This scoping review was conducted in accordance with the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) and JBI methodological guidance for scoping reviews [19,20].

A comprehensive literature search was conducted through PubMed, IEEE Xplore, Scopus, and Web of Science. Studies published up to April 2025 were considered, and all databases were last searched on June 23, 2025. The search strategy combined terms related to “digital twin,” “diabetes,” and “healthcare” using Boolean operators. The detailed search strategy is provided in the Multimedia Appendix 1. Reference lists of included studies and relevant reviews were also manually screened to identify additional records.

The search strategies developed for PubMed, Web of Science, IEEE Xplore, and Scopus were imported into the Triple-A (Article Analysis Assistant) software [21]. The tool was used to integrate bibliographic metadata, automatically remove duplicate records based on DOI, and perform additional deduplication using title, publication year, and author names. Reviewer decisions were subsequently imported into the platform, and the finalized dataset was prepared for downstream analysis and thematic synthesis.

Eligibility criteria were established a priori to ensure consistency and reproducibility during screening.

The selection process involved 3 stages—identification, screening, and eligibility assessment. We initially identified 208 studies from 4 major databases—PubMed (n=47), IEEE Xplore (n=1), Scopus (n=107), and Web of Science (n=53). During identification, 85 articles were excluded due to duplication, lack of an abstract, absence of original data, or being published in a language other than English.

Following this step, 123 articles proceeded to screening. At the screening stage, 84 articles were excluded according to the predetermined exclusion criteria. As a result, 39 articles advanced to eligibility assessment, and 28 were included in the final review [3,7-18,22-36]. The 11 full-text articles excluded at the eligibility stage and the reasons for exclusion are listed in Multimedia Appendix 2.

Screening was conducted in 2 stages:

To ensure consistent inclusion decisions, the following screening questions were applied, reflecting the key characteristics of DTs and their application in diabetes care (Table 1). For the purposes of this review, a study was considered to describe a DT if it included a patient-specific virtual representation or individualized computational model linked to diabetes-related data and intended for prediction, simulation, monitoring, or decision support. Studies using terms such as “virtual patient” or “simulation model” were included only if these DT-defining characteristics were present. Generic population-level models without individualized representation or diabetes-specific application were excluded.

Data from the 28 included studies were charted using a structured framework guided by the 13 predefined research questions introduced in the Introduction section. These research questions were organized into seven thematic domains to facilitate systematic synthesis: (1) system design and modeling foundations (RQ1, RQ2, and RQ3), (2) target conditions and use context (RQ4 and RQ5), (3) data sources and personalization mechanisms (RQ6 and RQ7), (4) intelligence and adaptability (RQ8), (5) evaluation and validation (RQ9 and RQ10), (6) implementation and governance (RQ11 and RQ12), and (7) research and development gaps (RQ13).

Each included study was analyzed systematically using this framework. Categories were not mutually exclusive, and individual studies could be charted under more than 1 category where appropriate. The full study characteristics and data charting table are provided in Multimedia Appendix 3.

Consistent with scoping review methodology, formal risk-of-bias, reporting bias, and certainty-of-evidence assessments were not performed because the aim was to map the breadth, characteristics, and gaps in a heterogeneous body of literature rather than to compare intervention effects or generate pooled estimates.

A review protocol and project materials for this scoping review were made available through the Open Science Framework (OSF) [37].

Across the 28 included studies (as shown in the PRISMA [Preferred Reporting Items for Systematic Reviews and Meta-Analyses] flow diagram in Figure 2) [3,7-18,22-36], DT systems for diabetes exhibited diverse architectures, data sources, and application goals. Most models were data-driven or hybrid (artificial intelligence [AI]+mechanistic), while purely mechanistic and conceptual designs were less common. Core system components included machine learning (ML) or AI modules, decision support layers, and real-time simulation engines. The majority of DTs leveraged CGM, wearables, and lifestyle data, with increasing use of patient-specific models to enable personalized therapy, behavioral nudges, and simulation-based feedback.

ML was the dominant modeling approach, while reinforcement learning, control theory, and signal processing appeared less frequently. Strategies for uncertainty management and interpretability were adopted inconsistently, with adaptive learning and explainable AI used in some studies, but with limited human-in-the-loop oversight. Reported outcomes most often focused on glycemic control (eg, hemoglobin A_1c [HbA_1c], time-in-range [TIR], and reduced hypoglycemia), alongside improvements in predictive accuracy, metabolic markers, and patient engagement. However, external clinical validation remained scarce, with most evaluations based on retrospective datasets or simulations.

Ethical considerations—mainly privacy and transparency, with occasional references to accountability and bias—were inconsistently addressed. Implementation barriers included validation limitations, data quality issues, model limitations, and workflow misalignment. Finally, the literature highlights persistent research gaps in integration with real-world systems, scalability, and methodological rigor that must be addressed to advance DT systems into clinical use.

Although DT systems for diabetes are showing technical feasibility, multiple areas require further investigation and refinement. From the 28 reviewed studies [3,7-18,22-36], seven major gap categories that emerged were (1) limited scope of application, (2) integration challenges, (3) lack of longitudinal data, (4) data quality and availability, (5) methodological limitations, (6) need for clinical validation, and (7) scalability or usability concerns. Table 14 summarizes the reported research and development gaps in diabetes DT systems.

Key findings included:

DT systems for diabetes use a wide range of modeling techniques, most commonly ML (eg, long short-term memory, gradient boosting, and reinforcement learning) and physiological simulation. Simulation engines and predictive ML modules were often integrated into layered architectures that also included personalization modules, decision support, and user-facing dashboards. Statistical and probabilistic methods (eg, regression and Bayesian inference) were also used in several studies, although less prominently. Few systems incorporated mechanistic control theory or signal-processing models. The inclusion of key components, such as simulation engines, control-feedback modules, and data integration pipelines, reflects a growing maturity in system design.

Most DTs targeted T1D or T2D, with limited applications in gestational diabetes or diabetes-related complications, such as retinopathy. Primary clinical goals included glycemic prediction, insulin-dose optimization, lifestyle guidance, and therapeutic planning. Several systems also addressed the diagnosis of complications or risk stratification for comorbidities. The breadth of clinical use cases suggests that DTs are evolving from simple simulators into multifunctional clinical-support tools.

Lifestyle data, wearable devices, and CGM were the dominant inputs, with hybrid combinations being common. EHRs and synthetic datasets were also widely used to provide historical or simulated information. Personalization was achieved through mechanisms such as real-time adaptation, individual model tuning, behavior-driven feedback (eg, nudges), and insulin titration. However, persistent challenges remain in data quality, sensor integration, and dataset heterogeneity.

Managing uncertainty and real-time updates is crucial for clinical reliability. Studies implemented adaptive learning, feedback loops, and explainable-AI methods (eg, attention mechanisms and knowledge graphs) to improve transparency and adaptability. Real-time CGM synchronization and, in some cases, human-in-the-loop oversight were used to enhance model responsiveness and safety.

Quantitative validation (eg, RMSE and AUC) was common, but real-world clinical trials were rare. Most studies validated systems via retrospective datasets or simulations. Reported clinical outcomes included improved TIR, fewer hypoglycemic events, and, in some cases, T2D remission. However, evidence on long-term effectiveness, generalizability, and cost-effectiveness remains limited.

A notable finding across the included studies is the mismatch between technical sophistication and clinical maturity. Although many DT systems incorporated adaptive learning, individualized simulation, and multimodal data integration, most were evaluated using retrospective datasets or in silico simulations rather than prospective clinical deployment. This likely reflects the high implementation burden of DTs in diabetes, including the need for reliable real-time data streams, safety safeguards, interoperability with devices and clinical systems, and acceptable workflow integration. It also reflects the regulatory complexity of systems that may influence insulin dosing or therapeutic decision-making.

Privacy and ethical considerations were addressed inconsistently, often limited to brief compliance mentions (eg, GDPR and HIPAA). A smaller subset of studies explicitly discussed accountability (eg, audit trails and governance mechanisms) or algorithmic bias and fairness, highlighting underexplored areas of governance. Implementation enablers included real-time feedback and sensor integration, whereas barriers included poor data quality, system complexity, lack of clinical workflow alignment, and limited scalability.

Another important finding is the limited attention to algorithmic bias and fairness. Despite the increasing use of AI-driven modeling and decision-support approaches, only a small subset of studies explicitly discussed bias, representativeness, or equity-related concerns. This suggests that the field is still focused primarily on technical feasibility and predictive performance rather than equitable deployment across diverse patient populations.

Key gaps include limited clinical validation, insufficient longitudinal data, a lack of standardized model architectures, and limited generalizability to diverse populations. Many studies emphasized the need for integration with EHRs, real-world testing, and regulatory alignment. Addressing these gaps will be essential to enable scalable, equitable, and clinically robust DT systems for diabetes management.

This review offers a panoramic view of the evolving DT landscape in diabetes. While notable technical advances are evident—particularly in data integration and personalization—the field remains formative, with substantial work needed in clinical validation, ethical governance, and system interoperability. Future research should emphasize not only algorithmic sophistication but also real-world applicability, safety, and equity to support the scalable and responsible deployment of DTs in diabetes care.

Taken together, the literature suggests that DT research in diabetes is progressing from conceptual and simulation-based work toward more clinically relevant systems, but the field remains early in real-world maturity. Future studies should prioritize prospective validation, broader demographic and clinical representation, transparent reporting, interoperability with routine care systems, and governance frameworks that address privacy, accountability, and fairness.

This scoping review has several limitations. First, only English-language studies with accessible full text were included, and gray literature was excluded, which may have led to the omission of some relevant studies. Second, formal risk-of-bias and certainty-of-evidence assessments were not performed because the aim was to map a heterogeneous body of literature rather than evaluate intervention effects. Third, the included studies differed substantially in design, terminology, validation methods, and outcomes, limiting direct comparison. Finally, many studies were early-phase, retrospective, or simulation-based, which limits conclusions about clinical effectiveness and real-world implementation.