Transfer-recursive-ensemble learning for multi-day COVID-19 prediction in India using recurrent neural networks

Area of Science:

• Epidemiology
• Artificial Intelligence
• Public Health

Background:

• The COVID-19 pandemic severely strained Indian healthcare infrastructure, particularly during the second wave.
• Overburdened hospitals faced shortages of supplies and oxygen, highlighting the need for predictive capabilities.

Purpose of the Study:

• To develop a predictive model for forecasting COVID-19 cases, deaths, and active cases in India.
• To improve medical resource allocation and inform pandemic-related decision-making through accurate, multi-day predictions.

Main Methods:

• Utilized gated recurrent unit (GRU) networks for COVID-19 case prediction.
• Employed transfer learning by pre-training models on data from the USA, Brazil, Spain, and Bangladesh.
• Fine-tuned pre-trained models on Indian COVID-19 data and generated 7-day ahead predictions using recursive learning.
• Developed an ensemble model combining predictions from individual pre-trained models.

Main Results:

• The ensemble model incorporating pre-training from Spain and Bangladesh demonstrated superior performance.
• This approach outperformed predictions from models pre-trained on other country data combinations.
• The proposed method showed better results compared to traditional regression models.

Conclusions:

• Transfer learning with diverse international COVID-19 data significantly enhances predictive accuracy for India.
• Ensemble GRU models pre-trained on specific country datasets (Spain, Bangladesh) offer a robust solution for short-term forecasting.
• The findings support the use of advanced AI techniques for proactive healthcare resource management during pandemics.