Artificial intelligence enabled rapid low-cost in-situ fluorescence sensor platform and excitation-emission matrices for faecal contamination detection in water

Abstract

Waterborne diseases such as Diarrhoea give rise to major concerns across humanity, as this disease alone can kill approximately 1.6 million people yearly. The knowledge of the microbial quality of water is therefore critical, which is generally achieved by detecting microbial contamination using E. coli through traditional methods such as plate counting or membrane filtration techniques as well as molecular, nucleic, or enzyme-based technologies. These technologies still require expertise, are time- consuming, and have portability issues. Hence, a low-cost, portable solution to detect faecal coliforms like E. coli quickly or in real-time is highly desirable.

TLF (Tryptophan-like Fluorescence), or natural fluorescence detection of bacteria, is an emerging method for low-cost and portable devices. However, the current TLF devices are either expensive or show a low-sensitivity for E. coli detection due to the use of a single-emission wavelength.

In this work, a novel, portable and low-cost LED-based device was developed that deploys feature selection and Machine Learning to predict E. coli detection in pure and river water. The UV LEDs utilised provide single excitation, and a mini
spectrometer detects a range of emissions as data features provide a detection limit of 1.29 μg/L of L-tryptophan in ultra-pure water.

The analysis between single feature vs multi-feature from the device is compared. Alternatively, a dataset with additional features, EEMs (Excitation Emission Matrices), was extracted from a commercial benchtop spectrometer, Deep Learning (CNN and BLSTM) was applied, and the results were compared to that of the developed device. The developed device’s multi-emission output in conjunction with ANNs provides better performance than using a single feature (emission) on a single-excitation for detecting lab-grown E. coli log10 (CFU/mL) in pure water with R2 of 0.731. For the multi-categorical classification of the WHO risk ratings, the multi-feature Machine Learning method (KNN) was more accurate (74.2%) than single-emission data, whilst another KNN model was able to detect Medium and below WHO risk (Accuracy of 87.1%). CNN with 2D EEMs data provided 82.3% overall accuracy while 1D data BLSTM had an accuracy of 88.7% when classified 'Low and Very Low combined' risk categories from 'Medium, High, and Very High combined' risk categories.

When classifying higher risk categories, the developed low-cost device coupled with traditional ML provided a similar accuracy of 87.1% as compared to the expensive benchtop fluorescence spectrometer coupled with DL (88.7%). In the future, this device could be used as a low-cost field-portable device capable of detecting various bacteria as per WHO guidelines.

Date of Award	Oct 2023
Original language	English
Sponsors	Global Challenges Research Fund (GCRF) UK Research and Innovation, EPSRC Grant Reference EP/P032427/1
Supervisor	Dewar Finlay (Supervisor), Mickey Keenan (Supervisor) & Jim McLaughlin (Supervisor)