Identifying COVID-19 cases in outpatient settings

Abstract

Case identification is an ongoing issue for the COVID-19 epidemic, in particular for outpatient care where physicians must decide which patients to prioritise for further testing. This paper reports tools to classify patients based on symptom profiles based on 236 severe acute respiratory syndrome coronavirus 2 positive cases and 564 controls, accounting for the time course of illness using generalised multivariate logistic regression. Significant symptoms included abdominal pain, cough, diarrhoea, fever, headache, muscle ache, runny nose, sore throat, temperature between 37.5 and 37.9 °C and temperature above 38 °C, but their importance varied by day of illness at assessment. With a high percentile threshold for specificity at 0.95, the baseline model had reasonable sensitivity at 0.67. To further evaluate accuracy of model predictions, leave-one-out cross-validation confirmed high classification accuracy with an area under the receiver operating characteristic curve of 0.92. For the baseline model, sensitivity decreased to 0.56. External validation datasets reported similar result. Our study provides a tool to discern COVID-19 patients from controls using symptoms and day from illness onset with good predictive performance. It could be considered as a framework to complement laboratory testing in order to differentiate COVID-19 from other patients presenting with acute symptoms in outpatient care.

Keywords: COVID-19; Classification; diagnosis model; online tool; respiratory symptoms.

Fig. 1.

Distribution of 12 symptoms across illness days separated by control and COVID-19 groups. Blue bars indicate the proportion of control group out of all controls having a symptom, and orange bars show the corresponding proportions of COVID-19 patients.

Fig. 2.

Estimated coefficients for adjusted ORs and their CIs for symptoms at illness days 1–2, 3–4, 5–7 and 8+. The effects of selected variables from LASSO are plotted as line segments to indicate CIs and dots as mean estimates; those did not enter the second round of modelling are marked as dots without line segments as place holders for aesthetic and contrast. CIs coloured in black indicate significant effects, and in grey indicate non-significant effects. Having nausea or vomiting, is omitted from the figure because all of its interaction effects with illness days are excluded by GLM LASSO. The scale of parameter effect on the OR is exponentially spaced for visualisation.

Fig. 3.

(a) ROC curve with LOOCV. AUC = 0.89. Using full data, AUC = 0.92. With a minimum specification threshold at 0.95 and 0.9, the cutoff points are found at 0.92 and 0.74 respectively as indicated by the orange and red stars on the curve. (b) ROC curve stratified by illness days 1–2, 3–4, 5–7 and 8+.

Fig. 4.

Comparisons of predicted risk to observations: (a) shows bars with the height indicating the percentage of cases in respective intervals of predicted risk, (b) plots the predicted risk grouped by case or control and (c) traces the calculated in-sample sensitivity, specificity, PPV and NPV for a grid of predicted risks as cutoffs spaced at 0.001 from 0 to 1.