Machine learning-powered wearable interface for distinguishable and predictable sweat sensing

Abstract

The constrained resources on wearable devices pose a challenge in meeting the demands for comprehensive sensing information, and current wearable non-enzymatic sensors face difficulties in achieving specific detection in biofluids. To address this issue, we have developed a highly selective non-enzymatic sweat sensor that seamlessly integrates with machine learning, ensuring reliable sensing and physiological monitoring of sweat biomarkers during exercise. The sensor consists of two electrodes supported by a microsystem that incorporates signal processing and wireless communication. The device generates four explainable features that can be used to accurately predict tyrosine and tryptophan concentrations, as well as sweat pH. The reliability of this device has been validated through rigorous statistical analysis, and its performance has been tested in subjects with and without supplemental amino acid intake during cycling trials. Notably, a robust linear relationship has been identified between tryptophan and tyrosine concentrations in the collected samples, irrespective of the pH dimension. This innovative sensing platform is highly portable and has significant potential to advance the biomedical applications of non-enzymatic sensors. It can markedly improve accuracy while decreasing costs.