Precision in Colorimetric Sweat Sensing through Machine Learning

Sweat contains diverse biochemicals that can serve as a source for detecting biomarkers, particularly in individuals with chronic conditions like diabetes. Non-invasive sweat sensors have gained popularity in various applications, promoting personalized healthcare and daily convenience. Different methods, including electrochemical, fluorescence, and colorimetric approaches, have been employed for real-time sweat analysis. Among these methods, colorimetric sensors stand out for their accuracy, stability, and portability in addition to their scalability in production processes. However, the main challenge in colorimetric sweat sensing currently lies in two areas: the need for improved sensor accuracy and the precise detection of color changes resulting from sweat exposure.<br/>In this presentation, we will showcase the development of our sensors, offering real-time, stable, and rapid monitoring of glucose concentration and pH value without causing discomfort. The detection process takes approximately five minutes. Most notably, we address the challenge of detecting subtle color changes, which are virtually imperceptible to the naked eye, by leveraging artificial intelligence. We will discuss how the utilization of a breathable, bio-compatible, acid-base-balanced cotton substrate is the key to achieving higher color differences. In a series of experiments, we cover the fabrication and testing of two generations of pH sensors and two types of glucose sensors. We will then demonstrate the application of machine learning to our sensors and how it significantly enhances prediction accuracy. Three machine learning algorithms, namely Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), and Convolutional Neural Network (CNN) are applied, demonstrating stable and excellent prediction accuracy of 90% for the test results. Finally, we will showcase the practical application of our combined pH and glucose sensors within a single substrate. This work contributes to the advancement of sensor preparation techniques and underscores the critical role of accurate machine-learning algorithms in the successful utilization of sweat sensors across various domains.