Estimating the feasibility of symptom based classification of COVID-19

Alison Callahan*, Ethan Steinberg*, Conor Corbin, Jason Fries and Nigam H. Shah (* = equal contributors)

Given limited testing capacity, symptom checker tools are increasingly employed to help decide who should be tested for SARS-CoV-2. Aside from indicating the severity of illness, the degree to which presenting symptoms are specific to COVID-19 is unclear.

To estimate the feasibility of using symptom based screening to assign a probability of having COVID-19, we attempted to predict positive test results of other respiratory viruses solely using the presenting symptoms. We did the analysis using de-identified data from our health system.

We trained logistic regression classifiers for predicting test results for six viruses — Adenovirus, Influenza Virus A, Metapneumovirus, Parainfluenza Virus, Respiratory Syncytial Virus (RSV), and Rhinovirus — using de-identified data of 60,394 patients (4,355+ve tests; 56,039 -ve tests). In order to emulate real-life usage of such symptom based classification, only the de-identified contents of the emergency department or urgent care admission note associated with a virus test order were used.

The ability to distinguish positive and negative test results using just presenting symptoms ranged from c-statistic of 0.62 to 0.77 for the six viruses tested. These results suggest that for common respiratory viral infections, the presenting symptoms may not have sufficient information to correctly anticipate whether a given patient will test positive for that virus infection.

While it is certainly possible that COVID-19 symptoms are highly specific, allowing accurate classification of positive cases; based on the results for other respiratory viruses it appears that symptom based screening may not be effective in assigning a probability of having COVID-19.