Leveraging Broad-Spectrum Fluorescence Data and Machine Learning for High-Accuracy Bacterial Species Identification

Rapid and accurate identification of bacterial species is essential for the effective treatment of infectious diseases and suppression of antibiotic-resistant strains. The unique autofluorescence properties of bacterial cells are exploited for rapid and cost-effective identification that is suitable for point-of-care applications. Fluorescence spectroscopy is combined with machine learning to improve the diagnostic accuracy. Good training data for machine learning can be obtained to achieve the same diagnostic accuracy for bacterial species as when each wavelength is measured in detail over a broad spectral width. Experiments were performed testing 14 bacterial strains. The excitation-emission matrix was analyzed, and Bayesian optimization was used to identify the most effective combinations of wavelengths. The results showed that fluorescence spectra using three specific excitation light regions or excitation spectra using two broad fluorescence detection regions could be used as supervised data to realize diagnostic accuracy comparable to that obtained with more complex instruments. The excitation-emission matrices of 10 bacterial species were measured to explore the spectral characteristics required for the diagnosis of bacterial species using autofluorescence and analyzed using Bayesian optimization for acceptable wavelength widths and wavelength combinations to maintain high diagnostic accuracy. Logistic regression analysis of fluorescence spectral data using three specific excitation light regions or excitation spectral data using two broad fluorescence detection regions classified groups with 98% accuracy.image