Machine Learning Models for Early Prediction of COVID-19 Infections Based on Clinical Signs

Nowadays, the appearance of common symptoms, such as cough, fever, and loss of smell and taste, is the starting point of a battle against the coronavirus. The first standard method of COVID-19 infection assertion has become the RT-PCR test, which is however an uncomfortable solution for both patients and medical staff due to its high cost, timeliness, and false-negative result issue. This has raised the need for reliable automatic detection systems that aid in the early prediction of the COVID-19 infections with a lower cost. In this work, we aim at profiting from the Machine Learning (ML) advances to provide a reliable and low-cost COVID-19 prediction system. This system is based on the disease starting point, which is the patients’ clinical symptoms, that are still under-explored. We developed seven predictive models using traditional ML classification algorithms using a public dataset of obvious high-risk factors from patients’ clinical signs. The dataset has first undergone a pre-processing phase consisting of feature engineering and dataset resampling to deal with imbalanced dataset issue. Our best classification model is able to detect true positives and true negatives and weed out false positive and false negatives with an accuracy of 93%.