High-resolution colorimetric detection on paper-based microfluidic devices via indicator merging and machine learning

Colorimetric analysis is being broadly applied in chemical sensing today; however, detection ranges and resolution limits are typically modest. In this paper, we introduce a methodology to quantify the colorimetric chemical response on a paper-based microfluidic device that enables high-resolution colorimetric detection over a broad pH range. We have achieved this by combining data from various indicators displaying sensitivity on partially overlapping small pH ranges and training machine learning classification models to the colorimetric output. The training dataset consists of images taken from the colorimetric response of three different pH indicators previously deposited on circular spots of a multilayer paper-based device, captured with a reference lab-grade camera. Instead of restricting the use of each pH indicator to their linear response regime within the RGB space, the models are trained against data spanning the entire range of pH values, from 3 to 9, in increments of 0.1, exploring the optimum combination of feature engineering and classification model to maximize the overall model accuracy. The combined analysis of image data captured simultaneously with the three indicators resulted in a pH detection accuracy above 85% with over the entire pH range with resolution down to 0.2 pH points. The demonstrated detection range and resolution are well-suited to support various applications in environmental and industrial analysis.