Machine-Learning-Based Colorimetric Sensor on Smartphone for Salivary Uric Acid Detection

Background & aim: Point-of-care sensors with colorimetric signal readout are attractive in providing efficient detection outcomes, but normally with qualitative results. Quantitative information is vital for deriving accurate and sensitive diagnosis towards precise medicine. Herein we developed a generic methodology for quantifying optical intensity on a microfluidic paper test strip to detect salivary uric acid using a machine learning-based colorimetric sensor on a smartphone. Methods: A colorimetric sensor adaptable to the smartphone was designed to image the paper test trip of uric acid. The complete algorithm associated with the sensor consists of four modules: the region of interest detection module for locating key areas in the image, the color calibration module for excluding interference from different lights, the feature extraction module that extracts multidimensional features from the reaction area on the test strip, and the feature analysis module built with machine learning models to provide prediction base on extracted features. Results: The performance of our algorithm was evaluated for quantifying uric acid in both artificial saliva and clinically collected saliva by the machine learning-based colorimetric sensor. The best-performing machine learning model, the decision tree model achieves a mean absolute error of 4.2 ppm on artificial saliva samples. For the clinical samples, we perform correlation analysis between predicted salivary uric acid concentration and actual blood uric acid level. Our method for the detection of salivary uric acid achieves an r-score of 0.6140 and a p-value < 0.0001, comparing the commercially available test trips for detection of uric acid in finger-prick blood.