A novel online multi-task learning for COVID-19 multi-output spatio-temporal prediction.

Area of Science:

• Epidemiology
• Data Science
• Machine Learning

Background:

• Predicting COVID-19 trends is crucial for public health decision-making but is complicated by temporal autocorrelation, spatial dependency, and concept drift.
• Existing machine learning methods struggle to simultaneously address these challenges in pandemic forecasting.

Purpose of the Study:

• To develop a novel online multi-task regression algorithm capable of handling temporal autocorrelation, spatial dependency, and concept drift in COVID-19 trend prediction.
• To improve the accuracy and adaptability of pandemic forecasting models.

Main Methods:

• Developed an online multi-task regression algorithm incorporating a chain structure for spatial dependency, ADWIN drift detector for concept drift adaptation, and lag time series features for temporal autocorrelation.
• Conducted comparative experiments using daily confirmed COVID-19 cases from 20 areas in California.

Main Results:

• The proposed model demonstrated superior adaptation to concept drift in COVID-19 data compared to existing methods.
• Effectively captured spatial dependencies across different regions, leading to enhanced prediction accuracy.
• Outperformed state-of-the-art batch machine learning models like N-Beats, DeepAR, TCN, and LSTM.

Conclusions:

• The novel algorithm successfully addresses the key challenges in COVID-19 trend prediction, offering improved accuracy and adaptability.
• This approach provides a more robust tool for pandemic forecasting and public health strategy development.