ABSTRACT
For the fight against the COVID-19 pandemic, it is particularly important to map the course of infection, in terms of patients who have currently tested SARS-CoV-2 positive, as accurately as possible. In hospitals, this is even more important because resources have become scarce. Although polymerase chain reaction (PCR) and point of care (POC) antigen testing capacities have been massively expanded, they are often very time-consuming and cost-intensive and, in some cases, lack appropriate performance. To meet these challenges, we propose the COVIDAL classifier for AI-based diagnosis of symptomatic COVID-19 subjects in hospitals based on laboratory parameters. We evaluate the algorithm's performance by unique multicenter data with approximately 4,000 patients and an extraordinary high ratio of SARS-CoV-2-positive patients. We analyze the influence of data preparation, flexibility in optimization targets, as well as the selection of the test set on the COVIDAL outcome. The algorithm is compared with standard AI, PCR, POC antigen testing and manual classifications of seven physicians by a decision theoretic scoring model including performance metrics, turnaround times and cost. Thereby, we define health care settings in which a certain classifier for COVID-19 diagnosis is to be applied. We find sensitivities, specificities, and accuracies of the COVIDAL algorithm of up to 90 percent. Our scoring model suggests using PCR testing for a focus on performance metrics. For turnaround times, POC antigen testing should be used. If balancing performance, turnaround times, and cost is of interest, as, for example, in the emergency department, COVIDAL is superior based on the scoring model. © 2023 Association for Computing Machinery.