Abstract
Background:
Diabetes is a risk factor for Alzheimer’s disease and related dementias (ADRD). Epidemiologic evidence shows an association between diabetes medications and ADRD risk; cell and mouse models show diabetes medication association with AD-related neuropathologic change (ADNC).
Objective:
This hypothesis-generating analysis aimed to describe autopsy-measured ADNC for individuals who used diabetes medications.
Methods:
Descriptive analysis of ADNC for Adult Changes in Thought (ACT) Study autopsy cohort who used diabetes medications, including sulfonylureas, insulin, and biguanides; total N = 118. ADNC included amyloid plaque distribution (Thal phasing), neurofibrillary tangle (NFT) distribution (Braak stage), and cortical neuritic plaque density (CERAD score). We also examined quantitative measures of ADNC using the means of standardized Histelide measures of cortical PHF-tau and Aβ1–42. Adjusted analyses control for age at death, sex, education, APOE genotype, and diabetes complication severity index.
Results:
Adjusted analyses showed no significant association between any drug class and traditional neuropathologic measures compared to nonusers of that class. In adjusted Histelide analyses, any insulin use was associated with lower mean levels of Aβ1–42 (–0.57 (CI: –1.12, –0.02)) compared to nonusers. Five years of sulfonylureas and of biguanides use was associated with lower levels of Aβ1–42 compared to nonusers (–0.15 (CI: –0.28, –0.02), –0.31 (CI: –0.54, –0.07), respectively).
Conclusion:
Some evidence exists that diabetes medications are associated with lower levels of Aβ1–42, but not traditional measures of neuropathology. Future studies are needed in larger samples to build understanding of the mechanisms between diabetes, its medications, and ADRD, and to potentially repurpose existing medications for prevention or delay of ADRD.
INTRODUCTION
The burden of Alzheimer’s disease and related dementias (ADRD) is large and growing, with substantial effects on patients, their caregivers, and the health system as a whole. In the United States, there are approximately 7 million individuals living with ADRD, and this number is projected to increase to 12 million by 2040 [1]. One promising strategy to alleviate this burden is to examine existing therapeutics, used for treatment of other chronic conditions, for potential effects on cognitive decline and ADRD risk [2]. Type 2 diabetes mellitus (hereafter “diabetes”) is a modifiable risk factor for ADRD [3, 4], with brain atrophy, reduced cerebral glucose metabolism, and CNS insulin resistance common to both diabetes and AD [5, 6]. Additionally, some comparisons across classes of diabetes medications have suggested lower risk of ADRD with metformin as compared to sulfonylureas [7, 8], suggesting a potential protective effect of specific diabetes management strategies.
The relationship between diabetes medications and ADRD has been examined in mouse models, in vitro studies, and human-based studies. There is consistent evidence showing a connection between diabetes and cerebral infarcts, but not AD neuropathologic change (ADNC), which supports a role for cerebrovascular disease in relating diabetes to ADRD [9 –15]. Since insulin modulates levels of amyloid-β (Aβ) in the brain and contributes to synaptogenesis, it is nevertheless possible that diabetes medications, which modify insulin levels, could affect AD pathology [16 –19]. Importantly, metformin has been hypothesized to affect AD pathology via specific mechanisms separate from insulin modification. Mouse models have shown that metformin decreases oxidative stress, corrects abnormal transport of Aβ through the blood-brain barrier, reduces neurodegeneration, and improves memory [20 –24]. In vitro studies also showed improved AD pathology with metformin treatment, as evidenced by reduction of elevated tau Ser396 and FAK tyrosine phosphorylation, restoration of Aβ production and acetylcholinesterase activity, and inhibition of tau kinase GSK3β and tau kinase ERK [25]. Other mouse models have shown that metformin reduces tau hyperphosphorylation [26, 27]. Preclinical studies of insulin showed protection against synaptotoxic effects of Aβ, and reduced tau hyperphosphorylation [28 –30].
Human-based studies on diabetes medications and ADRD have largely focused on cognitive decline and clinical dementia, rather than neuropathology. For metformin, most of these studies have shown protective benefits, including two randomized clinical trials [31, 32], and epidemiologic studies that showed association with lower ADRD risk compared to sulfonylureas [7, 8]. Sulfonylureas lack a direct effect that is hypothesized to alter AD pathology [8], but effects via insulin stimulation from the pancreas could still occur. The epidemiologic studies of sulfonylureas show mixed results, depending on the choice of comparator medication, but the associations generally suggest less protection than metformin for clinical ADRD [7 , 34]. Finally, hypothesis driven randomized trials of short-term intranasal insulin in individuals without T2DM have suggested protective effects on cognition, but the effect varied across sex, APOE genotype, and duration of insulin exposure [17 , 36].
Despite these important advances, few studies have examined relationships between diabetes medications and neuropathological outcomes in humans. These efforts have been constrained by difficulties of adjusting for diabetes severity and duration. Existing work featured limited examination of oral anti-diabetic agents, and did not disaggregate them into classes, which would allow translation to therapeutic implications [37]. This is an important distinction, because the aforementioned evidence suggests possible effects of specific classes of oral agents on AD-related neuropathology in animal models. Most work on subclasses of diabetes medications have examined clinical outcomes, which is a notable shortcoming in the literature, because there are often inconsistencies between clinical presentation of dementia and underlying neuropathology [38]. Accordingly, there have been increasing calls for pharmacoepidemiological studies using neuropathological data, especially in relation to the repurposing of existing drugs for potential use against AD, to more fully understand drug-brain associations [39].
In this study, we examined a community-based autopsy cohort to characterize the ADNC of individuals who used diabetes medications, while adjusting for diabetes severity and other confounders. In these hypothesis-generating analyses, we aim to provide an important reference point to support further inquiry in two main areas: 1) mechanistic research on the possible biologic links between diabetes medications, diabetes, and ADNC, and 2) research on repurposing existing chronic disease medications for the prevention and treatment of ADRD.
MATERIALS AND METHOD
Data and study population
We examined the Adult Changes in Thought (ACT) Study autopsy cohort, restricted to individuals who ever used a diabetes medication before death (N = 124). We confirmed that none of them had type 1 diabetes. Thus, the entire sample had type 2 diabetes, and we only make comparisons between users of different types of diabetes medications. The ACT Study is a prospective cohort study, with an original cohort of 2,581 randomly selected dementia-free members of Kaiser Permanente Washington (KPWA, formerly Group Health) [40]. Participants were required to be age 65 or older and not have dementia at time of enrollment, which occurred from 1994 to 1996 for the original cohort. An additional 811 participants were enrolled between 2000 and 2003. In 2004, ACT began using continuous enrollment to add 10–15 participants per month, in order to keep at least 2,000 people enrolled. As of September 30, 2018, the ACT Study compiled an autopsy sample of 815 individuals, in which AD-related neuropathology was measured. Lack of dementia at enrollment is determined, first, based on random sampling which eliminated persons with a dementia diagnosis in their electronic medical record and then by eliminating those who at screening score less than the cut-point (≤85) of the Cognitive Abilities Screening Instrument (CASI) and are determined to have pre-existing dementia based on standardized diagnostic evaluation of dementia, neuropsychological testing, and subsequent multidisciplinary consensus conference [41]. The lifetime exposure to prescription medications is observable for these participants from electronic pharmacy data and chart review. Within the autopsy cohort with diabetes medication utilization, we removed 6 individuals with unmeasured APOE genotype to arrive at an analytic sample of 118 individuals. We included any individual who had used the following classes of medications: biguanides, thiazolidinediones, sulfonylureas, DPP-IV inhibitors, GLP-1 analogs (i.e., incretin mimetics), alpha-glucosidase inhibitors, SGLT-2 inhibitors, meglitinides, amylin analogs, and insulin. While metformin is sometimes used for polycystic ovary syndrome (PCOS), the utilization observed in our sample occurred outside probable childbearing ages, strongly suggesting that it was not for PCOS.
Measures and outcomes
The exposure of interest was the use of diabetes medications, defined in binary use (any use ever) and duration (measured in five-year intervals of exposure years). These two exposure definitions were included to examine associations at both the extensive and intensive margins. One exposure year was defined as filling at least two prescriptions in a calendar year (1977-present), or any documented use in the medical record (pre-1977). We focused on the three most commonly observed classes of diabetes medications: sulfonylureas (N = 90), insulin (N = 75), and biguanides (N = 47). The full list of sulfonylureas and biguanides used by individuals in our sample is included in the Supplementary Material. Diabetes severity was measured using the diabetes complication severity index, which is based on past diagnoses of retinopathy, nephropathy (measured with both diagnoses and lab tests), neuropathy, cerebrovascular complications, cardiovascular complications, peripheral vascular disease, and metabolic complications [42].
Neuropathologic outcomes were measured at autopsy by a board-certified neuropathologist according to published guidelines, blinded to the participants’ medication exposure and clinical dementia status [43]. The vast majority of autopsies were performed with less than 48 h postmortem interval (PMI), with 40%less than 8 h. The outcomes of interest included three traditional measures of AD-related neuropathology. All three measures were dichotomized into binary variables from their original categorical form: amyloid plaque distribution from Thal phasing (binary variable for amyloid plaque score A2-A3 (equivalent to Thal phase 3–5)), Braak stage for neurofibrillary tangle (NFT) distribution (binary variable for NFT score B3 (equivalent to Braak stage V/VI)), and CERAD (Consortium to Establish a Registry for AD) score for cortical neuritic plaque density (binary variable for neuritic plaque score C2 or C3 (equivalent to CERAD moderate to frequent)). We also used a dichotomized version of the ABC score, a composite measure of “high AD-neuropathology,” defined as having Thal amyloid plaque score A3, Braak NFT score B3, and CERAD neuritic plaque score C2 or C3 [44]. Amyloid plaque score was only measured in 48 individuals; when missing, we defined the ABC score according to the aforementioned NFT and neuritic plaque scores.
We examined two Histelide (immuno
Statistical analyses
Descriptive statistics describe the use of diabetes medications by class, and the presence of ADNC across the most common classes of diabetes medications. For the traditional ADNC outcomes, we used Poisson regressions with robust standard errors, to adjust for the following covariates: age at death, sex, education (years), APOE genotype defined as presence of ≥1 ɛ4 allele, and diabetes severity. Regressions of Histelide outcomes that adjusted for the same covariates were estimated using OLS with robust standard errors. All adjusted analyses compare users of a drug class to non-users of the same drug class (e.g., insulin users to insulin non-users); since all members of the sample used diabetes medication, the reference groups include users of the other drugs (e.g., all insulin non-users in the sample used either biguanides or sulfonylureas). Sensitivity analyses examined if our models were robust to the inclusions of additional covariates: race, use of cholesterol medications, hypertension, and body mass index (BMI). Analyses were conducted using Stata version 16.1.
RESULTS
Sample description and prevalence of AD-related neuropathology
The characteristics of the ACT Study autopsy sample who used diabetes medications and non-missing APOE genotype (N = 118) are described in Table 1. These individuals used sulfonylureas (N = 90, for mean 8.2 years), insulin (N = 75, for mean 6.3 years), and biguanides (N = 47, for mean 5.9 years). Also, 4 participants used thiazolidinediones, and 1 used meglitinides. Many individuals used multiple classes of medications: among sulfonylureas users, 64%used insulin, and 40%used biguanides; among insulin users, 77%used sulfonylureas, and 47%used biguanides; among biguanides users, 77%used sulfonylureas, and 74%used insulin (Supplementary Table 1). Prescriptions for oral medications were typically 90 days in length, and 30 days in length for insulin.
Characteristics of analytics sample (N (%), unless otherwise noted)
Analytic sample of ACT Study autopsy cohort who used diabetes medications and had non-missing APOE status. Amyloid plaque A2-A3 uses 48 as the denominator. Exposure years defined as 2 claims in a calendar year, except before 1977 when any observed use was considered an exposure year. ABC high is defined as Thal amyloid plaque score A3, Braak NFT B3, and CERAD neuritic plaque C2 or C3 (when Thal was missing, ABC high was defined with Braak NFT and CERAD neuritic plaques). Histelide, immunohistochemistry and ELISA performed on a glass slide; APOE, apolipoprotein E; SD, standard deviation; DSM, Diagnostic and Statistical Manual; AD, Alzheimer’s disease; CERAD, Consortium to Establish a Registry for AD; NINCDS, National Institute of Neurological and Communicative Disorders and Stroke; ACT, Adult Changes in Thought; NFT, neurofibrillary tangle; DCSI, diabetes complication severity index.
The sample was 53%female, with mean education attainment of 14.3 years, and mean age at death of 86.1 years. Within this sample, 60 (51%) had clinical dementia (defined by DSM-IV (Diagnostic and Statistical Manual IV)) at time of death. According to the NINCDS-ADRDA criteria (National Institute of Neurological and Communicative Disorders and Stroke –Alzheimer’s Disease and Related Disorders Association), 25 (21%) exhibited probable AD, and 17 (14%) exhibited possible AD. Autopsies for the sample of people who died with treated diabetes (N = 118) showed that 21%had high AD-related neuropathology, 60%had Thal amyloid plaque score A2-A3, 25%had Braak NFT score B3, and 50%had CERAD neuritic plaque score C2-C3. Their mean (standard deviation) Histelide measures were –0.05 (0.64) for PHF-τ, and 0.09 (1.04) for Aβ1–42. For reference, the entire ACT Study autopsy sample (not restricted to diabetes medication users) had mean (standard deviation) of 0.00 (0.92) for Aβ1–42, and 0.00 (0.90) for PHF-τ.
Association of diabetes medications and AD-related neuropathology
The results of adjusted analyses of associations between binary diabetes medication use and traditional measures of AD-related neuropathology are reported in Table 2. The comparison group for each risk ratio is individuals who did not use the drug of interest in that row, but who used one of the other two types of drugs. The same associations with medication use measured in five-year increments of exposure are reported in Table 3; here, each risk ratio is the association between 5 additional years of use of the row drug and the column outcome. None of the binary measures show a statistically significant association with Thal amyloid A2-A3, Braak NFT B3, CERAD neuritic plaque C2-C3, or ABC high. This was consistent when the use of medications was defined in years.
Risk ratios of the association between AD neuropathologic change and diabetes medication use (binary), with 95%CIs
Results from Poisson regressions with robust standard errors. Each risk ratio corresponds to a separate regression of the dependent variable in the column, on the independent variable in the row. All regressions adjusted for age at death, sex, education, APOE ɛ4 genotype, and diabetes complication severity index. ABC high is defined as Thal amyloid plaque score A3, Braak NFT B3, and CERAD neuritic plaque C2 or C3 (when Thal was missing, ABC high was defined with Braak NFT and CERAD neuritic plaques). NFT, neurofibrillary tangle; AD, Alzheimer’s disease; APOE, apolipoprotein E; CERAD, Consortium to Establish a Registry for AD.
Risk ratios of the association between AD neuropathologic change and 5-year intervals of diabetes medication use, with 95%CIs
Results from Poisson regressions with robust standard errors. Each risk ratio corresponds to a separate regression of the dependent variable in the column, on the independent variable in the row. All regressions adjusted for age at death, sex, education, APOE ɛ4 genotype, and diabetes complication severity index. ABC high is defined as Thal amyloid plaque score A3, Braak NFT B3, and CERAD neuritic plaque C2 or C3 (when Thal was missing, ABC high was defined with Braak NFT and CERAD neuritic plaques). NFT, neurofibrillary tangle; AD, Alzheimer’s disease; APOE, apolipoprotein E; CERAD, Consortium to Establish a Registry for AD.
We report the associations between diabetes medication use and Histelide measures of AD neuropathology in Table 4 (binary use of medication) and Table 5 (five-year increments of medication exposure). The use of insulin was associated with lower levels of Aβ1–42 (–0.57, CI: –1.12, –0.02), compared to those who did not use insulin. This reduction is equivalent to 55%of a standard deviation. These results did not remain significant in Table 5, when insulin use was measured in five-year units of exposure, although the direction of the point estimate for the association was the same. Five years of biguanides use was associated with lower levels of Aβ1–42 (–0.31, CI: –0.54, –0.07) (equivalent to 30%of a standard deviation), and this result is loosely supported by the association of binary biguanides use and Aβ1–42 (–0.38, CI: –0.77, 0.01). Five years of sulfonylureas use was associated with lower levels of Aβ1–42 (–0.15 CI: –0.28, –0.02) (equivalent to 14%of a standard deviation), but this finding was not supported in analyses of binary sulfonylureas use. Sensitivity analyses were used to test if our main results were robust to the inclusion of variables for race, use of cholesterol medications, hypertension, and BMI; these analyses showed no meaningful variation from the main results (Supplementary Tables 2 and 3).
Coefficients of the association between Histelide measures and diabetes medication use (binary), with 95%CIs
Results from OLS regressions with robust standard errors. Each coefficient corresponds to a separate regression of the dependent variable in the column, on the independent variable in the row. All regressions adjusted for age at death, sex, education, APOE ɛ4 genotype, and diabetes complication severity index. APOE, apolipoprotein E.
Coefficients of the association between Histelide measures and 5-year intervals of diabetes medication use, with 95%CIs
Results from OLS regressions with robust standard errors. Each coefficient corresponds to a separate regression of the dependent variable in the column, on the independent variable in the row. All regressions adjusted for age at death, sex, education, APOE ɛ4 genotype, and diabetes complication severity index. APOE, apolipoprotein E.
DISCUSSION
In this hypothesis-generating study, we examined the ADNC of participants with diabetes according to medication class. Using lifelong diabetes medication utilization data in the ACT Study autopsy sample, we examined both traditional neuropathologic measures and quantitative solution-phase immunoassay (Histelide) measures, and their association with use of these medications. Our results showed weak evidence of a negative relationship between use of all three drugs and Aβ1–42, a quantified measure of Aβ plaques. Specifically, five years of biguanides use was associated with a 0.31 unit lower Aβ1–42 level, and five years of sulfonylureas use was associated with a 0.15 unit lower Aβ1–42 level, suggesting less Aβ plaques. Insulin use was associated with lower Aβ1–42 levels (mean 0.57 units lower). However, the results were not robust to changes in the measurement of drug exposure. Additionally, there were no significant associations observed in the traditional measures of AD-neuropathology (Thal amyloid plaque phasing, Braak NFT distribution, and CERAD neuritic plaque density). It is important to note that everyone in our sample had diabetes, as evidenced by their use of diabetes medications. Therefore, we are not making any comparisons to individuals without diabetes; we only describe differences in ADNC across users of different drugs. By focusing only on participants using diabetes medications, we implicitly control for unobserved differences in health and healthcare utilization tendencies. Another important strength of our analyses is that they control for diabetes severity, which would otherwise be likely to confound the observed associations.
With a small sample and limited strength of findings, our results on the connection between specific diabetes drugs and neuropathology should be considered hypothesis-generating. However, the description of these neuropathologic characteristics in the ACT Study autopsy sample, where lifelong observation of medication utilization is possible, provides an important reference point in evidence that seeks to understand the relationships between diabetes, its treatments, and AD. Additionally, these rich data include all past diagnoses, which allowed our analyses to adjust for a detailed measure of diabetes severity [42]. These results can inform subsequent research on the mechanisms of the relationships between diabetes medications and ADRD, and therapeutic research seeking to repurpose existing drugs for potential ADRD treatment. There is unique potential of neuropathological data to improve dementia research and expedite efforts in the drug repurposing pipeline [39].
None of the associations in our Histelide results were supported by analyses of traditional neuropathologic measures. Prior studies have shown that the Histelide measures have explanatory power independent of the traditional measures [46], which lends importance to studies that report on both types of outcomes. The traditional measures, which are identified via consensus histopathologic scoring, have certain weaknesses compared to the Histelide measures, including that they are semi-quantitative, generate a single score for the entire brain, have significant ceiling effects, and lack molecular specificity [47]. Despite these weaknesses, the lack of a signal in the amyloid plaque phasing reduces confidence in our findings related to Aβ1–42, and the lack of signal in NFT stage reduces confidence in our findings related to PHF-τ.
It is difficult to compare our study to human studies of the relationship between diabetes medications and AD-related outcomes, which vary in their findings according to methodologies and comparator group. That said, recent high quality epidemiologic studies found that metformin is associated with lower ADRD risk than sulfonylureas [7]. Our study found that both of these classes were associated with lower levels of Aβ1–42, but the magnitude of the biguanides association was larger.
An important consideration is that the relationships identified in this study are mostly independent of the connection between diabetes and neuropathology. That is, each of our reported associations compares users and non-users of a medication class, but both groups have diabetes, and we adjust for diabetes severity. As discussed in the introduction, diabetes is significantly associated with cerebral infarcts, and it is possible that these and stroke-related factors mediate the role of diabetes in dementia risk. Yet, substantial evidence has found a lack of association between diabetes and ADNC [9 –15]. Our Histelide results, if verified in additional analyses, could therefore suggest protective benefits of diabetes medication that exist in channels aside from diabetes control.
Limitations
Although this is one of the largest autopsy samples to examine associations between diabetes medications and ADNC, we were underpowered to find differences in outcomes because of the small sample size. We offer these descriptions of ADNC in diabetes medication users as a reference for future research that examines neuropathology of diabetes medication users. One limitation is unobserved confounding. For example, some characteristics of individuals in the study, such as longer duration of diabetes beyond the duration of treatment, may have a relationship with severity of neurodegenerative changes. Due to our small sample size, we are unable to control for the use of other diabetes medications in each analysis, and therefore we are unable to determine the independent association of each medication class with the outcomes. Some medications, like insulin, are typically used for more severe diabetes, which would confound the observed association with ADNC. However, our adjusted analyses control for diabetes severity, therefore at least partially addressing this limitation. Like many pharmacoepidemiologic studies, our data captures prescription fills, which do not necessarily mean that the medication was consumed by the patient. However, in general, pharmacy data from cohorts set in comprehensive care settings like the ACT study are considered highly reliable [48]. Additionally, reverse causality could impact results of this study, if cognitive function prior to death (potentially related to underlying neuropathology) affected utilization of medication in the years prior to death. Our small sample size restricted our ability to include additional controls for these other confounders. Another limitation is that our sample is limited to members of the ACT Study, based in KPWA, which has a predominantly White and highly educated patient population, who reached relatively old age despite diabetes. These factors could influence study outcomes, and our results may not generalize to other diabetes medication users or younger populations. An additional limitation is that newer diabetes drugs, including some that have been associated with AD outcomes, were not used in the care of this cohort so we could not report on any associations with them. Finally, this study did not examine outcomes related to vascular pathology, which often is co-morbid with AD-related pathology and a major area of study.
CONCLUSIONS
The ACT Study has one of the largest autopsy samples in the world linked to decades of medication use that capture the entire lifelong treatment of diabetes in our cohort; reporting on the ADNC of these individuals provides an important reference for future research examining neuropathology related to diabetes and diabetes treatments. This future work should harmonize samples in the ACT Study and other brain banks that feature increasing numbers of autopsied individuals. These larger samples will allow testing of specific hypotheses on the connection between diabetes medications and AD-related neuropathology, while controlling for dementia severity, diabetes control, blood pressure control, and other possible confounders. Ideally, this work will serve as a building block to understanding the mechanisms between diabetes, its medications, and ADRD, and to potentially repurpose existing medications for the prevention or delay of ADRD.
Footnotes
ACKNOWLEDGMENTS
Dr. Barthold was supported by the University of Washington Alzheimer’s Disease Research Center (ADRC) Pilot Award (parent grant P30 AG066509); National Heart, Lung, and Blood Institute (1OT3HL152448-01); and National Institute of Mental Health (R01MH121424). Drs. Gibbons, Keene, and Grabowski were supported by the University of Washington Alzheimer’s Disease Research Center (P30 AG066509). Drs. Gray, Keene, Larson, and Crane were supported by the National Institute on Aging (UO1AG0006781). Dr. Keene was supported by Nancy and Buster Alvord endowment. Dr. Marcum was supported by the National Institute on Aging of the National Institutes of Health (K76AG059929). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. Dr. Gray was supported by the Centers for Disease Control and Prevention U01CE002967.
