Submicro-Optical Spectroscopy Bioassay System for COVID-19 Neutralizing Antibody Immunity Response

COVID-19 led to widespread infections and severe outcomes, especially among vulnerable populations. Global vaccination efforts were coordinated to curb the virus's spread. However, vaccine efficacy varied based on factors like age, individual’s health status, living condition, type of infection and vaccine brand.

Assessing immunity requires monitoring neutralizing antibodies and involves the plaque-reduction neutralization test (PRNT), a costly procedure requiring biosafety laboratories. Alternatively, enzyme-linked immunosorbent assays (ELISAs) and FDA-approved cPass kits proved effective for antibody detection. While lateral flow immunoassays (LFIAs) offer rapid, equipment-free testing, their traditional reliance on visual inspection limits precision.

This study utilizes optical spectroscopy with gold nanoparticles (AuNPs) in LFIAs to enhance the quantitative accuracy of detecting neutralizing antibodies. The experimental setup includes a microprocessor that controls a stepper motor, an LED light source, and a micro spectrometer, enabling spectrum capture and wireless data transmission to cloud servers for analysis. Typically, spectrometers require large spaces, but this study applied a spectral chip to miniaturize the entire system, making the device portable and suitable for various clinical environments.

The serum samples from 265 vaccinated individuals were tested for neutralizing antibodies using a commercial LFIA kit, with an ELISA kit used for validation. Spectral data were captured in the range from 300 to 800 nm. Calculated reflectance spectra of AuNPs on the LFIA were processed using a Savitzky-Golay filter, allowing machine learning models to predict antibody concentrations.

Machine learning models were developed to predict COVID-19 neutralizing antibody levels from LFIA reflectance spectra. Serum samples with antibody levels measured via standard ELISA were categorized into three ranges to balance data distribution. Data standardization and principal component analysis (PCA) reduced complexity while preserving essential variance and enhancing model training.

Support Vector Machine (SVM), Random Forest, and Multilayer Perceptron (MLP) models were trained on 212 samples and tested on 53, achieving prediction accuracies of 90.5%, 92.4%, and 94.3%, respectively. The MLP model performed best, particularly in predicting extreme antibody concentrations, although some misclassifications occurred with mid-range levels due to sample distribution imbalance.

For quantitative prediction, three regression models—Support Vector Regression (SVR), Random Forest Regression (RFR), and Multilayer Perceptron Regression (MLPR) were used alongside a grayscale intensity calibration from the reaction area of LFIA strips. MLPR and RFR showed higher predictive accuracy and lower mean absolute error (MAE), though image processing predictions varied at high grayscale intensities due to an exponential fitting function.

Additionally, a hybrid regression model incorporating MLP classification and regression enhanced performance, achieving an R² of 0.923 and an MAE of 15.018 IU/mL (with the ELISA antibody detection range being 0–250 IU/mL). This model mitigated prediction biases from imbalanced sample distributions, providing LFIA results comparable to ELISA tests.

The experiments demonstrate that optical spectroscopy combined with machine learning enables rapid, precise antibody quantification for efficient diagnostics and offer a user-friendly and economical diagnostic approach. The system also has uses in other biological detections using AuNPs and has the potential to analyze complex biological samples and development of innovative diagnostic tools in biomedical research.