Longitudinal Parkinson’s Disease Progression Model using Item-Response-Theory Utilized to Predict Treatment Effect of Levodopa
Souvik Bhattacharya (1), Timothy Nicholas (2), Lawrence J. Lesko (1), Mirjam N. Trame (1)
(1) Center for Pharmacometrics and Systems Pharmacology, University of Florida, Orlando, FL; (2) Global Clinical Pharmacology, Pfizer Inc, Groton, CT
Objectives: To evaluate and understand the natural history of early and long-term disease progression in Parkinson’s Disease (PD) and to predict the effect of Levodopa treatment of the longitudinal change of item-level data from the Unified Parkinson Disease Rating Scale (UPDRS) using Item-Response-Theory (IRT).
Methods: Item-level UPDRS data from 1990 subjects, from five different National Institute of Neurological Disorders and Stroke (NINDS) trials, early and advanced stages of PD, was used to develop a Bayesian longitudinal IRT model with uninformative prior. The model was developed in R 3.2.3 to predict the patient specific latent traits of 44 individual subscores of the UPDRS at each study visit and estimate the change and severity of each subscore over time for placebo and treatment (levodopa) effect. An initial Bootstrap clustering was implemented on the UPDRS subscores in order to obtain a hierarchical structure identifying the most sensitive subscore and to determine the pattern of linkage between the UPDRS subscores. Prediction of subscores were implemented using a logistic regression algorithm (“nnet” package) in R. A time-varying function (including inter-individual variability) was developed to study the longitudinal trajectory for the placebo and treatment effects (“brms” package in R) based on the latent score of the most sensitive subscore. External model evaluation using data from two NINDS trials was performed over continuous (Visual Predictive Check) and at discrete (Boxplot) times.
Results: Bootstrap clustering identified “Rapid/Alter Movements (RAM)” and “Hand Movements (HM)” to be the most influential and predictive subscores within the UPDRS, hence the estimated combined population mean latent score of RAM & HM (LS=0.5792+random|ID) were used to predict subscores higher in hierarchy over time. Including an additional dose dependent function into the combined IRT logistic regression placebo model describing the treatment effect allowed for adequate predictions of the change and severity for each subscore within the UPDRS over time. External model evaluation showed good agreement with the observed subscores for both continuous and discrete times.
Conclusions: A Bayesian longitudinal IRT regression model was successfully developed to predict the overall disease progression and the levodopa treatment effect in PD based on early disease progression information. The developed longitudinal IRT model is embedded in a Shiny application.
