A deep learning approach for enhancing pandemic prediction: A retrospective evaluation of transformer neural networks and multi-source data fusion for infectious disease forecasting

Abstract

This paper introduces a deep learning model for county-level Covid-19 forecasting, presenting it as a retrospective case study. We utilize a transformer neural network with multi-source data fusion, incorporating historical case data, death data, and social media sentiment to capture complex temporal and spatial dynamics. Additionally, we develop multi-level and multi-scale attention mechanisms for adaptive time-frequency analysis. In a retrospective evaluation across three Omicron variant waves (December 2021 through February 2023), the model demonstrated strong performance in predicting county-level Covid-19 cases and deaths, with median county agreement accuracy ranging from 74.0 % to 82.6 % for one-week case forecasts and 68.7–79.6 % for 5-week case forecasts. While these historical results are promising, prospective validation is needed to assess the model's utility under live, evolving data conditions. Median county agreement accuracy for deaths ranged from 83.2 % to 86.3 % for one-week forecasts and 84.3–87.2 % for five-week forecasts. Incorporating social media data yielded mild to moderate improvement in forecasting accuracy. Overall, the proposed model yielded substantial improvements compared to a baseline persistence model utilizing the last observation carried forward. By integrating real-time data and capturing complex pandemic dynamics, this approach surpasses traditional methods. The results demonstrate the model's strong performance in a retrospective setting, highlighting the utility of multi-source data fusion and attention mechanisms for fine-grained epidemiological forecasting. This work serves as a case study on the application of advanced deep learning techniques to local-level pandemic data, offering a methodological framework for future research.