Reproduce “Physics-Informed Neural Networks for Chemotherapy Pharmacokinetics: Benchmarking the Clinical Estimator and Exposing Parameter Identifiability” on OpenResearch

About this paper

Riya Bisht, Dhruv Agarwal

This research benchmarks Physics-Informed Neural Networks (PINNs) against the clinical nonlinear least-squares (NLS) estimator for chemotherapy pharmacokinetics, demonstrating PINNs' comparable performance on linear models and their ability to transparently expose parameter non-identifiability in nonlinear systems. The study shows PINNs can resolve these identifiability issues by readily incorporating sparse, heterogeneous data, such as tissue observations.

Key insights

• PINNs provide performance statistically comparable to the established NLS method on well-specified linear two-compartment pharmacokinetic models.
• Unlike the NLS method, PINNs transparently expose structural or practical non-identifiability issues in nonlinear models (e.g., Michaelis-Menten kinetics) by converging to physically interpretable degenerate solutions (e.g., $k_{12} \approx 0$) when data are insufficient.
• The flexible PINN framework can readily integrate sparse, heterogeneous data (e.g., two tissue samples) to significantly resolve parameter identifiability and improve the accuracy of both observed and unobserved concentration predictions.

Summary by alphaXiv.