Development of a longitudinal model for characterizing adverse events of psychiatric medications in Korean patients

Abstract

In routine clinical care of psychiatric patients, early treatment is important because adverse events (AEs) in this period often lead to noncompliance to a drug and lowering the therapeutic effect. Quantitative understanding of such AEs, in terms of incidence, severity and dropout over the course of a treatment, is therefore needed to provide better treatment guidelines for patients. This study aimed to develop a longitudinal model to describe early-phase AEs in Korean psychiatric patients in an effort to be used as a guide to improve medication compliance and drug efficacy. The study consists of developing the models using clinical trial data and applying the developed models to characterization of routine clinical data. The clinical trial data was obtained from 1,630 patients with generalized anxiety disorder (GAD) in six phase III clinical trials that were used to investigate the dose-AEs-dropout relationship of pregabalin with drowsiness as an AE, in terms of incidence and severity. Treatment was up to five to seven weeks and ranged from the dose of 150 to 600 mg/day given with a one-week dose titration and a one-week taper period. Since most of the patients did not experience drowsiness during the study period, the AE was modeled in two separate steps: modeling the incidence first and then modeling the conditional severity. Using the incidence and the conditional severity models, the unconditional severity probability was then computed. A proportional odds model was developed for the severity, and a constant hazard model was implemented to evaluate the dropout effect, which was found to be insignificant when constant hazard was assumed at each severity level. Routine clinical data were collected retrospectively from the medical records of the outpatient clinic in Severance hospital, in Seoul, Korea. The records involved 150 patients with GAD or major depressive disorder (MDD) who were treated with anxiolytics or antidepressants. Data were censored on day 60 from their first visit. Three different longitudinal models were developed within a mixed-effect model framework to describe the incidence, the time to event of AEs (TTE), and the count of AEs where the TTE model was used to characterize dropout, also. The hazard function to describe censored data or dropout was chosen to be a constant for the incidence model and the Weibull function was employed for TTE and count models. To evaluate the model performances, a visual predictive check (VPC) was performed using simulated datasets from each model using estimated parameters.For the clinical trial data, the Emax model adequately described the incidence of drowsiness, and the probability of incidence increased with the dose in pregabalin treatments. Dose titration was an important factor for decreasing the incidence. Using a mono-exponential function for the placebo effect in the logit, the dose proportional model for the drug effect and AE attenuation for the time-dependent effect best described the severity of drowsiness. The visual inspection of severity probability versus time computed from the above choice of model revealed that after reaching the peak probability in about 8 days the incidence and the severity of drowsiness declined over 5 weeks, as expected from the estimated half-life of attenuation effect of 7.7 days. The dropout was not found influential.For the routine clinical data, the most frequently observed AE was drowsiness. About 70% of the patients reported AE more than once during the observation period. For the incidence model, a Markov element added in the baseline logit adequately described the data. Incorporating a mono-exponential function as a time effect further improved the model. VPC showed the good performance of the model. For the TTE models, Weibull hazard model dropped the objective function values (OFV) most significantly and the estimated shape parameters of hazard function were less than 1.0 (0.00147 for the AEs and 0.001 for the dropout), indicating that hazard rates in both models were decreasing with time. Median time to AEs was about day 20 from patients’ first visit. The VPC result showed a good agreement with the prediction. For the count model, the constant hazard model was applied and predicted mean counts of AEs were 1.20 for the BZD treated patients, 1.28 for the SSRI treated patients and 1.21 for the other CNS-drug treated patients. Despite lack of information, the count model well matched with the original data.