Leveraging universal and transfer learning models for influenza prediction in Thailand

Influenza is a major respiratory disease that causes significant morbidity and mortality worldwide. Accurate predictions of influenza incidence enable public health organizations to monitor and prepare for outbreaks, ultimately reducing mortality and optimizing resource allocation. However, many countries, including Thailand, face challenges in generating accurate forecasts due to limited feature data in certain regions. To address this, we developed universal deep learning (DL)-based models to predict influenza incidence across multiple provinces in Thailand from 2010 to 2019. We evaluated various model configurations and implemented a feature selection process to enhance model generalizability and performance by ensuring equal contributions from multiple time series features. Our findings indicate that single hidden layer models with 128 nodes performed the best in the universal framework. To extend predictions to provinces without meteorological and PM10 data, we applied transfer learning (TL) using pre-trained models. The TL-based model, fine-tuned for each province, significantly outperformed baseline models trained solely on previous incidence, achieving the highest accuracy. Our results demonstrate the potential of universal DL and TL frameworks in forecasting influenza trends, even in limited data regions, and highlight the importance of incorporating domain-specific knowledge for robust epidemic management strategies.