High-Speed Diagnosis of Bacterial Pathogens at the Single Cell Level by Raman Microspectroscopy with Machine Learning Filters and Denoising Autoencoders

Accurate and rapid identification of infectious bacteria is important in medicine. Raman microspectroscopy holds great promise in performing label-free identification at the single-cell level. However, due to the naturally weak Raman signal, it is a challenge to build extensive databases and achieve both accurate and fast identification. Here, we used signal-to-noise ratio (SNR) as a standard indicator for Raman data quality and performed bacterial identification using 11, 141 single-cell Raman spectra from nine bacterial strains. Subsequently, using two machine learning methods, a simple filter, and a neural network-based denoising autoencoder (DAE), we demonstrated 92% (simple filter using 1 s/cell spectra) and 84% (DAE using 0.1 s/cell spectra) identification accuracy. Our machine learning-aided Raman analysis paves the way for high-speed Raman microspectroscopic clinical diagnostics.