A Hybrid Epidemiological Model Approach to Improvement of Predictive Accuracy in Zambian Infectious Diseases Modelling

Abstract

Recurrent infectious disease outbreaks, including cholera and influenza, as well as recent global pandemics like COVID-19, pose persistent public health challenges in Zambia. Traditional compartmental models based on Ordinary Differential Equations (ODEs), particularly Susceptible-Exposed-Infectious-Recovered (SEIR) frameworks, have long been used to predict disease spread. While these models are relatively simple and require fewer data, they often lack the flexibility to capture non-linear and stochastic factors—such as environmental variables and abrupt policy shifts—that can critically influence epidemic trajectories in resource-limited settings. In contrast, Artificial Neural Network (ANN) approaches excel at learning complex, non-linear relationships directly from data. By incorporating diverse inputs (e.g., climatic variables, demographic distributions), ANNs can adapt to evolving outbreak patterns more effectively than traditional ODE-based methods. However, their reliance on large, high-quality datasets and considerable computational resources can hinder adoption in places with fragmented surveillance systems. To address these complementary strengths and weaknesses, this study explores a hybrid modelling strategy that integrates a parameter-optimised SEIR model with a Transformer-based ANN. Historical COVID-19 data from 2020 to 2024 and environmental data (temperature, rainfall, humidity) were used to develop and validate three models: (1) an SEIR model whose parameters were estimated via curve fitting, (2) a standalone Transformer ANN, and (3) a combined SEIR-ANN ensemble. Model performance was assessed using metrics such as Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and R-squared (R²). Results indicate that the hybrid model consistently outperformed the individual SEIR and ANN models, exhibiting the lowest RMSE and MAE. Furthermore, integrating environmental factors into the ANN substantially improved predictive accuracy. These findings highlight the promise of hybrid frameworks in capturing the multifaceted dynamics of infectious diseases in Zambia. By leveraging SEIR’s mechanistic insights alongside the ANN’s capacity to learn from diverse datasets, public health practitioners can improve outbreak predictions and resource allocation. Nevertheless, barriers—such as limited data availability, computational infrastructure, and model interpretability—must be addressed to foster broader implementation. Strengthened data collection systems, increased investment in computational tools, and targeted capacity-building programs are recommended to fully realise the benefits of hybrid epidemiological modelling in Zambia.