Hybrid Epidemiological Forecasting with AI, Environmental and Policy Insights: A Case Study from Zambia

Abstract

Zambia has, in recent years, confronted overlapping outbreaks of cholera, measles, anthrax, mpox and COVID-19, revealing weaknesses in forecasting tools that rely solely on classical modelling. Conventional Susceptible–Exposed–Infectious–Recovered (SEIR) formulations generally apply fixed parameters and treat each disease separately, while contemporary artificial intelligence (AI) approaches, such as long short-term memory networks (LSTM), gated recurrent units (GRU) and Transformers, are better suited to non-linear behaviour yet often lack clarity in how predictions are produced. In this study, we design and assess a combined SEIR–AI forecasting strategy that integrates environmental indicators and policy actions. Using weekly COVID-19 case reports from 2020 to 2023, together with climate variables (temperature, humidity and rainfall) and public health measures (including lockdowns and vaccination efforts), we compare the performance of SEIR, AI-only and hybrid models. The merged SEIR–AI model offers the most accurate forecasts, achieving a reduced root mean square error (RMSE) of 372.84 and a modest improvement in the coefficient of determination (R² = 0.02) when evaluated on unscaled case counts. SHAP (SHapley Additive Explanations) analysis further shows that rainfall (0.47) and the timing of interventions (0.43) were the most influential predictors. These findings indicate that incorporating environmental and policy information within a hybrid SEIR–AI framework enhances both predictive reliability and interpretability, offering a practical tool for epidemic management in Zambia and other resource-constrained settings.