Automated insulin delivery in type 2 diabetes: a meta-analysis of randomized trials protocol Authors: Luis Gordillo-Rosas, Cindy Lee Cajachagua-Pucuhuaranga, Julián Santa Cruz-Venegas, Luis Mateus Urrutia-Osorio, Samir Pantoja, Jimena Lozano-Núñez, Oriana Rivera-Lozada, Cesar Bonilla-Asalde, Joshuan J. Barboza Published: 2026 ABSTRACT BackgroundType 2 diabetes mellitus (T2DM) is a highly prevalent chronic metabolic disorder, and a substantial proportion of patients eventually require insulin therapy, which is frequently complicated by therapeutic inertia, hypoglycemia and glycemic variability. Automated insulin delivery (AID) systems combine continuous glucose monitoring with algorithm-driven insulin delivery, but their efficacy and safety in adults with T2DM remain uncertain. The objective of this systematic review and meta-analysis is to evaluate the efficacy and safety of AID systems compared with conventional insulin-based strategies in adults with T2DM.MethodsThe conduct of this review will follow the Cochrane Handbook for Systematic Reviews of Interventions (version 6.5) and reporting will adhere to the PRISMA 2020 and PRISMA-P 2015 statements. Randomized controlled trials (parallel-group and cross-over) evaluating AID systems in adults with T2DM will be identified through a systematic search of MEDLINE (via PubMed), EMBASE, Scopus, the Cochrane Central Register of Controlled Trials (CENTRAL), and Web of Science from inception to 30 November 2025, without language restrictions. The strategy will combine controlled vocabulary, free-text terms, and the Cochrane Highly Sensitive Search Strategy for identifying RCTs; will be developed with input from an expert health-sciences librarian; and will be peer-reviewed using the PRESS 2015 checklist. Two reviewers will independently screen, extract data, and assess risk of bias using the Cochrane Risk of Bias 2 (RoB 2) tool. Pooled estimates will be calculated using random-effects inverse-variance meta-analysis; certainty of evidence will be rated with GRADE.DiscussionThis review will provide a comprehensive synthesis of randomized evidence on AID systems in T2DM and will inform clinicians, patients, guideline developers, and policymakers about the magnitude of benefit and the safety profile of this technology in routine practice.Systematic review registrationPROSPERO (CRD420261302666)Supplementary InformationThe online version contains supplementary material available at 10.1186/s13643-026-03206-w. Background Type 2 diabetes mellitus (T2DM) remains one of the most prevalent chronic metabolic disorders worldwide. Around 589 million adults live with diabetes globally, over 90% of whom are diagnosed with T2DM [1]. T2DM is responsible for a substantial burden of microvascular and macrovascular complications, which constitute the leading cause of death in this population [2]. Although lifestyle modification and oral antihyperglycemic therapies represent the initial therapeutic approach, a considerable proportion of individuals with T2DM eventually require insulin therapy to achieve and maintain glycemic targets [3]. Despite progressive improvements in insulin formulations and delivery systems, insulin therapy in T2DM is frequently characterized by therapeutic inertia, fear of hypoglycemia, glycemic variability, and suboptimal achievement of individualized glycemic goals [4]. These challenges highlight an urgent need for strategies that can optimize insulin dosing, reduce treatment complexity, and improve day-to-day metabolic stability. Automated insulin delivery (AID) systems—also referred to as hybrid closed-loop, fully closed-loop, or algorithm-driven insulin delivery systems—represent a significant technological advancement in diabetes management [5]. By integrating continuous glucose monitoring with algorithm-based insulin adjustment, AID systems can modulate insulin delivery in real time, potentially improving time in range, reducing hypoglycemia, and minimizing glucose fluctuations [6]. Hybrid closed-loop systems have demonstrated substantial improvements in glycemic control in individuals with type 1 diabetes [7]. Pivotal randomized clinical trials indicate that these systems significantly increase time in range and reduce hyperglycemia compared with conventional insulin therapy, as exemplified by the PEDAP trial [8]. Similar metabolic benefits were observed in specialized populations, such as pregnant women with preexisting diabetes, in whom hybrid closed-loop therapy resulted in markedly improved maternal glycemic control under stringent physiological targets [9]. Given these consistent benefits across multiple populations with high glycemic variability, there is growing interest in whether AID systems might offer comparable advantages for adults living with T2DM—a group characterized by insulin resistance, heterogeneous insulin requirements, and complex comorbidity profiles [10]. Unlike type 1 diabetes, the pathophysiology of T2DM involves preserved but impaired endogenous insulin secretion, variable degrees of β-cell dysfunction, and a strong influence of lifestyle, obesity, and cardiometabolic factors [11]. These distinctions raise essential questions regarding the effectiveness, safety, and clinical applicability of AID in T2DM. Early proof-of-concept trials suggest that AID systems may reduce mean glucose levels, improve time in range, and potentially reduce the burden of hypoglycemia in insulin-treated T2DM [12]. However, the available evidence remains scattered across small randomized trials with heterogeneous intervention designs and diverse devices and algorithms. To date, only a limited number of systematic reviews and meta-analyses have synthesized the evidence on AID systems specifically in T2DM. The lack of a consolidated appraisal limits the ability of clinicians, researchers and policymakers to understand the magnitude of benefit, the safety profile, and the potential role of AID within the therapeutic continuum for insulin-treated T2DM. A rigorous and up-to-date synthesis is therefore needed. Therefore, this systematic review and meta-analysis aims to evaluate the efficacy and safety of AID systems in adults with T2DM, by quantifying their effects on time in range, glycated hemoglobin, hypoglycemia, glycemic variability, insulin dosing, and device-related adverse events. The findings will help clarify the role of AID as an emerging therapeutic tool in T2DM and support evidence-based decision-making in clinical practice. Methods Methodological guidance and reporting The conduct of this systematic review and meta-analysis will be guided by the recommendations of the Cochrane Handbook for Systematic Reviews of Interventions, version 6.5 (Higgins JPT, Thomas J, Chandler J, Cumpston M, Li T, Page MJ, Welch VA, editors. Cochrane, 2024) [13]. Reporting of the protocol will follow the Preferred Reporting Items for Systematic Review and Meta-Analysis Protocols (PRISMA-P) 2015 statement [14], and reporting of the final review will follow the PRISMA 2020 statement [15]. The completed PRISMA-P 2015 checklist for the present protocol is provided as Additional File 2. The protocol was prospectively registered in the International Prospective Register of Systematic Reviews (PROSPERO; CRD420261302666). Any amendments to the protocol after registration will be documented in PROSPERO with the date, justification and a description of the change, and reported in the final manuscript. Eligibility criteria This systematic review will include randomized controlled trials (RCTs), including both parallel-group and cross-over designs, that evaluate AID systems in adults (≥ 18 years) with a confirmed diagnosis of T2DM. AID systems will be defined as algorithm-driven insulin delivery technologies that integrate continuous glucose monitoring with automated insulin infusion, including hybrid closed-loop, fully closed-loop, and open-source AID systems. Eligible comparators will include conventional insulin-based strategies, such as multiple daily injections (MDI), standard insulin pump therapy, sensor-augmented pump therapy, or any non-automated glucose-lowering regimen. Studies will be eligible if they report at least one predefined glycemic or safety outcome and include an intervention period of ≥ 20 days, ensuring sufficient exposure to AID. Studies exclusively including individuals with type 1 diabetes, gestational diabetes, or mixed populations without extractable T2DM-specific data will be excluded. Non-randomized studies, observational designs, conference abstracts without subsequent peer-reviewed publication, editorials, narrative reviews, and case series will also be excluded to preserve internal validity. Information sources and search strategy A comprehensive and systematic literature search will be conducted in MEDLINE (via PubMed), EMBASE, Scopus, the Cochrane Central Register of Controlled Trials (CENTRAL) and Web of Science from database inception to 30 November 2025. To capture ongoing and unpublished trials, ClinicalTrials.gov and the WHO International Clinical Trials Registry Platform (ICTRP) will be searched. The reference lists of all included studies and of relevant systematic reviews and clinical practice guidelines will be hand-searched to identify additional eligible records, and Google Scholar will be screened (first 200 records sorted by relevance) to retrieve grey literature. No restrictions on language or publication status will be applied; non-English records will be translated using a combination of native-speaking collaborators and validated translation tools. The search strategy will be developed by the lead reviewer (L.G.-R.) in collaboration with an experienced health-sciences librarian/information specialist, who will also advise on database-specific syntax and translation between platforms. The strategy will combine three concept blocks: (i) the population (T2DM and related terms), (ii) the intervention (AID and closed-loop systems, continuous glucose monitoring), and (iii) the study design. For the design block, the Cochrane Highly Sensitive Search Strategy (CHSSS) for identifying randomized trials in MEDLINE—sensitivity- and precision-maximizing version (2008 revision)—will be used, as recommended in Chapter 4 of the Cochrane Handbook [13], and adapted for EMBASE, Scopus, CENTRAL and Web of Science. Controlled vocabulary (MeSH and Emtree) will be combined with free-text terms in title/abstract fields, with appropriate truncation, proximity operators and wildcards. Before the search is run, the MEDLINE strategy will be peer-reviewed by a second information specialist using the Peer Review of Electronic Search Strategies (PRESS) 2015 evidence-based checklist [16], and the validated strategy will then be translated to the remaining databases. The complete electronic search strategy for each database, including the date the search was run and the number of records retrieved, is provided in Additional File 1. All retrieved records will be exported to EndNote 21 (Clarivate Analytics, Philadelphia, PA, USA) for de-duplication, and the de-duplicated library will be uploaded to Rayyan (Rayyan Systems Inc., Cambridge, MA, USA) for blinded title/abstract and full-text screening [17]. The search will be re-run within 12 months of submission of the final review to incorporate the most recent evidence, as recommended by the Cochrane Handbook [13]. Outcomes The primary outcome will be time in range (TIR), defined as the percentage of time with sensor glucose between 70 and 180 mg/dL (3.9–10.0 mmol/L), measured during the final follow-up period reported in each study. TIR was selected as the primary outcome owing to its clinical relevance as an integrative measure of glycemic control and its growing use in diabetes-technology research and practice. Secondary outcomes will include time in hyperglycemia (> 180 mg/dL), time in hypoglycemia (< 70 mg/dL), mean sensor glucose, glycemic variability (standard deviation or coefficient of variation), glycated hemoglobin (HbA1c), total daily insulin dose, body weight, treatment satisfaction, device utilization metrics (e.g., percentage time in automated mode) and adverse events, including severe hypoglycemia and diabetic ketoacidosis. Study selection Study selection will be performed in two sequential phases (title/abstract screening and full-text assessment) by six independent reviewers organized into pairs, using Rayyan in blinded mode. Discrepancies will be resolved through discussion and, when necessary, adjudication by a senior reviewer (J.J.B.). Reasons for exclusion at full-text stage will be documented systematically and reported in a PRISMA 2020 flow diagram [15]. Data extraction Data will be extracted into a standardized, pilot-tested form developed in Microsoft Excel (Microsoft Corporation, Redmond, WA, USA) and piloted on three included studies before full deployment. Extracted variables will include study characteristics (authors, year, country, design, funding source), participant characteristics (sample size, age, sex, baseline HbA1c, BMI, diabetes duration, prior insulin regimen), intervention details (type of AID system, control algorithm, glucose sensor, infusion set), comparator details, follow-up duration, all predefined outcomes (mean and dispersion at end-point and change from baseline) and risk-of-bias judgments. Two reviewers will independently extract data, and any discrepancy will be resolved by discussion or by a third reviewer. Corresponding authors of included studies will be contacted by e-mail (up to two reminders, 2 weeks apart) to obtain missing or clarifying data. Risk of bias assessment The risk of bias of included studies will be assessed using the Cochrane Risk of Bias 2 (RoB 2) tool for randomized trials (Sterne JAC, Savović J, Page MJ, et al. RoB 2: a revised tool for assessing risk of bias in randomized trials. BMJ. 2019;366:l4898) [18]. For cross-over trials, the RoB 2 variant for cross-over trials will be used, which adds a domain on bias arising from period and carryover effects [13, 18]. Each study will be evaluated across the relevant domains and assigned an overall judgement of low risk, some concerns or high risk of bias, on a per-outcome basis. Two reviewers will independently assess each study; disagreements will be resolved through consensus or, if necessary, by a third reviewer. Data synthesis All quantitative analyses will be performed in R version 4.4 or higher, using the meta and metafor packages, and in Review Manager (RevMan) version 5.4 or 6.0 (The Cochrane Collaboration). A meta-analysis will be performed when at least two studies report a comparable outcome, intervention and comparator. Pooled effect estimates will be obtained using a random-effects model with the inverse-variance method, and between-study variance (τ2) will be estimated using the restricted maximum likelihood (REML) estimator with the Hartung–Knapp–Sidik–Jonkman adjustment of confidence intervals. Continuous outcomes will be summarized as mean differences (MD) with 95% confidence intervals (CIs) when measured on the same scale, or as standardized mean differences (SMD) when scales differ, with corresponding 95% CIs. Dichotomous outcomes will be expressed as risk ratios (RR) with 95% CIs. Where studies report medians with interquartile ranges or ranges, means and standard deviations will be estimated using the methods of Wan et al. [19] and Luo et al., as recommended in Chapter 6 of the Cochrane Handbook [13]. Cross-over trials and unit-of-analysis issues. For cross-over RCTs, the preferred approach will follow the recommendations of Chapter 23 of the Cochrane Handbook [13]. When a within-participant analysis is reported (paired mean difference and its standard error or 95% CI, or a within-participant correlation coefficient), this paired estimate will be incorporated using the generic inverse-variance method. When only first-period data are reported, the trial will be analyzed as a parallel-group RCT using only the first-period results, to avoid carryover bias. When neither paired nor first-period data are available, an imputed within-participant correlation coefficient (default 0.5; sensitivity analyses with 0.3 and 0.7) will be used to derive the standard error of the paired difference, as described by Elbourne et al. and the Cochrane Handbook [13]. Cross-over trials with a wash-out period of less than 2 weeks will be considered to be at higher risk of carryover bias and explored in sensitivity analyses. To prevent unit-of-analysis errors, participants in cross-over and cluster-randomized designs will be counted only once in any given comparison; for trials with multiple intervention arms (e.g., two AID arms versus a common comparator), arms will be combined into a single pairwise comparison following the formulas in Chapter 23 of the Cochrane Handbook [13], or shared comparator groups will be split proportionally to avoid double-counting. Heterogeneity. Statistical heterogeneity will be assessed using the I2 statistic, the τ2 estimate, and 95% prediction intervals when at least three studies are pooled. I2 values will be interpreted in line with Section 10.10.2 of the Cochrane Handbook [13]: 0–40% may not be important, 30–60% may represent moderate heterogeneity, 50–90% may represent substantial heterogeneity, and 75–100% considerable heterogeneity, with interpretation always guided by direction and magnitude of effects and clinical heterogeneity. Subgroup and sensitivity analyses. Pre-specified subgroup analyses will be performed, when data permit, by: baseline HbA1c (< 8% versus ≥ 8%), age (< 65 versus ≥ 65 years), type of AID system (hybrid versus fully closed-loop; proprietary versus open-source), type of comparator (MDI versus pump-based therapy), and follow-up duration (< 12 weeks versus ≥ 12 weeks). Subgroup differences will be tested using the Q-test for interaction. Sensitivity analyses will be conducted by (i) excluding studies at high risk of bias on RoB 2, (ii) leave-one-out analyses, (iii) restricting to parallel-group RCTs, and (iv) varying the imputed within-participant correlation coefficient for cross-over trials. Reporting bias and small-study effects. In line with Section 13.3 of the Cochrane Handbook [13], funnel-plot asymmetry and Egger’s regression test will be used to investigate small-study effects and possible reporting bias only when at least 10 studies are included in a given meta-analysis; with fewer studies, these tests have insufficient power to distinguish chance from real asymmetry. Selective outcome reporting will additionally be assessed at the study level via comparison with PROSPERO/trial-registry entries, published protocols and the RoB 2 domain on selection of the reported result. Assessment of certainty of the evidence The certainty of evidence for each pre-specified outcome will be assessed using the Grading of Recommendations Assessment, Development and Evaluation (GRADE) approach (Guyatt GH, Oxman AD, Vist GE, et al. GRADE: an emerging consensus on rating quality of evidence and strengths of recommendations. BMJ. 2008;336(7650):924–926) [20]. Two reviewers will independently rate the certainty of evidence (high, moderate, low or very low) considering the five GRADE domains for downgrading—risk of bias, inconsistency, indirectness, imprecision, and publication bias—and the three domains for upgrading where applicable. Disagreements will be resolved by consensus. Summary of Findings tables will be produced using GRADEpro Guideline Development Tool (GRADEpro GDT; McMaster University and Evidence Prime, 2024) to present, for each outcome, the absolute and relative magnitudes of effect, the corresponding 95% CI and the overall certainty of evidence, in order to facilitate interpretation for clinical and policy decision-making. Discussion This systematic review and meta-analysis will provide a comprehensive synthesis of randomized-controlled trials evaluating AID systems in adults with T2DM. The anticipated findings will help clarify the efficacy and safety profile of this technology in a population in which the available evidence is still limited and heterogeneous. This protocol has several methodological strengths. Its conduct will follow the Cochrane Handbook for Systematic Reviews of Interventions [13], and reporting will adhere to the PRISMA-P 2015 [14] and PRISMA 2020 [15] statements; the protocol has been prospectively registered in PROSPERO (CRD420261302666). The search strategy will combine controlled vocabulary, free-text terms and the Cochrane Highly Sensitive Search Strategy for identifying RCTs across five major databases plus trial registries and grey literature, will be developed with input from an expert health-sciences librarian, and will be peer-reviewed using the PRESS 2015 checklist [16]. The selection of TIR as the primary outcome aligns with contemporary standards in diabetes-technology research, as established in recent landmark trials. The inclusion of both parallel and cross-over RCTs, with explicit unit-of-analysis handling following the Cochrane Handbook [13], and a minimum intervention duration of 20 days ensures that only studies with adequate exposure to AID therapy are pooled. The planned use of RoB 2 [18] and GRADE [20], together with the broad scope of secondary outcomes, will provide a rigorous and multidimensional appraisal of the evidence. Several limitations should be anticipated. The relatively recent emergence of AID research in T2DM may result in a limited number of eligible trials, potentially constraining the precision of pooled estimates and the feasibility of some pre-planned subgroup analyses. Heterogeneity in AID system configurations—including differences in control algorithms, insulin formulations, glucose-sensor accuracy, and hybrid versus fully closed-loop designs—may introduce clinical and methodological diversity that complicates pooling. The inclusion of open-label trials, unavoidable given the nature of the device intervention, introduces inherent performance bias that cannot be fully mitigated. Cross-over designs, while efficient, may be susceptible to carryover effects when wash-out periods are insufficient; this will be examined in pre-planned sensitivity analyses. Finally, with fewer than 10 included studies per outcome, formal tests for small-study effects will not be performed, in line with current methodological guidance [13]. This review is expected to make several important contributions. First, it will provide a dedicated meta-analytic synthesis of randomized evidence specifically examining AID systems in T2DM, addressing a notable gap in the current evidence base. Second, by quantifying the magnitude of effect on TIR, HbA1c and hypoglycemia, the findings will inform clinicians, patients and policymakers about the expected benefits and risks of adopting this technology in routine practice. Third, by highlighting evidence gaps and methodological limitations in existing trials, the review will inform the design of future studies, including the need for longer-term trials, head-to-head comparisons of AID systems and investigations in underrepresented subgroups (older adults, individuals with advanced complications, and diverse ethnic populations). Finally, the findings may contribute to the development of clinical practice guidelines and health-technology assessments, supporting evidence-based coverage decisions and resource allocation in diabetes care. Supplementary Information Additional file 1: Full electronic search strategies for MEDLINE (via PubMed), EMBASE, Scopus, CENTRAL and Web of Science, including the date of the search, the number of records retrieved and the search-strategy peer-review report (PRESS 2015 checklist).Additional file 2: Completed PRISMA-P 2015 checklist for the present protocol (14). Additional file 1: Full electronic search strategies for MEDLINE (via PubMed), EMBASE, Scopus, CENTRAL and Web of Science, including the date of the search, the number of records retrieved and the search-strategy peer-review report (PRESS 2015 checklist). Additional file 2: Completed PRISMA-P 2015 checklist for the present protocol (14).