From Regression to Deep Learning: Advances, Challenges and Future Prospects in Epidemic Prediction

Abstract

: To mitigate and prevent the negative global impacts of epidemics like COVID-19, humanity urgently requires reliable pandemic models to support containment strategies. This paper reviews recent advances in statistical, traditional machine learning, and deep learning models applied to epidemic forecasting and risk assessment. Statistical models such as OLS regression, GWR, and logistic regression have been employed to capture spatial heterogeneity and evaluate factors including population density, mobility, and environmental conditions. Traditional Machine learning models like random forests and support vector machines extend predictive capabilities by handling nonlinear relationships and diverse data sources, while deep learning models such as LSTM, CNN, hybrid Inception ResNet, and spatio-temporal attention networks offer advanced techniques for modeling complex temporal and spatial dependencies. The paper also highlights three major challenges: heavy reliance on high-quality data, limited model applicability across regions, and weak interpretability of deep learning models. Approaches to address these issues include integrating multi-source data, employing federated learning to enhance adaptability while safeguarding data privacy, and developing interpretable AI frameworks in the future. Overall, this paper provides a comprehensive overview of current methodologies and identifies avenues for improving epidemic modeling, further positioning it as a practical and trustworthy tool for global public health preparedness.