Answers are grounded in this paper and cite their sources.
Estimating epidemiological parameters from noisy and limited surveillance data is a critical challenge in mathematical epidemiology and outbreak response. This study presents a rigorous comparative benchmark evaluating Physics-Informed Neural Networks (PINNs), which embed an SIR compartmental model into the neural learning objective via automatic differentiation, against classical Non-linear Least Squares (NLS) calibration with numerical Runge-Kutta integration. Both approaches are tested on synthetic epidemic data across observation counts of \(\:{N}ₜ=10\) and \(\:{N}ₜ=30\) and Gaussian noise levels from 1% to 20% over multiple stochastic realizations. Results show that for fully observed, correctly specified compartmental models, classical Least Squares recovers parameters more accurately and reliably in preliminary runs, achieving median transmission rate errors below 5% (0.089%–4.160%) and 100% compliance in capturing the basic reproduction number (\(\:{R}₀\)) within a 10% clinical tolerance margin. In contrast, preliminary PINNs yield median transmission errors of 7.76% to 34.21% alongside a systematic negative parameter bias (up to \(\:-30.43%\)) caused by optimization stiffness and gradient scale imbalances between data and physical residual losses. However, extended training experiments (10,000–20,000 iterations) demonstrate that PINN errors reduce asymptotically below 2% to 5%, confirming that early limitations stem from optimization stalling rather than representational incapacity. Classical calibration remains the optimal baseline for fully observed low-dimensional systems, while PINNs offer significant scientific potential for partially observed states and non-parametric functional dynamics.
No takes yet. Share an insight, caveat, or question.
Adeyemi et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: