Discussions
We found that TZD monotherapy was associated with reductions in risk of all-cause dementia compared with use of MET or SU among participants with T2D. The use of TZD with MET or SU showed a lower risk for all-cause dementia than MET monotherapy. In addition, TZD-related treatments were associated with much lower risks of VaD. This is consistent with the prior reports that TZDs can reduce carotid atherosclerosis and incident strokes.26 27 Vascular diseases increase the risk of AD,28 so TZD’s reduction in VaD may also reduce AD development. Some studies comparing TZD with either placebo or standard care within patients with T2D have reported reduced risk of AD.6 7 Comparing with MET monotherapy helps provide a relevant active comparator as it is the most used ADM, is a middle-of-the-road glucose-lowering drug, and has not been linked with increased incident dementia. In contrast, comparisons with untreated patients with diabetes would be complicated by their small numbers and healthier phenotype.
Subgroup analyses show that participants aged ≤75 years benefited more from TZD use than older participants, perhaps highlighting the difficulty of successfully intervening in more advanced disease stages and the importance of early prevention for dementia. TZD use also appeared to be more protective in overweight or obese participants. This may result as TZD reduces central obesity,29 a recognized risk factor for dementia.13
Our results add substantially to the literature concerning the effects of ADM on dementia where previous findings have been inconsistent.5–7 11 Studies with follow-up time less than 3 years have mainly reported null associations, while studies with longer follow-up time typically yielded protective findings.5–7 11 12 With a mean follow-up time of 6.8 years, we had a sufficient duration to detect treatment differences. Another strength of this study was that we required ADM supplies for more than one-third of the time per exposure year and observed similar patterns of dementia risks as studies that controlled for treatment doses or frequency.7 11 We also found that treatment effects increased with treatment durations and drug supply days, which implies a dose–response relationship. Finally, we controlled the treatment misclassification rate by setting a 6-month drug adaptative period.15
The complex pathways linking T2D with incident dementia require studies to control many confounders.30 31 Traditional multivariate regression is ineffective and even invalid in this situation.32 We used stabilized IPTW with a comprehensive list of covariates to address the issue. Orkaby et al compared MET with SU uses with similar statistical strategies as our study and reported findings consistent with ours.11 To assess surveillance bias in dementia detection during the long follow-up period, we examined patterns of HbA1c during follow-up, clinical visit frequencies, and hypoglycemia event frequencies. These factors did not explain the observed differences in treatment effects.
Dementia misclassification is possible using EHR-based observational studies because dementia is commonly underdiagnosed.1 Since IPTW balanced participants’ characteristics, we assumed equal rates of dementia underestimation across each group. To strengthen the robustness of our results, we broadened the dementia diagnosis to include the use of antidementia medications. With dementia cases increased by up to 8%, this did not affect the results. However, phenotyping algorithms using EHRs to distinguish dementia subtypes remain challenging. In our study, less than 20% of all-cause dementia could be coded as AD and VaD, lower than the real-world rate.3 This may reflect the higher frequency of initial dementia diagnoses by primary care providers and their tendency to provide more non-specific codes for dementia than neurologists and geriatricians.33 Newcomer et al reported that 80% of AD cases were underdiagnosed when using a single year of medical claims,34 but the percent declined to 13% after extending the claims-extraction period to 5 years.35 For VaD detection, a 75% sensitivity and a 74% specificity of ICD-10 codes were reported.36 In our study, the misclassification rate of AD could be higher than VaD, but the long follow-up time may partially offset the difference. Additionally, we excluded participants diagnosed with cognitive conditions or brain injury before follow-up, making results less biased by pre-existing conditions. We observed a high loss to follow-up rate among MET monotherapy users, which may lead to underestimating dementia rates in this population. Combined with the relatively low loss to follow-up rate in TZD users, our results provided a conservative estimation of the TZD treatment effects for dementia.
The study is also subject to residual confounding due to missing data (eg, kidney function) or unavailable information (eg, risk genes) in the database. The Food and Drug Administration (FDA) restricted TZD use in 2010 and again eased its use later, which changed clinicians’ prescription patterns.37 Although we did not find TZD’s treatment effects were mediated by year of treatment initiation (estimated by year of baseline), future studies may consider stratifying the analysis by the timing of FDA’s announcement. Although our use of MET as the comparator does not allow us to identify the specific relationship of MET use with dementia, its common use, mid-range effects on glucose control, apparently relatively neutral impact on dementia provides important advantages for comparison of early-stage diabetes medications. Given the predominately white and male VA population, future studies among more diverse populations are needed to confirm the findings.
In summary, TZD users had a lower risk of dementia, and SU users had a higher risk of dementia than MET users among T2D participants. The protective effects of TZD were more substantial for overweight or obese patients. Our findings provide additional information to aid clinicians’ selection of ADMs for patients with mild or moderate T2D and are at high risk of dementia.