3175 - Quantitative Evaluation of Parameter Identifiability in a Physics-Informed Radio-Immune Model under Limited Clinical Observability

Purpose/Objective(s):

Mechanistic radio-immune models characterize tumor–immune interactions, yet the feasibility of identifying specific biological parameters from sparse clinical data remains unquantified. We utilized physics-informed neural networks (PINNs) to evaluate the recovery of tumor growth rate (µ), rate of cell killing (w), tumor infiltration rate (?), lymphocyte decay constant (?_L), and doomed-cell decay constant (?_D) under varying observability conditions.

Materials/Methods:

A radio-immune ODE model was used to generate synthetic longitudinal data for viable tumor cells (T), lymphocytes (L), and doomed cells (D), assuming no use of radiation and immunotherapy. PINNs (DeepXDE) were trained under three scenarios: (1) full-state supervision (T, L, D) without noise; (2) full-state supervision with 10% Gaussian noise; and (3) tumor volume only supervision (T+D) with 10% noise, fixing ?_L and ?_D to mitigate identifiability loss. Parameter error was computed relative to ground truth.

Results:

Under noiseless full-state supervision, four of five parameters were recovered within 10% error (2.2–5.0%), while ?_D showed 22% error, indicating partial parameter coupling even under ideal conditions. With 10% noise, four parameters remained within 10% error (2.2–9.1%), with ?_D improving to 14% error. In contrast, when only T+D was observable, the remaining free parameters (µ, w, ?) exhibited ~49–50% error despite accurate trajectory fitting, demonstrating structural non-identifiability.

Conclusion:

PINNs enable robust recovery of most radio-immune parameters under complete observability but reveal intrinsic identifiability limitations, particularly for ?_D. When restricted to tumor-only measurements, parameter estimates exhibit substantial (~50%) bias despite accurate trajectory fitting, consistent with structural non-identifiability. Future work will evaluate adaptive loss balancing, residual-based collocation refinement, and multi-start optimization, along with extension of the model to incorporate immune suppression dynamics and radiation survival effects, to improve parameter identifiability in clinically representative settings.