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

Abstract

Bloodstream infection (BSI) is characterized by the presence of viable microorganisms in the bloodstream and may induce systemic immune responses. Early and appropriate antibiotic usage is crucial to effectively treating BSI. However, conventional culture-based microbiological diagnostics are timeconsuming and cannot provide timely bacterial identification for subsequent antimicrobial susceptibility test (AST) and clinical decision-making. To address this issue, modern microbiological diagnostics have been developed, such as surface-enhanced Raman scattering (SERS), which is a sensitive, label-free, and quick bacterial detection method measuring specific bacterial metabolites. In this study, we aim to integrate a new deep learning (DL) method, Vision Transformer (ViT), with bacterial SERS spectral analysis to build the SERS-DL model for rapid identification of Gram type, species, and resistant strains. To demonstrate the feasibility of our approach, we used 11,774 SERS spectra obtained directly from eight common bacterial species in clinical blood samples without artificial introduction as the training dataset for the SERS-DL model. Our results showed that ViT achieved excellent identification accuracy of 99.30% for Gram type and 97.56% for species. Moreover, we employed transfer learning by using the Gram-positive species identifier as a pre-trained model to perform the antibiotic-resistant strain task. The identification accuracy of methicillin-resistant and -susceptible Staphylococcus aureus (MRSA and MSSA) can reach 98.5% with only 200-dataset requirement. In summary, our SERS-DL model has great potential to provide a quick clinical reference to determine the bacterial Gram type, species, and even resistant strains, which can guide early antibiotic usage in BSI.

Artificial Intelligence for Surface‐Enhanced Raman Spectroscopy

“…When encountering insufficient training data, transfer learning is a desirable strategy, [ 301 ] which has exhibited its effectiveness in SERS domain. [ 128,129,136,302 ] For example, Tseng et al. reused the Gram‐positive species classifier (pretrained with ≈4k data) to perform antibiotic‐resistant strain identification with 200 data only, which achieved a high accuracy of 98.5%.…”

“…reused the Gram‐positive species classifier (pretrained with ≈4k data) to perform antibiotic‐resistant strain identification with 200 data only, which achieved a high accuracy of 98.5%. [ 302 ] The main idea is to transfer knowledge from task A to task B. The pre‐trained AI model for task A with abundant data serves as a starting point for task B and is then fine‐tuned to improve its performance with data available in task B.…”

Patch-Based Convolutional Encoder: A Deep Learning Algorithm for Spectral Classification Balancing the Local and Global Information

“…In particular, a transformer-based algorithm incorporates a self-attention mechanism and global receptive field 19 and allows the algorithm to identify dependencies between each part of the spectra in parallel. 20,21 However, both LSTM and transformer-based algorithms 22 show poor performance for the first task (Table 1). Derived from AlterNet, 23 ConvMSANet 24 fused the local and global information by combining 1D CNN and multihead selfattention (MSA) mechanism.…”

“…For example, the long short-term memory (LSTM) can comprehensively consider the whole Raman spectrum of different marine pathogens and achieve an accuracy of 94%, which is significantly better than that of CNN (89%). In particular, a transformer-based algorithm incorporates a self-attention mechanism and global receptive field and allows the algorithm to identify dependencies between each part of the spectra in parallel. , However, both LSTM and transformer-based algorithms show poor performance for the first task (Table ). Derived from AlterNet, ConvMSANet fused the local and global information by combining 1D CNN and multihead self-attention (MSA) mechanism.…”

Surface-Enhanced Raman Scattering-Based Surface Chemotaxonomy: Combining Bacteria Extracellular Matrices and Machine Learning for Rapid and Universal Species Identification

“…Existing SERS platforms primarily probe the bacterial cell wall. This approach has its limitations because it can be overly ideal and can only be derived in controlled, isolated laboratory conditions, thus not accurately representing bacteria in their natural habitats. − In their natural habitats, bacteria are often encapsulated by a complex extracellular matrix (ECM) comprising exopolysaccharides, proteins, and DNA. This matrix forms around and between individual cells in response to its native environment.…”