Marked improvements in glycaemic outcomes following insulin pump therapy initiation in people with type 1 diabetes: a nationwide observational study in Scotland

Data and data linkage were originally set up with approval from the Scottish A Research Ethics Committee (ref 11/AL/0225), Caldicott Guardians and the Privacy Advisory Committee (PAC - ref. 33/11), now running with approval from the Public Benefit and Privacy Panel for Health and Social Care (PBPP - reference 1617-0147).

We used anonymised data from the Scottish Care Information – Diabetes Collaboration database (SCI-Diabetes), which holds extensive electronic healthcare records of >99% of people with diabetes in Scotland, both adults and children, since 2004. SCI-Diabetes covers extensive clinical and prescription information including CSII start dates. These data are linked to hospital admissions records and mortality data from the Information Services Division (ISD) of NHS Scotland and the National Records of Scotland. SCI-Diabetes and associated linkage have previously been described [8, 9].

Among those alive and observable with type 1 diabetes in SCI-Diabetes at any time from 2004 to 2019, we analysed glycaemic outcomes among those who initiated CSII therapy before June 2019. Type of diabetes was assessed from our validated algorithm [10]. Individuals contributed person-time from the latest of: 1 January 2004, first observability date (based on prescribing, hospital admission or clinical measurements) or date of diabetes diagnosis. Person-time was censored at the earliest of: 12 December 2019 (administrative censoring), date of death, last date of observability, first stop date of CSII (if no restart within a year). Person-time was additionally censored when starting flash monitoring or continuous glucose monitoring (CGM) device after CSII to isolate the effect of CSII initiation from that of other technologies (see electronic supplementary material [ESM] Methods).

Exposure of interest is CSII therapy. Since CSII users in Scotland are selected based on criteria that are not measurable in our data (e.g. motivation, perceived ability to use a pump), we chose to conduct analyses within CSII users. Glycaemic outcomes were compared among CSII users before and after CSII initiation, using data up to a maximum of 5 years prior to initiation. We used a 1:1 non-user control group, matched by age, sex and diabetes duration (ESM Methods) to check for any trend occurring in HbA_1c among non-users in similar calendar time, and to inform results obtained in adolescents given the deterioration in glycaemic control in this age group [2, 11].

We defined DKA and SHH as any hospital admission or death involving the ICD-10 (http://​apps.​who.​int/​classifications/​icd10/​browse/​2016/​en) codes detailed in ESM Methods. Each unique hospital admission was considered as an event. Person-time for DKA/SHH analyses was censored at the last date of available data for hospital admissions: 30 April 2019.

Demographic and clinical information were obtained from SCI-Diabetes. For clinical covariates, baseline values were defined as median value over the 2 years prior to device initiation for continuous variables or the most clinically severe status over this time-window for categorical variables. We used Scottish Index of Multiple Deprivation (SIMD) quintiles as a measure of area-level deprivation, quintile (Q)1 being the most deprived. Age at CSII initiation (years) was categorised into six age bands: <13, ≥13–≤18, ≥19–≤24, ≥25–≤44, ≥45–≤64 and >64; and baseline HbA_1c levels (mmol/mol) into five groups, representative of different glycaemic control levels: <54 (7.1%), ≥54–≤68 (8.4%), ≥69–≤74 (8.5–8.9%), ≥75–≤84 (9.0–9.8%) and >84.

All analyses were conducted using R version 3.6.0–64 bit [12] at significance level p = 0.05. Missing data were not imputed.

We presented the crude prevalence of ever-CSII users overall and stratified by baseline variable. In order to study the effect of CSII on HbA_1c, we divided observation time into blocks of 1 year periods before and after CSII initiation, centred on the date of CSII initiation [7], and compared the median HbA_1c value of each individual in each time-block. HbA_1c levels and absolute within-person differences with respect to baseline were described before and after CSII initiation, overall and stratified by age, baseline HbA_1c, sex, SIMD and prior flash monitoring/CGM use. Significance of within-person differences were assessed at each time point after CSII initiation, using a one-sided Wilcoxon signed-rank test; p values were adjusted for multiple comparison using Bonferroni’s correction.

Crude pre-/post-exposure changes can under/overestimate an intervention’s effect when there is a pre-intervention trend in the outcome which is not taken into account. Log-transformed HbA_1c was modelled using linear mixed models [13] to assess whether changes were similar and statistically significant when accounting for pre-exposure trajectory [14]. CSII initiation in Scotland is a diffuse intervention. It is typically preceded by more frequent clinical encounters and attendance at a structured education programme, such as Dose Adjustment For Normal Eating (DAFNE), in the 2 years prior to initiation. In modelling the impact of CSII, we wished to ensure that any within-person trend over calendar time unrelated to CSII could be accounted for in the analysis. In order to evaluate this trend, the HbA_1c measurements within the 2 years prior to CSII initiation were excluded since they are related to CSII (see ESM Methods for more details).

Therefore the models estimated the ‘full package’ effect of CSII plus related prior interventions in the 2 year run-up to CSII as they reflect change in HbA_1c from before the package started to after it was fully implemented. Models were fitted using R package nlme version 3.1–143 [15]. Differential effect of CSII by sociodemographic variable was tested using interaction terms (ESM Methods). Where significant, stratified model results were presented.

Due to the sparse nature of DKA/SHH events, we described the crude event-rate before and after CSII, by dividing the number of events observed by the total number of person-years under observation pre- or post-CSII exposure. To account for repeated occurrence of events within individuals and the background pre-package time-trend, we used generalised linear mixed models with a Poisson likelihood with a random intercept, excluding the 2 years prior to initiation. CSII exposure was included as a binary, time-varying covariate (non-exposed/exposed). All models were adjusted for sex, age and diabetes duration at CSII initiation, baseline HbA_1c and pre-package time-trend. Classical likelihood inference for this model relies on approximations of intractable integrals, therefore we used a Bayesian approach [16]. We presented the back-transformed posterior mean (rate ratio) and associated 95% credible intervals (CrI), alongside the posterior probability (pp) of the β coefficient associated to CSII exposure being positive or negative in order to quantify the uncertainty around the direction of the effect [17]. Models were implemented using R package rstan version 2.19.3 (see ESM Methods for more details).

A total of 4941 people ever used CSII between 2004 and the end of 2019.The crude prevalence of ever-CSII users has increased over the past decade from 0.1% in 2004 to 15.1% in 2019. Disparities in CSII usage by the end of 2019 were observed across sociodemographic strata (ESM Fig. 1). Prevalence of ever-CSII users was highest in younger age bands and decreased with age (48.1% in the under 13 age band vs 4.3% in those aged 65 years or older), higher in female (19.9%) vs male participants (11.3%) and twice as high among those living in the least vs the most deprived areas (19.6% in SIMD Q5 vs 10.4% in Q1).

We analysed glycaemic outcomes in 4684 CSII users who initiated CSII before June 2019. Their baseline characteristics are described in Table 1, alongside those of the non-user sample. Their median overall observation time was 12.4 years; median post-exposure follow-up time was 3.6 years; and 3108 individuals started using flash monitoring/CGM after starting CSII.

HbA_1c decreased after CSII initiation (ESM Fig.2) from median 66.0 mmol/mol (IQR 58.0, 74.0) (8.2% [IQR 7.5, 8.9]) in the year before initiation to 60.0 mmol/l (54.0, 67.0) (7.6% [7.1, 8.3]) during the first year of CSII exposure and 63.0 (56.0, 71.0) (7.9% [7.3, 8.6]) after ≥5 years. HbA_1c levels started decreasing around 2 years prior to CSII initiation, and levels were stable in the preceding years.

Crude within-person differences in HbA_1c levels with respect to baseline HbA_1c are illustrated in Fig. 1 overall and stratified by age at CSII initiation, sex, SIMD quintile and baseline HbA_1c. At an individual level, there was a reduction in HbA_1c levels post-CSII initiation sustained through time (ESM Table 1): median −5.5 mmol/mol (IQR −12.0, 0.0) (−0.5% [IQR −1.1, 0.0]) (p < 0.01) within the first year post-exposure and −5.0 mmol/mol (−12.5, 2.8) (−0.5% [−1.1, 0.3]) (p < 0.01) for ≥5 years. Mixed-model results showed no time-trend trajectory of HbA_1c prior to CSII-package initiation. CSII initiation was associated with a significant and sustained long-term reduction in HbA_1c levels compared with pre-package levels (Table 2). The estimated fold-change was 0.88 overall. For someone with pre-package HbA_1c levels of 69 mmol/mol (8.5%), this would correspond to a change of around −8 mmol/mol (−0.7%).

Within-person HbA_1c remained stable in the calendar time following pump initiation in the non-user group, with a within-person change of 0.0 mmol/mol (IQR −6.0, 5.5) (0.0% [−0.5, 0.5]) in year 1 to −1.0 mmol/mol (−10.0, 9.0) (−0.1% [−0.9, 0.8]) at ≥5 years (ESM Fig. 3).

CSII exposure effect varied significantly across baseline HbA_1c, age and sex, and prior flash monitoring/CGM use (p_interaction <0.01 for all), but not across SIMD (p_interaction = 0.25).

Reductions in HbA_1c from baseline were observed in all bands ≥54 mmol/mol (7.1%) (Fig. 1, ESM Table 1), whereas HbA_1c levels increased slightly in the <54 mmol/mol band. Reductions were significant and sustained through time, with the greatest fall corresponding to the highest baseline HbA_1c bands: −21.0 mmol/mol (IQR −30.0, −11.0) (−1.9% [IQR −2.7, −1.0]) (p < 0.01) in the first year post-exposure and −19.0 mmol/mol (−27.6, −6.5) (−1.7% [−2.5, −0.6]) at ≥5 years in those with baseline HbA_1c >84 mmol/mol (9.8%) (p < 0.01). Results from the mixed models were similar (Table 2), with an estimated 0.8-fold-change sustained through time in this group.

Within-person reductions in HbA_1c were slightly higher in female than male participants (ESM Tables 2, 3).

Across age, within-person HbA_1c reductions were highest among those aged 19–24. Significant reductions were observed in all age bands and sustained through time except in the 13–18 years group. In these adolescents, after an initial decrease, there was no significant difference from baseline beyond the second year. Model estimates showed marginally higher values compared with the expected pre-package HbA_1c trajectory (ESM Tables 4, 5). When compared with the matched controls (ESM Fig. 4), HbA_1c was lower in the adolescent group in users both before and after CSII initiation. In non-users, HbA_1c rose from baseline levels until 3 years post-index: median within-person change from baseline 5.0 mmol/mol (IQR −6.0, 16.5) (0.5% [IQR 0.5, 1.5]) vs −0.0 mmol/mol (−8.4, 8.0) (0.0% [−0.8, 0.7]) in non-users vs users respectively (p < 0.01). Levels returned close to baseline in subsequent years.

Within-person reductions in HbA_1c did not vary across SIMD (ESM Table 6). They were slightly higher in those with prior flash monitoring/CGM use (ESM Tables 7, 8).

Overall, 1187 DKA events were observed over 17,479.8 person-years prior to CSII initiation, and 827 DKA events over 16,675.0 person-years after initiation. The crude DKA rate increased prior to package initiation and started decreasing thereafter (ESM Fig. 5, ESM Table 9).

The crude DKA event-rate was significantly lower in CSII-exposed person-time than in non-exposed person-time: 49.6 events (95% CI 46.3, 53.1) per 1000 person-years vs 67.9 (64.1, 71.9).

The Bayesian mixed models confirmed the increasing trend in DKA frequency prior to package initiation and demonstrated that CSII exposure was associated with a reduction in DKA rate compared with the counterfactual: rate ratio 0.61 (95% CrI 0.47, 0.77), with pp of β < 0 (i.e. reduction) = 1.00 (ESM Table 10). Hence, we can be very confident that there is a true association.

Crude DKA rates by CSII exposure status, stratified by baseline HbA_1c band, sex, age and SIMD are described in Fig. 2.

Stratified by baseline HbA_1c band, significant reductions in DKA rates were observed post-exposure in the 54–68 mmol/mol and ≥ 75 mmol/mol groups (ESM Table 11). Model results were consistent with an association of CSII exposure with a reduction in DKA rate when controlling for pre-intervention trend in the 54–68 mmol/mol and the >84 mmol/mol bands (ESM Table 12).

Crude DKA rates were higher in female than male participants (ESM Table 13) but lower in exposed person-time in both groups. Model results were similar, with CSII being associated with a rate reduction in both groups (ESM Table 14).

Across age band, crude rates were lower in CSII-exposed time, except in the 13–18 years and 45–64 years groups (Fig. 2, ESM Table 15). Model results were generally similar (ESM Table 16). Data were too sparse in the 65 years and older group to draw meaningful conclusions. DKA rates in the 13–18 years group were higher in non-users than users both before and after pump start-date. The crude event-rate in non-users increased from 157.5 (95% CI 139.9, 176.7) pre-index to 244.4 (224.7, 265.4) in the post-index person-time, i.e. as the adolescents get older (ESM Fig. 6). The post-/pre-index increase was similar to that observed in pump users: crude rate ratio 1.55 (95% CI 1.35, 1.79) in non-users vs rate ratio 1.43 (1.16, 1.77) in users.

Higher deprivation levels were associated with increased crude DKA rates. Rates were lower in exposed vs unexposed person-time, though the difference was not always significant (Fig. 2, ESM Table 17). Model results were similar (ESM Table 18).

Crude DKA rates were lower in CSII-exposed time in both those with and without prior flash monitoring/CGM use. Sample size was too low for meaningful inference in the prior flash monitoring/CGM group (ESM Tables 19, 20).

Overall, 451 SHH events were observed in the unexposed person-years, and 296 events in the exposed person-years. The crude SHH rate increased in all years prior to CSII initiation. The crude SHH event-rate was lower in CSII-exposed than non-exposed person-time: 17.8 events per 1000 person-years (95% CI 15.8, 19.9) vs 25.8 (23.5, 28.3). Model estimates from the Bayesian mixed models were consistent with an association of CSII with a reduction in SHH rates compared with pre-package levels: rate ratio 0.67 (95% CrI 0.45, 1.01), pp = 0.97 (ESM Table 21). Hence, we have good confidence that there is a true association. Event numbers were too low to perform stratified analyses.

This is one of the largest real-world, contemporary, nationwide assessments of the prevalence of CSII and its effect on glycaemic outcomes, combining near-complete capture of data on people with type 1 diabetes nationally, long follow-up time and extensive analyses.

The main limitations of our study are the lack of data in older people, incomplete capture of self-funded flash monitoring use and non-capture of non-hospitalised severe hypoglycaemia [31]. CSIIs can be prescribed for debilitating hypoglycaemia in adults or children older than 12, as recommended by NICE [6], but we could not evaluate hypoglycaemia as an effectiveness outcome for CSII therapy.

In Scotland, eligibility for receiving CSII on the NHS has evolved over time: from 2011 there was a major push to widen provision, and the clinical and sociodemographic profile of those receiving pumps has changed. However, the vast majority of CSII initiation occurred since 2011, with most of our data reflecting those initiating pump therapy in the period 2013–2017 and followed up to the present day (median start date 5 December 2014 [IQR 16 June 2013, 27 February 2017]). Therefore, our results reflect the contemporary profile and outcomes of those in receipt of CSII therapy which is most relevant to current effectiveness and safety estimation.

The majority of CSII users had their person-time censored for start of flash monitoring/CGM. The person-time lost to follow-up was relatively low and unlikely to affect the analyses since the step-change in HbA_1c occurs within the first year post-CSII initiation.

Generalisability of findings to all people with type 1 diabetes is limited by the allocation bias resulting from the selection process for pump therapy.

Our HbA_1c analyses are susceptible to regression to the mean. However, the magnitude of pre–post change among never-users indicates that only a small proportion of the decrease observed in users following CSII initiation would actually be attributable to this artefact. Finally, as with all observational analyses, our estimates may be affected by unmeasured confounding and measurement error, although the self-controlled nature of some analyses eliminates time-invariant confounding.

Our study shows that CSII therapy is associated with improvements in HbA_1c, DKA and SHH among pump users with type 1 diabetes in Scotland. The improvements associated with CSII therapy suggest that a more widespread use of pump therapy has the potential to reduce HbA_1c levels and DKA rates further. Our study highlights the importance of CSII technology reaching those with suboptimal glycaemic control and those from most deprived areas. This latter group, with historically poorer glycaemic control [10], stands to benefit from CSII usage, as we have not found evidence of CSII effects being worse in people from more deprived areas (in contrast to their CSII usage rates). Future research is needed to assess the effect of SIMD on outcomes in specific subgroups. Generalisation of our findings to all people with type 1 diabetes in Scotland is limited by the specific eligibility criteria to CSII. Hence, it is crucial to identify the determinants of good response to CSII to optimise a targeted wider roll-out. Moreover, other outcomes that we cannot directly quantify need to be considered, such as improvement in quality of life and reduction in glucose variability. Motivation is critical to successful CSII usage [23], and it would be interesting to examine how outcomes differ in a known-to-be-motivated group (pregnant women). Finally, future research is warranted to examine how the rapid evolution of technology (integrated pump-and-CGM systems with alarms and/or suspension systems, hybrid/DIY closed loops) affects DKA and severe hypoglycaemia rates.