Biomarkers for COVID-19 – a detailed analysis of feature selection methods on longitudinal data

As the COVID-19 pandemic evolved over time since its initial outbreak in December 2019, there has been an increasing need to provide appropriate treatments and medications for infected patients by developing predictive systems for potential changes in disease status. This study provides a detailed analysis and comparison of performance of various feature selection techniques using a longitudinal gene expression dataset in terms of the gene enrichment analysis results of selected subset of features (i.e., biomarkers) for COVID-19. The feature selection methods were based on hypothesis tests, statistical scores, and machine learning models. The subset selected using unsupervised Laplacian score was found to have the best quality by annotating the relevant pathways with the lowest p-values. The subsets selected by one-way ANOVA tests and an autoencoder-based feature selection method were also performing well. The selected subsets of features were proved to be able to introduce a boost in machine learning model performance, as compared to the original full set of features. The results have demonstrated that certain feature selection methods are well-suited for detecting biomarkers for COVID-19 and the selected biomarkers can help with early diagnosis and prognosis of COVID-19.