Signals From Out of the Blue (Light Wavelengths): Portable, Low-Cost Camera-Based Optical Chemical/Bio-Sensors Utilizing Fluorescence and Machine Learning Techniques

Abstract

Optical chemical/bio-sensors have the capacity to rapidly respond to challenges in our society such as oil spills, viral diseases, and medical conditions affected by bacterial imbalance. These powerful sensors often have many advantages including ease of use, low limits of detection, inexpensive implementation, and fast results. Such features are made possible using widely available cameras (such as the Raspberry Pi camera module and smartphone cameras), fluorescent particles or inherent molecular autofluorescence, and machine learning algorithms such as support vector machine (SVM) and neural networks. Four examples of optical chemical/bio-sensor methods will be demonstrated in this work. First, we created a field-ready, Raspberry Pi-powered autofluorescence sensor for analyzing ocean oil samples using SVM, to assist cleanup efforts after a spill. This device successfully classified oil samples as light fuel (F1 score 95.7%), lubricant (F1 score 100%), or heavy fuel (F1 score 85.7%), and achieved 94% accuracy in classifying the level of asphaltene in a sample. Then, when COVID-19 arrived in the US, we developed two different smartphone biosensor methods for SARS-CoV-2 antigen detection. The first uses a custom-built device for quantifying fluorescent particle immunoagglutination from smartphone images to determine if a saline gargle sample is positive or negative for COVID-19. This device achieved a low limit of detection (LOD) of 10 ag/µL for spiked samples and high performance metrics when tested on clinical saline gargle samples, although it requires some skilled handling of the smartphone microscope attachment. The second device simply requires a smartphone video to analyze the flow rate profile of particles moving along a paper microfluidic channel that is pre-loaded with a saline gargle sample, relying on changes in surface tension during flow to determine if the sample contains SARS-CoV-2. This method has somewhat inferior performance compared to the first, with an accuracy of 89% only when turbid samples are excluded; however, an in-depth analysis of turbid samples revealed that following some simple guidelines may improve the performance of this easy-to-run assay. Finally, we designed a custom-built autofluorescence device that uses smartphone images and a convolutional neural network (CNN) to determine whether or not a bacterial sample contains Staphylococcus aureus, as is common in patients with atopic dermatitis or eczema. This novel device and method could distinguish between “healthy” and “dysbiotic” bacterial images with an F1 score of 86%. These projects highlight the adaptability and usefulness of optical chemical/bio-sensors for shedding blue or UV light on the microscopic elements affecting our daily lives.