JMIR Publications

ABSTRACT

Background:

controlling the COVID-19 outbreak in Brazil is considered a challenge of continental proportions due to the high population and urban density, weak implementation and maintenance of social distancing strategies, and limited testing capabilities.

Objective:

to contribute to addressing such a challenge, we present the implementation and evaluation of supervised Machine Learning (ML) models to assist the COVID-19 detection in Brazil based on early-stage symptoms.

Methods:

firstly, we conducted data preprocessing and applied the Chi-squared test in a Brazilian dataset, mainly composed of early-stage symptoms, to perform statistical analyses. Afterward, we implemented ML models using the Random Forest (RF), Support Vector Machine (SVM), Multilayer Perceptron (MLP), K-Nearest Neighbors (KNN), Decision Tree (DT), Gradient Boosting Machine (GBM), and Extreme Gradient Boosting (XGBoost) algorithms. We evaluated the ML models using precision, accuracy score, recall, the area under the curve, and the Friedman and Nemenyi tests. Based on the comparison, we grouped the top five ML models and measured feature importance.

Results:

the MLP model presented the highest mean accuracy score, with more than 97.85%, when compared to GBM (> 97.39%), RF (> 97.36%), DT (> 97.07%), XGBoost (> 97.06%), KNN (> 95.14%), and SVM (> 94.27%). Based on the statistical comparison, we grouped MLP, GBM, DT, RF, and XGBoost, as the top five ML models, because the evaluation results are statistically indistinguishable. The ML models` importance of features used during predictions varies from gender, profession, fever, sore throat, dyspnea, olfactory disorder, cough, runny nose, taste disorder, and headache.

Conclusions:

supervised ML models effectively assist the decision making in medical diagnosis and public administration (e.g., testing strategies), based on early-stage symptoms that do not require advanced and expensive exams.