Physics-informed neural networks (PINNs) offer a promising approach for epidemic modeling by integrating mechanistic disease dynamics with flexible function approximation. Yet, hyperparameter choices—especially physics regularization weights—remain largely heuristic. This study develops a systematic data-adaptive approach for hyperparameter selection within a SUC-PINNs (Susceptible–Unconfirmed–Confirmed PINNs) applied to COVID-19 data from six countries. Synthetic experiments confirmed methodological validity, with parameter recovery errors below fifteen percent and all true values captured by bootstrap confidence intervals, but hyperparameters optimal for synthetic data did not generalize to real-world conditions, underscoring the need for data-adaptive tuning. A comprehensive grid search revealed substantial cross-country heterogeneity, with optimal physics regularization varying by a factor of four hundred, reflecting differing epidemic complexities. Higher regularization improved parameter stability but introduced accuracy loss when excessive, although estimated reproduction numbers remained consistent across settings. Beyond COVID-19, the findings highlight the broader importance of data-adaptive hyperparameter selection for applying PINNs to diverse infectious disease systems. The results provide practical guidance for regularization selection, favoring lower regularization for moderate-quality data, stronger regularization for complex dynamics, and bootstrap procedures for uncertainty quantification. • Demonstrate that data-adaptive hyperparameter selection improves epidemic model reliability across diverse healthcare systems. • Quantify the stability accuracy tradeoff in healthcare epidemic modeling using systematic analytics. • Establish evidence-based regularization ranges aligned with epidemic complexity and data quality. • Validate robust epidemic parameter estimation despite significant variability in model configuration choices. • Provide an analytics-driven template for uncertainty-aware epidemic modeling in healthcare practice.
Susyanto et al. (Wed,) studied this question.