The outbreak of infectious diseases can have profound impacts on socio-economic balances globally. Accurate short-term forecasting of infectious diseases is crucial for policymakers and healthcare systems. This study proposes a novel deep learning approach for short-term forecasting of infectious disease trends, using COVID-19 confirmed cases and hospitalizations in Japan as a case study.
This method provides weekly updates and forecasts outcomes over 104 weeks. The proposed model combines long short-term memory (LSTM) networks and a multi-head attention mechanism and is trained on public data sourced from open-access platforms. The study conducts a comprehensive and rigorous evaluation of the performance of the model, assessing its weekly predictive capabilities over a long period of time by employing multiple error metrics.
Furthermore, the study explores how the model's performance varies over time and across geographical locations. The results demonstrate that the proposed model outperforms baseline approaches, particularly in short-term forecasts, achieving lower error rates across multiple metrics. Additionally, the inclusion of mobility data improves the predictive accuracy of the model, especially for longer-term forecasts, by capturing spatio-temporal dynamics more effectively.
The proposed model has the potential to assist in decision-making processes, help develop strategies for controlling the spread of infectious diseases, and mitigate the pandemic's impact. Since the early 21st century, several outbreaks of infectious diseases have posed significant threats to human health globally, with measures like lockdowns, social distancing, and travel restrictions reshaping how people interact.
Given the profound impacts of infectious diseases, developing predictive models to forecast the pandemic's progression is essential. Reliability in forecasts allows for timely decision-making, thus enabling proactive measures to mitigate public health consequences. This study integrates relevant input data streams with advanced modeling techniques to enhance accuracy and reliability in forecasting disease dynamics.
The framework combines an LSTM layer to capture temporal dependencies with a transformer encoder layer that aggregates relevant information. Models were trained using time-series data from Japan, focusing on prefecture-level data modeling and analysis between December 6, 2020, and October 16, 2021. The time series consisted of confirmed case data and hospitalization data sourced from the Ministry of Health, Labour and Welfare.
Additionally, mobility data obtained from Google’s mobility reports categorized by location and type of place were integrated as input features in the model to further improve its predictive power.
The results reveal that incorporating mobility data significantly enhances forecasting performance, particularly over longer time horizons. The model successfully accommodates the complexities of disease dynamics, demonstrating consistency in predictive performance across various prefectures. This highlights the utility of the model in informing policymakers for effective interventions during infectious disease outbreaks.
Despite its promising capabilities, the model's accuracy may decline during peaks in the outbreak or with underrepresentation of certain groups. The incorporation of other data sources and factors into future iterations may further build on its robustness, ultimately guiding more effective public health responses to infectious diseases.
