Abstract
Background Diagnostic testing using PCR is a fundamental component of COVID-19 pandemic control. Criteria for determining who should be tested by PCR vary between countries, and ultimately depend on resource constraints and public health objectives. Decisions are often based on sets of symptoms in individuals presenting to health services, as well as demographic variables, such as age, and travel history. The objective of this study was to determine the sensitivity and specificity of sets of symptoms used for triaging individuals for confirmatory testing, with the aim of optimising public health decision making under different scenarios. Methods Data from the first wave of COVID-19 in New Zealand were analysed; comprising 1153 PCR-confirmed and 4750 symptomatic PCR negative individuals. Data were analysed using Multiple Correspondence Analysis (MCA), automated search algorithms, Bayesian Latent Class Analysis, Decision Tree Analysis and Random Forest (RF) machine learning. Results Clinical criteria used to guide who should be tested by PCR were based on a set of mostly respiratory symptoms: a new or worsening cough, sore throat, shortness of breath, coryza, anosmia, with or without fever. This set has relatively high sensitivity (> 90%) but low specificity (< 10%), using PCR as a quasi-gold standard. In contrast, a group of mostly non-respiratory symptoms, including weakness, muscle pain, joint pain, headache, anosmia and ageusia, explained more variance in the MCA and were associated with higher specificity, at the cost of reduced sensitivity. Using RF models, the incorporation of 15 common symptoms, age, sex and prioritised ethnicity provided algorithms that were both sensitive and specific (> 85% for both) for predicting PCR outcomes. Conclusions If predominantly respiratory symptoms are used for test-triaging, a large proportion of the individuals being tested may not have COVID-19. This could overwhelm testing capacity and hinder attempts to trace and eliminate infection. Specificity can be increased using alternative rules based on sets of symptoms informed by multivariate analysis and automated search algorithms, albeit at the cost of sensitivity. Both sensitivity and specificity can be improved through machine learning algorithms, incorporating symptom and demographic data, and hence may provide an alternative approach to test-triaging that can be optimised according to prevailing conditions.