Machine-learning models for the detection of bacterial contamination in the food industry

In this project, a system was developed for rapid microbiological quality control in the food industry. With the help of Raman spectroscopy, which uses light to recognise the chemical composition of bacteria, and machine learning, bacteria such as E. coli and salmonella are identified in just a few hours. In the future, sample preparation is to be further simplified for easier integration into production.

This project focused on developing a system for microbiological quality control in the food industry using Raman spectroscopy and machine learning. The objective was to achieve rapid detection and species-level identification of bacterial contaminants like E. coli, Salmonella, and Listeria, offering an alternative to the current lengthy tests, reducing the time to results from a few days down to a few hours.

To achieve our goals, extensive data sets were acquired on model organisms (E. coli and Pseudomonas strains, as a proof-of-concept) to train, validate and test the model in various environmental parameters. A convolutional neural network (CNN) machine learning model was developed and achieved excellent classification performance in randomized testing with a >98% accuracy. This model distinguishes between different bacterial species, and even specific strains, which can provide information relevant to contamination tracking and risk assessment.

Surface-Enhanced Raman Spectroscopy (SERS) was also employed, using gold nanoparticles to increase signal intensity and reduce measurement times to milliseconds per cell. Additionally, a microfluidic chip prototype was developed to facilitate sample preparation in view of industrial applications. Both aspects further increased the sensitivity and the identification speed of our solution.  

In addition, key connections were created with industrial stakeholders to ensure that our development was directly aimed at solving existing pain points. These connections involved several large Swiss food producers, including Migros Industrie and Eisberg and continue after the completion of this grant.

In summary, the project produced: a CNN machine learning model for bacterial strain identification using Raman spectroscopy; a supporting spectral database of tested bacteria and a microfluidic chip prototype for rapid sample preparation. This work provides a proof-of-concept for a rapid quality control tool enhanced with machine learning capabilities. Future implementations of our solution could potentially reduce quality control times, associated cost savings, decrease food waste, and enhance food safety, contributing to the capabilities and high standards of the Swiss food industry.

In future work, we will continue to refine our approach to make it more portable and further simplify sample preparation steps to ensure easy integration in production settings.