Combining label-free Raman spectroscopy and machine learning to identify early biomarkers of COVID-19 disease severity and mortality.

Abstract

Significance: Early prediction of COVID-19 severity and mortality is crucial for optimizing clinical care and patient outcomes, but remains challenging.

Aim: We aim to develop a screening tool combining label-free Raman spectroscopy and machine learning modeling to predict COVID-19 severity and mortality.

Approach: Patients infected by SARS-CoV-2 ( $N=58$ ) were recruited during the first wave of COVID-19 and stratified based on respiratory support. Blood samples were collected during hospitalization and analyzed using Raman spectroscopy and metabolomics. Machine learning models based on Raman spectra were developed to classify (1) survivors versus nonsurvivors, (2) critical patients with noninvasive versus invasive ventilation, and (3) noncritical (no respiratory support or oxygen via nasal cannula) versus critical patients.

Results: Raman peaks assigned to proteins, glucose, lactic acid, fatty acids, urea, and lipids were extracted by the models. Area under the receiver operating characteristic curve ranged between 0.83 and 0.94, with sensitivities and specificities ranging between 80% and 83% and 75% and 92%, respectively. Accuracy for detecting mortality, invasive ventilation, and critical disease was 90%, 87%, and 78%. A complementary metabolomic analysis confirmed some molecular differences between groups.

Conclusions: These results suggest the potential of Raman spectroscopy and machine learning modeling to stratify COVID-19 patients at admission, individualize care, and improve survival rates.