Insulin, hypoglycaemia and dementia: A causal mediation analysis showcasing challenges and potential opportunities

In a recent cohort study, insulin use was not associated with an increased risk of all-cause dementia after adjusting for confounding.1 The study showed how confounding by disease severity because of the lack of a clinically appropriate comparator could account for the higher risk of dementia with insulin use observed in previous observational studies.2, 3 However, questions surrounding the role of hypoglycaemia as a potential mediator of the insulin-dementia association remain.1-4 Serious hypoglycaemia episodes, defined based on hospitalization or a physician visit, have been consistently linked to an increased risk of dementia in observational studies using data from various routine-clinical settings.5 Given the plausible pathophysiological and pharmacological pathway by which insulin use may lead to hypoglycaemia, which in turn is linked to an increased risk of dementia, we believe a mediation analysis is worth conducting despite the absence of a total effect (i.e. the lack of an overall association between insulin use and dementia). It has been argued that mediation analyses should be informed by clinical hypotheses that are based on a priori biological knowledge rather than by the total effect.6 The risk of hypoglycaemia with insulin use is established and the reduction of glucose supply from the peripheral circulation to the brain negatively affects cognitive function. While this is often reversed, it has been hypothesized that more severe hypoglycaemia episodes can increase platelet aggregation and fibrinogen formation and lead to irreversible damage, including neuronal cell death.7, 8 Herein, we apply recent advances in mediation analysis to detail the direct (not through hypoglycaemia) and the indirect (through hypoglycaemia) effect of insulin use on the risk of dementia. We leveraged data from a previous cohort study of patients with newly diagnosed type 2 diabetes without a history of dementia (n = 414 089), wherein we used administrative health care data from British Columbia, Canada (1996-2018).1 In that study, diabetes was defined based on the validated case-defining algorithm from the Canadian Chronic Disease Surveillance System, whereby diabetes is defined as the earliest occurrence of two physician claims or one hospitalization captured by relevant ICD codes (Table S1) within a 2-year period.9 To ensure diabetes was newly diagnosed, we used a 2-year washout period during which patients did not receive any diabetes diagnosis codes or any antihyperglycaemic medications.1 We minimized confounding by diabetes severity by first restricting the cohort to those who received two distinct non-insulin antihyperglycaemic classes of medications, from whom we identified 40-70-year-old new users of insulin (n = 7863) or a non-insulin class (n = 25 230) between 1 January 1998 and 31 December 2016.1 The outcome of interest was all-cause dementia defined based on validated definition as one hospitalization code, three physician claims codes (at least 30 days apart in a 2-year period), or a prescription filled for a cholinesterase inhibitor or memantine (Table S1).10 The index date was the date of initiating insulin or a non-insulin class and patients were followed until dementia, death, emigration, switching between exposure groups or 31 December 2018. Switching between exposure groups was defined as discontinuation of insulin (>180-day gap) among insulin users or initiating insulin among non-insulin users. We assessed hypoglycaemia episodes that required medical attention between the index date and the end of follow-up. Hypoglycaemia was defined based on primary or non-primary hospitalization codes using ICD-10-CA codes (E15, E11.63, E13.63, E14.63, E16.0, E16.1, E16.2) and physician visit codes using ICD-9-CM (251.0, 251.1, 251.2). Hypoglycaemia episodes occurring within 1 year before dementia were not counted given the possibility of reverse causality, whereby undiagnosed dementia may increase the risk of hypoglycaemia. We used inverse probability of treatment weighting based on the high-dimensional propensity score algorithm to adjust for >500 potential confounders, including demographics, indicators of diabetes severity, previous medication use and comorbidities (Table S2). To assess the natural direct effect (NDE) and natural indirect effect (NIE) of insulin on the risk of dementia, we used the causal mediation analysis approach based on concepts of the potential outcomes framework for causal inference.11, 12 The use of this approach has increased dramatically over the last decade because of its advantages over naïve methods, such as the multiplicative and additive method.12 To estimate the direct and indirect effects, we combined parameters from (a) a logistic regression model for post index date hypoglycaemia conditional on insulin exposure, and (b) a Cox survival regression model for all-cause dementia conditional on treatment and post-index hypoglycaemia allowing for exposure-mediator interaction. We used the delta method to obtain standard errors and 95% confidence interval (CI). Given the large sample, the use of delta method standard errors may be preferred over bootstrapping because of computational efficiency.13 The causal mediation analysis SAS 9.4 macro was used for this analysis.14 We conducted two sensitivity analyses, i.e. (a) we varied the 'as-treated' exposure definition to allow any gaps in treatment, and (b) we excluded those who received sulphonylurea as the third non-insulin class comparator and censored those who received sulphonylurea during follow-up to minimize underestimating the insulin effect. Like insulin, sulphonylurea is also a class of diabetes medications with a high risk of hypoglycaemia and has been reported to increase the risk of dementia.2, 3 In the original study, a total of 78 dementia events occurred over a median (IQR) follow-up period of 3.9 (5.8) years among insulin users (weighted incidence rate: 1.61; 95% CI 1.24-2.09 per 1000 person-years) and 179 events over 4.6 (4.4) years among non-insulin users (1.43; 95% CI 1.24-1.65 per 1000 person-years).1 In this primary mediation analysis, the unadjusted total effect of insulin on dementia, as shown by the hazard ratio (95% CI), was 1.58 (1.21-2.07) with an NDE = 1.49 (1.14-1.97) and an NIE = 1.06 (1.02-1.09). After inverse probability of treatment weighting, the total effect of insulin on dementia, as shown by the hazard ratio (95% CI) was 1.06 (0.78-1.44) with an NDE = 1.02 (0.74-1.40) and an NIE = 1.04 (1.01-1.08) with a proportion mediated of 0.66 (Table 1). Regression coefficients are reported in Figure 1. Our results show a potential small indirect effect of insulin use on dementia through hypoglycaemia. This overall conclusion was consistent across the different exposure contrasts and definitions. This finding is clinically plausible given the existing evidence on the increased risk of all-cause dementia with hypoglycaemia episodes that required medical attention; nonetheless, caution in interpretation is warranted. First, a small NIE is observed. Given the use of administrative health care data, only hypoglycaemia episodes recorded in a physician visit or hospitalization claim were captured, while mild and moderate episodes were not captured. Hence, we expect an underestimation of the mediator among those exposed to insulin and, thereby, an underestimation of the NIE. Conversely, there is a higher level of awareness and concern of hypoglycaemia among patients receiving insulin and their health care providers, potentially leading to detection bias and therefore an overestimation of the NIE. Prospective longitudinal data on plasma glucose measurements that span over multiple decades would be a valuable source to ascertain all hypoglycaemia episodes. Second, a significant NIE in the absence of a total effect indicates 'inconsistent mediation', which suggests that the mediation effect has a different direction than other mediated or direct effects in the model.6, 15 Statistical power needed to detect a significant NIE is less than that needed for the test of total effect.16 Therefore, it is more probable to find a significant NIE than a significant total effect, particularly when these effects are not large.16 Although our study was population-based at the provincial level, our restrictions to adjust for confounding by disease severity diminished the final sample size. Conducting a multicentre analysis can help confirm these findings. Third, despite using longitudinal data and excluding hypoglycaemia episodes occurring within a year of dementia, reverse causality cannot be ruled out. Cognitive impairment before a diagnosis of dementia increases the risk of experiencing hypoglycaemia episodes.17 Further longitudinal data that include detailed cognitive assessments along with neuroimaging results are necessary to assess better the possibility of reverse causality. Fourth, an overinterpretation of the proportion mediated is not recommended as this measure is not robust.18 Nonetheless, the proportions reported can provide valuable information on other potential protective pathways by which insulin use is associated with dementia. Indeed, this is evident by findings from one of the sensitivity analyses, wherein the proportion mediated was >1 when we excluded sulphonylurea from the comparator group. This occurs when the size of the mediation effect is larger than the total effect, that is, when the directions of the main and mediated effects are opposite.15, 18 Pharmacological evidence to support a protective role of insulin on cognition exists.19, 20 Besides its effect on reducing glycated haemoglobin, insulin has neuromodulatory actions in the brain that are hypothesized to improve cognition, including synaptic formation and remodelling, regulation of neurotransmitters and amyloid clearance.19 Multiple clinical trials have been conducted to assess the role of intranasal administration of insulin, whereby hypoglycaemia is avoided, in improving cognition.20 Fifth, in the presence of an unmeasured confounder that affects only the hypoglycaemia-dementia relationship, the total effect of insulin use on dementia will remain unbiased, but the indirect effect of insulin use on dementia through hypoglycaemia will be biased.21 Clinically, a confounder that can affect the hypoglycaemia-dementia relationship but not the insulin-dementia relationship is unlikely; however, it can never be discounted. Despite adjusting for a wide range of variables, we did not have access to all potential confounders that can affect the insulin-hypoglycaemia, hypoglycaemia-dementia and insulin-dementia relationships, including laboratory measures (e.g. blood pressure, lipid levels), health indicators and/or behaviours (e.g. BMI, alcohol consumption, frailty). Sixth, given the lack of a clinically appropriate active comparator that is used in a similar disease stage as insulin, this study focused on comparing the use of insulin to the use of non-insulin antidiabetic medications as a third class to minimize confounding by disease severity. Therefore, the comparator group combines several antidiabetic classes that may have different effects on the risk of dementia. Indeed, evidence from recent observational studies suggests that thiazolidinediones and sodium-glucose cotransporter-2 inhibitors may have a protective role on cognitive impairment.2, 4 Randomized controlled trials designed to study the direct and indirect effect of insulin on the risk of dementia compared with each class of antidiabetic medications are necessary. Seventh, although our study provides a real-world analysis to signal future work, it is limited to data from one Canadian province and a population aged 40-70 years old. Replication studies in diverse populations are necessary to generalize these findings. Last, diabetes is a complex condition whereby the presence of time-varying covariates affected by previous exposure is expected. This includes changes to glycated haemoglobin and cardio- and cerebrovascular events. However, methods to conduct such complex analyses are currently computationally obscure, hindering their use by most clinical researchers.22 Despite the added value of detailing pathways by which medications and outcomes are connected, the use of mediation analysis in pharmacoepidemiological studies, particularly those relating to diabetes, is rare. This is highly consequential when these pathways are through a preventable side effect, such as hypoglycaemia. Notwithstanding the abovementioned challenges in implementing mediation analysis to detail the potential role of hypoglycaemia, findings from this analysis provide further evidence for clinicians, patients with type 2 diabetes and caregivers of the importance of preventing hypoglycaemia. A small indirect association between insulin use and dementia through the mediating effect of hypoglycaemia was observed, providing further evidence on the need to prevent hypoglycaemia in patients with type 2 diabetes. Importantly, several challenges in the implementation of mediation analysis to detail the complex association between insulin, hypoglycaemia and dementia are highlighted, signalling many opportunities for future research. WA conceived the study idea, designed the study, conducted all analyses and wrote the first draft of the manuscript. All authors contributed to design, methodology and the final draft of the manuscript. All authors approved the submitted version of the manuscript. This project was funded by the Mike & Valeria Rosenbloom Foundation Research Award at the Alzheimer's Society of Canada. The authors have no conflicts of interest to declare. The peer review history for this article is available at https://www.webofscience.com/api/gateway/wos/peer-review/10.1111/dom.15385. The data that support the findings of this study are available from Population Data BC. Restrictions apply to the availability of these data, which were used under license for this study. All data were deidentified and no personal information was available at any point of the study. Access to data provided by the Data Steward(s) is subject to approval, but can be requested for research projects through the Data Steward(s) of their designated service providers. All inferences, opinions, and conclusions drawn in this publication are those of the author(s), and do not reflect the opinions or policies of the Data Steward(s). Data S1. Supporting Information. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.