This large diabetes cohort study (i.e. n = 204,323 in the 2000–2011 cohort; Fig. 1) with long-term follow-up (median follow-up 5.59 and 4.97 years, respectively, for those aged ≥18 years and ≥65 years) provides empirical evidence of the comparative effects of pioglitazone vs other second-line glucose-lowering drugs on the risk of dementia among individuals with type 2 diabetes. The effects of TZDs (combined pioglitazone and rosiglitazone) on dementia risk have been investigated previously. Preclinical data suggest that long-term use of peroxisome proliferator-activated receptor gamma (PPAR-γ) agonists such as TZDs prevents the neuropathological and behavioural changes associated with different types of dementia (e.g. Alzheimer’s disease and vascular dementia) . Specifically, PPAR-γ is expressed in the brain at low levels under physiological conditions and exhibits a range of activities that positively influence the pathology of Alzheimer’s disease, including ameliorating the inflammatory status of the Alzheimer’s disease brain by repressing the secretion of proinflammatory molecules, enhancing mitochondrial function and involving the processing of the amyloid β peptide . Also, a previous animal study showed that PPAR-γ agonists significantly reversed diabetes-induced vascular dementia through their multiple actions, including antioxidative and anti-acetylcholinesterase activity and anti-inflammatory and neuroprotective actions . In the present study, we found that patients taking TZDs had a lower risk of dementia compared with those taking other second-line glucose-lowering drugs (ESM Table 5). By contrast, Cheng et al found that sulfonylurea or metformin use conferred a lower dementia risk compared with TZDs, and that users of TZDs had a significantly higher risk of dementia compared with those who had never used TZDs . However, the number of patients using TZDs in that study were relatively few (n = 28) , which may threaten the internal validity of the findings.
Pioglitazone, a TZD medication, could be a viable therapeutic agent for dementia, but it is still under investigation, especially in humans. Animal studies have shown that pioglitazone attenuates streptozotocin-induced memory deficits in mice  and improves cognitive impairment in 1-methyl-4-phenyl-1,2,3,6-tetrahydropyridine (MPTP)-treated monkeys . The first human research on the cognitive effects of pioglitazone showed that daily use of pioglitazone for 6 months among individuals with comorbid dementia and type 2 diabetes increased their logical memory  and improved cognitive performance  compared with individuals who did not receive pioglitazone. However, these studies included a small number of subjects (i.e. 15 pioglitazone users and 17 non-users ; 21 pioglitazone users and 21 non-users ) and assessed type 2 diabetes individuals with existing cognitive impairment. A recent large German cohort study of 145,928 individuals aged ≥60 years who were initially free of dementia and diabetes found that pioglitazone use was associated with a reduced risk of dementia . Longer pioglitazone use (i.e. eight or more calendar quarters) led to a lower dementia risk . Compared with diabetic individuals not on pioglitazone, pioglitazone users had a 23% lower risk of dementia (RR 1.23, non-use vs use of pioglitazone; p < 0.001) . Notably, the present study extends existing knowledge on the cognitive effects of pioglitazone use vs non-use [2, 31, 32] by comparing the cognitive effects of pioglitazone use with those of various second-line glucose-lowering drugs.
The German study also found that rosiglitazone use vs non-use led to a reduced risk of dementia in diabetic individuals ≥60 years, but this cognitive protective effect did not reach the level of statistical significance and was lower than that of pioglitazone use (RR 0.84, rosiglitazone use vs non-use, p = 0.328; RR 0.53, pioglitazone use vs non-use, p = 0.029) . Similarly, the present study found that patients receiving pioglitazone had a lower incidence of dementia compared with propensity score-matched rosiglitazone-treated patients, implying a better protective effect of pioglitazone over rosiglitazone.
The present study has the following strengths. First, it was a large nationwide population-based study highly representative of individuals with type 2 diabetes in Taiwan and it had a long follow-up period (2000–2013). Second, its design included active treatment comparisons (rather than comparing drug use with non-use). The study identified patients who were newly starting second-line glucose-lowering treatment, and excluded those who had a history of dementia before starting second-line glucose-lowering treatment as well as those who developed dementia in the first year of second-line glucose-lowering treatment. These efforts are expected to mitigate potential confounding by indication, as observed in previous studies [2, 10], and prevent the bias due to including immediate events that might not have been associated with glucose-lowering treatment. Third, propensity score matching was applied to achieve balance in subject baseline characteristics, thus reducing potential selection bias. Fourth, different identification periods for study cases (i.e. 2009–2011) were modelled to assess the cognitive effects of newer glucose-lowering drugs (i.e. DPP-4 inhibitors), which verifies the validity of the study findings. Finally, considering non-adherence or low adherence to glucose-lowering drugs commonly observed in practice (i.e. discontinuing or switching to or adding another glucose-lowering drug), we applied the as-treated scenario, where subjects were censored when the treatment regimen changed. The results from the as-treated scenario were consistent with our primary analyses based on the ITT scenario, which supports the validity of the study.
Several limitations should be addressed. First, some laboratory data (e.g. HbA1c) were not available in the NHIRD claims data. However, we used surrogate indicators to adjust for the diabetes severity of study subjects, including the adapted Diabetes Complication Severity Index, diabetes duration, and cumulative exposure to metformin before the index date (i.e. before the initiation of second-line glucose-lowering treatment). Second, we used 180 days as a washout period to define new users/initiators of glucose-lowering treatments, based on clinical practice in Taiwan where diabetic individuals who are stable on glucose-lowering treatments usually return for follow-up visits every month, and the prescription refill period is unlikely to be longer than 3 months. Doctors must evaluate patients’ clinical profiles (e.g. HbA1c and fasting blood sugar) every 30–90 days and then adjust their glucose-lowering treatment regimens if needed. This means that, even if a patient receives the same glucose-lowering drug after a 90 day interval, s/he would be considered a new user of that particular drug. Although a washout period of 180 days is clinically and practically plausible, we were unable to check the entire history of medication use for each individual subject because the claims data available to this study can only be traced back to 1999. Similarly, we used 365 days as a stable treatment period for glucose-lowering medication (e.g. metformin), which is approximately four times sequential 3 month glucose-lowering prescription refills (i.e. 360 days). However, we were also unable to check whether the subjects did indeed persistently take metformin based on the claims data we used.
There have been no validation studies on the ICD-9 diagnosis codes for the definition of dementia in Taiwan. Taiwan’s NHI, however, issues a catastrophic illness certificate (CIC) for dementia, which is classified as ‘senile and pre-senile organic psychotic conditions’. The CIC in Taiwan is typically considered to be valid in disease classification, because strict conditions have to be met before a CIC is issued (e.g. several clinical profile/data documents are required) by Taiwan’s NHI. Patients with a CIC do not have to pay for their medical care. However, a CIC application for dementia is strict: only neurologists or psychiatrists are allowed to submit a CIC application for dementia; furthermore, these patients usually also have other severe psychotic disorders. As a result, not all individuals with dementia in Taiwan have a CIC for dementia. We have tried to use both disease diagnosis codes (i.e. ICD-9 or A codes) and CIC approaches to classify dementia events. However, the number of cases of dementia confirmed by both disease diagnosis codes and CIC approaches (i.e. a patient having an ICD-9 code and also a CIC for dementia) was much smaller than that identified by disease diagnosis codes only (e.g. the number of dementia cases among metformin + pioglitazone users aged ≥18 years was 1 vs 39, respectively, according to disease diagnosis codes and a CIC or according to disease diagnosis codes only). This implies that only a few dementia cases can be identified by a CIC, and these individuals may not be representative of all those with dementia in Taiwan. This is why previous Taiwanese studies on dementia [4, 16, 33,34,35] only relied on disease diagnosis codes to capture dementia events when the NHIRD was used.
Although we had a long follow-up period in which to observe dementia events, we did not evaluate the long-term use of pioglitazone on dementia prevention. Treatment dose or duration or length of response to pioglitazone deserves future research. Moreover, the results of protective HRs for dementia when metformin + pioglitazone compared with other dual regimens from the as-treated analyses should be interpreted with caution due to the small number of dementia cases in the analyses. As such, the point estimates of HRs may be unduly influenced by the small number of event cases in the analysis, which can be further evidenced by wide CIs corresponding to these point estimates. Finally, the effect of glucose-lowering treatment may vary according to the type of dementia (e.g. Alzheimer’s disease or vascular dementia). However, this study only treated dementia as a single entity.