and extrapolation. However, pediatric parameterization is difficult, misspecified biological assumptions can be misleading, and repeated calibration may be computationally demanding[12,15-17].

Statistical longitudinal models

Mixed-effects, joint, landmark, and state-space models estimate change from repeated measurements. They can be transparent, data-efficient, and well calibrated, but they require prespecified relationships and may miss complex nonlinear interactions[12,20].

Machine learning models

Machine learning can integrate high-dimensional imaging, omics, and electronic health record data. It may capture nonlinear patterns, but it requires large datasets, is vulnerable to site-specific distribution shifts and bias, and can be difficult to explain. A static machine learning score is not itself a digital twin[15-17].

Hybrid models

Hybrid twins combine physiological constraints with statistical or machine learning components. This approach may balance interpretability and predictive flexibility, but it increases implementation and validation burden. For pediatric MASLD, modular hybrid development is more realistic than immediate construction of a fully comprehensive twin[15-17,20].

NON-INVASIVE AND NON-IMAGING DIAGNOSTICS IN PEDIATRIC MASLD

Digital-twin inputs should be clinically meaningful, repeatable, and feasible. Liver biopsy remains the reference standard for staging steatohepatitis and fibrosis, but it is invasive, subject to sampling variability, and unsuitable for frequent monitoring[3,9,10]. ALT is widely available but does not reliably stage histology. In a pediatric biopsy cohort, children with strictly normal ALT could have mild histologic abnormalities, although none of 17 had advanced fibrosis. Advanced fibrosis occurred in 7 of 74 children (9%) with mildly elevated ALT[33]. An ALT threshold of 80 U/L identified advanced fibrosis with 76% sensitivity and 59% specificity in one referral cohort[10]. Thus, a low ALT value does not exclude disease, and an isolated ALT value should not determine risk.

Imaging provides the most quantitative liver-specific measurements. Magnetic resonance imaging proton density fat fraction (MRI-PDFF) and magnetic resonance elastography estimate hepatic fat and stiffness, while ultrasound-based elastography and controlled attenuation parameter offer more accessible alternatives[34-36]. Performance varies with equipment, thresholds, body size, fasting status, and disease spectrum. Simple fibrosis scores may add context but should not replace pediatric imaging or clinical judgment[37]. A twin should preserve the method, device, quality metric, and uncertainty for every measurement.

Extrahepatic signals may help explain heterogeneity and prevent diagnostic anchoring. In one biopsy-confirmed cohort, 17 of 228 children (8%) used psychotropic medications. Exposure was associated with higher steatosis and activity scores but not fibrosis, so it should be treated as an observational modifier rather than a causal label[38]. Among 900 referred children, 19 (2%) had an alternative liver diagnosis, supporting explicit flags for diagnostic provenance[39]. Vitamin D associations remain inconsistent and context dependent[40,41]. In a multicenter pediatric cohort of 1164 children with nonalcoholic fatty liver disease/MASLD, 27% had hyperfiltration, which was independently associated with significant liver fibrosis[42]. Renal data may therefore identify a multisystem metabolic phenotype, but they do not establish future kidney or liver outcomes.

BODY COMPOSITION AND SARCOPENIA AS DETERMINANTS OF OUTCOME

Body composition is a core data domain because body mass index does not capture the distribution or function of adipose and lean tissue. Pediatric studies associate lower skeletal muscle mass and less favorable impedance-derived muscle measures with greater steatosis or disease severity, although these associations require prospective validation[43,44].

Standardized bioelectrical impedance analysis (BIA) and handgrip dynamometry are inexpensive, repeatable options for routine longitudinal phenotyping. Phase angle is a raw impedance measure that may serve as a proxy for body cell mass, membrane properties, and fluid distribution. It is not a direct measure of cellular health, and its interpretation requires reference values specific to age, sex, pubertal stage, device, and protocol[45].

Sarcopenia is characterized by low muscle strength, with low muscle quantity or quality used to confirm the diagnosis[46]. Sarcopenic obesity is the coexistence of excess adiposity and low muscle mass or function[47]. In adults with cirrhosis, sarcopenia, sarcopenic obesity, and myosteatosis are associated with poor survival, although these adult data should not be extrapolated directly to children[48]. Pediatric liver-transplant studies also show that low muscle status can persist and have clinical consequences[49]. These concepts are relevant during glucagon-like peptide-1 receptor agonist therapy or bariatric surgery because weight, liver fat, stiffness, and lean mass may change at different rates. Longitudinal monitoring should distinguish beneficial adipose loss from clinically important lean-tissue loss[2,3].

Dual-energy X-ray absorptiometry provides regional body-composition estimates in specialty and research settings, but access, cost, and low-dose radiation limit frequent use[50]. Deuterated-creatine dilution more directly estimates skeletal muscle mass, yet it remains primarily a research method and requires additional pediatric validation[51,52]. A practical twin could use frequent BIA and strength measurements, with selected time points anchored by more precise methods.

Skeletal muscle also communicates with the liver through myokines that influence lipid and glucose metabolism[53]. Proteomic and transcriptomic studies further show that steatotic liver disease comprises biologically distinct pathways that could eventually complement clinical phenotyping[54]. In addition to body composition, inherited and acquired molecular factors therefore contribute to pediatric MASLD heterogeneity.

GENETIC, METABOLIC, MICROBIAL, AND LIFESTYLE DRIVERS OF HETEROGENEITY

Inherited susceptibility partly explains why children with similar clinical profiles diverge. The PNPLA3 I148M variant impairs triglyceride hydrolysis and promotes hepatic triglyceride accumulation[55]. It is associated with hepatic fat and progressive liver risk, although its hepatic effects may be partly dissociated from systemic metabolic traits. The variant should therefore not be described as protective against metabolic risk[55,56]. Variants in PNPLA3 and GCKR influence hepatic lipid handling, while loss-of-function variation in PNPLA3 may reduce progression from steatosis to steatohepatitis[56]. Pediatric trajectory data remain limited, but PNPLA3 may help stratify baseline susceptibility when interpreted alongside ancestry, phenotype, and longitudinal exposures[57].

Metabolic and microbial factors contribute further to heterogeneity. Insulin resistance is central to pediatric MASLD, and affected children have an increased risk of incident T2D and other cardiometabolic complications[58]. The gut microbiome changes across childhood and has been associated with steatohepatitis and fibrosis phenotypes[59,60]. Microbial and multi-omic measures remain exploratory. Their analytic variability, cost, and temporal instability currently limit routine use.

Lifestyle and social context are more readily measurable in current practice. Diet quality, physical activity, sleep, screen time, and treatment exposure influence metabolic health. Community deprivation and barriers to care also affect disease severity and the ability to complete monitoring[61]. These variables should inform interpretation rather than encode structural disadvantage as biological risk. The documented risk of biased health algorithms makes subgroup performance and equity auditing essential[62].

THE PEDIATRIC REQUIREMENT FOR LONGITUDINAL MODELING

These domains converge on one pediatric requirement: A MASLD digital twin cannot be an adult model scaled to a smaller body. Growth changes the meaning of nearly every input. Anthropometric and body-composition values depend on age, sex, and pubertal stage; insulin sensitivity changes during puberty; and medication dosing, behavior, and family context evolve over time.

The model should therefore normalize inputs against contemporaneous developmental references rather than fixed adult thresholds[45,50]. It should model within-child trajectories, distinguish expected maturation from pathological change, and allow reference distributions to shift with growth. It should also represent irregular visits, measurement error, treatment exposure, and missingness explicitly. This longitudinal approach is the central feature that distinguishes a pediatric twin from static risk classification.

AUTHOR-PROPOSED, INVESTIGATIONAL ARCHITECTURE FOR PEDIATRIC MASLD DIGITAL TWINS

Layer 1: Longitudinal multimodal inputs

Layer 1 should be modular. The proposed minimum viable core uses data likely to be available in routine pediatric care: Age and sex; pubertal stage when documented; serial height, weight, body mass index z score, waist measures, and blood pressure; standard liver and metabolic laboratory tests; treatment and medication exposure; and routine clinical imaging or elastography when obtained (Figure 1). Advanced magnetic resonance imaging, formal body composition, genetics, microbiome, multi-omics, and passive or wearable-derived behavioral data are optional research modules. Each item should retain units, pediatric reference ranges, acquisition method, quality indicators, and provenance. Optional modules should add information only after incremental value and fairness are demonstrated.

Figure 1 Author-proposed, investigational layered architecture for a developmentally referenced pediatric metabolic dysfunction-associated steatotic liver disease digital twin. The figure distinguishes a minimum viable core of routinely available inputs from optional research modules and separates association-based prognostic prediction from a causal simulation module that would require independent validation. Candidate outputs are research targets and do not constitute validated clinical guidance. MASLD: Metabolic dysfunction-associated steatotic liver disease; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; GGT: Gamma-glutamyl transferase; BMI: Body mass index; HbA1c: Hemoglobin A1c; GLP-1 RA: Glucagon-like peptide-1 receptor agonist; MRI-PDFF: Magnetic resonance imaging proton density fat fraction; MRE: Magnetic resonance elastography; BIA: Bioelectrical impedance analysis; DXA: Dual-energy X-ray absorptiometry; MNAR: Missing not at random; SHAP: Shapley additive explanations.

Layer 2: Developmentally referenced twin engine

The second layer harmonizes data and estimates the child's evolving state. Federated learning can train models across institutions without transferring patient-level data, which may enlarge pediatric datasets while preserving local control[63]. The engine may combine trajectory models, calibrated probabilities, uncertainty estimates, and explainability methods such as feature attribution[64,65]. Initial risk estimates should be interpreted as association-based prognosis under observed care. A separate causal module would be required for counterfactual treatment-effect simulation. Any model should update only within prespecified, validated boundaries and should record every model version and decision-relevant output.

Layer 3: Clinician-facing outputs

The third layer contains candidate investigational outputs rather than validated clinical recommendations. These could include a current monitoring category, an uncertainty-bounded risk range, the main drivers of change, and prompts for clinician review, repeat imaging, referral, or diagnostic reconsideration. Adaptive follow-up and alerts should not alter care until prospective evaluation demonstrates acceptable benefit, error rates, and workflow burden. Scenario comparisons may illustrate plausible trajectories, but they should not be labeled individualized treatment effects unless causal validity has been established. The twin supports rather than replaces clinician and family decisions.

Governance links all three layers. Role-based access, encryption, audit logs, data minimization, and privacy-by-design protect longitudinal pediatric information[66,67]. Federated infrastructure reduces transfer of patient-level data but does not eliminate local bias, cybersecurity risk, or the need for transparent oversight. Table 1 contrasts conventional monitoring with this author-proposed investigational approach, and Table 2 summarizes the evidence hierarchy underlying the framework.

Figure 1 and Tables 1, 2, and 3 are author-proposed conceptual research frameworks. They are investigational, do not constitute validated clinical guidance, and should support rather than replace clinician and family decisions.

Table 3 Author-proposed evidence pathway for evaluating pediatric metabolic dysfunction-associated steatotic liver disease digital twins before clinical use.

Phase	Purpose	Evidence focus	Readiness signal
1a	Model specification and retrospective development	Define intended use, target population, outcome, minimum viable core, label provenance, biopsy-selection bias, missing-data assumptions, and subgroup analysis before modeling	Model and outputs are fully specified; internal validation and subgroup calibration justify prospective observation, not clinical use
2b	Silent prospective evaluation	Run a locked model in real time without influencing care; measure calculability, missingness, calibration, false alerts, false reassurance, equity, and workflow burden	Prospective error profile and workflow burden are characterized, with no care decisions driven by the model
3c	External and temporal validation	Test independent sites and later time periods; assess recalibration, subgroup performance, and incremental value of optional modules	Performance remains calibrated and clinically interpretable across settings and time
4d	Prospective decision-support evaluation	Test clinician-supervised outputs; assess net benefit, safety, downstream testing, alert fatigue, clinician/family understanding, consent, and equity	The tool improves or safely supports decisions beyond usual care without unacceptable harms or disparities
5e	Controlled updating and scaling	Use version control and governed updates; monitor drift, bias, cybersecurity, safety, performance, and model cards; federated scaling may be used when appropriate	Performance remains stable during monitored updates under an appropriate regulatory and governance process

Phase 1a: TRIPOD + AI[70], PROBAST + AI[75], MI-CLAIM[77], and MINIMAR[78] inform model specification, transparent reporting, and risk-of-bias assessment. Phase 2b: DECIDE-AI[81] and FUTURE-AI[76] inform early prospective workflow evaluation; FHIR[74] supports interoperable data exchange where applicable. Phase 3c: External-validation principles are informed by Riley et al[68], TRIPOD + AI[70], and PROBAST + AI[75]. Phase 4d: SPIRIT-AI[79], CONSORT-AI[80], DECIDE-AI[81], and STARD-AI[82] apply according to study design and intended use. Phase 5e: FDA predetermined change control plan guidance[71], IMDRF good machine learning practice[72], FUTURE-AI[76], and the European Union Artificial Intelligence Act[73] inform controlled updating and oversight as applicable to jurisdiction and product classification. This phased pathway is an author-proposed investigational framework synthesized from existing methodological, reporting, regulatory, and interoperability guidance. It is not a validated staging system or a regulatory pathway.

CLINICAL AND RESEARCH USE CASES

Intended initial use and outcomes. The first clinical research use should be clinician-supervised prognostic monitoring in children already diagnosed with or being evaluated for MASLD, not autonomous diagnosis or treatment selection. Candidate outcomes for early model development should be prespecified and separately validated, such as 12- month to 24-month worsening in liver stiffness or MRI-PDFF when obtained, persistent aminotransferase worsening, incident T2D, or a clinician-confirmed need for diagnostic escalation. These examples are research targets rather than validated surrogate endpoints, and models should report which outcome is being predicted rather than collapsing distinct endpoints into an opaque composite.

The first proposed use case is association-based risk stratification. Consider a 10-year-old with MASLD whose ALT rises across several visits while liver stiffness and waist-to-height ratio also increase. A validated prognostic model could flag the worsening trajectory, show which inputs drove the alert, and prompt earlier imaging or specialist review. It should not generate a precise two-year probability until that estimate has been calibrated and externally validated in comparable children. Such a probability estimates future risk under the observed care pattern; it does not estimate the effect of changing treatment.

Adaptive follow-up is a second proposed use case. Stable children might eventually avoid unnecessary testing, while a child with deteriorating liver, metabolic, or functional trajectories could return sooner. This strategy requires prospective evaluation because false alerts can increase visits and testing, whereas false reassurance can delay needed evaluation. The third proposed use case is decision support. A model might alert the clinician that interpretation should change because of a newly prescribed psychotropic medication, an increase in glycated hemoglobin, or a flag suggesting an alternative diagnosis. Outputs should remain concise, allow clinician override, suppress duplicate alerts, and quantify alert burden to reduce alert fatigue.

The fourth proposed use case is in silico scenario comparison. Prediction and causal treatment-effect estimation are not interchangeable. A counterfactual claim-what would happen to this child if therapy A rather than therapy B were chosen-requires a causal model with clearly defined interventions and outcomes, appropriate handling of time-varying confounding and treatment selection, and independent validation. Existing adult metabolic digital-twin studies can support feasibility but do not validate pediatric MASLD treatment effects[17,20,21]. Until such evidence exists, scenario displays should be labeled hypothesis-generating and should not be used as individualized treatment recommendations.

TRANSLATIONAL BARRIERS

Data availability and quality

Pediatric MASLD cohorts are small relative to the variation they must capture. Follow-up is irregular, and paired histology, imaging, body composition, and molecular data are scarce. Label quality is also imperfect. Biopsy is subject to sampling variability and is preferentially performed in children with greater clinical concern, creating biopsy-selection and verification bias; imaging measurements and thresholds vary by device, protocol, body size, and disease spectrum[3,9,10,33-36]. A model trained on such labels may learn referral and testing patterns rather than disease biology. Missingness is plausibly missing not at random because disease severity, access barriers, clinician concern, and treatment exposure all influence whether visits, imaging, or biopsies occur. Models should represent missingness explicitly, test sensitivity to alternative missing-data assumptions, and widen uncertainty when optional domains are absent. Near-term development should focus on modular components with defined clinical uses.

Model errors and alert burden

False-positive alerts may lead to unnecessary imaging, referrals, anxiety, and cost, whereas false reassurance may delay evaluation of progressive disease. Evaluation should therefore report false-positive and false-negative rates, calibration, threshold-specific predictive values, downstream testing, clinician override, and alert frequency in addition to discrimination[65,68-70]. Duplicate or low-value alerts should be suppressed, and thresholds should be tuned to the intended use so that alert fatigue does not erode clinician attention.

Validation and quality evaluation

Clinical prediction models are easy to build and difficult to trust. Sample sizes for model development and external validation should be justified, performance should be reported transparently, and calibration, discrimination, clinical utility, subgroup performance, and temporal stability should be tested[68-70]. Silent prospective evaluation should precede decision support. Reporting guidance improves transparency, but it does not prove safety or effectiveness.

Governance, consent, and regulation

A pediatric twin may follow a child through changing capacity to assent. Governance should specify ownership, permitted secondary uses, retention, access, correction, withdrawal, cybersecurity response, and re-consent as adolescents mature[26,67]. Regulation depends on intended use and jurisdiction. In the United States, final Food and Drug Administration guidance describes predetermined change control plans for bounded updates to artificial intelligence (AI)-enabled device software[71]. International good machine learning practice emphasizes representative data, independent testing, human-AI team performance, and real-world monitoring[72]. Under the European Union AI Act, high-risk classification depends on intended purpose and the applicable product route; not every clinical algorithm is automatically high risk[73].

Implementation, interoperability, and cost

Electronic health record integration requires more than an interface. Systems must reconcile units, pediatric reference ranges, timestamps, device metadata, provenance, duplicate records, and local coding. Health Level Seven Fast Healthcare Interoperability Resources can support data exchange[74]. Digital Imaging and Communications in Medicine, Logical Observation Identifiers Names and Codes, and Systematized Nomenclature of Medicine Clinical Terms can reduce semantic mismatch. Costs include data engineering, secure computing, validation, monitoring, regulatory quality systems, clinician training, and governance time. Sustainable deployment requires a clear clinical benefit that justifies these recurring costs.

Patient and family perspectives

Children and families should help define acceptable data use, meaningful outcomes, and understandable displays. Explanations should show what changed, why the model is concerned, what uncertainty remains, and what choices are available. Families need a way to correct errors and question recommendations. Co-design can also identify stigma, surveillance burden, language barriers, and problems with digital access that technical validation may miss[22,26].

Low-resource settings

A tiered design may improve access. The same minimum viable core shown in Figure 1 should remain usable when advanced modules are absent: Routine anthropometrics, standard liver and metabolic tests, blood pressure, treatment history, and clinically obtained imaging can form the base model, while advanced imaging, formal body composition, genetics, microbiome, multi-omics, and passive digital measures remain optional. The system should widen uncertainty when data are sparse rather than penalize children for missing expensive tests. Offline data capture, low-bandwidth interfaces, local calibration, and shared regional validation may be more practical than a uniform high-technology platform.

AUTHOR-PROPOSED EVIDENCE PATHWAY FOR TRANSLATION

Progress from concept to clinical use should follow an evidence pathway rather than an implementation mandate (Table 3). Table 3 is an author-proposed investigational sequence synthesized from existing prediction-model, artificial-intelligence reporting, clinical-evaluation, interoperability, and regulatory frameworks; it is not a validated staging system or regulatory pathway.

Progress from concept to clinical use should follow an evidence pathway rather than an implementation mandate (Table 3). This author-proposed sequence is informed by prediction-model, artificial-intelligence reporting, clinical-evaluation, interoperability, and regulatory frameworks and is not a validated staging or regulatory pathway. Phase 1 defines the intended use, population, outcome, minimum viable core, label provenance, missing-data assumptions, and subgroup analyses before retrospective modeling[70,75-78]. Phase 2 evaluates a locked model silently in routine workflows, assessing missingness, calibration, equity, errors, and workflow burden, with relevant guidance from FUTURE-AI, DECIDE-AI, and FHIR.

Phase 3 assesses external and temporal validity, subgroup calibration, and the incremental value of optional modules. Phase 4 prospectively evaluates clinician-supervised decision support, using SPIRIT-AI, CONSORT-AI, DECIDE-AI, or STARD-AI as appropriate to the study design and intended use[79-82]. Phase 5 allows only controlled, versioned updating with ongoing monitoring of drift, bias, safety, cybersecurity, and performance[71-73,76]. These frameworks are complementary and do not replace prospective evidence of clinical benefit.

CONCLUSION

Pediatric MASLD requires developmentally referenced longitudinal modeling because growth and puberty change the meaning of clinical data. A useful digital twin would integrate repeated liver measures with body composition, treatment, molecular and multi-omic susceptibility, lifestyle, and social context while preserving uncertainty. No validated pediatric MASLD digital twin currently exists, and available evidence does not support an evidence-based timeline for clinical deployment. The appropriate near-term endpoint is a validated minimum viable core and modular components that can undergo external validation, silent prospective testing, and controlled decision-support evaluation. Association-based risk prediction should remain distinct from causal treatment-effect simulation until the latter has been specifically validated. Achieving eventual clinical use will require multicenter collaboration, interoperable data, sustainable implementation, family-centered governance, and rigorous predeployment evidence.