Diabetes mellitus, prediabetes and the risk of Parkinson’s disease: a systematic review and meta-analysis of 15 cohort studies with 29.9 million participants and 86,345 cases

This systematic review and meta-analysis of fifteen cohort studies with 86,345 PD cases and 29.9 million participants, provides strong epidemiological evidence of a 27% increase in the relative risk of developing PD among DM patients, compared to persons without DM. Persons with prediabetes had a 4% increase in risk, though this latter finding was based on two studies only. The increased risk of PD in persons with DM was observed in men and women, across geographic regions, and strata of other study characteristics (number of cases, type of diabetes, assessment of diabetes, timing of diabetes diagnosis, study quality, and adjustment for most confounding factors).

These findings are in line with previous meta-analyses of five and seven cohort studies, of much smaller sample sizes, which reported a 37–38% increase in risk of PD among diabetes patients [22, 23], and two more recent meta-analyses including 7 and 10 cohort studies which both reported a 29% increase in risk of PD among diabetes patients [24, 25]. Therefore, our findings provide more robust epidemiological evidence on this association, as our analysis was based on roughly twice the number of studies and a much larger sample size and number of PD cases. The findings are also consistent with a recent Mendelian randomization study which found an increased risk of Parkinson’s disease with genetically determined diabetes mellitus [24], providing some support for causality. Our results are not concordant with the findings of case-control studies [23], which are known to be more prone to survival, recall and selection biases, as well as reverse causality. Increased mid-life and early late-life mortality of DM patients may, indeed, underpin the discrepancies between case-control and cohort studies. A “reverse causation” bias may occur in observational studies, related to the long prodromal period of PD, potentially spanning over two decades. This period is known to be associated with progressive accumulation of PD neuropathology, including inclusions of Lewy bodies (LB) and dopaminergic nigrostriatal neuronal loss, known to potentially affect glycaemic control [10], as well as the occurrence of subtle motor and non-motor clinical manifestations that may alter physical activity and diet, both key lifestyle factors in the management of DM. The observation of considerable weight loss in PD patients, seen several years before diagnosis and persisting for several years after diagnosis [36], may also impact diabetes risk favorably, and could potentially explain the observed protective association of DM on PD in case-control studies. Interestingly, in the current analysis based on cohort studies, the positive association between DM and PD persisted in studies with ≥ 15 or ≥ 20 years of follow-up, and was if anything slightly stronger among studies with long durations of follow-up than short follow-up, suggesting that reverse causation is less likely to explain the findings.

The main limitations of studies employing meta-analysis methods involve factors that may potentially affect results, such as potentially unaccounted and residual confounders, sub-optimal study quality, heterogeneity, misclassification or misdiagnosis of diabetes status, as well as publication bias. For instance, pesticide exposure and head trauma are rarely accounted for, in observational cohort studies; additionally diabetes patients have a higher prevalence of overweight/obesity and smoking, and lower levels of physical activity [37], compared to persons without diabetes; therefore, confounding from some of these factors could have impacted the observed results. In our stratified analyses, the associations persisted among studies which adjusted for these and several other factors, making this a less likely explanation for the observed results. In addition, there was no evidence of heterogeneity between the subgroups stratified by various adjustments with meta-regression analyses, however, given the few studies in some of the subgroups it is also possible that we may have been underpowered to detect significant differences between some subgroups. The calculated E-values suggest that such a confounder would have to be relatively strongly associated (RR = 1.86, lower CI: 1.69) with both DM and PD to fully explain away the observed association. However, the E-value for the association between prediabetes and PD was weaker (RR = 1.24, lower CI: 1.16), suggesting the association may have been more vulnerable to confounding.

In terms of study quality, although the mean study quality across studies was graded as moderate, the summary estimates were similar to the subgroup of studies with high study quality. One common contributor to lower than optimal study quality scores was a lack of reporting on adequacy or completeness of follow-up. However, a low score on this point may not necessarily have been a source of bias, but rather due to poor reporting, because many studies have nearly complete follow-up data, facilitated by linkages to health or mortality registries. Other contributors to non-optimal quality scores included features such as non-optimal assessment of exposure and outcome (e.g. using only self-report or registry linkages without independent assessment or validation) and inadequate adjustment for confounders. Nevertheless, the summary estimates persisted in most subgroup analyses stratified by adjustment for various confounding factors.

Although there was high heterogeneity in the overall analysis as measured by the I²-value, 12 of 14 risk estimates were in the direction of increased risk and 11 of the risk estimates were statistically significant (95% CIs excluding the null value). Moreover, none of the cohort studies reported a statistically significant reduction in relative risk. Hence, the observed heterogeneity is more likely driven by differences in the strength of the association, than by differences in the direction of the association. Also, several of the included studies were very large and had rather narrow 95% CIs around the risk estimates, but with different sizes of risk estimates, and thus, the 95% CIs for the different studies did not always overlap. This likely explains the high I²-value in spite of the relative consistency of the direction of the association, thus less problematic than the presence of heterogeneity related to the direction of an association. In our subgroup analyses, there was less heterogeneity in studies of men or men and women combined, and among US studies.

Diabetes was self-reported in some studies and this may have led to misclassification of diabetes status that could, potentially, result in an underestimation of the association between DM and PD. However, we found little difference in the summary estimates between subgroups of studies that only relied on self-report and those that used a combination of self-report and linkages to medical records or had a measure of fasting blood glucose at baseline. Subgroup analyses stratified by whether studies only used prevalent (baseline) diabetes cases or also included incident diabetes cases, did not show significant heterogeneity between subgroups, although the summary estimate was slightly higher for the former than the latter (1.30 vs. 1.17). Misclassification of diabetes diagnoses by using only prevalent cases could result in regression dilution bias (bias to the null) as the lack of information on incident diabetes cases could lead to misclassification and potential underestimation of the association. However, this would depend on a recent diabetes diagnosis being most relevant to Parkinson’s disease development. Given that we observed the strongest association in the subgroup using prevalent diabetes cases only, it is possible that long-standing diabetes may be of greater importance. Survival bias is less likely to explain our findings because if such bias was present one would expect a weaker association in the studies using only prevalent diabetes cases, than in those also using incident cases. Although we found no significant difference in the results when stratified by the reported diabetes duration, this analysis was based on only two studies. The summary estimates were somewhat stronger for participants with diabetes complications than among those without such complications (summary RR, 95% CI = 1.54, 1.32–1.80 vs. 1.26, 1.16–1.38). As diabetes complications typically occur in patients with longer disease duration or poor glycaemic control, it is plausible that a longer disease duration and poor diabetes management are key driving factors for the stronger association with PD risk. Given the limited number of studies in these subgroup analyses and the lack of consistent cut-offs reported for diabetes duration, further epidemiological and mechanistic studies are warranted to elucidate the precise role of disease duration and glycaemic control on PD risk. In addition, the available studies did not investigate the role of specific diabetes medications on the observed risk. The majority of the cohort studies included in our meta-analysis only reported on DM overall, three studies reported on type 2 diabetes (T2D) only and none on type 1 diabetes. Given that the vast majority (over 90%) of diabetes cases are of type 2 (1), the current findings likely reflect the impact of insulin resistance and T2D. Lastly, we explored potential publication bias, but found no evidence of such bias with the statistical tests or by inspection of the funnel plot.

Several biological mechanisms may contribute to the increased risk of PD in patients with diabetes. Hyperglycemia, resulting from hypoinsulinaemia in type 1 diabetes or insulin resistance (IR) in T2D, exposes neurons to increased metabolic stress, neuronal dysfunction and death, thus directly contributing to PD pathogenesis [38]. Experiments in diabetic mice showed reduced dopamine transporters [39] and dopamine levels in the striatum [40], thus increasing the vulnerability of nigro-striatal neurons. Recent experiments, using diabetes-induced MitoPark mice, showed that the acquisition of IR phenotype in these animals results in mitochondrial dysfunction by suppressing PGC-1α expression, promoting the upregulated ROS production and oxidative stress, as well as the upregulated expression of phosphorylated α-synuclein (SNCA) [41], a key constituent of Lewy bodies (LB) [42]. Mitochondrial dysfunction, leading to neuronal death was also a main finding of studies using knockout of insulin receptor (NIRKO) mice [43] and diabetic db/db mice [44]. Increased accumulation and phosphorylation of α-synuclein was also observed within the cortex, pre-commissural putamen and dopaminergic neurons in the substantia nigra of cynomolgus monkeys, with spontaneous T2D-like pathology [45]. Of note, abnormal SNCA and LB burden are key pathological features in PD.

A direct effect of hyperglycemia is the increase of advanced glycation end-products (AGEs) and glycation agents, such as the highly reactive methylglyoxal (MGO), that may reach particularly high levels in the substantia nigra [46]. The interactions of AGEs with their receptors (RAGE) may lead to oxidative stress, inflammation and cell death [47]. AGEs have been found in LBs and have been shown, in vitro studies, to cross-link with SNCA to induce its aggregation and formation of SNCA oligomers, of higher neurotoxicity [48]. Furthermore, MGO may inhibit SNCA degradation and increase its accumulation and also react with dopamine to form 1-acetyl-6,7 dihysroxy-1.2.3.4-tetrahysroisoquinoline (ADTIQ), that may further contribute to dopaminergic degeneration [48].

In view of the potential increase of PD risk in diabetes patients, repurposing of antidiabetic medications for the treatment of PD is gaining increasing research interest, in the absence of effective disease modifying therapies. Recently, Mor et al. have demonstrated the neurotoxic effect of bcat-1 knockdown, in an animal model known to recapitulate PD features [49]. Neurotoxicity was found to be mediated through increased mitochondrial respiration (or “hyperactivity”) and oxidative damage. The authors showed that administration of metformin, a first- line anti-T2D medication, reduced mitochondrial respiration to control levels and significantly improved both motor function and neuronal viability [49]. Metformin was also shown to improve motor functions in 6-hydroxydopamine (6-OHDA)-lesioned mice, by activating the AMPK and BDNF signaling pathways and regulating or suppressing genes in reactive astrocytes [50]. In human trials, a double blind placebo-controlled study of Exenatide, a GLP-1 receptor agonist, showed a beneficial effect in PD patients [51]. In a subsequent study, evaluating target engagement through neuronal derived exosomal vesicles (NEVs) isolated from serum samples, it has been shown that patients with exenatide had an increased protein activation of Akt and mTOR cascades, compared to placebo [50].

Strengths of this meta-analysis include the population-based cohort design of the included studies, mitigating the potential for survival bias, recall bias, selection bias, and reverse causation, which may affect case-control studies to a larger degree; secondly it’s large sample size of 29.9 million participants, including over 86,000 PD cases, providing statistical power to detect a moderate association; finally the robustness of the findings in multiple subgroup and sensitivity analyses.

As the ever increasing numbers of persons with diabetes are projected to reach 700 million by 2025 worldwide [1], increased rates of debilitating late-onset neurodegenerative diseases such as AD, PD, as well as the associated forms of dementia with Lewy bodies, and Parkinson’s disease dementia are set to become another consequence of the diabetes epidemic, adding significant healthcare and socio-economic burden worldwide. Although the observed association between DM and PD is of moderate size, the findings are still likely to have important public health implications because of the large number of persons who live with diabetes worldwide. In this context, our findings strongly support the need for urgent global public health measures to effectively address the diabetes epidemic worldwide, that may have the added significant benefit in preventing PD, AD and related late-onset neurodegenerative diseases.