Open Access

ARTICLE

A Machine Learning Approach for Early COVID-19 Symptoms Identification

by Omer Ali^1,2, Mohamad Khairi Ishak^1,*, Muhammad Kamran Liaquat Bhatti²

1 School of Electrical and Electronic Engineering, Universiti Sains Malaysia (USM), Nibong Tebal, 14300, Penang, Malaysia
2 Department of Electrical Engineering, NFC Institute of Engineering and Technology (NFC IET), Multan, 60000, Pakistan

* Corresponding Author: Mohamad Khairi Ishak. Email:

(This article belongs to the Special Issue: AI for Wearable Sensing--Smartphone / Smartwatch User Identification / Authentication)

Computers, Materials & Continua 2022, 70(2), 3803-3820. https://doi.org/10.32604/cmc.2022.019797

Received 26 April 2021; Accepted 01 July 2021; Issue published 27 September 2021

Abstract

Symptom identification and early detection are the first steps towards a health condition diagnosis. The COVID-19 virus causes pneumonia-like symptoms such as fever, cough, and shortness of breath. Many COVID-19 contraction tests necessitate extensive clinical protocols in medical settings. Clinical studies help with the accurate analysis of COVID-19, where the virus has already spread to the lungs in most patients. The majority of existing supervised machine learning-based disease detection techniques are based on clinical data like x-rays and computerized tomography. This is heavily reliant on a larger clinical study and does not emphasize early symptom detection. The aim of this study is to investigate anomalies in patient physiological data for early COVID-19 symptoms identification. In this context, two of the most prevalent symptoms, fever and cough, were examined in a two-fold manner utilizing an unsupervised machine learning model. To examine disease progression, physiological features from a chest-worn device were analyzed. First, a Single Vector Activity Index (SVAI) parameter is proposed to monitor the breathing and cough patterns. Second, the dataset's variance is examined using the DBSCAN method for clustering and outlier detection. Finally, the model accuracy is evaluated to identify outliers on real-time data based on feature dissimilarities, yielding an overall detection accuracy of 90.34%.

---

Keywords

---