Infectious diseases caused by foodborne pathogenic bacteria are always one of the most severe public health problems. Accurate detection of pathogenic microorganisms in food is necessary to guarantee food safety and to contain bacterial infection. Microbial culture-based methods and biochemical tests are still the golden standard in bacterial detection; however, these methods are time-consuming, taking about 2-3 days to carry out, and follow more than ten operation steps. In addition, new diagnostic technologies, such as conventional polymerase chain reaction, mass spectrometry, and DNA sequencing, suffer from many disadvantages including long processing time, laborious operation steps, limited sensitivity, and high cost; thus, they still cannot meet the requirements for clinical diagnosis and point-of-care testing. In recent years, bacterial detection methods based on surface enhanced Raman scattering (SERS) have achieved significant success and performed excellently on high-sensitivity, easy-to-operate, and fingerprint-based detection methods. In this paper, four major foodborne bacteria, namely, Staphylococcus aureus (S. aureus), Escherichia coli (E. coli), Vibrio parahaemolyticus (V. parahaemolyticus), and Listeria monocytogenes (L. monocytogenes), are used as research objects. Furthermore, a novel SERS method, which combines positively charged Ag nanoparticles (AgNPs⁺ ) and convolutional neural networks (CNN), is proposed in this paper for accurate and rapid detection of the above four bacteria.

Clinical isolates including 10 strains from each of S. aureus, E. coli, V. parahaemolyticus, and L. monocytogenes are collected from the laboratory department of the Affiliated Hospital of Xuzhou Medical University. First, AgNPs⁺ are prepared via reduction method of NaBH₄ and are fabricated in a buffer solution as substrate for SERS. Then, AgNPs@bacteria complexes are formed via electrostatic interactions, and high-quality SERS signals in a shift range of 400-1800 cm^-1 of pathogenic bacteria are measured from the forming complexes. Finally, a residual network consisting of 11 one-dimensional convolutional layers (ResNet11) is established and trained on these signals as the spectral classifier. In the spectral identification process, while a SERS spectrum collected from unknown samples is inputted to the trained classifier, the classification probability corresponding to the above four bacteria is calculated, and the label of the maximum value is taken as the predicted label. Based on this strategy, the accurate and precise laboratory testing of bacteria is realized by a high-performance optical analysis technique.

It can be observed from the transmission electron microscope images that AgNPs⁺ are closely binding onto the cell walls of S. aureus and E. coli in mixed solution. Zeta potential measurement results of AgNPs⁺ and four bacteria represent the mechanism of the closely combined strong electrostatic attraction between bacteria and AgNPs⁺ . The spectral measurement results of four types of AgNPs@bacteria complex show that AgNPs⁺ are an excellent SERS substrate. Mainly in bands of 624 cm^-1, 730 cm^-1, etc., obvious Raman peaks of the four pathogens with strong intensity are enhanced. By comparing and identifying the functional groups corresponding to the main Raman peaks in SERS fingerprint spectrum, it is confirmed that the SERS measurement results are consistent with the reported literature. In addition, the average relative standard deviation of SERS measurements of ten times is about 7%, which presents well reproducibility. Meanwhile, the differences between SERS spectra from the four bacteria are studied and identified. For these four types of approximate spectra, the trained classifier ResNet11 achieves average accuracies of 99.30% for SERS fingerprints, with bacteria solution molecular concentration of 10⁷ mL^-1, and also achieved average accuracies of 98.00% for low-intensity SERS fingerprints with the molecular concentration of 10³ mL^-1. ResNet11 performs better accuracy and stability compared with other commonly used classification methods, such as Logistic regression, SVM, random forest, and KNN. Furthermore, ResNet11 may promote the practical application level of SERS technology.

SERS technology can promise sensitive and label-free detection for bacteria and antibiotic susceptibility testing in single steps. However, achieving practical application remains challenging due to the unstable signal intensity and similar spectral curve. In addition, due to the large size of bacteria (0.5-10 μm), which is far beyond the nano gap that can produce stable SERS hot spots, the SERS enhancement effect of bacteria has always at a lower level. There are two ways to improve the quality of bacterial SERS signal: 1) by improving the stability of the substrate in the experiment; 2) by exploring the excellent combination between the substrate and bacteria. In this paper, AgNPs⁺ are prepared as SERS sensing substrate. The close and dense combination with bacteria that are negatively charged in solution can provide high-quality signal with high intensity and good reproducibility. On the other hand, novel algorithm of CNN based on residual structure can guarantee the identification accuracy of SERS spectra. It can reduce the opposite impact caused by spectral quality indirectly and can promote the practicability of SERS in bacterial detection. In summary, we propose an accurate and sensitive method for foodborne bacteria detection by using AgNPs⁺ and CNN. Considering the other advantages including good stability and easy operability, we believe that our approach will become a promising tool for bacterial detection in the laboratory.