Deep learning via LSTM models for COVID-19 infection forecasting in India

Area of Science:

• Epidemiology
• Computational Biology
• Data Science

Background:

• The COVID-19 pandemic significantly impacted global health, economies, and infrastructure.
• Existing computational and mathematical models for infection spread have shown limitations due to complexity and data scarcity.
• Reliable data and innovative forecasting models are crucial for understanding and managing the pandemic.

Purpose of the Study:

• To apply deep learning models, specifically recurrent neural networks (RNNs), for multi-step COVID-19 infection forecasting.
• To evaluate the performance of Long Short-Term Memory (LSTM), Bidirectional LSTM, and Encoder-Decoder LSTM models.
• To provide short-term (two months ahead) infection forecasts for Indian states during the first and second waves of the pandemic.

Main Methods:

• Utilized recurrent neural networks, including LSTM, Bidirectional LSTM, and Encoder-Decoder LSTM architectures.
• Focused on Indian states identified as COVID-19 hotspots.
• Modeled spatiotemporal sequences of infection data from 2020 and 2021 to generate forecasts.

Main Results:

• The models accurately predicted short-term COVID-19 infection trends.
• Forecasts indicated a low likelihood of a major infection wave in October and November 2021.
• The study highlighted the potential for applying these deep learning methods to other regions.

Conclusions:

• Deep learning, particularly LSTM-based models, offers a promising approach for reliable COVID-19 infection forecasting.
• Continued vigilance is necessary due to the emergence of new variants.
• Challenges remain in data reliability and incorporating socio-demographic factors into models.