Glycemic Control During Continuous Subcutaneous Insulin Infusion Versus Multiple Daily Insulin Injections in Type 2 Diabetes: Individual Patient Data Meta-analysis and Meta-regression of Randomized Controlled Trials

There is now a well-established evidence base for the routine clinical use of continuous subcutaneous insulin infusion (CSII; insulin pump therapy) in selected people with type 1 diabetes who have failed to achieve target levels of glycemic control with the best insulin injection regimens (multiple daily insulin injections [MDI]) and structured diabetes education (1,2). In many such patients, there can be a clinically valuable, sometimes substantial, reduction in HbA_1c level and all grades of hypoglycemia on switching to CSII.

The value of insulin pump therapy in people type 2 diabetes is less certain. Some national guidelines for insulin pump therapy (e.g., the U.K. National Institute for Health and Care Excellence Technology Appraisal of CSII) (2) do not recommend this treatment for type 2 diabetes. This is despite the fact that at least one-quarter of people with type 2 diabetes receiving insulin injections have very poor glycemic control, say an HbA_1c level ≥9% (75 mmol/mol) (3), and new options for improving control are urgently needed. Present guidance on not using CSII in people with type 2 diabetes is largely based on the limited and variable evidence of the efficacy of insulin pump therapy in this type of diabetes in the relatively small-scale randomized controlled trials (RCTs) published over the last several years (4–7), where there is support both for and against the superiority of CSII versus MDI in reducing HbA_1c. A meta-analysis (8) of aggregate data from four trials of CSII versus MDI in type 2 diabetes (including one in newly diagnosed type 2 diabetes) reported no difference in HbA_1c levels between treatments.

A recent large-scale, multicenter RCT (9) reported that people with type 2 diabetes with poor glycemic control that persisted after a period of optimized MDI achieved a substantially better HbA_1c level while receiving treatment with CSII than while receiving MDI (mean difference between groups 0.7% [8 mmol/mol]), with 20% less total daily insulin dosage and without an increase in hypoglycemia or weight gain. With data from this new trial, it is likely that a meta-analysis of all available RCTs may now provide a more robust view of the comparative effectiveness of CSII and MDI in individuals with type 2 diabetes. However, since the differing efficacy of CSII in trials may be due to the different characteristics of the participants at baseline, such as level of glycemic control, rather than estimating just the overall pooled effect size (reduction in HbA_1c), it is more important to explore how the effectiveness of CSII depends on patient characteristics such as age, baseline quality of control, and insulin requirements, a strategy that has the potential to inform patient-centered therapeutic decision-making (10). The analysis approach that enables this most reliably and with most power is the use of individual patient data (11).

The purpose of this study, therefore, was to perform an individual patient data meta-analysis and meta-regression of RCTs that have compared glycemic control with CSII with MDI in patients with type 2 diabetes in order to test the hypothesis that CSII achieves significantly lower HbA_1c levels than MDI in identifiable patient groups. We aimed to model the determinants of final HbA_1c level and insulin requirements on these therapies, as this may identify the individuals with type 2 diabetes who are most likely to benefit from CSII and, in due course, allow appropriate cost-effectiveness analyses to be performed if insulin pump therapy is found to be clinically valuable.

We followed recent guidance on the conduct and reporting of individual patient data meta-analyses (12,13). The protocol was predefined, and the meta-analysis registered with ClinicalTrials.gov (clinical trial reg. no. NCT02910141). All original trials selected for analysis operated under the supervision of an appropriate human ethics committee. The current analysis involved anonymized data only.

Trials were identified without language restriction as those published up to January 2016 that met the inclusion criteria. We searched the Cochrane Database for RCTs, Ovid Medline, Embase, and Google Scholar (search terms “diabetes mellitus,” “diabetes mellitus type 2,” “CSII,” “MDI,” “insulin pump therapy,” and “randomized controlled trial”). We also searched literature cited in retrieved articles, previous meta-analyses, and lists of articles supplied by the manufacturers of insulin pumps.

Two independent reviewers (J.C.P. and Y.R.) decided on trial eligibility. We selected for inclusion RCTs comparing glycemic control during CSII and MDI in participants with type 2 diabetes who had been studied for at least 2 months. We excluded observational studies; reviews, surveys, and meta-analyses; cost-effectiveness analyses; trials of CSII in type 1 diabetes, pregnant women, and newly diagnosed diabetes; studies that were short term (<2 months) or where MDI was not the comparator; studies where participants had not previously been treated by insulin; duplicate reports; and extensions of previous trials. Differences concerning trial eligibility or data interpretation were resolved by consensus after discussion.

Trial quality was assessed by the components of a 6-point scale, according to the method of Jadad et al. (14) (based on the study being randomized, the randomization scheme being described and being appropriate, whether the study was double blind, a description of the method and appropriateness of blinding, and a description of withdrawals and dropouts) but with an additional item for reporting allocation concealment (the person randomizing is blinded to next treatment allocation). A score of ≥3 was considered appropriate quality for inclusion in the meta-analysis.

Data on individual participants in the trials that met the criteria for meta-analysis were obtained from the original research team or the funding sponsors who held the trial data. We asked the sources to provide information on individual trial participants, including age, sex, duration of diabetes, treatment group (CSII or MDI), and baseline and final HbA_1c levels, insulin dose, weight, and BMI. We recontacted authors for further clarification when there were issues over the interpretation of data or when additional data were required. For aggregate data meta-analysis, summary information was extracted from text, tables, and graphs in published articles.

The primary outcome was glycemic control at study completion, as measured by HbA_1c level. The secondary outcomes were insulin dose (total units per day and units per kilogram) and weight at study completion. We did not analyze data on hypoglycemia because we could not obtain complete information on this for all trials; BMI at study completion was also analyzed but was recorded in only four studies. Center information was not available for multicenter trials.

We included data from the participants who had completed each trial. In the two crossover trials, we analyzed data from the first period only in the primary analysis because of evidence of a carryover effect and the absence of a washout period in one study.

We checked the consistency of data by comparing major participant characteristics and results in published reports of the trials with analyses of files from individual patient data. Clarifications were sought and discrepancies were resolved when possible by contacting investigators. Assessment of potential study bias included baseline imbalances, design of crossover studies (presence of washout period, evidence of carryover), and inclusion of patients of one type in studies (e.g., only elderly or obese participants).

Patients with missing data were excluded from analysis. To explore the effect of patient-level covariates on outcome, we carried out a one-step meta-regression analysis (modeling the impact of potential effect size modifiers on the effect size) by creating a single, large data set from the individual patient data. In this way, we explored the determinants of final HbA_1c level and insulin dosage using a hierarchical random-effects regression model considering the covariates sex, age, study duration, diabetes duration, baseline HbA_1c level, insulin dose, and weight. Initially, we fitted all covariates separately, and then we created best-fit models considering all covariates of interest for each outcome. The deviance information criterion (a Bayesian method for model comparison that balances “goodness-of-fit” with model complexity, calculated by WinBUGS) was used for choosing between models, with differences in the criterion of three or more considered to be important (15).

We assessed the potential impact of ecological/aggregation biases by decomposing the variability accounted for by covariates into between- and within-study variability to check covariate effects were not being influenced by between-study information (16).

Deviance residuals were examined for extreme points and models recalculated to ensure robustness to the exclusion of the most influential points.

To explore overall mean effect sizes, we carried out a meta-analysis of individual patient data for HbA_1c, insulin dose, weight, and BMI using a two-step approach. Initially, we modeled individual patient data for each trial using a linear regression model including terms that distinguished between CSII and MDI treatment groups and baseline measurements to produce a treatment effect estimate and associated SE for each trial. Using a random-effects meta-analysis, we then combined these to calculate an overall effect size for the difference in means between treatments. In sensitivity analyses, we explored robustness by also carrying out meta-analysis with a fixed-effect model.

We did not explore potential publication bias using a funnel plot because the number of trials analyzed was <10 (17); however, we quantified heterogeneity between trials by the I² statistic (the percentage of variability in effect due to heterogeneity rather than sample error), with >50% representing substantial heterogeneity and >75% representing considerable heterogeneity.

Stata version 11 was used for the meta-analysis of aggregate data and the two-stage individual patient data meta-analysis. We used the Bayesian Markov Chain Monte Carlo software in WinBUGS version 1.4.3 to carry out the one-stage regression analyses on all the individual patient data. Prior distributions for all model parameters were specified as vague. For all models, we used a minimum burn-in of 10,000 and sample size of 30,000. All models were checked for the convergence of all variables using the history and density plots available in WinBUGS version 1.4.3. To confirm the convergence of the best-fitting regression models, we fitted these with two different sets of initial values and produced a Gelman-Rubin plot to check convergence. Further details of statistical procedures are available from the authors on request.

The initial literature search identified 90 publications, of which 70 concerned insulin pump therapy in persons with type 2 diabetes (Fig. 1). We then excluded 22 reviews, surveys, cost-effectiveness analyses, or methods articles; 10 short-term studies; 3 trials of CSII in pregnancy; 5 trials in newly diagnosed diabetes; 6 trials where the comparator was not MDI; 13 trials that were not RCTs; and 6 duplicate articles. Five RCTs (4–7,9) that compared glycemic control during CSII or MDI were selected as being eligible for meta-analysis. Individual patient data were obtained from all five eligible trials, consisting of 590 participants with type 2 diabetes who were randomly allocated to receive MDI (n = 287) or CSII (n = 303).

Table 1 shows the characteristics of the trials included in the meta-analysis. Three studies were parallel RCTs (one multicenter) (4,5,9), and two studies were crossover RCTs (6,7). The study duration ranged from 3 to 24 months, and the dropout rate ranged from zero to 27.5%. The baseline HbA_1c level varied between trials from 8.1% (65 mmol/mol) to 9.6% (81 mmol/mol), and the baseline insulin dose ranged from 0.72 to 1.16 units/kg. Oral antihyperglycemic agents were either not used or discontinued in three trials (4,5,7) or prior use of metformin was continued (6,9) in the others. All five trials scored 3 of 6 on the study quality scale because of the absence of double blinding (not possible with trials of insulin pump therapy) and because of the lack of allocation concealment or information on this.

Notable features of the design, the risk of bias and the interpretation of trials are also shown in Table 1. These include an absence of a washout period in one crossover trial (7); evidence of carryover effect in a crossover trial (7); baseline imbalances in age (7), sex (5), insulin dosages (5), and weight (5); missing BMI data in one trial (6); inclusion of only older participants in one study (5) or obese participants in another (6); and some discrepancies between the individual patient data and the published information (6). Most agreement between individual participant data (IPD) and reported data were found for the OpT2mise trial (9). Two studies used long-acting insulin analogs as the basal insulin with short-acting insulin analogs before meals, two studies used types of isophane insulin as the basal insulin, and one study used three daily injections of premixed isophane/short-acting insulin. In one study (9), participants underwent a prerandomization run-in period designed to optimize glycemic control with MDI, and only those who still had poor control (HbA_1c 8–12% [64–108 mmol/mol]) and were insulin resistant (0.7–1.8 units/kg) were then randomized to receive CSII or continue treatment with MDI.

Using the entire data set of 590 patients in the five trials, we explored the independent determinants of the HbA_1c difference between MDI and CSII at study completion by including a range of covariates in regression models. The best-fit model included only baseline HbA_1c as a predictor of final HbA_1c; age, sex, baseline weight, and baseline insulin requirements did not affect outcome. For all trials combined, HbA_1c difference was best described by the equation:

The effect of baseline HbA_1c on HbA_1c treatment difference varied between trials. Figure 2 illustrates the increase in treatment difference as baseline HbA_1c levels increase for all trials and for the OpT2mise trial, which were calculated from the best-fit models. The largest effect was in the OpT2mise trial (9), where there was prerandomization optimization of glycemic control on MDI; here, the regression equation for HbA_1c difference = 5.39 − (0.669 [0.326–1.011] × baseline HbA_1c). With an example baseline HbA_1c of 10% (86 mmol/mol), the effect size in this study is expected to be −1.3% (14 mmol/mol) compared with −0.59% (7 mmol/mol) for all participants combined. The effect size was minimal below a baseline HbA_1c of ∼8.0% (64 mmol/mol) (Fig. 2).

Best-fit models showed that baseline insulin dose was the best predictor of the final insulin dose and the difference in dose between treatments. The best-fit model for difference in insulin dose (units/kg), CSII versus MDI was:

With an example baseline dose of 100 units/day, the effect size is predicted to be a difference of −23.6 units/day insulin dose; with a baseline of 150 units/day, the effect size would be 35.5 units/day.

Figure 3A shows a forest plot (a graphical representation of the results of aggregate meta-analysis with the effect size of all studies, associated confidence intervals, and the summary/overall effect measure) for the mean HbA_1c difference between MDI and CSII using a random-effects model and with covariate adjustment for potential baseline imbalance between treatment groups in each study. The overall mean HbA_1c difference for the five trials combined was −0.40% (95% CI −0.86 to 0.05 [−4.4 mmol/mol (−9.4 to 0.6)]), favoring CSII. The I² statistic was 81%, indicating considerable heterogeneity between trials. Using a fixed-effect model, the overall HbA_1c difference was lower at −0.30% but with a narrower CI (−0.47 to −0.13 [−3.3 mmol/mol (−5.2 to −1.5)]), favoring CSII. Similar results for HbA_1c difference, favoring CSII, were obtained with an aggregate data meta-analysis of summary effect sizes extracted from the published articles rather than individual patient data: random-effects model, −0.45% (−0.81 to −0.09 [−5.0 mmol/mol (−8.9 to −1.0)]); and fixed-effect model, −0.47% (−0.65 to −0.29 [−5.2 mmol/mol (−7.1 to −3.2)]).

Figure 3B shows a forest plot of the difference in insulin requirements on MDI and CSII using individual patient data meta-analysis (random-effects model). The overall insulin dose was reduced by −0.25 units/kg (−0.31 to −0.19) on CSII versus MDI (26% reduction of the baseline insulin requirements), with an I² statistic of 4.7%, indicating little heterogeneity between trials. The fixed-effect model gave an almost identical effect size (−0.26 units/kg [−0.31 to −0.20]). Meta-analysis (random-effects and fixed-effect models) showed that the total daily insulin dose was reduced by −24.0 units/day (−30.6 to −17.5) on CSII versus MDI (27% of baseline daily insulin) with an I² of 0%, indicating little heterogeneity (forest plot not shown).

A two-stage individual patient data meta-analysis with baseline adjustment indicated that the mean weight at study completion did not differ between treatments (0.08 kg [95% CI −0.33 to 0.48], I² 0%, random-effects and fixed-effect models) (Fig. 3C). BMI data were available for four trials, and a two-stage individual patient data meta-analysis with baseline adjustment also indicated no difference in the final mean BMI between treatment groups (0.00 kg/m² [−0.23 to 0.24]), random-effects and fixed-effect models (forest plot not shown).

We show in this individual patient data meta-analysis and meta-regression of five RCTs involving 590 participants with type 2 diabetes that insulin pump therapy achieves better glycemic control than MDI in participants with poor diabetes control at baseline. Using Bayesian statistical best-fit models, where we explored a wide range of potential effect modulators, we found that the HbA_1c treatment difference was dependent on prerandomization HbA_1c level with MDI, increasing as baseline HbA_1c level increases. For example, the expected all-study difference increases from −0.15% (2 mmol/mol) with a baseline of 8.0% (64 mmol/mol) to −0.59% (6 mmol/mol) with a baseline HbA_1c level of 10% (86 mmol/mol). We also found that the reduction in insulin requirements on switching to CSII was dependent on the baseline insulin dose: the treatment difference would increase to −35.5 units/day, for example, when the baseline insulin dose is 150 units/day compared with a treatment difference of −23.6 units/day for a baseline insulin dose of 100 units/day. The percentage insulin dose reduction with CSII at the end of the study period was about −25% irrespective of baseline insulin dose.

The overall mean difference in HbA_1c level for the meta-analysis of all trials was −0.40% (4 mmol/mol), for a mean HbA_1c baseline level of 8.8% (73 mmol/mol), with the mean difference varying between −0.3% (3 mmol/mol) and −0.47% (5 mmol/mol), depending on the meta-analysis model and individual patient data versus aggregate data meta-analysis. This improvement was accompanied by a reduction in insulin requirements (mean 24 units/day), but no weight change. The large degree of heterogeneity in HbA_1c effect size between trials (I² = 81%) was likely due to the wide variation in baseline HbA_1c, from 8.1% (65 mmol/mol) to 9.6% (81 mmol/mol), which we show is a major determinant of CSII efficacy.

We found that the HbA_1c difference varied markedly between trials and was greatest for the study (9) in which patients underwent prerandomization optimization of control and only those who continued to have poor control were entered into the trial. For this study, the mean HbA_1c difference from individual patient data analysis was 0.63% (6 mmol/mol) and the expected effect at 10% (86 mmol/mol) baseline HbA_1c was 1.3% (14 mmol/mol). One may speculate that the greater treatment difference for this trial was due to the fact that, in the nonoptimized trials, glycemic control with MDI continued to improve after randomization, thus minimizing the difference between CSII and MDI. This highlights the notion that CSII in type 2 diabetes may be best targeted at those who have failed to achieve target HbA_1c levels after best attempts with MDI, including dose titration, optimization of dietary counseling, and physical activity (18). A likely mechanism for improved control on CSII is that the traditional large bolus injections of insulin required during MDI in type 2 diabetes are absorbed less well than the slow basal insulin infusion of CSII. In this respect, Parkner et al. (19) found that the same dose of insulin given via the basal rate of CSII achieved better glycemic control and higher circulating insulin concentrations than when given as an injection of long-acting (glargine) insulin. Another reason for improved control with CSII may be the increased treatment satisfaction of insulin pump therapy in persons with type 2 diabetes (4), which may improve adherence to treatment compared with MDI. A lack of treatment adherence may also be more easily detected with CSII than MDI because of the computer download function of modern insulin pumps that allows the survey of events such as the number of meal boluses given per day and the detection of basal-rate suspends.

In our analysis, the improved glycemic control associated with CSII was not associated with a greater weight gain in this group, and one may speculate that any decrease in glycosuria and retained calories with better control (favoring weight gain) was balanced by lower insulin dosages and therefore less anabolic insulin effect in the insulin pump-treated participants.

The strengths of our study include the fact that we were able to obtain individual data from investigators on all patients in all eligible RCTs and were thus able to explore patient-level covariates as effect size modulators in a way that is not possible with a conventional summary meta-analysis. A further strength is that we used Bayesian statistical methods including Markov chain Monte Carlo simulation that enables highly complex, multiparameter probability models to be analyzed. Advantages and further details and discussion of Bayesian methods in meta-analysis and evidence synthesis are reviewed elsewhere (20,21).

Our study also has some considerations in its interpretation. First, the best MDI regimen for type 2 diabetes is debatable (22) (e.g., the choice of long-acting insulin that will offer the best glycemic control), and participants in the trials studied used a variety of regimens. The extent to which participants underwent a structured diabetes education program varied between trials. A range of adjunctive agents are currently being investigated for improving control in insulin-treated type 2 diabetes, including glucagon-like peptide 1 agonists (23,24), dipeptidyl peptidase 4 inhibitors and sodium–glucose cotransporter 2 inhibitors (25), although none are widely established in clinical practice. Possibly then, further improvement in glycemic control with MDI might be achieved in some patients by additional therapeutic approaches before switching to CSII, and this needs further study.

It is a limitation that we were unable to analyze data on hypoglycemia frequency in this meta-analysis because there was incomplete information on this for all trials, but, in the largest RCT, the lower mean HbA_1c level in the CSII versus MDI group was obtained without an increase in hypoglycemia (9). It should also be noted that there was a relatively small number of studies in the meta-analysis and two of the trials (6,7) had small numbers of participants (n = 17 and 40), while there was a large number of participants in one trial (n = 331) (9).

In the individual patient data meta-analysis, and as recommended by the Cochrane Collaboration (17), we analyzed the first period only from crossover trials because we detected evidence of carryover, one trial had no washout period between periods (7), and the lower baseline HbA_1c level for both treatments at the start of the second period might result in a lower effect size if baseline HbA_1c was found to be a determinant of glycemic outcome.

The implication of our meta-analysis for clinical practice is that insulin pump therapy in type 2 diabetes is effective at lowering HbA_1c levels, but, as with type 1 diabetes, it should be targeted at those persons with worst glycemic control and highest insulin dose after best attempts with MDI (26). In a previous meta-analysis of treatment with CSII versus MDI in persons with type 1 diabetes (27), we also found that the mean HbA_1c level is reduced with insulin pump therapy and that the greatest effect was in those with the highest baseline HbA_1c level with MDI. In the current study, the effect size for HbA_1c and insulin dose reduction was greatest in poorly controlled insulin-resistant participants, and therefore the cost-effectiveness will likely be best in these patients (28). We found that the difference in glycemic control is small below a baseline HbA_1c level of ∼8% (64 mmol/mol), although treatment satisfaction may still be superior for CSII versus MDI at this level of control (4,9); this topic needs investigation.

Despite treatment targeting, there are likely to be many people with type 2 diabetes with continued poor control after best attempts with MDI who are thus candidates for a trial of CSII, and that will present notable logistic and economic issues. The technologically sophisticated and relatively costly infusion pumps traditionally used for type 1 diabetes, with flexible rate adjustments and meal-bolus calculators, which were used in the trials of type 2 diabetes analyzed here, are probably not required for type 2 diabetes. There is emerging evidence that pumps with a limited number of fixed basal rate options and with simple meal bolusing will be adequate for most people with type 2 diabetes (18,29,30), indicating that smaller, cheaper, and more cost-effective devices might eventually be used for this type of diabetes.

In conclusion, we have found that insulin pump therapy achieves better glycemic control than MDI in poorly controlled type 2 diabetes with a substantial reduction in insulin requirements and no change in weight. The best effect of CSII in type 2 diabetes is in those patients with the worst control and with the highest insulin dose at baseline.

Acknowledgments. The authors thank all of the investigators, trial sponsors, and colleagues for generously giving permission for the use of individual patient data from their trials to be analyzed, for supplying files, and for helping to interpret data and resolve discrepancies. The authors particularly thank Drs. M. Boaz (Ariel University, Ariel, Israel), J. Castaneda (Medtronic, Maastricht, the Netherlands), O. Cohen (Sheba Medical Center, Tel Hashomer, Israel), W. Herman (University of Michigan, Ann Arbor, MI), S. de Portu (Medtronic, Tolochenaz, Switzerland), P. Raskin (University of Texas Southwestern Medical Center, Dallas, TX), I. Raz (Hebrew University of Jerusalem, Jerusalem, Israel), J. Wainstein (Wolfson Medical Center, Holon, Israel), and the Independent Review Board for Clinical Trials Data Access of Novo Nordisk (Copenhagen, Denmark).

Duality of Interest. J.C.P. has received speaker and/or consultancy fees from Medtronic, Roche, Insulet, Eli Lilly, Novo Nordisk, and CeQur that were not connected with the present work. Y.R. has received research funds from Medtronic, Eli Lilly, and Novo Nordisk and consultancy and/or speaker fees from Medtronic, Abbott, and Eli Lilly that were not connected with the present work. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. J.C.P. initiated and designed the study; performed the literature search, review, and data extraction from published articles; collected individual patient data from trialists and other sources; and wrote the first draft of the manuscript. Y.R. performed the literature search, review, and data extraction from published articles. A.J.S. designed the statistical plan, performed the statistical analyses, and contributed to the study design. All authors read, revised, and approved the final version of the manuscript.