Dynamics-Informed Neural Network Modeling of COVID-19 Transmission in Afghanistan Using the SEIR-V Framework

Background: The COVID-19 pandemic posed significant challenges for public health systems globally, particularly in resource-limited settings such as Afghanistan. Limitations in diagnostic capacity, inconsistent data reporting, and low vaccination coverage hindered timely public health responses. To support real-time decision-making, accurate and adaptive modeling frameworks are essential. Methods: This study presents a hybrid modeling approach that integrates the classical SEIR-V (Susceptible–Exposed–Infectious–Recovered–Vaccinated) compartmental model with Dynamics-Informed Neural Networks (DINNs). The model embeds the SEIR-V system of differential equations into the loss function of a deep neural network to enable dynamic estimation of time-varying parameters. Epidemiological data from Feb 2020 to Apr 2024 were collected from multiple publicly available sources, including Worldometer, Our World in Data, the World Health Organization and the Johns Hopkins University COVID-19 repository. Results: The proposed DINNs-SEIRV model effectively reconstructed multiple epidemic waves and generated accurate forecasts of COVID-19 transmission dynamics in Afghanistan. The model achieved high predictive performance, particularly for the infectious (I) compartment, with a coefficient of determination R² = 0.9973. It also demonstrated strong capacity in capturing vaccination trends and maintaining robustness in the presence of incomplete or noisy data. Conclusion: The DINNs-SEIRV framework offers a powerful and flexible tool for modeling infectious disease dynamics in low-resource settings. Its ability to learn and update time-varying parameters in response to real-world data makes it valuable for informing public health strategy, forecasting outbreaks, and evaluating vaccination efforts in environments like Afghanistan