Abstract
Since the World Health Organization declared the Coronavirus Disease 2019 (COVID-19) pandemic as an international concern of public health emergency in the early 2020, several strategies have been initiated in many countries to prevent healthcare services breakdown and collapse of healthcare structures. The most important strategy was the increased testing, diagnosis, isolation, contact tracing to identify, quarantine and test close contacts. In this context, finding a rapid, reliable and affordable tool for COVID-19 screening was the main challenge to address the pandemic. Molecular diagnosis by reverse transcriptase polymerase chain reaction (RT-PCR), even though considered as the gold standard in the diagnosis of COVID-19, was time consuming and therefore does not fit the objective of rapid screening. In addition, serological tests to detect anti-severe acute respiratory syndrome coronavirus 2 (SARS-COV-2) antibodies suffered from low sensitivity. Prediction models based on machine-learning (ML) that combined several clinical features to estimate the risk of COVID-19 have been developed. To address these screening challenges, we created a ML model (MLM) based on gradient boosting method. We included several clinical features and the daily geographic prevalence of COVID-19 cases in the MLM. The MLM was trained on 1554 cases (757 COVID-19), and tested on 547 cases (169 COVID-19). Our MLM successfully predicted RT-PCR positivity with an accuracy of 97.06%. Moreover, the variable sensitivity and specificity of our MLM depending on the disease geographic prevalence has introduced the concept of "dynamic" disease screening. In the context of future world pandemic emergencies, we believe that this MLM method can be very useful as a rapid, reliable and dynamic screening tool for contagious diseases, especially in the developing countries.