QUANTIFICATION OF BACTERIA IN WATER USING FLUORESCENCE SPECTROSCOPY AND MACHINE LEARNING

Rapid detection of bacteria in drinking water is a major public health challenge, with waterborne bacteria causing >40,000 hospitalizations yearly in the US. Most cases are avoidable via early detection of the contamination and prompt treatment or redirection of the system to different sources. The major challenge of early detection of microbial contamination in water is that current methods are slow (24-72 hours), laborious, and require experienced technicians. Newer methods, based on DNA hybridization or enzyme-linked immuno-absorbance assay, pose technical difficulties; require expensive reagents and complicated sample preparation. Conversely, spectral methods enable quick, non-destructive scanning of samples without sample preparation. These techniques use light emission and detection to identify the chemical or biological content of a sample based on light transmission, scatter and fluorescence. In this work, we used fluorescence spectroscopy to quantify bacteria in water. We scanned bacteria-enriched and natural drinking water samples using a spectrofluorometer at various ultra-violet and visible wavelengths. No specific peak that correlates to the concentration of waterborne bacteria was identified from fluorescence data. In order to search for potential latent information, we have applied a machine learning approach and used an iterative algorithm, which takes into account a large number of excitation-emission wavelength combinations. This approach enabled us to successfully quantify total heterotrophic bacteria in water, based on 3D excitation-emission data, in concentrations as low as 100 colony-forming-units per ml. These findings suggest that fluorescence spectroscopy, coupled with a machine-learning algorithm can potentially be applied to quantify bacteria in the drinking water and food industries.