Handling frequent observations of composite scores: Application to PROs in COPD
Eva Germovsek (1), Claire Ambery (2), Shuying Yang (2), Misba Beerahee (2), Mats O Karlsson (1), Elodie Plan (1)
(1) Department of Pharmaceutical Biosciences, Uppsala University, Uppsala, Sweden; (2) GlaxoSmithKline, London, UK
Objectives: Chronic obstructive pulmonary disease (COPD) is an inflammatory disease of the lung, characterised by not-fully reversible airflow obstruction that progresses with time [1]. To record how well a disease is managed from the patient’s point of view, daily questionnaires are often used; however, their analysis is challenging since the scale is composed of several individual elements, and the results contain a strong memory component. The aim of this study was to develop a model able to describe and learn from such patient-reported outcome (PRO) data.
Methods: Data were collected over a 1 year period using daily electronic diaries with 14 items (9 with 5 categories, and 5 with 4 categories) from patients enrolled in a prospective, observational study; the Acute Exacerbation and Respiratory InfectionS in COPD (AERIS) study, conducted at Southampton General Hospital, UK [2]. An item response theory (IRT) model [3] was used to relate the response data from each individual item to the underlying disease state. On the item level, Markov models (MM) were required to account for the dependence of an observation on the preceding observation. Minimal MM were used (i.e. the mean equilibrium time was assumed not to differ between compartments) and runtime reduction was explored by finding an analytical solution (AS) for the ordinary differential equations (ODEs). Since an AS for 3 compartments was obtained, some scores were initially merged; however, different combinations of scores were tested to allow for acquiring of good initial estimates for the full 5-compartment model (parameterised with ODEs).
Results: We analysed data from 127 COPD patients (median age 67 years, 54% male, 39% current smokers), providing approximately 40,000 observations per item. Parameterisation with the AS gave equivalent results to ODEs, and was about 6-times faster, therefore it was used for further model development. In this preliminary assessment mean (standard error) equilibrium time was estimated as 2.85 (0.22) days, and the slope on disease progression 0.035 (0.26) per year. Visual predictive checks on the individual item level showed satisfying fit to the data.
Conclusions: A preliminary IRT model was linked to Markov models for the first time to our knowledge and applied to real data from an observational study.
References:
[1] Wedzicha et al. Respir Care. 2003; 48(12):1204-13.
[2] Bourne et al. BMJ open. 2014; 4(3):1-8.
[3] Ueckert et al. Pharm Res. 2014; 31(8):2152-65.
Acknowledgments: GSK funded (114378/NCT01360398).
Disclosures: EG, MOK and EP declare no conflicts of interest; CA, SY and MB are GSK employees and hold GSK shares.
