Extracting the SERS characteristic features of SARS-CoV-2 and using machine learning to quickly screen positive specimens among hospitalized patients.

The emergence of widespread viral transmission has prompted further development of early diagnostic methods, especially in situations where rapid and accurate testing is urgently needed. Label-free surface enhanced Raman spectroscopy (SERS) specimen classification through analytical chemistry combined with machine learning (ML) was shown to address such issue. However, the complexity of clinical specimens limit the qualitative interpretation, posing a challenge to the fast screening of clinical specimens. SERS-active substrates consisting of Au NPs/pZrO2 or Au NPs/fZrO2 were used to screen for SARS CoV-2 BA.2 variant among hospitalized inpatients. Nasopharyngeal and throat swabs were used, and their SERS spectra were distinguished. Full SERS spectra were extracted to distinguish CoV (+) and CoV (-) specimens to investigate the differences between ML of full SERS spectra and that of extracted spectral features. Nasopharyngeal specimens have shown to provide identifiable distinct features. Au NPs/fZrO2 exhibited better signal detection than Au NPs/pZrO2 owing to the filtering capability of the fibers, retaining only the virus particles on the surface. Full SERS spectra as input data in ML resulted in a sensitivity and specificity of 60 %; on the other hand, when extracted spectral features were used, a sensitivity of 100 % was exhibited by both substrates and specificities of 43 % and 57 % with Au NPs/pZrO2 and Au NPs/fZrO2, respectively. This study can quickly screen CoV (+) inpatients but is limited by the complexity of spectral profile affected by comorbidities, the number and diversity of inpatient specimens, and the need to further distinguish spectral profiles of CoV (-).