The use of sodium-glucose co-transporter-2 inhibitors or glucagon-like peptide-1 receptor agonists versus sulfonylureas and the risk of lower limb amputations: a nation-wide cohort study

Lower limb amputation (LLA) is a highly feared outcome of type 2 diabetes (T2D) [1]. LLAs not only have a profound impact on the patient’s life, but are also associated with great economic burden [2] and high mortality rates [3]. LLAs are approximately ten times more common among individuals with diabetes compared to those without [4]. Particularly at risk of amputation are patients with diabetic foot ulcer (DFU) and peripheral arterial disease (PAD) [5].

Sodium-glucose co-transporter-2-inhibitors (SGLT2-Is) form a new class of drugs used in treating T2D. These drugs exert their glucose-lowering effect by preventing the reabsorption of plasma glucose and sodium in the proximal renal tubule [6]. The use of SGLT2-Is has been associated with lower mortality as well as positive cardiac and renal outcomes [7‐9]. However, they have also been associated with several safety concerns, such as a higher risk of LLA compared to placebo [10].

The first report of a higher risk of LLA associated with the use of the SGLT2-I canagliflozin in the CANagliflozin CardioVascular Assessment Study (CANVAS) Program [10] led to substantial concern. Subsequently, numerous studies were performed to investigate this potential association. Individual observational studies produced conflicting results, whereas meta-analyses reported no higher risk of LLA associated with SGLT2-I use [9, 11]. The wide range of contrasting findings may have resulted from differences in study population and study design, especially the use of different reference groups. Interestingly, two studies in which SGLT2-Is were compared to glucagon-like peptide-1 receptor agonists (GLP1-RAs) (another new class of glucose-lowering drugs) both reported a higher risk of LLA associated with SGLT2-I use [12, 13]. These findings raise the question whether the results of these studies could have been driven by a GLP1-RA-effect rather than an SGLT2-I-effect. Indeed, the GLP1-RA liraglutide has tentatively been associated with a lower risk of LLA compared to placebo [14]. The use of the GLP1-RA exenatide, however, was not associated with LLA risk versus placebo [15]. As a result, the association between SGLT2-I use and LLA remains uncertain.

SGLT2-I induced hypovolemia with reduced tissue perfusion has been proposed as a potential underlying pathway leading to LLA [16], but this mechanism has not been confirmed. A previous study reported no association between SGLT2-I use and LLA, irrespective of the presence of signs hypovolemia [17]. However, this study was limited by a low number of events, possibly because of the use of a primary care database in which LLAs might have been under recorded. LLA registration is likely to be more accurate in hospital registries using amputation procedure codes, such as the Danish National Patient registry [18].

Ultimately, the presence of an association between SGLT2-Is and LLA, the potential role of hypovolemia, and the influence of the use GLP1-RAs as a reference group remain unclear. Therefore, the primary aim of the current study was to investigate the risk of LLA with the use of SGLT2-Is or GLP1-RAs versus other antidiabetic drugs, using the Danish National Patient registry. The secondary aim was to study whether the presence of signs of hypovolemia or PAD is associated with a higher risk of LLA among SGLT2-I users, in order to assess whether these characteristics could be involved in the underlying pathway.

We conducted a retrospective cohort study using data from the Danish National Health Service. The Danish Civil Registry System assigns a unique 10-digit civil registry (CPR) number to all Danish residents at birth or immigration. This number is consistently used across all Danish registries and allows for linkage between the population-based registries [19]. Data on changes in vital status, including date of death and changes of address are registered in the Civil Registration System since 1968. Data on hospital admissions have been recorded since 1977 in the Danish National Hospital Registry. This registry covers 99.4% of all discharge records from Danish hospitals and includes all inpatient contacts, outpatient visits to the hospital, and outpatient visits to clinics and emergency rooms [20, 21]. Diagnoses are registered according to the Danish version of the International Classification of Diseases 10th revision (ICD-10) since 1994. Surgical procedures are registered according to the Danish version of the Nordic Medico-Statistical Committee Classification of Surgical Procedures since 1996. Data on refundable drug prescriptions have been recorded since 1996 in the register of Medicinal Products Statistics of the Danish Medicines Agency [22]. The prescription data provide information on the type and amount of drug prescribed according to the Anatomical Therapeutic Chemical (ATC) classification system, as well as the date on which the prescription was filled.

Figure 1 shows the selection of patients from the original data extraction, yielding a study cohort of 74,475 individuals who started treatment with the exposure(s) of interest. Table 1 shows the baseline characteristics of this study cohort, grouped by drug class of each individual’s index prescription (i.e. the first prescription of a study drug, marking cohort entry). Based on these index prescriptions, 13,736 individuals started treatment with an SGLT2-I, 11,512 with a GLP1-RA, and 12,851 with an SU. The median follow-up time was 1.6 years [inter quartile range (IQR) 0.7–2.9] for SGLT2-I users, 3.8 years (IQR 1.7–5.9) for GLP1-RA users, and 3.8 (IQR 2.1–5.6) for SU users. Median diabetes duration was highest in SGLT2-I users (6.6 years, IQR 2.8–10.1) and lowest in SU users (2.9 years, IQR 0.8–5.6). There were relatively more women among the GLP1-RA users, and their mean age was lower compared to the other groups. Overall, both SGLT2-I users and GLP1-RA users appeared to suffer more from comorbidities (hypertension, hyperlipidaemia, retinopathy, neuropathy) compared to SU users, and insulin use was higher in these two groups. However, renal disease was less common among SGLT2-I users compared to the other groups. There were few recordings of signs of hypovolemia and LLAs in the patients’ history.

In total we observed 564 LLAs (367 of which were forefoot amputations, i.e. through or below the metatarsal bones) during follow-up. The current SGLT2-I exposure time comprised of 49.8% empagliflozin, 47.3% dapagliflozin, 2.9% canagliflozin, and < 0.1% ertugliflozin. The current GLP1-RA exposure time comprised of 93.2% liraglutide, 3.3% dulaglutide, 1.7% exenatide, 1.6% semaglutide, and 0.3% lixisenatide. Table 2 shows that the risk of LLA was similar with current SGLT2-I use compared to current SU use (adjusted HR 1.10; 95% confidence interval [CI] 0.71–1.70). This finding remained consistent after stratification by sex and age groups. Similarly, the risk of forefoot amputations was similar with SGLT2-I use compared to SU use (adjusted HR 0.90 95%CI 0.52–1.57). From Table 3 it becomes apparent that the risk of LLA remains similar after stratification by concomitant antihypertensive use. Wald tests confirmed there were no differences between the groups of these stratifications. There were no LLAs with both SGLT2-I use exposure and a record of signs of hypovolaemia (volume depletion, anuria/oliguria, hypotension), so we were unable to study the potential effects of these conditions (data not shown). The risk of LLA was markedly elevated with the presence of PAD. Table 4 indicates that a higher cumulative dose but the duration of use of an SGLT2-I is associated with a higher risk of LLA.

The risk of LLAs was lower (adjusted HR 0.57; 95%CI 0.39–0.84) with current GLP1-RA use compared to current SU use (Table 5). After stratification, this finding remained apparent in females (9 LLAs) but not in males (46 LLAs) nor in any of the age categories (number of LLAs ranging from 6 to 18 per category). The lower risk was also observed when limiting the outcome to forefoot amputations only (adjusted HR 0.57; 95%CI 0.36–0.92). Table 6 shows that the lower risk of LLA remained consistent after stratifying by concomitant antihypertensive use. There was a limited number of LLAs with both GLP1-RA use and a record of signs of hypovolaemia (volume depletion, anuria/oliguria, hypotension), so we were unable to study the potential effects of these conditions (data not shown). Stratification by history of PAD showed a large difference in IR (yes 9.86, no 1.10 LLAs /1000 PY). Table 7 shows the association between longer duration of GLP1-RA use and the risk of LLA, which was mostly apparent in the highest continuous duration of use category (> 365 days, adjusted HR 0.44; 95%CI 0.27–0.72).

We observed 1,524 DFUs during follow-up. Neither SGLT2-I use nor GLP1-RA use were associated with the risk of DFU (adjusted HR 1.18; 95%CI 0.90–1.55 and adjusted HR 1.08; 95%CI 0.87–1.35, respectively).

In the first sensitivity analysis in which we changed the reference group from SU to DPP4-I, we obtained similar results to our main analyses. There was no significant association of SGLT2-I use with the risk of LLA (adjusted HR 1.23; 95%CI 0.85–1.79), forefoot amputation (adjusted HR 0.98; 95%CI 0.61–1.58) and DFU (adjusted HR 1.06; 95%CI 0.85–1.33) compared to SU use. GLP1 use was associated with a lower risk of LLA (adjusted HR 0.64; 95%CI 0.47–0.88) and forefoot amputation (adjusted HR 0.63; 95%CI 0.43–0.92), and a similar risk of DFU (adjusted HR 1.00; 95%CI 0.84–1.17) compared to SU use (Additional file 1: Table S1). In the second sensitivity analysis in which we excluded individuals with a history of LLA at baseline, we observed 470 LLAs during follow up and the findings remained consistent. There were 29 LLAs with current SGLT2-I use (adjusted HR 1.05; 95%CI 0.65–1.69) and 43 with current GLP1-RA use (adjusted HR 0.51; 95%CI 0.33–0.78) (Additional file 1: Table S2). Finally, the addition of number of concomitant glucose-lowering drugs and number of bed days as confounders, did not alter the results. The resulting HRs were 0.97 (95%CI 0.62–1.52) with current SGLT2-I use and 0.51 (0.34–0.76) with current GLP1-RA use.

The results of this nation-wide cohort study indicated that SGLT2-I use was not associated with a higher risk of LLA compared to SU. This finding remained consistent after stratification by age, sex, concomitant antihypertensive use, history of PAD, and continuous duration of use categories. GLP1-RA use, however, was associated with a lower risk of LLA. This risk reduction was particularly apparent in men, but also remained consistent after stratification by concomitant antihypertensive use and history of PAD. Stratification by continuous duration of use indicated that increasing duration of GLP1-RA use was associated with a lower risk of LLA. Both SGLT2-Is and GLP1-RAs did not alter the risk of DFU.

The findings of previous studies might have been influenced by choosing GLP1-RAs as a reference group when evaluating the risk of LLA. In the current study, we took another approach at evaluating the association between SGLT2-Is and GLP1-RAs and LLA, by separately comparing them to an active comparator. The results of this study are in line with meta-analyses in which no higher risk of LLA was observed with SGLT2-I use compared to other glucose-lowering drugs and placebo [9, 11, 28, 29]. One meta-analysis of RCTs observed a slightly higher risk of LLA with SGLT2-I versus placebo [30], but this risk was driven by canagliflozin, the SGLT2-I which was hardly used by the participants in the current study. Another meta-analysis reported no evidence was found for an effect of SGLT2-I or GLP1-RA on LLA, but there was high uncertainty regarding these results based on Grading of Recommendations, Assessment, Development and Evaluations (GRADE) [31]. Our results are also in line with a recent meta-analysis of observational studies, with a hypothesis about SGLT2-I versus GLP1-RA effects that was similar to ours [32]. Indeed, the author reported a lower risk of LLA with GLP1-RA compared to SGLT2-I. Unfortunately none of the studies included were designed to investigate LLAs and both direct and indirect comparisons were made. Moreover, in contrast to our study, the comparison of SGLT2-I with GLP1-RA did not allow for distinction between the effects exerted by the drug classes separately. Finally, our results are in line with several cohort studies which reported no higher risk of LLA with SGLT2-Is [17, 33, 34] and a lower risk of LLA with GLP1-RAs [30, 35].

Overall, the available evidence for the association between SGLT2-I use and GLP1-RA use and the risk of LLA is limited, has high uncertainty [31], and the studies are highly heterogeneous [36]. It is important to note that the choice of reference group varies among the abovementioned studies. These differences could play a large role in the conflicting nature of the reported results. Some studies investigating the risk of LLA with SGLT2-I have chosen GLP1-RAs as the reference group. Several of these studies report no difference [33, 34, 37], whereas others report a higher risk with SGLT2-I [12, 13, 38, 39]. In order to compare the current study with these previous findings, we performed a post-hoc analysis comparing SGLT2-I directly to GLP1-RAs. We observed an increased risk with current SGLT2-I use versus current GLP1-RA use (adjusted HR 1.91; 95%CI 1.25–2.92), which is in line with our expectations, as well as the previous studies. The comparison of SGLT2-I with SU has only been reported in two cohort studies; one reporting a similar risk of LLA [17] and one reporting a lower risk with SGLT2-Is [40]. The discordant findings between the current study and the study of Dawwas et al. [40], may be explained by a difference in populations included in the used data sources. The population included in the study by Dawwas et al. consisted of privately insured individuals with employer-sponsored health insurance programs and retirees whereas our study includes the complete Danish population. Furthermore, a shorter average follow-up time (median of 131 days) was reported potentially explaining the lower IRs found in the study by Dawwas et al. As a result of these differences a part of the population at risk of LLA is possibly not captured and potential harmful or protective effects on LLA may not have been visible yet. The comparison of GLP1-RA with SU has not been reported before, most likely because in most countries these drugs are not used in a similar disease stage. In Denmark, however, both these drug classes, as well as SGLT2-Is, are recommended as second-line treatment [24, 41].

The effect size of current GLP1-RA on the risk of LLA was remarkably large in this study, with approximately 40% risk reduction in our analyses. In order to rule out potential issues causing such an unexpectedly large effect, we further explored this finding. We found no disproportionality regarding person time, IRs and unadjusted HRs. Since the large effect size was particularly apparent in the fully adjusted HRs, we evaluated various aspects of the confounders included in the models. We ruled out multicollinearity and that individual confounders had an unexpected effect on the beta coefficient in the model, paying special attention to insulin use. The results remained consistent in models with and without correction for concomitant insulin use. Furthermore, we explored the possibility of distortion by unmeasured confounding. We calculated the e-value, which estimates the strength of unmeasured confounding required to nullify the observed effect [42, 43]. Using the observed HR of 0.57(95%CI 0.39–0.84) (Table 5), the e-value for the HR point estimate was 2.9 and the e-value for the upper bound of the CI was 1.7. This means that the unmeasured confounding would have to threefold the risk in order to nullify the HR, or have a strength of 1.7 to shift the CI to overlap the null. Although some highly intensive treatment interventions have shown to reduce the risk of LLA up to threefold compared to usual care [44], we don’t expect these types of strategies would be implemented differentially between the evaluated drug classes in this study. So, although we are unsure of the actual size of the effect, substantial unmeasured confounding is required to nullify the observed effect. Therefore, it is safe to say our data suggest GLP1-RA use has a protective effect against LLA.

Many mechanisms through which drugs could alter the risk of LLA have been proposed. Although the effect of SGLT2-Is on LLA is highly debated, the most commonly proposed mechanism for this potential effect is hypovolaemia. It has been proposed that the increased sodium excretion and osmotic diuresis caused by increased glucose excretion due to SGLT2-I use, could lead to reduced peripheral tissue perfusion, necrosis and finally requiring amputation [16]. This hypovolaemia-based mechanism is supported by the finding of an higher risk of LLA in T2D patients using diuretics compared to those not using diuretics [45] and the association of a reduction in body weight and blood pressure with lower limb complications in patients using SGLT2-Is [46]. However, in the current study we did not observe a higher risk of LLA with SGLT2-Is, nor did we find LLAs with current SGLT2-I exposure and the presence of signs of hypovolaemia. Moreover, we did not observe a change in the risk of LLA with or without concomitant diuretic or RAAS inhibitor use.

The evidence regarding the role of GLP1 in the processes leading to LLA is limited. We speculate the pathways involved could be either via effects on wound healing or PAD. Findings from basic research have suggested that GLP1-RAs may promote DFU healing by boosting various physiological processes involved in diabetic wound healing [47, 48]. This would prevent a non-healing DFU to require amputation as treatment. The other possible pathway is via PAD, since GLP1 exhibits positive vascular effects [49, 50]. In the current study, we did not observe a lower risk of LLA in patients with a history of PAD; the risk was elevated in this high risk group with both SGLT2-I and GLP1-RA use. Based on this finding, it is more likely that wound healing might play a role in the protective effect of GLP1-RAs. We observed no change in risk of DFU with GLP1-RA use, but we were unable to study the healing of DFUs, the phase in which GLP1-RAs might potentially exert their effect. Therefore, further study is required to look into the mechanism underlying the protective effect of GLP1-RAs on LLA.

This study has several strengths. The availability of nation-wide real-world data and the completeness of the Danish Health Registries allowed for more accurate recording of LLAs. Moreover, the database allowed us to correct for a comprehensive set of confounders, such as socio-economic variables, which play an important role in the development of diabetic foot issues [51]. Another strength is the consistency of the results across various stratifications, showing that the models are robust and the findings reliable. Additionally, the findings remained consistent in the sensitivity analyses in which we changed the reference group to another second-line treatment option, and another in which we excluded individuals with LLA before index date. This further supports the fact that our models are robust and produced reliable findings. Moreover, we critically assessed potential causes for overestimation of the effect size observed with current GLP1-RA exposure but found no indications of augmentation of the effect size. Finally, a post-hoc analysis comparing current SGLT2-I use directly to current GLP1-RA use, confirmed that the use of GLP1-RA use as a reference group indeed leads to the observation of an increased risk of LLA.

Despite the availability of nation-wide data, the observational study design requires us to consider several limitations when interpreting the results. The results might have been influenced by unmeasured confounding by factors which were not available in the dataset, such as glycosylated haemoglobin A1c (HbA1c) values, body mass index, and smoking status. However, the e-value we calculated indicated a considerable amount of unmeasured confounding would be required to nullify the observed effect. Another limitation to take into account is the limited number of events. Due to the small number of events, we were unable to study various signs of hypovolaemia, and to evaluate substance-specific effects instead of drug-class effects. The power issue caused by the limited number of events is also reflected in some stratifications. The point estimate of the protective effect of GLP1-RAs is similar in men and women, but the 95% CI indicates significance only in men. This is also the case in the highest age stratum (70 +) of current GLP1-RA use. Although the point estimate for the risk of LLA with SGLT2-I versus SU was close to 1, the width of the CI also allows for a true risk to be either higher or lower.

Moreover, records of DFU might be less accurate then records of LLA, since amputation procedures are recorded in the hospital database, whereas a DFU record relies on a registration of a diagnosis. Therefore the data might suffer from underrecording of DFU diagnoses.

In conclusion, this nation-wide cohort study indicates that SGLT2-I use is not associated with a higher risk of LLA compared to SU use. However, GLP1-RA use was associated with a lower risk of LLA. The risk of LLA was higher in patients with PAD, irrespective of SGLT2-I or GLP1-RA use. The results of this study add to the growing body of evidence that SGLT2-I use is not associated with a higher risk of LLA, and help to explain the observed harmful effect when comparing this drug class to GLP1-RAs. Previous studies reporting a higher risk of LLA with SGLT2-I use compared to GLP1-RA use might have been looking at a protective GLP1-RA-effect, instead of a harmful SGLT2-I-effect. The current findings put forward GLP1-RAs as a potential treatment strategy to reduce the risk of LLA in patients with T2D, but further study is required to confirm this potential effect and the underlying mechanisms.