Crisp: A Modeling Framework for Identifiability and Interpretation In PBPK-Based Rpt Dosimetry
Abstract
Purpose
Physiologically based pharmacokinetic (PBPK) models used in radiopharmaceutical therapy (RPT) and theranostic digital twins are rapidly increasing in size and mechanistic detail, often containing tens to hundreds of parameters. While such models fit imaging and dosimetric data well, their high dimensionality pose fundamental challenges for parameter identifiability, uncertainty quantification, and mechanistic interpretation. This work introduces Contextualized Reduction for Identifiability and Scientific Precision (CRISP), an analysis framework developed in quantitative systems pharmacology (QSP), as a transferable approach for addressing these challenges in PBPK-based theranostics.
Methods
CRISP was developed to analyze “sloppy” models, i.e., models whose predictions are controlled by a few effective parameter combinations while remaining insensitive to many others. The framework uses Fisher Information analysis, information geometry, and systematic model reduction using the Manifold Boundary Approximation Method (MBAM). CRISP has been successfully applied in QSP, where sloppiness similarly bottlenecks traditional analysis. Here, PBPK-based theranostics (TAC) is fitting is similarly examined. We consider a 177Lu-PSMA PBPK model with approximately 150 parameters, representative of emerging theranostic digital twins, to illustrate properties of sloppiness under realistic data constraints.
Results
Large PBPK models for RPTs exhibit wide separation of parameter sensitivities, leading to poor identifiability of individual parameters despite excellent agreement with data. Consequently, conventional uncertainty quantification, sensitivity analysis, and parameter estimation can become unstable or misleading, and prior assumptions may dominate posterior inferences. CRISP reframes model analysis around preserving information for decision-relevant predictions by targeting the few stiff directions controlling clinically meaningful outcomes.
Conclusion
CRISP provides a unifying, cross-domain framework for reasoning about identifiability, uncertainty, and mechanisms in large PBPK models used for digital twins and predictive dosimetry. By treating sloppiness as a structural feature rather than a modeling failure, this approach supports more credible, auditable, and interpretable use of complex mechanistic models as they continue to grow in scale and scope.