Performance and explainability of feature selection-boosted tree-based classifiers for COVID-19 detection

Area of Science:

• Computational epidemiology
• Machine learning applications in public health
• Infectious disease modeling

Background:

• Self-reported data offers a vast resource for tracking infectious diseases like COVID-19.
• Previous methods often used limited symptom sets, potentially missing broader influencing factors.
• The University of Maryland Global COVID-19 Trends and Impact Survey (UMD-CTIS) collected extensive relevant information.

Purpose of the Study:

• To develop and evaluate a machine learning methodology for predicting COVID-19-positive cases using comprehensive self-reported data.
• To enhance the explainability and performance of predictive models through feature selection.
• To identify key factors contributing to COVID-19 spread across different countries and time periods.

Main Methods:

• Employed recursive feature elimination (RFE) for feature selection to optimize model input.
• Utilized tree-based supervised machine learning classifiers: Random Forests (RF), Light Gradient Boosting (LGB), and Extreme Gradient Boosting (XGB).
• Evaluated model performance using metrics like F1-score, sensitivity, specificity, precision, ROC, and AUC across Brazil, Canada, Japan, and South Africa for 2020-2021.

Main Results:

• The methodology successfully predicted COVID-19 cases with high accuracy across diverse countries and timeframes.
• Feature selection effectively reduced input variables without compromising detection performance.
• Shapley values and feature importance analyses provided insights into the drivers of COVID-19 incidence.

Conclusions:

• Machine learning models, enhanced by feature selection and broad data inputs, offer a robust approach for COVID-19 prediction.
• The methodology demonstrates the utility of diverse data sources, including community information and vaccination status, for epidemiological surveillance.
• Explainability analysis confirmed the model's ability to identify significant predictors of COVID-19 spread.