Predicting local COVID-19 emergences: A time-series classification approach and value of data from social media, search engines, and neighbouring regions

Erin E Rees; Mani Sotoodeh; José Denis-Robichaud; Hélène Carabin; Simon de de Montigny

doi:10.1016/j.epidem.2026.100891

Predicting local COVID-19 emergences: A time-series classification approach and value of data from social media, search engines, and neighbouring regions

Epidemics. 2026 Jan 22:54:100891. doi: 10.1016/j.epidem.2026.100891. Online ahead of print.

Authors

Erin E Rees¹, Mani Sotoodeh², José Denis-Robichaud³, Hélène Carabin⁴, Simon de de Montigny⁵

Affiliations

¹ Public Health Agency of Canada (PHAC), Saint-Hyacinthe, Québec, Canada; Université de Montréal, Montréal, Québec, Canada; Groupe de recherche en épidémiologie des zoonoses et santé publique (GREZOSP), Saint-Hyacinthe, Québec, Canada. Electronic address: erin.rees@phac-aspc.gc.ca.
² Université de Montréal, Montréal, Québec, Canada.
³ Université de Montréal, Montréal, Québec, Canada; Groupe de recherche en épidémiologie des zoonoses et santé publique (GREZOSP), Saint-Hyacinthe, Québec, Canada; Independent Researcher, Amqui, Québec, Canada.
⁴ Université de Montréal, Montréal, Québec, Canada; Groupe de recherche en épidémiologie des zoonoses et santé publique (GREZOSP), Saint-Hyacinthe, Québec, Canada; Centre de recherche en santé publique (CReSP), Montréal, Québec, Canada.
⁵ Public Health Agency of Canada (PHAC), Saint-Hyacinthe, Québec, Canada; Université de Montréal, Montréal, Québec, Canada.

PMID: 41604747
DOI: 10.1016/j.epidem.2026.100891

Abstract

Background: Early warning for known infectious disease threats use methods that focus on detection of outbreaks, often at large geographical scales. However, earlier warning, specifically at the onset of disease emergence (i.e., first case(s)) and at finer spatial scales could significantly improve timeliness and targeting of prevention and control efforts. As a proof-of-concept, we demonstrate that a early classification time-series approach can predict COVID-19 emergence at a local jurisdictional level with a 10-day lead time.

Methods: To predict emergence with a 10-day lead time in Canadian health regions (HRs) during January to November 2020, we developed three classification models. Predictor variables were restricted to information about COVID-19 and included daily metrics at the HR level for social media and traditional EBS data (i.e., news media), and at the provincial/territorial (P/T) level for search engine data. Predictor contributions from neighbouring areas additionally included reported case data (with the other predictors) from the nearest region, or weighted by distance and/or population size of all adjacent regions.

Results: Using the highest performing model, Deep Gated Recurrent Unit, the classification balanced accuracy was higher for distance- and population-based spatial weighting (0.78), than for nearest neighbour data only (0.64). It was also higher when open-access information was included with traditional EBS information (0.78), compared to excluding open-access information (0.63).

Conclusions: In a Canadian context for COVID-19, using a retrospective approach, study results demonstrate classification models can predict emergence with a 10-day lead time at the finest spatial scale of health governance (i.e., HRs) used by P/Ts. Furthermore, prediction accuracy improves with information from neighbouring regions and open-access data (social media, search engine). Implications for operationalizing our method in event-based surveillance systems are discussed.

Keywords: Deep Learning; Early Warning Systems; Public Health; Search Engine; Social Media.