COVID-19 Diagnosis at Early Stage Based on Smartwatches and Machine Learning Techniques

Early detection of COVID-19 positive people are now extremely needed and considered to be one of the most effective ways how to limit spreading the infection. Commonly used screening methods are reverse transcription polymerase chain reaction (RT-PCR) or antigen tests, which need to be periodically repeated. This paper proposes a methodology for detecting the disease in non-invasive way using wearable devices and for the analysis of bio-markers using artificial intelligence. This paper have reused a publicly available dataset containing COVID-19, influenza, and Healthy control data. In total 27 COVID-19 positive and 27 healthy control were pre-selected for the experiment, and several feature extraction methods were applied to the data. This paper have experimented with several machine learning methods, such as XGBoost, k-nearest neighbour k-NN, support vector machine, logistic regression, decision tree, and random forest, and statistically evaluated their perfomance using various metrics, including accuracy, sensitivity and specificity. The proposed experiment reached 78 % accuracy using the k-NN algorithm which is significantly higher than reported for state-of-the-art methods. For the cohort containing influenza, the accuracy was 73 % for k-NN. Additionally, we identified the most relevant features that could indicate the changes between the healthy and infected state. The proposed methodology can complement the existing RT-PCR or antigen screening tests, and it can help to limit the spreading of the viral diseases, not only COVID-19, in the non-invasive way.