Predictive and interpretable machine learning for COVID-19 resurgences: the role of SARS-CoV-2 variants in the post-pandemic era

Traditional COVID-19 forecasting often misses the rapid dynamics of viral competition, limiting timely ublic health responses. This study demonstrates the value of incorporating SARS-CoV-2 variant data into recurrent neural networks, using an interpretable, data-driven approach to improve accuracy in the current pandemic phase. We validated our approach on post-pandemic data (2022–2025) from New York City and the United Kingdom, integrating epidemiological time series with genomic surveillance of variants. We implemented and compared several neural network structures, with LSTM achieving the best performance. To assess the contribution of variant-specific data, we compared models with and without variant inputs. For interpretability and understanding model decisions, we applied XAI techniques to quantify variant influence on predictions. Incorporating variant data markedly improved forecasting accuracy across all horizons. In New York City, MAPE dropped from 32.15% to 7.35% during periods of rapid variant change, while in the UK it fell from 35.62% to 7.73%. XAI analyses revealed the dominant role of specific variants and captured their competitive displacement dynamics, with model explanations closely matching observed epidemiological trends. This study introduces a variant-aware methodology that improves COVID-19 prediction in the current endemic phase. The main contributions are: (i) ablation studies demonstrating the value of incorporating variant data to model case resurgences and declines; (ii) interpretable results into variant-driven dynamics via XAI; and (iii) validation across multiple geographical scales. Our approach establishes a scalable paradigm for genomic-informed epidemic forecasting, adaptable to evolving respiratory viruses.