Machine learning-based SARS-CoV-2 viral detection using surface-enhanced raman spectroscopy optimized with genetic algorithm

The gold standard method for COVID-19 diagnosis is the real-time polymerase chain reaction (RT-PCR) test on nasal swabs. Due to the high false-negative results, especially at low viral loads, and complex test processes, including the sample collection, pre-processing, and analysis, a highly sensitive and quick detection method based on easily obtained fluids such as saliva is needed. In this thesis, we present surface-enhanced Raman spectroscopy (SERS) metasurfaces for directly measuring COVID-19 from unprocessed saliva. The SERS substrates were optimized using a genetic algorithm (GA) to computationally screen gold metasurfaces from a pattern space of 2100 possibilities for strongly enhanced light-virus interaction, and we used the optimized SERS chip design to evaluate the presence and concentration of SARS-CoV-2. The electromagnetic field profiles were used to determine the field enhancement of each structure and used as a fitness value to sort the generated designs during the GA optimization. The optimized structure was fabricated with an electron-beam lithography and tested first with inactivated viruses. Since the Raman measurements contain far too many peaks to be assigned manually, the collected spectra were classified with machine-learning algorithms to show the system's performance. Then, on 36 PCR positive and 33 negative clinical samples (n=69), we used linear discriminant analysis and support vector machines model to identify the virus from Raman spectra with 95.2% sensitivity and 95.2% specificity and to discriminate the wild-type, alpha, and beta variants of SARS-CoV-2. Our findings may pave the way for more effective, safe, and quantitative preventive screening and variant identification.