Harnessing Participatory Surveillance Cohorts and Proxy Indicators to Dynamically Track Epidemic Trends and Undiagnosed COVID-19 Infections in Singapore: Longitudinal Observational Study.

Abstract

Accurate COVID-19 incidence estimates, including undiagnosed cases, are vital for epidemic management but are often unavailable in real time. Participatory surveillance can capture community illness episodes; however, quantifying undiagnosed infections remains difficult. We assessed a Singaporean cohort to estimate medically unattended COVID-19 infections by combining symptom models with proxy epidemic indicators. This study aims to estimate COVID-19 incidence and medically attended fractions using participatory surveillance data and to evaluate the consistency of these estimates against independently derived serological measures of infection in a community cohort in Singapore. We analyzed 11 survey waves (September 2021 to November 2022) from the SOCRATEs (Strengthening Our Community's Resilience Against Threats from Emerging Infections) community cohort (n=1899), spanning Delta and Omicron variant waves. Respondents reported recent illness, symptoms, health care use, and COVID-19 diagnoses. Multilevel logistic regression of medically attended episodes estimated the probability of COVID-19 in unattended episodes, incorporating symptoms and external indicators-wastewater viral-load index and health care staff surveillance. The estimates of total infections and medically attended fractions were validated against independent serological survey results. Among 2284 illness episodes, 756 were diagnosed with COVID-19, of which 62.4% (472/756) were medically attended. Health care-seeking declined from 83.9% (26/31) of COVID-19 episodes early in 2022 to 55.3% (47/85) by late 2022. Regression models demonstrated strong associations between COVID-19 infection and key symptoms and epidemic activity indicators. Estimated total infections were substantially higher than notified cases, reaching 1.0 to 2.9 times the reported incidence across successive variant waves. Model-based estimates of medically attended fractions were broadly consistent with serological benchmarks. Incidence estimates closely matched serological estimates in earlier intervals, with slight overestimation in later periods due to reinfection. Participatory surveillance, when combined with probabilistic modeling and external indicators of epidemic activity, can generate robust estimates of infection incidence and health care use. Agreement with serological data supports the validity of this integrated framework, although discrepancies persist in later epidemic phases due to reinfection dynamics. This approach provides a scalable and timely complement to traditional and serological surveillance systems.