Raman spectrum combined with deep learning for precise recognition of Carbapenem-resistant Enterobacteriaceae

Area of Science:

• Microbiology
• Spectroscopy
• Artificial Intelligence

Background:

• Carbapenem-resistant Enterobacteriaceae (CRE) present a significant global health threat due to limited effective treatments.
• Rapid and accurate identification of CRE and its resistance mechanisms is crucial for effective clinical management.

Purpose of the Study:

• To develop a rapid, culture-independent method for identifying CRE strains.
• To classify CRE based on enzyme-producing subtypes and predict sensitive antibiotics using SERS and deep learning.

Main Methods:

• Utilized surface-enhanced Raman spectroscopy (SERS) with a gold-silver composite substrate to analyze 22 CRE strains.
• Employed residual networks with an attention mechanism for classifying SERS spectra.
• Evaluated classification accuracy for bacterial type, enzyme-producing subtype, and antibiotic sensitivity.

Main Results:

• SERS spectra were repeatable and consistent, showing species-specific differences related to enzyme-producing subtypes.
• The attention mechanism improved ResNet model accuracy for CRE classification (94.0% for strains, 96.13% for subtypes).
• Achieved 93.9% accuracy in predicting sensitive antibiotic combinations for CRE.

Conclusions:

• SERS combined with deep learning offers a promising approach for rapid CRE identification without culture.
• This method can guide antibiotic selection for CRE infections, aiding in resistance management.
• Demonstrates potential for new clinical diagnostic tools for CRE detection and treatment guidance.