Colorimetric detection of glucose with smartphone-coupled μPADs: harnessing machine learning algorithms in variable lighting environments

Recent advances in microfluidic paper-based analytical devices (mu PADs) have shown immense potential for point-of-care testing (POCT) in resource-constrained settings, especially for monitoring glucose levels in patients with diabetes. However, the practical adoption of smartphone-coupled mu PADs has been limited due to the variability in ambient lighting conditions, which affects the colorimetric detection. In this study, we have developed a machine learning-based algorithm together with a flash/no-flash technic, adaptable to various mu PADs and capable of accurately monitoring plasma glucose levels across a wide range (25 mu M to 30 mM), addressing issues related to environmental illumination variations, and shooting angles.The algorithm is implemented in an Android application, Gluco estimator. It utilizes images captured by smartphones and incorporates intelligent feature engineering. The XYZ color space is adopted for device-independent analysis. KI and TMB color indicators are used to enhance resolution for low and high glucose concentrations, respectively. The EBC classifier, combined with a handcrafted feature set, exhibits outstanding performance, achieving an accuracy of 95% and 91% for TMB and KI, respectively. Additionally, multiple linear regression (MLR) yields high reliability among regression models, with R2 values of 0.95 at 0.25 mu M to 3 mM and 0.97 at 3-30 mM. This work represents a significant stride toward realizing a portable, high-resolution, and reliable smartphone-coupled mu PAD in uncontrolled environments.

Automated summary

Recent advances in microfluidic paper-based analytical devices have shown great potential for point-of-care testing, particularly for monitoring glucose levels in diabetes patients. However, the practical application of smartphone-coupled devices has been hindered by variability in ambient lighting conditions. In this study, a machine learning-based algorithm and flash/no-flash technic were developed to address the issue, resulting in accurate monitoring of plasma glucose levels across a wide range and in various lighting conditions.