Practical considerations for empirical Bayes estimation using a pharmacokinetic/pharmacodynamic model developed from summary-level data of ustekinumab therapy in Crohn’s disease

OBJECTIVES: This study aimed to develop a pharmacokinetic/pharmacodynamic model using summary-level data (SLD-PK/PD model) for ustekinumab in Crohn's disease and to propose practical considerations and an operating framework for empirical Bayes estimation (EBE) when individual patient data (IPD) models are unavailable. METHODS: An SLD-PK/PD model was constructed using arm-level Crohn's Disease Activity Index (CDAI) data from phase IIb/III trials. To evaluate EBE performance, simulated IPD (simIPD) was generated using a reported IPD-PK/PD model. Pseudo-SLD was derived from simIPD to refit the SLD-PK/PD model. EBE was performed at Week 8 to predict CDAI up to Week 20. Because individual-level variability is not identifiable from SLD, uncertainty was propagated using a prespecified σ2 grid and ω2 scaling, with η-shrinkage-guided bounds. KEY FINDINGS: Based on the variance-range framework, patients were classified as consistently predicted remission (6.9%), consistently predicted non-remission (66.2%), or indeterminate (26.9%). Predictive accuracy varied across σ2 assumptions, demonstrating the practical impact of variance specification. CONCLUSIONS: Although EBE using an SLD-PK/PD model is practically useful when IPD-PK/PD models are unavailable, predictions depend on variance assumptions. Therefore, σ2 and ω2 should be treated as uncertainty ranges, and decision stability should be evaluated across these ranges to ensure robust clinical applications.