Paper-based SERS chip with adaptive attention neural network for pathogen identification.

High-speed and accuracy identification of pathogens has become increasingly critical in both individual patient care and public health. Artificial intelligence (AI)-assisted surface-enhanced Raman scattering (SERS) biosensors enable simultaneous identification of multiple pathogens. However, there are still problems such as low accuracy and limited diversity in bacterial fingerprints. To this end, we present a novel multi-branch adaptive attention convolutional neural network (MBAA-CNN)-assisted paper-based SERS chip for prompt and reliable pathogen discrimination. In the approach, we employed a dual-function molecule 4-mercaptophenylboronic acid (4-MPBA) to capture bacteria and enhance Raman spectra diversity, referring as 4-MPBA labeled mode (label mode). Meanwhile, we utilized the K-means algorithm to identify pathogens in the label mode, producing much higher accuracy compared to label-free mode (n = 2000). Furthermore, we acquired 98.6 % accuracy at all pathogen species and 99.5 % accuracy at the antibiotic-resistant and sensitive strains (n = 10,000) using MBAA-CNN. The superior performance of MBAA-CNN was further validated through comparisons with traditional machine learning models, particularly in terms of loss value, speed and accuracy. We envision the developed approach has potential for early culture-free diagnosis of pathogens and real-time monitoring of microbial contamination in water environment.