Data are defined as the brain of the DT and are essential for virtualization [45]. Cross-domain literature reveals that the data layer encompasses data collection, organization, and management, and represents domain knowledge [45,51,52]. Furthermore, data compatibility with AI-driven predictions and decision-making processes should be ensured [23]. AI-readiness involves transparent data collection, preparation, quality assurance, and documentation, providing data as machine-readable metadata [75,76].

This section explores existing standards, collected data, organizational strategies, and storage solutions specifically tailored to diabetes research. Through an integrative synthesis of DT-related approaches, we illustrate how the data layer of a holistic DT for diabetes could be structured.

Combining the methods and use cases identified in the literature, a holistic DT for diabetes should integrate data from diverse sources, including routinely available clinical test results, electronic health records (EHRs), and wearable sensor data (eg, glucometers, CGM devices, smart scales, sphygmomanometers, activity trackers, calorie calculators, and insulin pumps) [9,14,25,27,47,48,53]. In particular, historical and real-time context (eg, location, activity, time, stress, and vital parameters) should be aligned. For a comprehensive patient profile, the literature emphasizes collecting data on diabetes history, FH, comorbidities, behavior, and medication [23], especially taken up to 3 years before diagnosis [8,30]. For diet management, studies commonly collect the type, amount, and frequency of consumed food, as well as metabolic, activity, and emotional data. To increase personalization, the importance of integrating these variables with genetic information is highlighted [53,54], alongside incorporating patient taste preferences, allergies, and individual goals [55,56]. Additionally, personalized, adaptive diet planning that considers comorbidities is enabled by representing knowledge of different diets, macronutrient combinations, and fasting schedules [54]. Finally, Vaskovsky et al [55] propose mapping food components to their own DTs, increasing semantic context.

Regarding the use of Internet of Things (IoT) and mobile devices, several studies propose mobile health care or virtual data platforms linked to smart devices and wearables, enabling automatic, continuous data acquisition and integrated analysis [9,27,39]. Lee et al [9] present a government-supported big data exchange platform in which data are stored on personal cloud services or smartphones and can be collected from health care apps, log data, and IoT devices. In addition, a platform to upload clinical and genetic data is provided. The platforms can also integrate different data modalities, such as images of meals or retinal data [39,57]. Retinal images can be collected using IoT devices and smartphones [57]. Moreover, Sai et al [39] enhance data exchange platforms with nonfungible token technology to incentivize data sharing through monetary rewards. To ensure privacy, patient data are anonymized using generative adversarial networks (GANs). These platforms enhance self-control over data, improve glucose management, and increase awareness about lifestyle impact. Moreover, clinicians can monitor patients within a comprehensive contextual framework [9,27,39].

Regarding input data, the distribution of input data across all studies is shown in Table S2 in Multimedia Appendix 2 summarizes commonly used data from the included studies. Glucose and carbohydrate values are the most frequently monitored features, with 69% (20/29) and 72% (21/29), respectively. Glucose data are categorized into CGM data, which are more often incorporated into data-driven frameworks, and self-glucose blood measurements obtained from blood tests and finger pricks, which are more common in mechanistic models. Only half of the studies using carbohydrate data collected protein (10/29, 34%) and fat components of the meal (9/29, 31%). Insulin and PA are regularly tracked, with 52% (15/29) and 38% (11/29), respectively. Often, demographics (12/29, 41%), EHRs (9/29, 31%), and vital signs (9/29, 31%) are considered for personalization, besides wearable data. However, drug use, images, genetic data, and laboratory test results are seldom explored, with less than 6 studies each, limiting the potential for personalized care and for models to learn contextual information.

Regarding data cleaning, one submodule of the DT needs to focus on data preprocessing, including data cleaning, outlier detection and correction, imputation of missing values, and extraction of reliable information [58,77]. For imputation, studies commonly use forward filling, averaging, mathematical approaches, or data-driven methods such as linear interpolation, Multiple Imputation by Chained Equations, or logistic regression [23,56,59]. For metabolic values, missforest imputation is proposed [25].

Literature also highlights the importance of addressing uncertainty and variance. Proposed methods include pooling data using Rubin rules [56,78] or correcting CGM data errors with calibration-dependent parameters and autoregressive noise modeling [79]. Data reliability can be further ensured through validation and error correction via user feedback and automated detection of historical trend deviations [56,78]. To mitigate the problem of small and imbalanced datasets, synthetic data generated with GANs are fused with real data [80].

Regarding feature extraction, after data cleaning, features can be extracted, and behavior can be automatically retrieved from measured IoT data. These features can improve the performance of subsequent models of the other representation levels. For instance, studies specify the activity type (eg, standing, sitting, walking, or unknown) [23,60] or categorize lifestyle into busy, reasonably active, lightly active, or lethargic based on a weighted threshold comparison, if more context is available (eg, calorie, consumption, step count, distance, sleep, weight, and active minutes) [39,60]. From time-series data, recent trends, long-term patterns, and cohort-based models can be extracted [56]. Thamotharan et al [23] use structural time-series analysis (TSA) to decompose time-series data. They also estimate the possible prediction performance of various features with an autocorrelation or partial autocorrelation function, showing that time-varying intermittent behaviors, causal variables, and exogenous factors influence predictions of blood glucose levels (BGLs) [23]. Additionally, they create matrix profiles, storing the distance between each subsequence within a time series and its neighbor, to discover intrapersonal and population-based patterns and anomalies that impact the trajectory of BGLs [23]. Text data from EHR can be processed with pretrained natural language models such as ClinicalBERT [8,61].

In particular, diet information and nutrition knowledge can be organized differently. Patients can be grouped by diet scheme (eg, ketogenic or low-fat and low-calorie) to generate different recommendations [62], or dietary habits, nutritional content, and patient profiles can be extracted from EHRs and lifestyle data [63]. A commercial example is Twin Health, a DT platform for T2D that enables patients to manage their diet and glucose levels. The platform integrates a food database, from which the user selects consumed food components, and the user must input the type and quantity to extract nutritional data (eg, calories, macronutrients, glycemic index, and glycemic load). Impactful extracted features are stressed to be carbohydrate-to-protein ratio, time since the last meal, and glucose trends [28,30,56,81]. Other approaches include automatic detection of meal and meal-related context, such as meal timing from raw glucose and insulin data [77] and estimation of carbohydrate and nutrient intake from food images (eg, using the VGG-16 (Visual Geometry Group 16 Layer) or a convolutional neural network) and meal quantity or recipe data [23,27,60,64]. These models can be refined by user feedback, behavioral inferences, alerts, and context to enhance personalized recommendations [23,56]. To streamline data collection, the literature proposes using heuristic-based event labeling to automatically categorize events and generate context [5]. Finally, food can be digitally replicated using mathematical models and embedded food sensors that extract nutritional data. These replications can interact with the HDT, enabling precise dietary recommendations while structuring food to align with health needs [63].

Regarding data fusion, after collecting all available data, the data are fused and integrated. Wang et al [61] further divide daily data into 7 time periods based on pre- and postmealtime and bedtime. Then, for each time period, feature vectors are created with the fused datasets. Transformers are further approaches to encoding patients as vectors.

To ensure holistic representations, contextual, historical, and stream data must be efficiently processed using robust infrastructures [35]. The data storage framework and architecture primarily rely on Internet of Medical Things (IoMT) technology, edge computing, and cloud computing. Edge computing reduces latency and enables real-time decisions. Cloud computing optimizes resources to achieve variable-cost-driven access and ensure effective service delivery [80]. A typical architecture using IoMT technology can comprise a communication, perception, transfer, and application module. The perception layer includes the sensors, while the transfer layer facilitates data exchange via edge nodes, storing information locally in an SQL database and syncing with a NoSQL cloud database. A web interface visualizes patterns, and the application layer implements the DT on the edge. Additionally, conceptual approaches emphasize 3 core requirements to aggregate real-time data. First, automatic detection and integration of available physical devices within a network is required, typically facilitated with mobile apps. The second is interoperability, enabling seamless detection, composition, and synchronization of DTs across domains for efficient data exchange. This can be enabled by using a cloud-based, multipurpose backend service that stores data in a permission-based SQL database. The third is fidelity, ensuring accurate representation of the physical entity through dynamic data collection, processing, and contextual updates. A context-monitoring feedback loop refines health status using adaptive structural models [5,34].

Regarding data modeling, data modeling defines the data concept, requiring semantic structuring and a shared vocabulary to represent domain knowledge [78]. For instance, ontologies represent entities in machine-readable form and can model domains using knowledge graphs (KGs) [45]. For DTs in diabetes, studies use ontologies and KGs to provide templates for mechanistic models, identify causal relationships, and extract disease-related features for ML applications. They can identify relationships between biomedical entities (eg, genes, proteins, metabolites, drugs, and clinical phenotypes) and diagnose health states. In addition, disease progression can be modeled by exploring indirect links among predictive features and diseases [25,47]. Across the literature, personal KGs are generated from EHR and continuous wearable data. Proposed frameworks support dynamic and bidirectional data mapping using the global-local-as-view architecture. Mapping accuracy can be further improved by refining data integration with conditional random fields [47].

Besides patient profile KGs, generalized metabolic flux (GMF) maps can be created, in which metabolites are grouped into generalized fluxes based on biological processes. Nodes represent key components such as glucose and lipid metabolism, metabolites, and physiological parameters, while edges represent the generalized fluxes. The framework can model metabolic conditions as outcomes of lifelong changes in metabolic fluxes driven by biomolecular pathways and gene networks. Surian et al [40] and Batagov et al [48] also use statistical methods, including the Wilcoxon-Mann-Whitney test, to analyze associations among metabolites, flux vectors, and diabetic complications. Then, path queries, profile comparisons, and correlation analysis identify features linked to specific complications and disease progression, such as retinopathy, ophthalmic complications, and chronic kidney disease (CKD). Time-to-event analysis further reveals differences in progression rates across risk categories (eg, low, moderate, and high severity).

KGs can also support explainability by revealing relationships between predictive factors and disease outcomes. For instance, the relationship between weighted features and disease progression can be explained using graph analysis. Zhang et al [25] derive a subgraph from the Scalable Precision Medicine Open Knowledge Engine to connect the most predictive features and the T2D-related node. Scalable Precision Medicine Open Knowledge Engine is an ontology integrating data from various biomedical databases across diverse domains [82]. Additionally, the Topic PageRank algorithm highlights graph features related to the target features of interest [25]. The network parsimony-based shortest-path method can further reveal insights into associated drugs and proteins in patients with T2D [25].

Finally, the framework should adopt unified medical standards, such as HL7 (Health Level Seven), the international standard for clinical software that enables the exchange and communication between medical devices or applications.

Process representation models physiological and biological processes using mathematical and statistical methods. Studies usually rely on mechanistic models that use mathematical functions and data assimilation to simulate metabolism, allowing parameter adjustments for therapy titration and the identification of changes in biological features [24,25,42]. Their key advantage is interpretability, as the relationship between input parameters and outputs can be displayed, improving awareness [42,54,83]. While most proposed models are hybrid and iteratively refined using data-driven optimization functions, we do not consider them data-driven because they primarily rely on mathematical functions.

This section synthesizes the process representations identified across the literature and illustrates how biological processes are modeled for diabetes-related DT systems. First, individual simulation modules are presented. The simulation and optimization methods are then depicted in detail. The synthesis is based on all the retrieved literature and does not separate between diabetes types. However, practical implementation of the process layer should be tailored to individual patient characteristics, disease state, and diabetes subtype.

The literature reveals that diabetes-related DT simulators model key processes of insulin, meal, and exercise impacts on glucose through various submodules, as shown in Table S3 in Multimedia Appendix 2. The modules can be classified into insulin, glucose, and glycogen, drug, exercise, the inflammatory system, the gastrointestinal tract, and metabolism. The following synthesis integrates the identified modules from the literature to illustrate the possible components of a holistic metabolism for a diabetes-related DT. The individual studies usually model only a subset.

Regarding the insulin module, it models insulin kinetics, dynamics, resistance, and sensitivity across tissues to simulate insulin requirements and their impact on glucose levels. Insulin kinetics determine plasma concentration, while dynamics describe its glucose-regulating effects, based on insulin mass, infusion rate, absorption time, sensitivity, and effectiveness [84]. This module can model basal and bolus insulin inflows separately, typically using nonlinear differential equations [85]. Across the reviewed literature, insulin processes are represented at multiple biological abstraction levels, ranging from physiological compartment models to cellular levels. Insulin inflow into the plasma is modeled from pancreatic secretion, while outflow is represented by liver clearance [54,85]. In particular, short-acting insulin is absorbed into the plasma, and time lags are estimated using coupled delay equations [85]. In the interstitial fluid, insulin inflow matches plasma outflow, while insulin outflow reflects cellular uptake [85]. Insulin resistance and diabetes progression are primarily modeled using ODEs at the tissue and cellular levels. These models incorporate organ-specific glucose uptake, intracellular signaling, protein expression, and adiposity-related changes [7]. A subcutaneous tissue compartment comprises absorption, dissociation, clearance, and distribution volume [79] and accounts for the time delay between insulin administration and its appearance in the blood [82]. Additionally, several studies model daily variations in insulin sensitivity using time-varying parameters, a variability control signal to capture circadian changes, or insulin-dependent glucose use and insulin action on endogenous glucose production [37,85,86]. To account for differences in insulin dynamics across diabetes types and disease stages, β-cell insulin production is set to 0 in models for T1D, reducing the dynamics to insulin absorption alone [65,85].

Regarding glucose and glucagon modules, this module simulates glucose dynamics integrating mechanisms of meal absorption rate, insulin-glucagon actions, and nonlinear glucose responses to hypoglycemia, including counterregulation [84,86]. Likewise, glucose inflows and outflows are modeled through different representations, including usage in fat, muscle, and adipose tissue. Moreover, in the gastrointestinal tract, studies simulate glucose appearance, renal and liver production, storage, and uptake [7,54,86]. In particular, nonlinear coupled differential equations can describe compartments such as the gut, plasma, interstitial fluid, and subcutaneous tissue [85]. Plasma glucose balance includes gut inflow, liver production, tissue uptake, and renal clearance, with insulin-dependent and independent uptake following Michaelis-Menten kinetics. Exogenous insulin inflow from short- and long-acting injections is also included by various studies [85,87]. The literature models hepatic glucose production using hyperbolic functions with asymptotic maximum rate or ODEs [65,87]. Context to improve CGM-based simulations can be further provided by an intestinal model that accounts for glucose diffusion from plasma into the interstitial space [66,79,86]. The interstitial glucose kinetics are described using a single-compartment linear model, adjusted for plasma-interstitial glucose gradients [79]. Glucose kinetics output the glucose absorption rate from meals, incorporating actions of glucose absorption, glucose-insulin kinetics, and glycogen kinetics and dynamics [77]. The glucagon submodule represents glucagon absorption, plasma action, clearance, secretion, and sensitivity [84,86].

Regarding the drug module, diabetes can be accompanied by multiple comorbidities, each with its own medication plan. For instance, weight-reducing drugs such as topiramate can be included in simulations of weight-related meal dynamics. One study proposes modeling topiramate response at tissue levels using ODEs within a compartmental pharmacokinetics model, in which energy intake is altered in response to the drug dose [7]. Other drug interactions remain underexplored.

Regarding the inflammatory system module, it can simulate insulin deficiency by modeling gene-regulatory mechanisms that lead to the destruction of insulin-producing beta cells [88]. For instance, it is proposed to simulate dynamic hormonal and immune responses using a multiscale discrete immune system model, in which key immune cells are represented as distinct tissues within a cellular immunology framework, each modeled as an agent. These are governed by rules describing responses to pathogens [88,89]. The module also examines the interplay between adipose tissue, inflammation, and diabetes, particularly in the context of high-calorie meal intake. It describes adipose tissue growth, leading to fat accumulation, the production of proinflammatory cytokines, and the development of an inflammatory state [88].

Regarding the activity module, it represents the personalized effects of PA on hormone regulation, interleukin-6 secretion, and weight changes [8]. Exercise is usually quantified by volume of oxygen, with the model simulating oxygen consumption, epinephrine, and glucagon-insulin secretion dynamics. Additionally, hormone and glucose reactions to different exercise types (eg, cycling, walking, running, and stepping) can be simulated [90]. In particular, the literature suggests that exercise types, such as resistance and cardiovascular training, should be modeled separately [24]. Plasma interleukin-6 dynamics, influenced by variations in oxygen uptake during exercise and by skeletal muscle and adipose tissue secretion, can be described using ODEs, with oxygen consumption serving as a measure of exercise intensity [91,92]. Studies also describe epinephrine release as it is correlated with pancreatic insulin and glucagon secretion [90]. Consequently, the exercise model can model changes in insulin sensitivity impacted by increases in peripheral insulin, liver and glucose production, active tissue glucose uptake, oxygen consumption, and active muscle mass [84].

Regarding the gastrointestinal tract module, this module models the effects of gastric emptying and meal composition on glucose metabolism and weight dynamics, which can be described using ODEs [7]. Stomach emptying is differentiated between glucose leaving the stomach, glucose in the jejunum, and glucose in the ileum [65,87]. For mixed meals, studies simulate the appearance rates of glucose, alanine, and triglycerides based on carbohydrate, fat, and protein intake [92]. Meal responses can be modeled as metabolic flows between organs, accounting for meal frequency, composition, ingestion rate, and body weight [54]. In particular, postprandial glucose dynamics incorporate delays between digestion and plasma glucose appearance [67,85]. The module also incorporates gut absorption, glucose-insulin and glucagon-insulin hormonal dynamics, glucose usage in different organs and tissues, hepatic glycogen storage, protein metabolism, and long-term metabolic dynamics [54,92]. The gastric emptying model delivers glucose to the gut, and the gut compartment accounts for nutrient absorption efficiency, appearance rates, and effective distribution volume [85,92].

Regarding the metabolism module, it simulates the glucose-insulin dynamics, integrating hepatic and renal responses. Across studies, key parameters include insulin action, glucose appearance rates, insulin sensitivity, and glucose-insulin meal interactions [37,54,79,87]. The liver compartment describing hepatic glucose metabolism includes a storage module that provides a continuous glucose supply, using glucose appearance rates from the gastrointestinal module [67]. Furthermore, studies model fasting responses, diet effects on glycogen, gluconeogenesis, and protein metabolism [54]. Additionally, relationships between metabolites and reactions can be described to estimate the evolution of biochemical pathways during disease progression [40,48]. Glucose-insulin-glucagon interactions, gut absorption, and insulin action can be represented using differential equations [66,68,84]. Single-hormone models, which can use ODEs, describe insulin kinetics and dynamics, as well as carbohydrate absorption, enabling the generation of virtual patients with varying insulin sensitivity. Dual-hormone models extend this by incorporating glucagon dynamics and kinetics [84]. The literature models exercise via insulin sensitivity adjustments based on oxygen consumption [84].

Process representation often involves metabolic simulators that model key physiological processes to educate patients and personalize treatment plans [37]. The literature shows that the core mechanism of the DT is grounded in mathematical functions fitted to individual patient data. Model parameters can be either global, representing fixed knowledge, or optimized using data-driven or mathematical approaches.

Commonly, disease progression, glucose-meal responses, glucose-insulin dynamics, and hypoglycemia risk are simulated. For instance, a metabolic network can be modeled using a generalized stoichiometry matrix, where linear systems can represent biochemical reaction fluxes, while scalar variables can identify time-dependent fluxes and metabolic changes [40,48]. Meal responses can be simulated, incorporating knowledge of human metabolism derived from clinical studies, to describe personalized responses to meal compositions and decrease the risk of diabetes onset [54]. Additionally, deterministic models based on linear discrete-time systems are proposed to simulate glucose-insulin dynamics. Such a model would require 4 interconnected modules representing the gastrointestinal tract, subcutaneous tissue, liver, and metabolism, and should also account for daytime insulin and glucose sensitivity [67,93]. Furthermore, a global ranking and collinearity analysis can be used to construct compact matrix representations of linear ODEs. After optimizing the main parameters and the measurement-error covariance matrix, the patient can be represented by a dynamic mathematical model [37]. Clinical applications would include optimizing insulin doses and diets using the virtual patient, time, and the meal plan [67,93]. Another use case is enhanced management of multiple daily insulin injections, especially relevant for children with T1D. Here, the literature simplifies the input parameter complexity of commonly used simulators, such as the UVA/Padova simulator, which models glucose-insulin-carbohydrate interactions [77,94].

Some simulators focus specifically on open-loop insulin optimization, simulating insulin dosing strategies (eg, dual-wave, split-bolus, dual-wave with fixed duration, and standard bolus) while estimating the probability of hypoglycemia [69]. For insulin optimization, glucose levels and response can be simulated iteratively, or multiple statistically described patient representations can be generated and simulated in parallel. These simulators should account for uncertainties in sensor errors, insulin pump delivery, and daily variability in food and insulin responses [5,69,77]. The insulin policy is commonly modeled using stochastic and convex optimization, designed to reduce the average peak BG and minimize constraint violations [69]. Dynamic adaptability is ensured by the model predictive control method, which computes optimal insulin infusion at each time step by minimizing BGL excursions and insulin costs while maintaining target glucose values [23].

Furthermore, simulators based on differential equations or linear affine models can enhance the performance of artificial pancreas (AP) systems. Studies have particularly emphasized the importance of incorporating a set of personalized, time-varying constants and multimodal inputs (eg, daytime, PA, food images, and patient records) [60,65]. Time-varying parameters should be used across all modules, as they describe intrapatient glucose variability [86]. Insulin adjustments should depend on the patient’s daily activity level [60]. Stability can be improved with a scaled error term to penalize implausible chromosome behavior or oscillations [65].

Finally, 1 study presents a metabolic language for a glucose simulator that can model multiweek scenarios. It enables physicians to create and refine physiological models based on medical terminology, which are then translated into mechanistic models. The presented framework initializes the simulated metabolism based on real patients and personalizes it using glucose time series, meal records, and medication data. Such a framework built on differential evolution algorithms can simulate individual real-world scenarios, personalized diet and drug treatments, and support disease management [70].

Parameters are generally grouped into static, which are extracted from validated studies; general, which are optimized worldwide; population-specific, which are tailored to the characteristics of the specific population or demographics; and individual parameters, which are optimized using mathematical or data-driven optimization algorithms [7,54]. Optimization methods can be categorized primarily as linear, nonlinear, statistical, or heuristic approaches [69]. Nonlinear optimization includes convex and quadratic optimization methods [40,48]. Statistical optimization encompasses Bayesian methods, expectation-maximization, semiparametric regression optimization [23], and probabilistic techniques such as Markov chain Monte Carlo [5,77]. Finally, heuristic or metaheuristic approaches include scatter search [7], genetic algorithms, and particle swarm optimization [65,70].

To improve prediction accuracy and prioritize recent glucose values, the literature incorporates a forgetting factor. Further optimization and stability are achieved through slack variables and constraints, such as recommended BGL and insulin infusion limits [23]. Studies also address variability and uncertainty through fuzzification and probabilistic or stochastic elements, since human biological processes are dynamic and less deterministic [67,93]. For instance, a Profile Likelihood method can assign variation to unmeasured factors like stress, exercise, or meal intake [85].

In conclusion, simulation processes are usually represented using linear, nonlinear, or statistical methods. Methods can also be classified as deterministic or stochastic [66]. Most studies adopt hybrid methods, in which nonfixed parameters are optimized and tailored to the data and context. Finally, many studies already integrate several physiological abstractions, including organ, tissue, cellular, and behavioral processes, to model a whole-body system [7,8,24].

Across the reviewed literature, data-driven approaches use statistical and ML models to detect patterns, predict anomalies, and assess disease progression. These models support decision-making systems, further enhanced by explainable AI (XAI) [47]. By continuously learning from historical and real-time data, they enable precision medicine through personalized therapy recommendations, refining predictions over time to improve accuracy and adaptability. Furthermore, complex relationships within the data can be uncovered [44]. A key characteristic is the automatic refinement of models with incoming data or user feedback, dynamically selecting the best-fitting model and integrating improved alternatives. This adaptability defines them as closed-loop DTs [5,43].

This section presents use cases of data-driven representations and categorizes existing approaches into disease diagnosis and progression prediction, glucose forecasting and simulation, and lifestyle recommendations. Common use cases and data-driven methods are summarized in Table S4 of Multimedia Appendix 2.

DTs can be used to estimate the risk of disease onset and progression to diabetes based on data trajectories over time. Notably, the temporal evolution of biomarkers associated with T2D onset has been described using multi-input, multioutput models. For instance, a multivariate Gaussian process model with an autoregressive structure, accounting for interdependencies between outputs, is proposed. It contains input parameters for a specific time window and output parameters for the next time window, enabling the comparison and characterization of biomarker dynamics. Additionally, patients can be categorized into high- and low-risk groups based on their biomarker evolution, enabling individual risk assessment years before onset [95]. If already diagnosed with diabetes, disease progression and the onset of comorbidities can be predicted. Binomial logistic regression models trained on demographics and metabolic fluxes can provide predictive insights into baseline identification and the future onset of complications such as CKD, diabetic retinopathy, and cataract. By incorporating correlation analysis, studies identify key parameters contributing to elevated GMF profiles. The literature further evaluates disease progression by analyzing correlations between health-state distance metrics and patient diagnoses, while distance-based classification methods enable stratification of patients into high- and low-risk groups for increased personalization [40,48]. Usually, regression-based models forecast changes in parameters. In contrast, binary classification can classify changes in value (eg, of at least 5%), leading to lower variance [25]. Moreover, statistical methods such as Kaplan-Meier analysis and Cox proportional-hazard models can estimate progression rates, showing the probability of remaining healthy or the risk of progression of a comorbidity, respectively [40,48].

To diagnose diseases using multitarget binary classification, a hybrid model consisting of ML and neural networks is proposed to predict health care metrics such as peripheral capillary oxygen saturation, body temperature, and diabetes [80]. By including multimodal inputs such as genomics, CGM device data, and wearable data, additional comorbidities can be predicted before onset, and interventions can be tailored. In particular, extending models with neuroimaging can forecast dementia risk, which is correlated with diabetes [96]. Image classification techniques can detect and categorize diabetic retinopathy based on severity (mild, moderate, or severe). Model confidence is increased with an integrated feedback layer, enabling HPs to provide input, facilitating continuous model refinement and improved diagnostic performance [57]. Another proposed use case for diabetes-related DT systems includes chronic wound management, a common complication of T2D, requiring periodic examinations and continuous care. Via image processing and AI models, the concept predicts healing trajectories to assist clinicians in determining treatment plans. Such a framework should support the entire model lifecycle, from creation through experimentation to continuous monitoring and refinement [58].

The main research and application domain is glucose forecasting and hypoglycemia prediction using neural networks and deep learning (DL). Across the literature, glucose values were forecasted using different prediction horizons of short-term (5 to 60 min) or long-term (24 h). For short-term predictions, models based on long short-term memory (LSTM) or its hybrid variants (eg, gated recurrent unit–LSTM and convolutional neural network–LSTM) are most often applied to multisensor inputs or multimodal data [14,23,64]. For daily glucose predictions, models based on artificial neural networks [62] or GANs using recurrent neural networks with a combination of supervised and unsupervised learning to learn temporal dynamics in latent spaces are presented [47,59]. Furthermore, studies apply multilayer perceptron models to predict glucose levels using transformer-based vector encodings of the input data [61].

Besides glucose forecasting, blood peaks can be forecasted with hybrid models that use a CatBoostRegressor for categorical variables, a random forest model to model the nonlinear relationship between meals and glucose, and an LSTM to capture temporal patterns [56]. Thus, DL-based glucose simulators can be generated to predict glucose concentrations for the next day using sensor input, lifestyle behaviors, and the patient’s individual characteristics. Simulators can be used to explore new treatments by modifying conditional inputs and predicting corresponding glucose responses, thereby enhancing decision support and awareness [14,59]. With threshold-based comparison, the forecasted glucose can alert to predictive adverse events, predict the risk for hyperglycemia or hypoglycemia, and the time spent in target ranges [14,61,95]. Moreover, structured TSA is used to detect hyperglycemic and hypoglycemic patterns [23]. Further insights are provided in reviews of different algorithms that do not incorporate a DT framework [51,97,98].

Based on predicted glucose data, the literature generates further lifestyle recommendations, augmenting decision support. For instance, data-driven insulin titration policies are proposed based on reinforcement learning. Methods can include the soft actor-critic with entropy-driven reward functions or reward functions based on target glucose levels [14,47,61]. Dosing strategies can be continuously refined and improved using data-driven methods that integrate patient representation, real-time data, and other relevant predictions [61]. Another research field is nutrition and diet suggestions. After analyzing food intake, studies outline unsuitable choices to optimize nutrient balance and prevent over- or underconsumption, such as adding vitamins or proteins [39]. Personalized dietary recommendations can be enhanced by integrating patient profiles, behavioral factors, and DTs of food generated through mathematical modeling and simulation of meal components. Notably, ML models are used to identify diet patterns, forecast optimal food choices, and reduce diet-related disease risks [63]. Moreover, a KG embedded with logic rules that provide the patient’s management plan, allergies, and diet preferences can reason whether a meal meets the requirements [47]. Methods can also predict weight with neural networks and be optimized to target weight with particle swarm optimization algorithms [62]. The predicted blood peaks can personalize diet recommendations using a multiobjective optimization strategy that considers glucose profiles, nutrition, activity, sleep, and stress, to minimize glycemic variability while maximizing nutritional quality [56]. Further use cases include drug recommendations, where the disease is first identified from symptoms and drugs are then suggested, aligned with those symptoms using ML models [39].

Lastly, XAI is introduced as a submodule. XAI can explain glucose trajectories or provide insights into the parameters that lead to adverse events. For instance, an XGBoost (Extreme Gradient Boosting) classifier, serving as a base learner, enhanced with the LIME (locally interpretable model-agnostic explanations) tool, is used to explain predicted adverse events to improve self-management of diabetes [23]. Furthermore, risk factors for the onset or progression of diabetes can be prevented with personalized countermeasures. Counterfactual explanations, which are part of local XAI methods, predict outcomes based on modified input parameters to explain why the model made a specific decision, identifying necessary changes leading to the optimal outcome. The counterfactual explanations can be assessed using biomarkers extracted from EHRs and clinician survey responses [8,99]. Another proposed method is based on a Gradient-Weighted Class Activation Mapping model. It identifies key factors contributing to adverse events and allows iterative adjustments to test intervention impacts virtually [14]. Finally, after generating predictions and recommendations, large language models based on a generative pretrained transformer can provide explanations for their predictions and suggestions in a user-friendly report [5].

This section highlights promising studies that can be defined as preliminary DTs or implemented DT approaches, whereas DT-enabling components are not considered. A detailed analysis of multilevel representation frameworks is provided in Table 1. Table S1 in Multimedia Appendix 2 summarizes all studies that present themselves as DTs, including single representations, but excluding concepts and simulators that do not provide new insights or have been previously discussed. Most studies do not propose holistic frameworks and cannot tailor their model to the patient. In particular, the generation of patient profiles and the representation of patients as virtual replicas are frequently neglected. Data representation accounts for only 25% (7/28), which requires data organization and structured management that goes beyond basic data collection and preprocessing. In addition, only 5 studies include model refinement for continuous adaptation and improvement. The distribution of process- and data-driven representations is balanced, with 64% (18/28) and 61% (17/28), respectively. Analyzing the individual studies, single-level abstractions are addressed in 64% (18/28) of studies, with the majority using methods for process representation, followed by data-driven approaches, whereas data-only representations are not found. Two-level abstractions, primarily integrating data and data-driven representations, account for 21% (6/28) of studies, while combinations of data and process representations are less common. In this context, mechanistic models optimized using data-driven techniques are not classified as 2-level representations. Finally, 3-level representations are covered by 14% (4/28) of the studies.

According to Table S1 in Multimedia Appendix 2, simulators are the most common components. Simulating processes and trajectories is essential to forecasting events and adjusting treatments. Nevertheless, they do not qualify as full DTs on their own. Data are usually collected in cloud storage and visualized through smartphone apps. However, these data are often not further processed, structured, and modeled for knowledge representation or AI-ready formats. Key use cases of all the retrieved literature include virtual trial simulators, which refine therapy parameters by predicting long-term outcomes, generating synthetic data, forecasting glucose trajectories or insulin responses, and aiding in selecting a healthy lifestyle. Additionally, DT models are frequently used to predict disease onset and progression. Table 1 reveals that, in multilevel DT approaches, insulin optimization is usually addressed.

Proposed 2-level architectures were usually classified as intermediate DTs because they do not fully implement and integrate multilevel representations. These approaches differ in maturity and clinical scope across the literature. Paglialonga et al [8] use complementary process- and data-driven abstractions to enhance functionality for a specific clinical use case. They extend the process representation toward a multiorgan, metaflammatory system. Data-driven approaches estimate the risk of T2D onset using personal data, forecast progression to T2D onset, and define personalized recommendations [8,95,99]. The process representation integrates a gastrointestinal tract module, a PA model, and hormonal responses to evaluate the impact of the individual’s metaflammatory status on developing T2D risk [8]. This work presents a framework in which subcomponents have been tested with real clinical data but not integrated into a single meta-architecture [73,95]. In contrast, Lee et al [9] integrate both representations for separate applications. A cloud-based data collection platform facilitates communication between patients and HPs, enabling feedback, titration, and education [9]. However, in the reported architecture, this layer primarily functions as a storage and data-sharing platform rather than as a structured virtual representation of the patient. As no explicit semantic, relational, or patient-specific data model is described, this layer was not classified as a data representation within our framework. They described a stochastic simulation framework that models diabetes-related conditions with separate transition probabilities and management strategies for T1D and T2D, which is notable because most reviewed approaches focus on a single diabetes subtype. The process representation estimates interactions among diabetes-related complications at the individual-patient level [100,101]. In parallel, a T1D-focused component uses wearable and personal data (eg, CGM, insulin pump, food-tag, and activity) to support insulin and lifestyle management [9]. Besides the data collection platform, the proposed architecture remains largely conceptual and is not presented as a fully implemented or clinically validated system. Furthermore, 1 study considers personalized multiple daily insulin adjustment in children with T1D. Cappon et al [5] model the glucose-insulin metabolism and generate therapy suggestions. The process representation involves twinning, simulation, and recommendation procedures using an open-source methodology [77] to represent a person’s physiology. They present a data layer focused on AI readiness, including data preprocessing, labeling, and context extraction, but do not report a patient representation layer, which is why we did not classify their method as a full data representation. The system integrates all modules to provide interaction with data and HPs. Similarly, the data-driven representation is not fully developed and is used only as an LLM to generate suggestions and provide user-friendly text for HPs and users [5]. However, as heterogeneous data is used and the system aims for personalized models with multiple subcomponents, it was not considered a single simulation framework. This system comprises partly implemented components, but most are not validated with real clinical data, and the meta-system remains a framework. Likewise, Keshary et al [60] present a personalized insulin optimization but for AP systems in geriatric patients with T2D. The system combines a patient simulator with data-driven meal recognition from images and activity-aware insulin optimization. Individual components were evaluated in silico, while the patient simulator and insulin optimization were further computed using real clinical data. Although the study is not explicitly framed as a DT system, it was included as a DT-related prototype because it combines multimedia data collection, patient-specific simulation, and individualized insulin optimization [60].

Data and data-driven representations typically structure information using ontologies and preprocess data for ML submodules. In this context, Zhang et al [25] develop a T2D-based DT model integrating demographic, clinical, and multiomic data structured as KGs, which should provide data for mechanistic and data-driven models. The KG maps biomedical relationships, identifies unknown parameters, and interprets predictions. Additionally, clinical trajectories and progression to CKD are predicted using ML. The models are validated with real clinical data. However, as the data structures provide domain-specific information rather than representing patient data, they do not yet qualify as a comprehensive data representation. Moreover, the process representation is only mentioned as a conceptual framework. In contrast, Rad et al [47] develop a hierarchical, top-down, personalized ontology tailored to diabetes (types 1 and 2) and use ML models with continuous refinement to uncover complex patterns, disease risk factors, and health trajectories. Proposed applications include personalized insulin adjustments to minimize hypoglycemia risk, 24-hour glucose forecasting, KG exploration, and logic-driven personalized meal recommendations. Nevertheless, this approach lacks a process representation. Furthermore, the predictive modules are only validated with virtual data, while the integrative meta-system was not fully implemented. Using the same representation levels, Vaskovsky et al [55] and Vaskovsky and Chvanova [74] differ in the data format and use case. They present a conceptual framework for personalized nutrition targeting individuals with a genetic predisposition to diabetes based on tabular data. The system leverages a nutrition knowledge base with machine-rule-based logical inferences, structured according to HL7 standards. First, the framework creates a twin based on individual data (eg, EHR data, taste preferences, and individual goals and restrictions), and synthetic patient groups are computed. Second, the twin of the food product is created from the given meal components. Third, an inference machine analyzes the effects of food items on digestion, absorption, and metabolism. Finally, personalized diets are recommended by modeling interactions among genetics, microbiomes, immunity, and metabolism, influenced by diet and activity.

Lastly, Shamanna et al [56] propose a fully implemented system that generates patient profiles and uses AI for prediction. In contrast to Vaskovsky et al [55,74], Shamanna et al [56] propose a DT system for T2D comprising a data management, prediction, and recommendation layer, with the data organization unspecified. A personalized metabolic model is created using wearable data, dietary logs, and biometric inputs. The data-driven representation consists of rule-based expert systems and AI models using vital signs, bloodwork, nutrition, and demographics [28,81]. Similarly, the prediction model evaluates the impact of a meal. Glucose peaks are predicted and penalized for medication use to account for artificially reduced levels. Whereas Paglialonga et al [8] use redundancy by combining process- and data-driven abstractions, this system introduces redundancy within the data-driven layer for glucose-response prediction. Personalized decision support is provided for medication, diet, and exercise [28,56,81]. A feedback layer enables continuous refinement, adapting to new data, trends, and cohort-based models [56]. Finally, the telemonitoring module enables HPs and coaches to monitor progress, manage medication, and provide support. This system has already been used in randomized controlled trials [28,56,81].

In general, the approaches remain conceptual frameworks, are mostly tested sparsely or in silico, and are neither fully implemented nor ready for clinical trials or clinical use cases, except for the system of Shamanna et al [56].

For three-layered representations, the included literature reports implemented (retrospective clinical) proof-of-concept studies with clinical data, whereas no included architecture was used in clinical trials. Moreover, across studies, the role of the representation layer differs. For instance, Wang et al [61] create an insulin policy model but do not incorporate biological processes. They present a DT model for dynamic insulin dosing for patients with T2D. A patient model encodes patient-specific data from EHRs, historical data, and continuous data using NLP and a transformer, which tracks evolving states, generates transitions, and estimates rewards from past trajectories. The module computes hidden states and internal patient conditions, and data-driven methods forecast glucose levels, target-range compliance, and the following patient state. Interacting with the patient model, the policy model optimizes the insulin dose based on the current state [61]. While Thamotharan et al [23] rely on sparse ML models as additional or complementary components rather than as the main simulator, they also focus on personalized insulin infusion. They present a DT adapted to older adult patients with T2D using contextual and geriatric data. The module assesses key indicators such as time in range, adverse events, and model accuracy. Mechanistic equations describe glucose-insulin interactions and are optimized to patient-specific parameters. Likewise, personalized insulin infusion is supported through recursive refinement of glucose, insulin, and carbohydrate dynamics. Furthermore, glucose is forecasted with data-driven models. In contrast to Wang et al [61], their system is expanded to multiple subcomponents. The prediction module integrates glucose forecasting and structured TSA. A matrix profile detects recurring motifs, anomalies, and inter- and intrapatient patterns affecting glucose trajectories. Then, anomalies, sensor errors, and influential activity patterns are detected. It further integrates XAI to interpret hyper- and hypoglycemic events. Finally, physicians’ recommendations and patient conditions are integrated for improved diabetes management [23]. Similarly, Surian et al and Batagov et al [40,48] are classified as mechanistic physiological-biological approaches supported by data-driven methods. Both systems rely on the same underlying representations and principles, but extend the DT to different use cases. The DT is based on a metabolic flux network, in which metabolic reactions and pathways are represented through GMF analysis. The model is initialized with patient-specific metabolite and physiological data, while metabolic states are formalized through a stoichiometric matrix fitted to clinical data. This enables personalized prediction of metabolic states and disease trajectories, as well as the identification of comorbidity-associated patterns. Health state distances measure disease progression and identify patient subgroups at high or low risk. In addition, data-driven methods incorporating demographic and metabolic flux data are used to predict disease onset and progression, quantifying risk using health-state distance metrics [48]. While Batagov et al [48] investigate progression to cataract and retinopathy within 3 years, Surian et al [40] predict CKD over 3, 5, and 10 years, classifying patients into high-, moderate-, and low-risk groups. They enhance the model with nutritional and respiratory data and use it to model microvascular complications leading to CKD, incorporating medication effects. Here, mathematical models structure the patient’s state and changes, while data-driven methods forecast disease risk.

In summary, 3-level representations aim either to optimize insulin or to predict disease onset and progression, while emphasizing personalization and incorporating context and person-specific information.