Explainable Deep-Learning-Assisted Sweat Assessment via a Programmable Colorimetric Chip.

Abstract

Multianalytes and individual differences of biofluids (such as blood, urine, or sweat) pose enormous complexity and challenges to rapid, facile, high-throughput, and accurate clinical analysis or health assessment. Deep-learning (DL)-assisted image analysis has been demonstrated to be an efficient big data process which shows accurate individual identification. However, the data-driven "black boxes" of current DL algorithms are suffering from the nontransparent inner working mechanism. In this work, we designed a programmable colorimetric chip with explainable DL to approach accurate classification and quantification analysis of sweat samples. Gel (sodium alginate) capsules with different indicators were adopted to combinate as designed programmable colorimetric chips. We collected 4600 colorimetric response images as the data set and assessed two DL algorithms and seven machine learning (ML) algorithms. Glucose, pH, and lactate in human sweat could be facilely and 100% accurately classified and quantified by the convolutional neural network (CNN) DL algorithm, and the testing results of actual sweat via the DL-assisted colorimetric approach match 91.0-99.7% with the laboratory measurements. Class activation mapping (CAM) was processed to visualize the inner working mechanism of CNN operation, which could help to verify and explicate the design rationality of colorimetric chips. The explainable DL-assisted programmable colorimetric chip provided an "end-to-end" strategy to ascertain the black box of the DL algorithm, promoted software design or principium optimization, and contributed facile indicators for clinical monitoring, disease prevention, and even new scientific discoveries.