Using influenza surveillance networks to estimate state-specific prevalence of SARS-CoV-2 in the United States

Abstract

Detection of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infections to date has relied heavily on reverse transcription polymerase chain reaction testing. However, limited test availability, high false-negative rates, and the existence of asymptomatic or subclinical infections have resulted in an undercounting of the true prevalence of SARS-CoV-2. Here, we show how influenza-like illness (ILI) outpatient surveillance data can be used to estimate the prevalence of SARS-CoV-2. We found a surge of non-influenza ILI above the seasonal average in March 2020 and showed that this surge correlated with coronavirus disease 2019 (COVID-19) case counts across states. If one-third of patients infected with SARS-CoV-2 in the United States sought care, this ILI surge would have corresponded to more than 8.7 million new SARS-CoV-2 infections across the United States during the 3-week period from 8 to 28 March 2020. Combining excess ILI counts with the date of onset of community transmission in the United States, we also show that the early epidemic in the United States was unlikely to have been doubling slower than every 4 days. Together, these results suggest a conceptual model for the COVID-19 epidemic in the United States characterized by rapid spread across the United States with more than 80% infected individuals remaining undetected. We emphasize the importance of testing these findings with seroprevalence data and discuss the broader potential to use syndromic surveillance for early detection and understanding of emerging infectious diseases.

Fig. 1. An early surge of ILI visits across the US.

The proportion of patients presenting with ILI that could not be explained by influenza or typical seasonal variation (that is, excess ILI) is shown for four states (blue line and ribbons represent the posterior median as well as 95% and 50% credible sets; results from all analyzed states are shown in fig. S1). ILI that could not be attributed to influenza was calculated based on influenza laboratory surveillance data (2019-2020 flu season shown in red, prior seasons are shown in black). A time-series model was used to infer seasonal variation of non-influenza ILI. Excess ILI was then calculated as the difference between non-influenza ILI from 2019-2020 and the seasonal baseline of non-influenza ILI. Excess ILI after March 7th is highlighted in darker blue as these data correlated strongly with observed COVID-19 case counts (fig. S2).

Fig. 2

The ILI surge imposes a dependence between growth rate and clinical rate in epidemiological models (A-B) SARS-CoV-2 prevalence estimates based on the ILI surge are consistent with an epidemiological model parameterized based on a January 15th epidemic start date and a doubling time equal to that observed for new deaths within the US (A) or Italy (B). Epidemiological models were either stochastic (simulated via tau-leaping) or deterministic (solved by numerical integration). In addition to our raw estimates of the ILI surge size (unadjusted), we provide adjusted prevalence estimates accounting for sub-clinical cases by assuming an 18% asymptomatic rate and a 40% rate of health-care seeking of symptomatic ILI patients (adjusted). Epidemic trajectories were simulated using an SEIR model (black lines). The increasing gap between ILI prevalence estimates and SEIR trajectories (orange) suggest the presence of additional factors such social distancing, changes in care-seeking behavior, or heterogeneity in susceptibility or transmission. (C) More generally, the size of the clinical population estimated from ILI data imposes a dependence between epidemic doubling time, the clinical rate, and the lag between onset of infectiousness and ILI reporting. Combinations of these three variables that are consistent (black) or inconsistent (gray) are shown as well as a smoothed estimate of clinical rate as a function of doubling time.