Identifiability, Effective Dimensionality, and Interpretability In PBPK-Based Alpha Rpt Dosimetry
Abstract
Purpose
Quantitative image-based dosimetry for radiopharmaceutical therapy (RPT) increasingly relies on large, multi-organ mechanistic models of physiologically based pharmacokinetics (PBPK). While PBPK models often fit clinical data well, the extent to which their internal biological distinctions are identifiable remains poorly understood. This work examines the identifiability, degeneracy, and effective dimensionality of an AlphaRPT dosimetry model across datasets, with the goal of clarifying what mechanistic inferences are supported by available image-based measurements.
Methods
We analyzed a comprehensive AlphaRPT dosimetry model of Ac-225-DOTATATE, incorporating competing transport, clearance, and redistribution mechanisms of the Bi-213 daughter. Using multiparameter inference followed by repeated applications of the Manifold Boundary Approximation Method (MBAM), we explored model behavior originating from approximately 3500 distinct parameter initializations across 18 independent datasets. This approach systematically collapses parameter combinations that are either unidentifiable from the data or insignificant for decision-making, thereby identifying the dominant effective mechanisms for explaining the response in each parameter regime.
Results
Despite excellent fits across datasets, the unreduced models exhibited extensive parameter degeneracy, with large regions of parameter space yielding indistinguishable predictions for time-integrated activity and absorbed dose. MBAM revealed that this degeneracy is highly structured: reductions consistently collapsed into low-dimensional effective parameter combinations, hierarchically stratified across datasets and reflecting nested levels of identifiability. While some core mechanisms are shared across all reductions, other biologically distinct mechanisms—particularly alternative Bi-213 redistribution pathways—were found to depend on the local context in parameter space.
Conclusion
These results demonstrate that the primary limitation for image-based AlphaRPT dosimetry with PBPK models is not predictive accuracy but mechanistic identifiability. Model reduction is a prerequisite for meaningful uncertainty quantification, hypothesis testing, and context-specific interpretation. By exposing which biological distinctions survive contact with data, this framework provides a principled foundation for auditable dosimetry modeling and guides both experimental design and responsible personalization in AlphaRPT.