Deep Learning-Assisted Surface-Enhanced Raman Scattering for Rapid Bacterial Identification

Area of Science:

• Microbiology
• Spectroscopy
• Artificial Intelligence

Background:

• Bloodstream infections (BSI) require rapid diagnosis for effective antibiotic treatment.
• Conventional methods are slow, delaying critical clinical decisions.
• New diagnostic approaches are needed for timely bacterial identification and antimicrobial susceptibility testing (AST).

Purpose of the Study:

• To develop a rapid diagnostic model integrating surface-enhanced Raman scattering (SERS) with deep learning (DL) for bacterial identification in BSI.
• To utilize Vision Transformer (ViT), a DL model, for analyzing SERS spectra to identify Gram type, species, and antibiotic resistance.
• To assess the model's accuracy and efficiency using clinical blood samples.

Main Methods:

• Collected 11,774 SERS spectra from eight common bacterial species in clinical blood samples.
• Developed a SERS-Deep Learning (SERS-DL) model using Vision Transformer (ViT).
• Employed transfer learning for antibiotic-resistant strain identification, pre-training on Gram-positive species.

Main Results:

• The SERS-DL model achieved 99.30% accuracy for Gram type identification and 97.56% for species identification.
• Demonstrated high accuracy (98.5%) in identifying methicillin-resistant *Staphylococcus aureus* (MRSA) and susceptible strains (MSSA) with a small dataset.
• The model successfully identified bacterial characteristics directly from clinical blood samples without artificial introduction.

Conclusions:

• The developed SERS-DL model offers a rapid and accurate method for identifying bacterial Gram type, species, and antibiotic resistance in BSI.
• This approach has significant potential to guide early and appropriate antibiotic usage, improving patient outcomes.
• The integration of SERS and ViT presents a promising advancement in clinical microbiological diagnostics.