Learning from pandemics: Feature selection and clustering for robust digital disease nowcasting

Disease now-casting increasingly relies on non-traditional data, such as online activity, which offer fast, inexpensive insights but are often noisy, unstructured, and influenced by behavioral confounders. Models using search engine data have frequently suffered from overfitting and unstable performance. However, pandemics may help distinguish true disease signals from behavioral noise, as both tend to surge during outbreaks. We propose a hybrid approach to improve prediction that combines model evaluation with behavioral insight. We apply simple supervised and unsupervised methods to different flu and COVID-19 datasets, using the 2009 H1N1 and 2020 COVID-19 pandemics as training periods. In both cases, selecting a curated subset of search terms improved model performance, reduced overfitting, and increased stability, consistently outperforming models that used search terms indiscriminately. These results suggest that feature curation informed by behavioral insights can enhance now-casting beyond purely statistical approaches. Our method offers a scalable and adaptable framework for integrating digital epidemiology into public health surveillance.