Leveraging Transfer Learning from SARS-CoV-2 Data for Forecasting HMPV and HCoV Using LSTM Networks

• Mohit Angurala,
• Deepak Kaul,
• Nidhi Chopra,

Vol. 19, No. 12, December 31, 2025
10.3837/tiis.2025.12.003, Download Paper (Free):

• Keywords: Human Coronavirus
• Human Metapneumovirus
• LSTM Models
• Respiratory viruses
• SARS-CoV-2
• Transfer Learning.

Abstract

Motivation: The accurate prediction of seasonal respiratory viruses is necessary for public health planning. Nevertheless, due to the absence of data for lesser studied pathogens, including HMPV (human metapneumovirus) and HCoV (human coronaviruses), routine prediction models are not so effective. Description: To address this, we employ transfer learning. We first pre-train an LSTM network on a large SARS-CoV-2 percent-positivity dataset from 10 U.S. HHS regions. We then retrain the model for HMPV, and HCoV. Our approach consists of preprocessing, segmenting the data into sliding windows sequences with fixed length, defining the model architecture and loss, validating and submitting. Results: Our experiments show that transfer learning greatly improves predictive accuracy compared to standalone LSTMs trained from scratch. For HMPV, the model achieves an MSE of 0.0044 with R² = 0.97. For HCoV, it achieves an MSE of 0.0051 with R² = 0.96. Additionally, the transfer-based models converge more quickly and need fewer training epochs than non-transfer baselines. Conclusion: These results show that cross-virus transfer successfully captures shared seasonal and regional patterns. This enables reliable forecasting even when data is limited. The approach provides a useful tool to support early warning systems and improve public health responses.

Statistics

Cite this article