Hybrid multi-disease SEIR and transformer models for infectious disease forecasting in Zambia

Abstract

Abstract Background Traditional infectious disease models like SIR and SEIR have long served epidemiologists in understanding outbreak dynamics. However, these models often fall short in capturing the realities of multiple concurrent infections, particularly in countries like Zambia where diseases such as COVID-19, cholera, influenza, measles, anthrax, and mumps often co-exist. Problem Most epidemiological models used in Zambia focus on single diseases, failing to capture disease interactions, co-infections, and immune suppression effects, which leads to inaccurate forecasting and inefficient allocation of public health resources. Purpose This study aimed to develop a hybrid disease prediction framework that integrates an extended SEIR model with machine learning (ML), particularly transformer-based artificial neural networks (ANNs), to improve multi-disease outbreak forecasting in Zambia. Literature backing Prior studies have proposed modifications to SEIR models to accommodate multiple infections, while others have demonstrated the effectiveness of integrating ML techniques such as ANNs to improve forecasting accuracy in non-linear, real-time epidemic contexts. However, few have applied these advances in resource-constrained settings like Zambia, where multiple epidemics often overlap. Methodology The study focused on six diseases prevalent in Zambia and combined historical infection data (2020–2023) with environmental variables such as temperature, rainfall, and humidity. An extended SEIR model was developed incorporating co-infection dynamics, interaction terms, and variable transmission rates. This model was integrated into a transformer-based ANN, trained using data from 2020 to 2022 and tested on 2023 data. The models were evaluated using RMSE, MAE, R², and MAPE, comparing single-disease and multi-disease versions across three approaches: baseline SEIR, ANN-only, and hybrid models. Results Results revealed that integrating multiple diseases improved forecasting accuracy. In dual-disease models, the multi-disease SEIR model reduced RMSE from 593.138 to 557.065. The hybrid model outperformed both SEIR-only and ANN-only models, reducing RMSE from 409.267 (single-disease hybrid) to 387.845 (multi-disease hybrid). When extended to all six diseases, hybrid models consistently outperformed single-disease models for COVID-19, mumps, measles, and cholera. For instance, the COVID-19 hybrid model showed a notable RMSE improvement from 0.541 (single-disease) to 0.210 (multi-disease). Conclusion This research demonstrates that hybrid models integrating extended SEIR and ANN techniques provide superior disease prediction accuracy, particularly when modelling co-infection and disease interactions. The SEIR component offers strong epidemiological grounding, while ML captures real-time non-linear dynamics. Consistent data significantly boosts model performance. Recommendations The study recommends adopting multi-disease hybrid forecasting models in Zambia and similar regions to improve disease preparedness. Collaborations between computer scientists and epidemiologists can bridge gaps in interpretability and technical expertise. Future studies should further explore the role of data quality in enhancing SEIR model accuracy and consider developing user-friendly ML tools for public health professionals. Clinical trial number Not applicable.