Role of automated insulin delivery in managing insulin-treated outpatients with type 2 diabetes: A systematic review and meta-analysis

Abstract

BACKGROUND

Automated insulin delivery (AID) systems have demonstrated benefits in managing patients with type 2 diabetes (T2D), but data are still limited. Moreover, the efficacy and safety of the AID systems in these patients have been inadequately explored by systematic reviews and meta-analyses.

AIM

To provide a comprehensive understanding of the optimal use of AID in managing insulin-treated outpatients with T2D.

METHODS

A systematic search of multiple databases and registries, including MEDLINE, Scopus, Web of Science, Cochrane Library, and ClinicalTrials.gov, was conducted from inception to May 15, 2025, to identify studies on AID use for outpatients with T2D. The co-primary outcomes were the change in glycated hemoglobin (HbA1c) and continuous glucose monitoring (CGM) metrics. Statistical analyses were conducted using Review Manager Web software with random-effects models and the inverse variance statistical method. The results were presented as mean differences (MDs) or risk ratios (RRs) with 95%CI.

RESULTS

A total of 15 studies with 28985 participants were identified, including 6 randomized trials (n = 748; 3 crossover and 3 parallel-group trials) and 9 single-arm studies. All included randomized trials raised some concerns, and the single-arm studies had serious risks of overall bias. Meta-analysis of randomized trials showed that AID is more effective than the control group in lowering HbA1c (MD: -0.89%, 95%CI: -1.32 to -0.46, P < 0.0001, I² = 82%). Compared to control interventions, AID use was linked to a higher percentage of time in range (MD: 19.25%, 95%CI: 11.43-27.06, P < 0.00001, I² = 74%) and a lower percentage of time above range > 10 mmol/L (MD: -19.48%, 95%CI: -27.14 to -11.82, P < 0.00001, I² = 73%); however, time below range remained similar between the two groups. The mean sensor glucose level was lower in the AID group; however, the coefficient of variation of glucose was the same in both groups. AID use also led to a reduction in insulin dose, but this is not a consistent finding across all study designs. The risks of serious adverse events (AEs) and severe hypoglycemia were similar in both groups; however, AID use raised the risk of device deficiency. Single-arm studies with participants using AID systems also demonstrated reductions in HbA1c (ranging from 0.7% to 2.07%) and improvements in CGM metrics, along with acceptable safety data.

CONCLUSION

Based on short-term study data, the use of AID systems in outpatients with T2D appears to improve glycemic outcomes and CGM metrics, with no significant AEs. Larger and longer-term randomized controlled trials involving diverse populations, along with a cost-benefit analysis, are needed to guide more informed clinical practice decisions.

Key Words: Automated insulin delivery; Type 2 diabetes; Time in range; Glycated hemoglobin; Meta-analysis

Core Tip: This is the first systematic review and meta-analysis specifically targeting the efficacy and safety of automated insulin delivery (AID) systems in patients with type 2 diabetes in the outpatient setting. Analyzing six randomized controlled trials and nine single-arm studies involving 28985 participants, we observed that AID systems enhance glycemic control, evidenced by larger decreases in glycated hemoglobin and improved continuous glucose monitoring metrics compared to standard treatment. AID use increases time in range, reduces time above target glucose, and maintains similar rates of hypoglycemia and serious adverse events. While AID slightly increases the risk of device deficiency, safety is still within acceptable limits. The results indicate short-term effectiveness, but larger and longer-term studies are necessary to confirm the benefits and assess cost-effectiveness.

INTRODUCTION

Despite a range of oral and non-insulin injectable glucose-lowering medications, including newer agents such as glucagon-like peptide-1 receptor agonists and sodium-glucose cotransporter-2 inhibitors, many patients with type 2 diabetes (T2D) still need insulin therapy as their disease progresses. Current guidelines recommend initiating insulin when oral and non-insulin therapies fail to achieve adequate glycemic control, particularly in cases of significant hyperglycemic symptoms. Subcutaneous basal insulin is typically initiated first; if glycemic control remains insufficient, intensification to a basal-bolus insulin (BBI) or premixed insulin regimen is considered[1,2]. However, the vast majority of insulin-treated patients with T2D do not achieve glycemic control; fewer than one-fourth (23.75%) achieved glycated hemoglobin (HbA1c) < 7%, according to a recent meta-analysis[3]. Even the BBI regimen with insulin analogs, regarded as the best form of multiple daily insulin injections (MDIs), achieves the HbA1c target of < 7% in 53.9% (95%CI: 43.5-64) of these patients, according to a meta-analysis of randomized controlled trials (RCTs)[4]. In real-world practice, this percentage may be lower, as some studies report that only about 20% reach the same target with BBI analogs[5]. For many individuals, BBI regimens pose challenges, such as the burden of administering MDIs, a relatively high risk of hypoglycemia, weight gain, and difficulties with treatment adherence[6]. Such limitations of MDI treatment necessitate the need for new therapies for this group of patients.

Automated insulin delivery (AID) systems, also known as closed-loop systems or artificial pancreas, are designed to improve blood sugar control by automatically adjusting insulin delivery according to real-time glucose levels. These systems integrate a continuous glucose monitoring (CGM), an insulin pump, and a control algorithm. Hybrid closed-loop (HCL) systems automatically adjust basal insulin using CGM data; however, users still need to deliver bolus insulin for meals manually. In contrast, fully automated systems adjust both basal and bolus insulin delivery, aiming to keep blood sugar consistently within a target range. These technologies reduce the burden of diabetes management and improve quality of life, although challenges such as cost and accessibility persist as they become more widely adopted[7]. AID systems benefit people with type 1 diabetes (T1D), demonstrating increased time spent in the target glucose range, reduced HbA1c levels, and improved safety profiles in both clinical trials and real-world studies[8,9]. The American Diabetes Association (ADA) recommends AID systems for diabetes management in youth and adults with T1D and other insulin-deficient types of diabetes, provided they can use the device safely[10]. AID systems are increasingly evaluated in T2D, showing safety and efficacy in improving glycemic variability and time in range (TIR) without raising insulin doses or causing hypoglycemia, particularly in monitored or hospital settings[11]. ADA advises against AID for T2D and recommends insulin pump therapy with CGM for youth and adults on MDI who can use it safely. Given the scarcity of data on AID use in T2D, Amer et al[12] conducted a meta-analysis that included seven RCTs investigating the efficacy and safety of fully closed-loop AID in patients with T2D. The meta-analysis has several shortcomings, including the combination of outpatient studies with longer durations and inpatient studies with shorter durations for pooled outcome estimates, as well as the pooling of data from randomized crossover and parallel-group studies[12]. Additionally, they did not incorporate studies without a non-AID control group, and some new studies have been published since then. Therefore, this systematic review and meta-analysis (SR/MA) of RCTs and single-arm studies was conducted to provide a comprehensive understanding of the optimal use of AID in managing insulin-treated ambulatory outpatients with T2D.

MATERIALS AND METHODS

This SR/MA was conducted following the procedures specified in the Cochrane Handbook for Systematic Reviews of Interventions and is reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) checklist[13,14]. The SR/MA has been registered with PROSPERO (CRD420251067987), and a protocol summary is available online. Since ethical approval already exists for the individual studies included in the metaanalysis, no separate approval was required for this study.

Search strategy

A systematic search was conducted across multiple databases and registers, including MEDLINE (via PubMed), Scopus, Web of Science, Cochrane Library, and ClinicalTrials.gov, from their inception to May 15, 2025. Using the Boolean operators “AND” and “OR”, the following terms were searched: “type 2 diabetes mellitus”, “type 2 diabetes”, “non-insulin dependent diabetes mellitus”, “T2DM”, “T2D”, “automated insulin delivery”, “closed loop insulin”, “artificial pancreas”, “hybrid closed loop”, “continuous subcutaneous insulin infusion”, and “insulin pump”. The search terms were applied to the titles and abstracts of the documents without the use of any language filters for articles. The goal was to identify both published and unpublished studies. Additionally, the search entailed reviewing the references within the published article retrieved for this study, as well as relevant journals. The complete search strategies are included in the Supplementary material.

Study selection

The Population, Intervention, Comparison, and Outcomes Study design served as a framework for establishing eligibility criteria for studies in this SR/MA. The patient population (P) consisted of adults of either sex with T2D on insulin therapy, either as subcutaneous MDI or continuous subcutaneous insulin infusion (CSII). The intervention (I) involved any form of AID, whether a hybrid or fully automated closed-loop system; meanwhile, the comparison or control (C) group (optional) included individuals who continued their previous MDI or CSII regimen. The outcomes (O) included HbA1c level and CGM metrics at the end of the studies. RCTs or observational studies (prospective or cohort) were considered the study type (S) for inclusion. The studies with a comparator arm were included in the meta-analysis, whereas those without an AID control arm were used for the qualitative review. Exclusion criteria included case reports or case series, commentaries, letters to the editor, studies conducted among patients with T2D in inpatient settings, studies conducted in T1D or other types of diabetes, studies not reporting the outcomes of interest, and secondary or post hoc analyses of the included studies. The study selection process involved four independent authors who, after removing duplicates, assessed titles and abstracts based on predefined inclusion and exclusion criteria. This screening eliminated irrelevant studies, allowing potentially eligible ones to proceed. After the authors reached a consensus, the full texts of the remaining articles were retrieved and independently reviewed in detail to determine eligibility. Exclusion reasons were recorded for each article, as shown in the PRISMA flow diagram. Any disagreements were resolved through consensus.

Outcomes analyzed

The co-primary outcomes of interest were the HbA1c level and CGM metrics at the end of the studies. The CGM metrics included the percentage of TIR of 3.9-10 mmol/L, time above range (TAR) > 10 mmol/L, TAR > 13.9 mmol/L, TAR > 20 mmol/L, time below range (TBR) < 3.9 mmol/L, TBR < 3 mmol/L, mean sensor glucose (SG), standard deviation of glucose, coefficient of variation (CV) for glucose as a measure of glucose variability, and the duration of CGM sensor usage. Additional outcomes included the total daily dose of insulin (TDDI) and safety parameters, such as device deficiencies and instances of severe hypoglycemia.

Data extraction

Four review authors independently conducted data extraction using standardized forms. The results were consolidated upon identifying multiple publications from the same study group, and relevant data from each article were incorporated into the analyses. The following data were extracted from all eligible studies and included in the review: (1) First author; (2) Year of publication; (3) Country where the study was conducted; (4) Study design; (5) Major inclusion criteria for study subjects; (6) Sample size; (7) Mean age; (8) Baseline HbA1c; (9) Diabetes duration; and (10) Specific interventions used in the AID and control arms, including the device model and manufacturer, and trial duration. Additionally, as mentioned earlier, data on primary and secondary outcomes were extracted. All disagreements were resolved through consensus.

Dealing with missing data

The necessary supplementary files for the articles were obtained from the websites of the publishing journals. Additional information, if needed, was collected from the corresponding authors of the relevant articles via email. All pertinent information gathered in this manner was included in the meta-analysis. A thorough analysis of key numerical data was conducted, encompassing the number of individuals screened and randomized, as well as a meticulous evaluation of the intention-to-treat, as-treated, and per-protocol populations. Moreover, attrition rates-including dropouts, losses to follow-up, and withdrawals-were carefully analyzed.

Statistical analysis

The Review Manager (RevMan) computer program, version 7.2.0, was used to conduct meta-analyses and generate forest plots[15]. The results of the outcomes were expressed as mean differences (MDs) for continuous variables and as risk ratios (RRs) for dichotomous variables with 95%CI. Random-effects analysis models were selected to account for the expected variability resulting from differences in population characteristics and variations in trial duration. The inverse variance statistical method was applied in all cases. The 95%CI was calculated using the Wald-type method, while χ² was computed using the DerSimonian and Laird method. Meta-analyses were conducted in cases where at least two trials reported a variable. All included crossover trials were incorporated into the meta-analysis by taking all measurements from the AID intervention periods and all measurements from the control intervention periods, analyzing these as if the trial were a parallel-group trial comparing AID to control. Consequently, we assumed that there was no correlation between the groups. Theoretically, this approach gives rise to a unit-of-analysis error[13]. However, the number of participants in each arm or sequence of the included crossover trials was equal. Moreover, the elimination half-life of insulin after CSII is relatively short (approximately 53 minutes), thereby reducing the likelihood of a carryover effect[16]. The relatively long treatment periods (at least 3 weeks each) in these crossover trials likely reduce the risk of carryover, as insulin effects and glycemic control stabilize over time. Additionally, specific crossover analysis was employed to examine the impact of AID and control interventions on the primary and secondary endpoints of the trials. For each outcome, subgroup analyses were performed based on the type of randomized trial (crossover or parallel-group). A significance level of P < 0.05 was used. Meta-analyses were not performed for the single-arm studies; however, they were included only in the systematic review.

Risk of bias assessment

Three authors independently assessed the risk of bias (RoB). The Cochrane risk-of-bias tool for randomized trials version 2 was used to evaluate the RoB of the RCTs, while the Risk of Bias in Non-randomized Studies of Interventions version 2 (ROBINS-I V2) was employed for prospective and retrospective intervention studies[17,18]. When applicable (i.e., a minimum of ten studies in a forest plot), publication bias was assessed using funnel plots generated by RevMan web software[19].

Assessment of heterogeneity

Heterogeneity assessment began by analyzing forest plots. Following this, χ² tests were employed with N-1 degrees of freedom and a significance level of 0.05 to evaluate statistical significance. Additionally, the I² test was applied in the subsequent analysis. The thresholds for I² values were set at 25% for low heterogeneity, 50% for moderate heterogeneity, and 75% for high heterogeneity[20].

Grading of the results

The Grading of Recommendations Assessment, Development, and Evaluation methodology was used to assess the quality of evidence for the primary outcomes of the meta-analysis[21]. The process of creating the summary of findings (SOF) table and evaluating the quality of evidence as “high”, “moderate”, “low”, or “very low” has been previously described[22].

RESULTS

Search results

The PRISMA flow diagram illustrating the steps involved in selecting the studies is presented in Figure 1. The initial search yielded 12362 articles, which were narrowed down to 17 following the screening of titles, abstracts, and full-text reviews. Finally, 15 studies (n = 28985) that met all prespecified criteria were included in the systematic review[22-36]. Meta-analyses were conducted for six studies (n = 748) that had a non-AID control group[23-28]. The systematic review included the remaining nine single-arm studies (n = 28237), but they were not part of the meta-analyses[29-37]. Two studies were excluded: (1) One was a letter to the editor and a post hoc analysis of an included study[38]; and (2) The other was a report on the extension phase of an included study[39].

Figure 1  Flowchart on study retrieval and inclusion in the meta-analysis.

Characteristics of included and excluded studies

Table 1[23-28] summarizes the characteristics of the studies included in the meta-analysis, while Table 2[29-37] presents the specifics of studies that are included only in the systematic review and not in the meta-analysis. Out of the six randomized trials in the meta-analysis, three were crossover trials[23-25], while the remaining three had a parallel-group design[26-28]. In the AID arm, Borel et al[23] used an HCL system, while the other five studies used fully automatic advanced closed-loop systems. In the control arm, subcutaneous MDI was utilized in all studies except Borel et al[23], who used a usual insulin pump therapy. The duration of the crossover trials ranged from 40 days to 6 months[23-25], while the parallel-group trials spanned 12 weeks to 6 months[26-28]. The nine studies included only in the systematic review (but not in the meta-analysis) lacked a non-AID control group[29-37]. Out of these studies, four were prospective[29,31,35,36], and the remaining five were retrospective[30,32-34,37]. Among the nine studies, three employed an HCL system[29,30,37], while the other six utilized fully automated, advanced closed-loop systems[31-36]. The duration of these single-arm studies varied from six weeks to approximately two years. Supplementary Table 1 summarizes the excluded studies[25,31,38,39].

Table 1 Baseline characteristics of individual randomized trials and study participants included in the meta-analysis.

Ref.	Trial design, Trial reg number	Major inclusion criteria	Groups	Interventions	Number	Age (years), mean ± SD	HbA1c (%), mean ± SD	Diabetes duration (years), mean ± SD/median (interquartile range)	Key outcomes	Duration
Borel et al[23], 2024, three hospitals in France	Cross-over, NCT05369871	Age > 18 years, body weight ≤ 150 kg, insulin pump for ≥ 6 months, equipped with a CGM or flash glucose monitoring system, TDDI < 160 U/24 hours, and HbA1c < 10%	AID	DBLG1 hybrid closed-loop system (Accu-Chek Insight insulin pump, DexcomG6 CGM system and Diabeloop application)	17	63 ± 9	7.9 ± 0.9	24 ± 9	TIR, TAR, TBR, coefficient of glucose variation, TDDI, glucose management indicator	13 weeks (including two therapy: 6 weeks therapy of hybrid closed loop and 6 weeks of CSII + CGM)
			Control	Usual insulin pump therapy equipped with a DexcomG6 CGM system (CSII + CGM)	17
Boughton et al[24], 2021, two centers in United Kingdom and Switzerland	Cross-over, NCT04025775	Age ≥ 18 years, subcutaneous insulin therapy and end-stage renal disease requiring maintenance dialysis (hemodialysis or peritoneal dialysis)	AID	CamAPS HX closed-loop app (CamDiab) (Dana Diabecare RS pump, Dexcom G6 CGM system and Cambridge adaptive model predictive control algorithm, version 0.3.71)	26	68.3 ± 11.2	7.2 ± 1.3	20 ± 10	TIR, TAR, TBR, coefficient of glucose variation, TDDI, AEs	40 days (two therapy: 20 days therapy of closed loop and standard insulin therapy)
			Control	Standard subcutaneous insulin therapy with masked CGM	26
Daly et al[25], 2023, two centers in United Kingdom	Cross-over, NCT04701424	Age ≥ 18 years, subcutaneous insulin therapy, HbA1c ≤ 11%	AID	CamAPS HX Closed-loop (Dana insulin pump, Dexcom G6 real-time CGM system, and CamAPS HX Application v.0.3.71)	26	59 ± 11	9.0 ± 1.4	17.5 ± 8.2	TIR, TAR, TBR, HbA1c, mean sensor glucose, coefficient of glucose variation, TDDI, AEs	16 weeks (two therapy: 8 weeks therapy of closed loop and standard insulin therapy)
			Control	MDI and Dexcom G6 CGM system	26
Kudva et al[26], 2025, 21 centers in the United States and Canada	Parallel group, NCT05785832	Age ≥ 18 years, T2D ≥ 6 months. On MDI or insulin pump for ≥ 3 months1	AID	T: Slim X2 insulin pump with CIQ + technology (Tandem) and Dexcom G6 CGM system	215	59 ± 12	8.2 ± 1.4	18 (11-26)	HbA1c, TIR, TAR, TBR, AEs	13 weeks
			Control	Pretrial insulin delivery regimen and Dexcom G6 CGM system	104	57 ± 12	8.1 ± 1.2	18 (11-24)
Reznik et al[27], 2014, 36 centers in Canada, Europe, Israel, South Africa, and the United States	Parallel group, NCT01182493	Age 30-75 years, on MDI, TDDI ≤ 220 U/24 hours, HbA1c 8%-12%, ≥ 2.5 SMBG/day on average	AID	Medtronic MiniMed Paradigm Veo system; Medtronic	168	55.5 ± 9.7	9.0 ± 0.8	14.9 ± 8	HbA1c, mean 24-hour glucose concentrations, area under the curve for hypoglycaemia and hyperglycaemia, TAR, TBR, AEs	6 months
			Control	MDI with SMBG	163	56.4 ± 9.5	9.0 ± 0.8	15.3 ± 8
Reznik et al[28], 2024, muti-center in France	Parallel group, NCT04233229	Age > 18 years, T2D ≥ 6 months. On MDI ≥ 6 months. Requiring nursing support at home, HbA1c 8%-12%	AID	T: Slim X2 insulin pump with CIQ technology (Tandem) and Dexcom G6 CGM system	14	69.3 ± 6.7	9.0 ± 1.2	20.4 ± 12.3	TIR, TAR, TBR, coefficient of glucose variation, HbA1c, TDDI, body weight, body mass index, AEs	12 weeks
			Control	MDI with SMBG (from day 70 onwards, Dexcom G6 CGM until the planned completion of the study at day 90)	15	69.7 ± 10.3	9.25 ± 1.0	17.0 ± 9.05

¹In both groups, 96% were on multiple daily insulin injection and 4% were on an insulin pump.

AEs: Adverse events; AID: Automated insulin delivery; CGM: Continuous glucose monitoring; CIQ: Control-IQ; CSII: Continuous subcutaneous insulin infusion; HbA1c: Glycated hemoglobin; MDI: Multiple daily insulin injection; SMBG: Self-monitoring of blood glucose; TAR: Time above range; TBR: Time below range; TDDI: Total daily dose of insulin; TIR: Time in range; T2D: Type 2 diabetes.

Table 2 Baseline characteristics of individual single-arm studies and study participants included in the meta-analysis.

Ref.	Key inclusion criteria	Study type	Number	Baseline HbA1c (%)	Previous insulin delivery method (%)	Closed-loop system used in the study	Study period	Key findings
Bhargava et al[29], 2025, NCT05238142, United States	T2D for ≥ 2 years; using either MDI or CSII pump therapy, with or without CGM for ≥ 3 months; HbA1c < 10%	Prospective	95	7.9 ± 1.0	MDI (61.1), CSII with CGM (20.0), CSII (9.5), AID pump (7.4), other (2.1)	MiniMed 780G advanced hybrid closed-loop	90 days	No severe hypoglycemia, severe hyperglycemia, diabetic ketoacidosis, or hyperglycemic hyperosmolar state events and no device-related serious or unanticipated adverse device effects. Significant CFB in HbA1c (-0.7%), TIR (7.2%), TAR > 10 mmol/L (-7.1%), TAR > 13.9 mmol/L (-1.7%), mean SG: -0.51 mmol/L, SD of SG: -0.08 mmol/L, TDDI: 13.9 U. No significant CFB in TBR < 3.9 mmol/L or < 3 mmol/L, CV of SG
Telci Caklili et al[30], 2024, Turkey	Age ≥ 60 years, brittle diabetes, and a follow-up of ≥ 3 months with an AID system	Retrospective cohort	34	9.4 ± 2.1	MDI (100)	Medtrum A7 + Touchcare patch pump and integrated A7 + CGM and Medtronic 780G	6 months	Significant CFB in HbA1c (-2.07%), TDDI: -10.77 U. At the end of the study after AID use: TIR (64.8%), TAR > 10 mmol/L (26.7%), TAR > 13.9 mmol/L (6.7%), TBR < 3.9 mmol/L (1.8%), TBR < 3 mmol/L (0.5%)
Davis et al[31], 2023, NCT04617795, United States	Age 18-75 years, on basal-bolus or basal only insulin regimen, has no insulin pump within 3 months of screening, HbA1c 8%-12%	Prospective	24	9.4 ± 0.9	BBI (50), basal only insulin (50)	Omnipod 5, Dexcom G6, and Omnipod 5 app on a locked-down. Android phone	8 weeks	Significant CFB in HbA1c (-1.3%), TIR (21.9%), TAR > 10 mmol/L (-21.4%), TAR > 13.9 mmol/L (-16.9%), TAR > 20 mmol/L (-9.2%), mean SG: -1.83 mmol/L, TBR < 3.9 mmol/L (-0.08%), SD of SG: -0.61 mmol/L. No significant CFB in TBR < 3 mmol/L, CV of SG, TDDI: -8.8 U. No SAE, one hypoglycemic episode
Fabris et al[32], 2024, United States	Adults with T2D transitioned from PLGS to AID	Retrospective	796	NR	PLGS system (Basal-IQ Technology, Tandem Diabetes Care) (100)	AID (CIQ Technology, Tandem Diabetes Care)	3 months	Significant CFB in TIR (9%), TAR > 10 mmol/L (-8.9%), TAR > 13.9 mmol/L (-4.4%), mean SG: -0.65 mmol/L, TDDI: 6.5 U. No significant CFB in TBR < 3.9 mmol/L or < 3 mmol/L, CV of SG
Forlenza et al[33], 2022, United States1	At least 12 consecutive months of data available on CIQ, and had at least 30 days of ≥ 75% CGM data availability before and after CIQ initiation	Retrospective	500	NR, GMI 7.3% (6.9%-7.7%)	NR	Tandem t: Slim X2 CIQ system	3 months	Significant CFB in GMI (-0.2%), TIR (8%), TAR > 10 mmol/L (-5%), TAR > 13.9 mmol/L (-2%), mean SG: -0.47 mmol/L. No significant CFB in TBR < 3.9 mmol/L
Kadiyala et al[34], 2024, NCT04977908 and NCT04701424, United Kingdom1	Age ≥ 18 years, on subcutaneous insulin, HbA1c ≥ 11%	Retrospective	26	9 ± 1.4	Subcutaneous insulin (100)	CamAPS HX fully closed-loop system	8 weeks	At the end of the study after AID use: TIR (66.3%), TAR > 10 mmol/L (32.2%), TAR > 16.7 mmol/L (1.8%), TBR < 3.9 mmol/L (0.43%), TBR < 3 mmol/L (0.04%), mean SG: 9.17 mmol/L, SD of SG: 2.98 mmol/L, CV of glucose (322%)
Levy et al[35], 2024, NCT05111301, United States	either BBI therapy (MDI or via an insulin pump) or basal-only insulin for at least 3 months, TDDI ≤ 200 U, HbA1c 7.5%-12.0%	Prospective	30	8.6 ± 1.2	Basal insulin only (43), MDI (50), insulin pump (7)	T: Slim X2 insulin pump with CIQ technology	6 weeks	Significant CFB in TIR (15%), TAR > 10 mmol/L (-15%), TAR > 13.9 mmol/L (-2.7%), mean SG: -1.22 mmol/L. No significant CFB in TBR < 3.9 mmol/L or <3 mmol/L, CV of SG
Pasquel et al[36], 2025, NCT05815342 United States	Age 18-75 years, treated with a stable insulin regimen for at least 3 months prior to screening, HbA1c < 12%	Prospective	305	8.2 ± 1.3	MDI (73), basal insulin only (21), insulin pump (5.6), premix insulin injections (< 1)	Omnipod 5 AID System	13 weeks	Significant CFB in HbA1c (-0.8%), TIR (20%), TAR > 10 mmol/L (-20%), TAR > 13.9 mmol/L (-12%), TAR > 20 mmol/L (-5%), mean SG: -1.78 mmol/L, TBR < 3.9 mmol/L (0.0%), TBR < 3 mmol/L (0.01%) SD of SG: -0.61 mmol/L, coefficient of variation of SG (-0.7%). 13 patients experienced SAE and one severe hypoglycemia
Thijs et al[37], 2025, multiple countries in Europe	MM780G data uploaded to CareLink Personal from January 2020 to April 2024	Retrospective	26427	NR	NR	MiniMed 780G system	Mean observation period: Cohort A: 213 days, cohort B: 148 days	Cohort A: At the end of the study after AID use-TIR (71.1%), TAR > 10 mmol/L (27.9%), TAR > 13.9 mmol/L (6.6%), TBR < 3.9 mmol/L (1%), TBR < 3 mmol/L (0.2%), mean SG: 8.72 mmol/L, SD of SG: 2.82 mmol/L, GMI (7.1%). Cohort B: At the end of the study after AID use-TIR (75.1%), TAR > 10 mmol/L (24.3%), TAR > 13.9 mmol/L (4.9%), TBR < 3.9 mmol/L (0.6%), TBR < 3 mmol/L (0.1%), mean SG: 8.52 mmol/L, SD of SG: 2.48 mmol/L, GMI (7%). Both cohort A and cohort B had significant CFB in GMI (-0.5% and -0.3%, respectively), TIR (15.9% and 12.1%, respectively)

¹These studies included separate cohorts of type 1 diabetes and type 2 diabetes (T2D). We have considered only T2D cohort for this systematic review.

Data in mean ± SD or median (interquartile range). AID: Automated insulin delivery; BBI: Basal-bolus insulin; CFB: Change from baseline; CIQ: Control-IQ; CGM: Continuous glucose monitoring; CSII: Continuous subcutaneous insulin infusion; GMI: Glucose management indicator; HbA1c: Glycated hemoglobin; MDI: Multiple daily insulin injection; NR: Not reported; PLGS: Predictive low glucose suspend; SAE: Serious adverse event; SG: Sensor glucose; TAR: Time above range; TBR: Time below range; TDDI: Total daily dose of insulin; TIR: Time in range; T2D: Type 2 diabetes.

RoB of the included studies

All three crossover RCTs (Supplementary Table 2) raised some concerns about bias due to deviations from the intended intervention in Borel et al[23], issues with the randomization process in Boughton et al[24], and period and carryover effects, as well as deviations from the intended intervention in Daly et al[25]. Some concerns about bias were also identified in all three parallel-group RCTs, which stemmed from deviations from the intended interventions (Supplementary Figure 1)[26-28]. All nine single-arm studies exhibited significant RoB, mainly due to confounding bias, as assessed by the ROBINS-I tool (Supplementary Figure 2)[29-37]. Publication bias was not assessed because there were not enough RCTs (at least 10) in the forest plots[19].

Grading of the results

The certainty of evidence grades for the primary outcomes of the meta-analysis are provided in the SOF table (Supplementary Table 3).

Change in HbA1c

AID outperformed control interventions in reducing HbA1c, as shown by lower HbA1c levels at the end of the study in the AID group (MD: -0.89%, 95%CI: -1.32 to -0.46, P < 0.0001, I² = 82%, low certainty of the evidence). Its superior HbA1c-lowering effect was evident in both parallel-group (MD: -0.74%, 95%CI: -1.14 to -0.33, P = 0.0004, I² = 78%) and crossover trials (MD: -1.4%, 95%CI: -1.96 to -0.84, P < 0.00001) (Figure 2A)[25-28].

Figure 2 Forest plot highlighting the mean difference in glycated hemoglobin levels and time in range at the end of the study in the automated insulin delivery group vs the control group. A: Glycated hemoglobin levels; B: Time in range. AID: Automated insulin delivery. ^aCI calculated by Wald-type method; ^bTau2 calculated by DerSimonian and Laird method.

TIR

At the end of the studies, TIR was higher in the AID group compared to the control group (MD: 19.25%, 95%CI: 11.43-27.06, P < 0.00001, I² = 74%, moderate certainty of the evidence). Subgroup analysis demonstrated the superiority of AID over the control group in both crossover (MD: 20.64%, 95%CI: 8.52-32.76, P = 0.0008, I² = 77%) and parallel-group (MD: 17.91%, 95%CI: 4.03-31.79, P = 0.01, I² = 77%) trials (Figure 2B)[23-26,28].

TAR

TAR > 10 mmol/L was lower in the AID group than in the control group, as shown in the pooled analysis of all trials (MD: -19.48%, 95%CI: -27.14 to -11.82, P < 0.00001, I² = 73%, moderate certainty of the evidence). This difference was also observed in the subgroups of crossover (MD: -20.69%, 95%CI: -33.35 to -8.03, P = 0.001, I² = 79%) and parallel-group (MD: -18.38%, 95%CI: -31.42 to -5.35, P = 0.006, I² = 73%) trials (Table 3). The AID group had a lower TAR > 13.9 mmol/L in all trials (MD: -8.33%, 95%CI: -12.89 to -3.77, P = 0.0003, I² = 71%, moderate certainty of the evidence). However, this difference was not observed in the subgroup analysis of crossover trials (MD: -8.05%, 95%CI: -17.70 to 1.61, P = 0.10, I² = 79%) and parallel-group trials (MD: -11.71%, 95%CI: -23.51 to 0.08, P = 0.05, I² = 73%). In crossover trials, TAR > 20 mmol/L was higher in the AID than in the control group (MD: -4.39%, 95%CI: -6.79 to -1.98, P = 0.0004, I² = 0%) (Table 3).

Table 3 Comparison of the efficacy outcomes in the automated insulin delivery vs the control arms at the end of the studies.

Outcome variables (continuous)	Type of the randomized trial	Number of study reports	Number of participants with outcome analyzed		Pooled effect size, mean differences (95%CI)	P value	I2 (%)
			Automated insulin delivery arm	Control arm
TAR > 10 mmol/L (%)	All	5	294	186	-19.48 (-27.14 to -11.82)	< 0.00001	73
	Crossover	3	69	68	-20.69 (-33.35 to -8.03)	0.001	79
	Parallel-group	2	225	118	-18.38 (-31.42 to -5.35)	0.006	73
TAR > 13.9 mmol/L (%)	All	4	268	160	-8.33 (-12.89 to -3.77)	0.0003	71
	Crossover	2	43	42	-8.05 (-17.70 to 1.61)	0.10	79
	Parallel-group	2	225	118	-11.71 (-23.51 to 0.08)	0.05	73
TAR > 20 mmol/L (%)	Crossover	2	52	51	-4.39 (-6.79 to -1.98)	0.0004	0
TBR < 3.9 mmol/L (%)	All	5	294	186	-0.07 (-0.21 to 0.08)	0.37	22
	Crossover	3	69	68	-0.17 (-0.42 to 0.09)	0.20	32
	Parallel-group	2	225	118	0.00 (-0.09 to 0.09)	0.99	0
TBR < 3 mmol/L (%)	All	5	294	186	-0.01 (-0.03 to 0.02)	0.57	39
	Crossover	3	69	68	-0.03 (-0.09 to 0.03)	0.32	69
	Parallel-group	2	225	118	0.00 (-0.02 to 0.02)	1.00	0
Mean SG (mmol/L)	All	5	451	335	-1.21 (-1.93 to -0.49)	0.001	83
	Crossover	3	69	68	-1.82 (-3.34 to -0.30)	0.02	84
	Parallel-group	2	382	267	-0.66 (-1.34 to 0.03)	0.06	81
SD of SG (mmol/L)	Crossover	3	69	68	-0.40 (-0.64 to -0.15)	0.001	0
Coefficient of variation of glucose (%)	All	5	294	186	0.91 (-1.04 to 2.87)	0.36	68
	Crossover	3	69	68	0.51 (-2.90 to 3.93)	0.77	77
	Parallel-group	2	225	118	1.35 (-0.27 to 2.97)	0.10	17
Time using continuous glucose monitoring sensor (%)	Crossover	3	69	68	2.80 (1.55-4.05)	< 0.0001	13
TDDI (U)	All	6	458	344	-5.45 (-22.20 to 11.29)	0.52	74
	Crossover	3	69	68	13.24 (-11.64 to 38.12)	0.30	65
	Parallel-group	3	389	276	-21.07 (-30.15 to -11.99)	< 0.00001	0
TDDI (U/kg)	All	3	262	150	-0.04 (-0.20 to 0.11)	0.59	53
	Crossover	2	52	51	0.06 (-0.20 to 0.32)	0.67	50
	Parallel-group	1	210	99	-0.13 (-0.24 to -0.02)	0.02	Not reported

SG: Sensor glucose; TAR: Time above range; TBR: Time below range; TDDI: Total daily dose of insulin.

TBR

The AID and control groups exhibited identical TBR < 3.9 mmol/L (MD: -0.07%, 95%CI: -0.21 to 0.08, P = 0.37, I² = 22%, high certainty of the evidence) and < 3 mmol/L (MD: -0.01%, 95%CI: -0.03 to 0.02, P = 0.57, I² = 39%, high certainty of the evidence) in the pooled analysis of all trials; as well as in subgroups of crossover trials [< 3.9 mmol/L (MD: -0.17%, 95%CI: -0.42 to 0.09, P = 0.20, I² = 32%); < 3 mmol/L (MD: -0.03%, 95%CI: -0.09 to 0.03, P = 0.32, I² = 69%)] and parallel-group trials [< 3.9 mmol/L (MD: 0%, 95%CI: -0.09 to 0.09, P = 0.99, I² = 0%); < 3 mmol/L (MD: 0%, 95%CI: -0.02 to 0.02, P = 1.00, I² = 0%)] (Table 3).

SG

At the end of the studies, the mean SG level was lower in the AID group compared to the control group across all studies (MD: -1.21 mmol/L, 95%CI: -1.93 to -0.49, P = 0.001, I² = 83%, low certainty of the evidence) and the crossover study subgroup (MD: -1.82 mmol/L, 95%CI: -3.34 to -0.30, P = 0.02, I² = 84%), but not in the parallel group studies (MD: -0.66 mmol/L, 95%CI: -1.34 to 0.03, P = 0.06, I² = 81%). In the crossover studies, the SD of glucose was lower in the AID group (MD: -0.4 mmol/L, 95%CI: -0.64 to -0.15, P = 0.001, I² = 0%). The CV of glucose was consistent across the AID and control groups in all studies (MD: 0.91, 95%CI: -1.04 to 2.87, P = 0.36, I² = 68%) and in the subgroups of crossover (MD: 0.51, 95%CI: -2.9 to 3.93, P = 0.77, I² = 77%) and parallel-group (MD: 1.35, 95%CI: -0.27 to 2.97, P = 0.10, I² = 17%) studies (Table 3).

Time using the CGM sensor

In crossover trials, the AID group spent more time using the CGM sensor compared to the control group (MD: 2.8%, 95%CI: 1.55-4.05, P < 0.0001, I² = 13%) (Table 3).

TDDI

In the subgroup of the parallel-group trials, the TDDI was lower in the AID group compared to the control group (MD: -21.07 U, 95%CI: -30.15 to -11.99, P < 0.00001, I² = 0% and MD: -0.13 U/kg, 95%CI: -0.24 to -0.02, P = 0.02). However, differences in TDDI between the two groups were not evident in the pooled analyses of all trials (MD: -5.45 U, 95%CI: -22.2 to -11.29, P = 0.52, I² = 74% and MD: -0.04 U/kg, 95%CI: -0.2 to 0.11, P = 0.59, I² = 53%) and the crossover trial subgroup (MD: 13.24 U, 95%CI: -11.64 to 38.12, P = 0.30, I² = 65% and MD: 0.06 U/kg, 95%CI: -0.2 to 0.32, P = 0.67, I² = 50%) (Table 3).

Safety outcomes

The results of meta-analyses regarding safety outcomes are summarized in Table 4. The AID and control groups had identical numbers of severe adverse events (SAEs) (RR: 1.58, 95%CI: 0.85-2.93, P = 0.15, I² = 0%) and participants with SAEs (RR: 1.58, 95%CI: 0.85-2.93, P = 0.15, I² = 0%). In the crossover and parallel-group trial subgroups, the numbers of SAEs (crossover: RR: 1.92, 95%CI: 0.5-7.37, P = 0.34, I² = 0%; parallel-group: RR: 1.5, 95%CI: 0.74-3.01, P = 0.26, I² = 0%) and the number of participants experiencing SAEs (crossover: RR: 1.92, 95%CI: 0.5-7.37, P = 0.34, I² = 0%; parallel-group: RR: 1.5, 95%CI: 0.74-3.01, P = 0.26, I² = 0%) were also identical. The number of study-related SAEs (all studies: RR: 1.85, 95%CI: 0.54-6.4, P = 0.33, I² = 0%; crossover: RR: 2.89, 95%CI: 0.12-67.75, P = 0.51; parallel-group: RR: 1.71, 95%CI: 0.44-6.58, P = 0.44, I² = 0%) and non-study-related SAEs (all studies: RR: 1.37, 95%CI: 0.81-2.33, P = 0.24, I² = 0%; crossover: RR: 1.47, 95%CI: 0.44-4.91, P = 0.53, I² = 0%; parallel-group: RR: 1.44, 95%CI: 0.67-3.06, P = 0.35, I² = 29%) was comparable in the two groups. Both groups experienced the identical number of severe hypoglycemia events (all studies: RR: 2.1, 95%CI: 0.22-19.81, P = 0.52, I² = 0%; crossover: RR: 3.0, 95%CI: 0.13-70.42, P = 0.50; parallel-group: RR: 1.46, 95%CI: 0.06-35.5, P = 0.82), as well as an equal number of participants affected by these events (all studies: RR: 1.13, 95%CI: 0.18-7.1, P = 0.90, I² = 0%; crossover: RR: 3.0, 95%CI: 0.13-70.42, P = 0.50; parallel-group: RR: 0.69, 95%CI: 0.07-6.57, P = 0.74, I² = 0%). In the pooled analysis of all studies (RR: 1.67, 95%CI: 1.07-2.63, P = 0.02, I² = 0%) and in the parallel-group subgroup (RR: 1.81, 95%CI: 1.07-3.04, P = 0.03), non-SAEs were higher in the AID group than in the control group; such events were equal in crossover subgroup (RR: 1.33, 95%CI: 0.54-3.28, P = 0.53). However, the number of participants experiencing non-SAEs was identical in the AID and control groups (all studies: RR: 1.08, 95%CI: 0.92-1.26, P = 0.35, I² = 0%; crossover: RR: 1.22, 95%CI: 0.49-3.04, P = 0.67, I² = 0%; parallel-group: RR: 1.07, 95%CI: 0.92-1.26, P = 0.38, I² = 0%). The number of device deficiencies (all studies: RR: 9.13, 95%CI: 2.17-38.37, P = 0.003, I² = 0%; crossover: RR: 6.9, 95%CI: 1.29-36.78, P = 0.02, I² = 0%; parallel-group: RR: 19.93, 95%CI: 1.22-326.34, P = 0.04) and participants experiencing device deficiencies (all studies: RR: 5.7, 95%CI: 2.11-15.36, P = 0.0006, I² = 0%; crossover: RR: 6.49, 95%CI: 1.21-34.74, P = 0.03, I² = 0%; parallel-group: RR: 5.31, 95%CI: 1.55-18.16, P = 0.008, I² = 0%) was higher in the AID group.

Table 4 Comparison of the safety outcomes in the automated insulin delivery vs the control arms.

Outcome variables	Type of the randomized trial	Number of study reports	No. of participants with outcome/participants analyzed		Pooled effect size, risk ratio (95%CI)	P value	I2 (%)
			Automated insulin delivery arm	Control arm
Number of SAEs	All	4	24/234	15/229	1.58 (0.85-2.93)	0.15	0
	Crossover	2	6/52	3/51	1.92 (0.50-7.37)	0.34	0
	Parallel-group	2	18/182	12/178	1.50 (0.74-3.01)	0.26	0
Number of participants with SAEs	All	4	24/234	15/229	1.58 (0.85-2.93)	0.15	0
	Crossover	2	6/52	3/51	1.92 (0.50-7.37)	0.34	0
	Parallel-group	2	18/182	12/178	1.50 (0.74-3.01)	0.26	0
Number of SAEs (study-related)	All	3	6/208	3/203	1.85 (0.54-6.40)	0.33	0
	Crossover	1	1/26	0/25	2.89 (0.12-67.75)	0.51	N/A
	Parallel-group	2	5/182	3/178	1.71 (0.44-6.58)	0.44	0
Number of SAEs (not study-related)	All	5	41/449	20/333	1.37 (0.81-2.33)	0.24	0
	Crossover	2	6/52	4/51	1.47 (0.44-4.91)	0.53	0
	Parallel-group	3	35/397	16/282	1.44 (0.67-3.06)	0.35	29
Number of severe hypoglycemia events	All	4	2/281	0/170	2.10 (0.22-19.81)	0.52	0
	Crossover	2	1/52	0/51	3.00 (0.13-70.42)	0.50	N/A
	Parallel-group	2	1/229	0/119	1.46 (0.06-35.50)	0.82	N/A
Number of participants with severe hypoglycemic events	All	5	2/449	1/333	1.13 (0.18-7.10)	0.90	0
	Crossover	2	1/52	0/51	3.00 (0.13-70.42)	0.50	N/A
	Parallel-group	3	1/397	1/282	0.69 (0.07-6.57)	0.74	0
Number of non-SAEs	All	3	66/267	22/155	1.67 (1.07-2.63)	0.02	0
	Crossover	2	10/52	7/51	1.33 (0.54-3.28)	0.53	0
	Parallel-group	1	56/215	15/104	1.81 (1.07-3.04)	0.03	N/A
Number of participants with non-SAEs	All	4	154/435	120/318	1.08 (0.92-1.26)	0.35	0
	Crossover	2	9/52	7/51	1.22 (0.49-3.04)	0.67	0
	Parallel-group	2	145/383	113/267	1.07 (0.92-1.26)	0.38	0
Number of device deficiencies	All	4	53/281	3/170	9.13 (2.17-38.37)	0.003	0
	Crossover	2	11/52	1/51	6.90 (1.29-36.78)	0.02	0
	Parallel-group	2	42/229	2/119	19.93 (1.22-326.34)	0.04	N/A
Number of participants with device deficiencies	All	4	31/281	3/170	5.70 (2.11-15.36)	0.0006	0
	Crossover	2	10/52	1/51	6.49 (1.21-34.74)	0.03	0
	Parallel-group	2	21/229	2/119	5.31 (1.55-18.16)	0.008	0

N/A: Not applicable; SAEs: Severe adverse events.

Outcomes of the single-arm studies

Four of the included single-arm studies reported significant reductions in HbA1c from baseline, ranging from 0.7% to 2.07%[28-30,35]. Two other studies found substantial declines in glucose management indicator (GMI) ranging from 0.2% to 0.5%[32,36]. Mean SG (-0.47 to -1.83 mmol/L) and SD of SG (-0.08 to -0.61 mmol/L) also improved significantly[28,30,31,34,35]. However, the change from baseline (CFB) in CV of SG[28,30,31,34,35] was insignificant except for the study by Pasquel et al[36]. An increase in TIR (7.2%-21.9%) was observed in six studies[28,30,31,34-36]. Improvements in TAR > 10 mmol/L (-5% to -21.4%), TIR > 13.9 mmol/L (-1.7% to -16.9%) in six studies[28,30-32,34,35], and TIR > 20 mmol/L (-5% to -9.2%)[30,35] were also reported. Two of the studies observed significant improvements in TBR < 3.9 mmol/L (0% to -0.08%)[30,35], while others did not[28,31,32,34]. Only one study[35] reported improvement in TBR < 3 mmol/L, while others found no CFB[28,30,31,34]. Significantly greater change in TDDI (6.5-13.9 U) was found in two studies[28,31], whereas TDDI did not change in the other one[30]. The occurrences of SAEs, severe hypoglycemia, diabetic ketoacidosis (DKA), hyperglycemic hyperosmolar state (HHS), and other adverse events (AEs) were infrequent in the studies (Table 2)[29-37].

Management of heterogeneity

Since the number of included trials is small, subgroup analyses beyond the type of the randomized trial and meta-regression were not performed. Leave-one-out sensitivity analyses were conducted for primary and important secondary outcomes with moderate to high heterogeneity (when I² > 50%) to detect changes in statistical significance and significant heterogeneity (at least a 2-step change) (Supplementary Table 4)[23-28]. For CFB in HbA1c, removing either of the studies did not change the statistical significance or heterogeneity among the studies. For TIR 3.9-10 mmol/L and TAR > 10 mmol/L, removing either of the studies did not alter the statistical significance; however, heterogeneity among the studies was reduced (from 74% to 31% and from 73% to 20%, respectively) when Daly et al[25] was removed. For TAR > 13.9 mmol/L and mean SG, removing either of the studies did not change the statistical significance or heterogeneity among the studies. For the CV of glucose, removing either of the studies did not alter the statistical significance; however, heterogeneity among the studies was reduced (from 68% to 30%) when Borel et al[23] was removed. For TDDI, removing either of the studies did not change the statistical significance or heterogeneity among the studies.

DISCUSSION

Main findings of this review

This SR/MA, involving fifteen studies (n = 28985) that met all predefined inclusion criteria, examined the role of AID in managing patients with T2D. Six of these studies [one used an HCL system, while the other five used fully automatic anterior cruciate ligament (ACL) systems] with a non-AID control group (n = 748; treatment duration from 6 weeks to 6 months) were included in meta-analyses. AID outperformed control interventions in reducing HbA1c (MD: -0.89%), increasing TIR (MD: 19.25%), and decreasing TAR > 10 mmol/L (MD: -19.48%) without increasing the risk of hypoglycemia. There were no differences in TBR between the AID and control groups. Although a lower TDDI was observed in a subset of the parallel-group trials, this difference was not seen in the combined analyses of all trials. AID treatment appears to be safe and does not increase the risk of SAEs. Additionally, the risk of non-SAEs (both related and unrelated to the study) was similar in the AID and control groups. However, device deficiencies were more common in the AID group, including in sub-group analyses, compared to the control group.

The remaining nine single-arm trials (n = 28237; six involving ACL and three involving HCL), with treatment durations from 6 weeks to 2 years, also demonstrated significant improvements in glycemic outcomes. Among these, four studies reported HbA1c reductions of 0.7%-2.07% from baseline, while two other studies showed a decline in GMI ranging from 0.2% to 0.5%. A significant increase in TIR (7.2%-21.9%) and reductions in TAR across all levels of hyperglycemia, along with two studies showing improvements in TBR < 3.9 mmol/L and one study with TBR < 3 mmol/L. The occurrence of SAEs, severe hypoglycemia, DKA, HHS, and other AEs was rarely reported in these studies, indicating that AID is a relatively safe approach for managing T2D that requires intensive insulin therapy.

Implications for clinical practice

This SR/MA examining the efficacy and safety of AID systems suggests a generally positive impact of this technology in managing T2D, although the number of participants in the meta-analyzed RCTs was limited. Our meta-analysis showed a significant decrease in HbA1c of 0.89%, carrying important clinical implications for lowering microvascular and macrovascular complications of T2D, as supported by many previous studies[40-42]. The United Kingdom Prospective Diabetes Study follow-up data demonstrated that reducing the average HbA1c level by 1% is associated with a 37% decrease in microvascular disease risk, a 21% drop in diabetes-related death, and a 14% reduction in myocardial infarction rate[43]. If we project this outcome data onto our study results, the expected risk reduction could be approximately 30% for microvascular disease, around 12% for myocardial infarction, and about 18% for diabetes-related deaths in the long-term follow-up. The larger data from single-arm studies in our SR showed an even better HbA1c reduction (-0.7% to -2.07%), suggesting that the benefits could be even greater in real-world settings.

Improvements in TIR were also associated with better cardiovascular and overall mortality outcomes in patients with T2D, as demonstrated by previous studies[44,45]. Bergenstal et al[46] reported a 6% reduction in major adverse cardiovascular event (MACE) and a 10% decrease in severe hypoglycemia risk with every 10-percentage point increase in TIR. Therefore, we could expect approximately a 12% reduction in MACE and a 20% reduction in the risk of hypoglycemia if we extrapolate the results of the above study to the TIR (19.3%) outcome observed in this meta-analysis. Improvements in microvascular disease, such as retinopathy, neuropathy, and albuminuria, are other important outcomes associated with a higher TIR, as observed in previous studies[47-49]. Thus, using AID systems in the long term would lead to a significant decrease in microvascular disease. Similarly, the noteworthy improvement in TAR (-19.5%) observed in this meta-analysis also aligns with better macro- and microvascular outcomes, as well as improved survival, as shown in previous studies[48,49]. The subgroup analysis for parallel group trials showed a notable reduction in TDDI (although this was not observed in the pooled data). This suggests that AID systems could be beneficial for managing diabetes in individuals with obesity, given the increased risk of weight gain associated with higher insulin doses[50]. A reduction of TDDI from 6.5-13.9 U, as reported in the single-arm trials, is especially noteworthy for improving diabesity outcomes in patients. A reduced risk of hypoglycemia is another important and positive outcome of a lower TDDI in AID users, as observed in single-arm trials (not in the meta-analysis of RCTs), which enhances quality of life and decreases other potential risks associated with hypoglycemia, such as accidents and falls.

This study suggests that AID systems generally have a reasonable safety profile, with risks of SAEs, severe hypoglycemia, DKA, HHS, and other AEs being similar. However, non-SAEs related to device issues were higher in the AID group compared to the control groups. Single-arm studies with larger participant numbers also indicated reasonable safety for AID systems. Thus, our study provides additional comprehensive data, including studies with a non-AID control group and some newer studies published after the previous meta-analysis by Amer et al[12], suggesting that AID systems are effective options for glycemic outcomes in patients with T2D, with reasonable efficacy and safety profiles.

As previously mentioned, AID systems are recommended for use only in T1D[10]. They are generally considered a cost-effective option for many individuals with T1D, offering a better quality of life and potentially leading to long-term cost savings by reducing the burden of diabetes-related complications[51,52]. AID systems can be cost-effective for managing T2D, particularly when considering the long-term benefits of improved glycemic control and reduced complications. While the initial cost of the technology can be a barrier, studies suggest that the improved health outcomes and potential cost savings from reduced hospitalizations and long-term complications can offset these costs over time[53]. However, future RCTs and real-world studies are needed to determine the long-term cost-effectiveness of AID in T2D.

Limitations and the strengths of this review

We acknowledge several limitations in our study, such as the limited number of available RCTs, the small sample sizes in the trials, and the brief follow-up periods. For a common and lifelong condition like T2D, larger datasets with extended follow-up periods are essential for generating more reliable and robust evidence. While AID enhances CGM metrics and HbA1c, and may lower TDDI, practical challenges such as availability and cost limit access for large populations who could benefit. Selecting appropriate patients, including those with longer diabetes duration and who are unable to reach glycemic targets despite optimal MDI use with high doses of insulin, or those at higher risk of hypoglycemia, especially if they can afford it, can help justify the use of these systems. None of the studies included in the SR/MA conducted a cost-benefit or effectiveness analysis, which is a crucial aspect to be considered in future research. While AID systems require substantial training and support for patients, it can be argued that the level of support needed is comparable to that for standard insulin pump therapy. The studies in this SR/MA might not represent a broader population due to limited diversity in the study samples, which could restrict how widely our findings can be applied. Certain key aspects of AID systems-like quality of life, user-friendliness, and psychological effects-remain unexamined because of limited data. Future research should focus on these areas. Despite these limitations, with 6 studies involving 748 participants and a very expensive intervention using the latest technology, we can generate reasonable evidence for the use of AID systems in guiding clinical practice. The single-arm studies with a large number of participants contribute to the evidence base, and their findings align with those of the studies included in our meta-analysis, which is another significant strength of our SR/MA.

CONCLUSION

Based on limited short-term study data, the use of AID systems in outpatients with T2D appears to improve glycemic outcomes, as evidenced by reduced HbA1c levels, improved TIR, and lower TAR; however, the levels of evidence were low to moderate. The use of AID in these individuals did not raise additional safety concerns, except for device-related issues. Larger, long-term RCTs and real-world studies involving diverse populations are essential to assess the efficacy of AID systems, patient satisfaction, impacts on quality of life, and long-term cost-effectiveness, thereby helping improve clinical practice guidelines.

ACKNOWLEDGEMENTS

We are thankful to Annlyn Vinu Thomas for providing us the audio clip for the core tip of this article.