Related Experiment Video
Updated: Jul 3, 2026
Dynamic Monitoring of Seroconversion using a Multianalyte Immunobead Assay for Covid-19
Published on: February 16, 2022
Fast screening of COVID-19 inpatient samples by integrating machine learning and label-free SERS methods
Jaya Sitjar1, Huey-Pin Tsai2, Han Lee1
1Engineered Materials for Biomedical Applications Laboratory, Department of Materials Science and Engineering, National Cheng Kung University, Tainan, 701, Taiwan.
Background:
Advances in bio-analyte detection demonstrate the need for innovation to overcome the limitations of traditional methods. Emerging viruses evolve into variants, driving the need for fast screening to minimize the time required for positive detection and establish standardized detection. In this study, a SERS-active substrate with Au NPs on a regularly arranged ZrO2 nanoporous structure was utilized to obtain the SERS spectrum of inpatient samples from COVID-19 patients. Two analytical approaches were applied to classify clinical samples - empirical method to identify peak assignments corresponding to the target SARS-CoV-2 BA.2 variant, and machine learning (ML) method to build classifier models.
Results:
Comparison of spectral profiles of pure BA.2 variant and inpatient samples showed significant differences in the occurrence of SERS peaks, requiring the selection of regions of interest for further analysis through the empirical method. SERS spectra are classified into CoV (+) and CoV (-) using both empirical and machine learning methods, each demonstrating a sensitivity of 85.7 % and a specificity of 60 %. The former method relies on peak assignment, which is performed manually relying on established and results in a longer turnaround time. In contrast, the latter method is based on the mathematical correlations between variables across the entire spectrum. The machine must continuously learn from larger datasets, and the response time for interpretation is short. Nonetheless, both methods demonstrated their suitability in classifying clinical samples.
Significance:
This study demonstrated that a more comprehensive discussion can be formed with the integration of machine learning classification with biochemical profiling with the empirical analysis approach. Further improvement is expected by combining these two methods by utilizing only the regions of interest instead of the entire spectrum as input for machine learning, as a feature extraction technique.
