Pharmacometric Evaluation of Item Response Modeling to Inform Clinical Drug Development

Drug development is the process of advancing novel therapeutics to market to improve patient outcomes. However, in hard-to-treat diseases like neurodegenerative disorders there is a high failure rate in late-stage trials, creating significant unmet needs. This highlights the need for more sensitive endpoints, improved trial designs, or analytical methods to optimize data utilization. In many diseases, clinical outcome assessments (COAs) serve as clinical endpoints and are often reported as a composite score, potentially losing important information present at the item level. Alternatively, item response theory (IRT) leverages item-level data to describe the relationship between a subject’s response on an item and their underlying ability, through item characteristic functions (ICFs), offering a more informed analysis of COAs. This thesis evaluates the robustness of IRT, estimation strategies and its applicability to model rating-scale-based COAs to facilitate model-informed drug development (MIDD). For single time point analysis, our findings suggest at least 100 subjects and 20 items are generally sufficient. Comparison of Laplace and Gaussian-hermite quadrature (GHQ-EM) for the estimation of item parameters, indicated similar accuracy and precision with slight improvement in accuracy for GHQ-EM. IRT models in reduced assessments were relatively stable up to ~40-60% information remaining. However, removing items shifts the measured disease construct, which can affect the accurate assessment of disease progression and drug effect. The trade-offs in information lost or gained should be considered when shortening assessments. Comparison of two common estimation strategies for determining ICFs indicated similar performance, each providing different advantages. IRT was also effective in classifying disease (Parkinson’s vs SWEDDs), showing comparable performance to artificial neural networks. Additionally, IRT demonstrated superior power for detecting symptomatic treatment effect in a short duration trial compared to traditional approaches, highlighting IRT’s potential not only for endpoint analyses but as a strategic tool to optimize trial design. Greater public disclosure of applied IRT in real-time drug development, such as inclusion in trial protocols or in regulatory milestones could foster broader acceptance and wider adoption beyond ad-hoc analyses. In conclusion, this thesis presents a methodological foundation for successful implementation of IRT in a pharmacometric framework to facilitate MIDD and inform clinical decision-making.