Refining COVID-19 retrospective diagnosis with continuous serological tests: a Bayesian mixture model

Abstract

AbstractCOVID-19 serological tests with a “positive”, “intermediate” or “negative” result according to predefined thresholds cannot be directly interpreted as a probability of having been infected with SARS-CoV-2. Based on 81,797 continuous anti-spike tests collected in France after the first wave, a Bayesian mixture model was developed to provide a tailored infection probability for each participant. Depending on the serological value and the context (age and administrative region), a negative or a positive test could correspond to a probability of infection as high as 61.9% or as low as 68.0%, respectively. In infected individuals, the model estimated a proportion of “non-responders” of 14.5% (95% CI, 11.2-18.1%), corresponding to a sub-group of persons who exhibited a weaker serological response to SARS-CoV-2. This model allows for an individual interpretation of serological results as a probability of infection, depending on the context and without any notion of threshold.