Use of thiazolidinedione and risk of Parkinson’s disease in patients with type 2 diabetes
The data source
Participants were drawn from the Yinzhou Regional Health Care Database (YRHCD), which integrated longitudinal information on electronic medical records, disease registry and management, death registry, and other health services in Yinzhou District, Ningbo City China.27.30. In 2008, a disease registry and management systems were established for diabetes mellitus, cancer, cardiovascular disease, hypertension and chronic obstructive pulmonary disease. Diabetes patients were logged into the disease surveillance system once diagnosed and would be monitored at least four times a year by community physicians, with routine health measures including blood pressure, fasting blood glucose (FPG), glycated hemoglobin (HbA1c) being measured or asked27. In this study, the disease registry system and electronic diagnostic records were linked and patients with T2D were included if they: (1) were registered in the diabetes registry system and had been diagnosed with T2D; or (2) had more than two T2DM diagnostic records and no type 1 diabetes records in electronic medical records. Longitudinal records of medication prescriptions, laboratory tests, and outpatient and inpatient visits were linked for information on drug exposure, covariates, and outcome capture. The data used in this study and their relationship have been presented in Fig. 1 additional.
Exhibition and cohort
We applied an ACNU design, using a comparator of alpha-glucosidase inhibitors (AGIs), which are another class of second-line oral hypoglycemic agents commonly used in the same stage of T2D as TZDs in China.31. A cohort of patients with T2D who were new users of TZD or IAG after January 1, 2009 was formed. Drug prescriptions were identified by the Anatomical, Therapeutic, and Chemical (ATC) system classification code. TZDs (ATC A10BG) used in the study population contained pioglitazone and rosiglitazone, and AGIs (ATC A10BF) included acarbose, miglitol, and voglibose (Supplementary Table 2). New users were identified using a baseline washout period of 6 months prior to first refill of TZD or AGI, during which participants could not be prescribed a drug of either. another class. The date of the first filling was defined as the index date.
We excluded participants who were less than 18 years old and who had started combined treatment of TZD and IAG at the index date, and who had been diagnosed with PD before the index date. We further excluded patients who had not taken at least two consecutive prescriptions of TZD or AGI within 6 months of the index date to ensure that participants had actually started taking these drugs.
The study was approved by the Peking University Health Sciences Center Ethics Review Board (approval number: IRB00001052-18013-Exempt). Informed consent was not required due to the use of anonymized routine data.
Result and follow-up
The primary outcome was incident diagnosis of PD, defined as (1) having at least two consecutive International Classification of Diseases, 10th Revision (ICD-10) diagnosis codes or G20 description, or (2) having at minus a Parkinson’s disease diagnostic record and antiparkinsonian agent prescription records (ATC N04, Supplementary Table 3). The date of the first diagnosis was defined as the date of the result.
Our primary analysis was to assess the effects of any intention-to-treat (ITT) exposure of TZD use on the risk of PD. Thus, participants were followed from the index date until the first occurrence of the following events: diagnosis of PD, death, last medical record in the database or end of the study period (December 31, 2021).
Covariates were measured during the initial withdrawal period and included demographic characteristics (age, gender, and education level); behavior and lifestyle (smoking, alcohol and regular exercise); duration of T2D; comorbidities measured in ICC (Supplementary Table 4)32; concurrent use of prescription medications, including other antidiabetic medications except TZDs and AGIs (insulins, metformin, sulfonylureas, glinides, and other oral hypoglycemic agents), common medications for cardiovascular disease (diuretics, beta-blockers, calcium channel blockers, angiotensin-converting enzyme inhibitors (ACEI), angiotensin receptor blockers (ARBs) and aspirin), lipid-modifying agents and pump inhibitors protons (IPP). In addition, blood glucose level (FPG and HbA1c), blood lipid level, blood pressure, body mass index (BMI) and health care utilization (hospital and outpatient visits) were also included.
Descriptive statistics (mean and standard deviation for continuous covariates and frequency and percentage for categorical variables) summarized the baseline covariates and the standardized mean difference (SMD) was used for comparisons between TZD and AGI initiators. , as suggested by Austin et al.33. SMD less than 0.1 was used to show comparable balance in covariates33. Inverse probability of treatment weights (IPTW) was applied to control for initial confusion and a Cox regression model was used to estimate HR with a 95% confidence interval (CI) for the association between use of TZD and PD. We fitted a logistic regression model to estimate stabilized IPTW for each subject. The denominator of the IPTW was the probability of receiving the actual drug treatment condition on all of the measured covariates listed above. Continuous variables such as FPG, HbA1c were modeled as a restricted cubic spline with nodes at the 5th, 25th, 50th, 75th and 95th percentiles. The numerator of the IPTW was the marginal probability of using TZD or AGI in the overall sample. We have also included the year of the index date in the model for the IPTW to account for changes in prescribing habits over time. The final stabilized IPTW was truncated at the 1st/99th percentile to mitigate the impacts of extreme weights. Two other standard Cox models with different confounding adjustment strategies were provided for comparison: unadjusted and multivariate regression adjusted. Multiple imputation was applied for imputation of missing data using conditional full specification method with five imputations according to the quadratic rule recommended by von Hippel34. The proportional hazards assumption was tested using the Schoenfeld residual method and no violation of this assumption was found. Robust variance was used to calculate 95% CIs of HR when applying weighted models.
Subgroup analyzes and cumulative duration of use of TZDs
We examined the association of TZD use and PD incidence within different subgroups to test for potential interactions between TZD use and baseline characteristics: age (≤60 and > 60 years old), sex (female and male), ICC (0 and ≥1), smoking and alcoholic behavior, FPG (≤7 mmol/L and >7 mmol/L), HbA1c (≤7% and >7%), BMI (≤24 kg/m2 and >24 kg/m2), and the duration of diabetes at the index date (≤2.5 years and >2.5 years).
The cumulative duration of use of TZD was calculated as the period between the index date and the date of the last prescription of TZD. Cumulative duration tertiles were used to define categorical cumulative TZD therapy. We then examined the association between different durations (≤0.5, 0.51–4 and >4 years) of treatment with TZD and the incidence of PD compared to the use of AGIs.
We emulated a per-protocol (PP) analysis to examine the effects of prolonged exposure, in which participants were further censored upon discontinuation of the initial medication in addition to the reasons for censoring in the ITT analyses. Discontinuation of treatment was defined as the absence of renewal of the initial drug within 6 months of the previous prescription. This artificial censoring could induce selection bias and could be influenced by time-varying confounders35. Thus, we applied a marginal structural model with time-varying inverse probability censoring weights (IPCW) to adjust for the selection bias introduced by artificial censoring. The IPCW was the inverse of the probability of remaining uncensored at each follow-up conditioned by invariant and time-varying factors, and was estimated using a pooled logistic model. The time-varying factors were all covariates listed above except for demographic, behavioral and lifestyle characteristics, duration of T2DM, and year of index date. We used a 6-month interval to assess time-updated exposure and covariates at the start of each new period in the PP analysis.
We additionally performed several sensitivity analyzes to examine the robustness of the results in our main analysis. First, alternative washout periods of 12, 18, and 24 months were applied to define new users of TZD and AGI. Second, we excluded all possible cases of secondary PD, who had received a diagnosis of ICD-10 code G21 or G22 after receiving the first diagnosis of G20. Third, stabilized unstabilized and untruncated IPTWs were applied. Fourth, the Fine-Gray under-distribution risk model was used to verify the possible competing risk per death from any cause. Fifth, we limited the study population to patients with T2DM aged over 40 at the index date. This could help rule out cases of early-onset PD since young people rarely develop this disease. Sixth, we used several alternative definitions of incident PD such as: (1) having at least two G20 diagnoses; (2) having more than two diagnoses of G20 and prescriptions of anti-Parkinson’s agents (ATC N04) after the first diagnosis; and (3) have a consecutive diagnosis of G20 or a prescription for antiparkinsonian agents within one year of the first diagnosis of PD. Seventh, we excluded patients with T2DM who had received antidiabetic drugs during the baseline washout period and assessed the effects of monotonic treatment of TZDs versus monotonic treatment of AGIs. Finally, we excluded a period after the index date to adjust for potential latency. For example, we identified participants with more than 6 months of baseline drug use and follow-up began 6 months after the index date. This analysis was repeated with successively longer minimum follow-up time requirements by adding 1 month to each analysis, up to a maximum of 24 months. This type of analysis could also help adjust for unmeasured confounding by undiagnosed disease.35.
Statistical analysis was performed using SAS 9.4 (SAS Institute Inc., Cary, NC, USA). All statistical tests were conducted bilaterally and one p– value