Interpretability-driven deep learning for SERS-based classification of respiratory viruses.

Area of Science:

• Nanotechnology
• Biomedical Engineering
• Machine Learning

Background:

• Respiratory viruses like influenza, RSV, and SARS-CoV-2 pose significant global health risks.
• Accurate and rapid variant-level diagnostics are crucial for managing outbreaks.
• Existing diagnostic methods may lack speed, accuracy, or the ability to differentiate specific variants.

Purpose of the Study:

• To develop an integrated diagnostic platform for the rapid detection and differentiation of multiple respiratory viruses.
• To leverage surface-enhanced Raman scattering (SERS) and 3D plasmonic nanopillars for enhanced viral detection.
• To apply interpretability-driven deep learning for accurate virus classification and model transparency.

Main Methods:

• Development of a diagnostic platform utilizing 3D plasmonic nanopillar substrates for enhanced SERS signal acquisition.
• Training a one-dimensional convolutional neural network (1D-CNN) on SERS spectra from 13 respiratory virus types, including SARS-CoV-2 variants.
• Application of gradient-weighted class activation mapping (Grad-CAM) to identify critical Raman shift regions for virus discrimination.

Main Results:

• Achieved over 98% classification accuracy in identifying 13 respiratory virus types using the 1D-CNN model.
• Demonstrated robust and reproducible capture of viral components, enhancing SERS signals for molecular fingerprinting.
• Validated reliable performance in complex clinical samples, confirming real-world applicability.

Conclusions:

• The developed platform provides a scalable, label-free solution for rapid, accurate, and variant-level respiratory virus detection.
• The integration of 3D SERS substrates and deep learning enhances diagnostic capabilities.
• The technology holds potential for point-of-care applications and improved epidemiological surveillance.