RESUMO
Detection of the physiological response to severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection is challenging in the absence of overt clinical signs but remains necessary to understand a full subclinical disease spectrum. In this study, our objective was to use radiomics (from computed tomography images) and blood biomarkers to predict SARS-CoV-2 infection in a nonhuman primate model (NHP) with inapparent clinical disease. To accomplish this aim, we built machine-learning models to predict SARS-CoV-2 infection in a NHP model of subclinical disease using baseline-normalized radiomic and blood sample analyses data from SARS-CoV-2-exposed and control (mock-exposed) crab-eating macaques. We applied a novel adaptation of the minimum redundancy maximum relevance (mRMR) feature-selection technique, called mRMR-permute, for statistically-thresholded and unbiased feature selection. Through performance comparison of eight machine-learning models trained on 14 feature sets, we demonstrated that a logistic regression model trained on the mRMR-permute feature set can predict SARS-CoV-2 infection with very high accuracy. Eighty-nine percent of mRMR-permute selected features had strong and significant class effects. Through this work, we identified a key set of radiomic and blood biomarkers that can be used to predict infection status even in the absence of clinical signs. Furthermore, we proposed and demonstrated the utility of a novel feature-selection technique called mRMR-permute. This work lays the foundation for the prediction and classification of SARS-CoV-2 disease severity.