Instrumentation, microfluidics and machine learning for rapid, ultrasensitive, and high throughput bioburden detection

Abstract

Thousands of lives are lost every year, and millions are infected due to food, water, ormedicine contamination. The lack of rapid low-cost tools for pathogen identification results in misdiagnoses that worsen the outcome and promote over-prescription of antibiotics as well. As the threat of antibiotic-resistant bacteria continues to rise, there is an increasing demand for diagnostic devices that can quickly identify bacteria and their antibiotic susceptibilities in a wide range of applications. Contamination detection often requires lengthy culturing steps to detect low-level bioburden. To increase the rate of detection and decrease the limit of detection (LOD), in this dissertation, a system featuring microfluidics and a multichannel fluorometer is presented. This low-cost system utilizes the slope of fluorescence intensity that serves as the criterion for bioburden detection. The sample under investigation is spiked with an indicator dye resazurin and loaded in a special-design microfluidic cassette and the rate of change was observed via the fluorometer. We have validated and tested the assay and the device for different laboratory-cultured samples and different real-world surface and water samples. The multichannel fluorometer successfully detects the bioburden with comparable performance with the spectrophotometer and the7 golden standard plate count method within 8 hours of incubation which is much faster than the conventional methods. Microfluidics is a critical and integral part of this application, but the requirement for complex fabrication techniques, cleanroom facilities, and advanced microfabrication equipment hinders the path toward mass production of microfluidic systems. As most of the thermoplastics are not biodegradable, it has resulted in large-scale pollution and an environmental threat. This thesis also aims to find out an easy and simple way of bonding microfluidic chips using green biodegradable thermoplastic material easily adaptable for bioburden detection. We introduce an environment-friendly, low-cost, and safe welding technology used in the fabrication of microcassettes from polymethyl methacrylate (PMMA) and biodegradable cellulose acetate (CA) thermoplastics. Also, microfluidic chips are designed and bonded with low-cost plywood materials. These microfluidic devices are adapted for various applications, including rapid detection of microbial contamination and monitoring vitamin C content which follows the same reduction mechanism as bacterial detection. Additionally, using the novel microwave-induced technique, a novel microfluidic device has been fabricated and bonded for bioburden detection which shows much higher sensitivity than the normal microfluidic cassettes and the standard device spectrophotometer. All biosensors along with the multichannel fluorometer exhibit some irregular signal noise. The accuracy and reliability of most biosensors caused by noise limit their commercialization. In this context, machine learning (ML) can provide novel strategies for overcoming the challenges faced by biosensors. In this thesis, various machine learning algorithms have been implemented on the data collected over different experiments using8 different primary culture cell samples for five different strains. As opposed to previously used only slope criterion for bioburden detection, ML is applied for the automatic qualitative detection of bacteria, prediction of bioburden level of the same strain of bacteria, and identification of different strains of bacteria. We have primarily applied different conventional ML algorithms along with the normal feedforward neural network (NN) and recurrent neural network (RNN) particularly suited for time-series data. Both classification and regression analysis are performed for the identification and prediction of bioburden levels present in the sample from the multichannel fluorometer data. Our classification and regression analysis show promising results in identifying and predicting bioburden strains and levels that have the potential to detect bioburden in real time for different field applications.