Physics-informed machine learning for robust calibration and physiological validation of wearable electrochemical sweat biosensors for metabolite monitoring

Wearable medical devices represent a most adequately investigated research paradigm inherent to their continuous, non-invasive health monitoring capabilities. Electrochemical biosensors are the basic units in such devices, and they are responsible for reliable biomarker measurements from non-invasive biofluids like sweat. However, such biosensors are often affected by calibration drift, which directly influences the reliability of their results. Biosensor responses are affected by numerous factors like device-dependent parameters, environmental variability, and complex physiological factors. Thus, a reliable calibration demands modeling all these influencing factors. This is a time-consuming process and is highly tedious. However, the introduction of domain-specific relationships while learning enables capturing non-linearities in sensor responses and could improve calibration performance. Thus, this study proposes a calibration framework that could improve calibration accuracy and enable capturing measurement variability under evaluated experimental conditions, considering a wearable sweat metabolite monitoring biosensor. The physics-guided machine learning (PGML) framework harvests the synergistic benefits of extracted electrochemical signal features and physics-driven learning from monotonicity constraints to make meaningful predictions. Experimental results supported the performance efficacy of the proposed PGML framework compared to the conventional linear regression model. Specifically, the proposed framework achieved an R-squared value of ~ 0.7940 and a reduced root mean square error of ~ 0.90. Cross-validation results supported the model’s generalization ability compared to baseline models. Cross-domain analysis between sweat– serum, branched-chain amino acids, and leucine revealed a moderate correlation with R-squared values of ~ 0.48 and ~ 0.57, respectively. Time-series evaluation and normalized trend analysis of measurements from multiple subjects revealed metabolite dynamics such as inter-individual metabolite variability and intra-individual metabolite-specific temporal variations. This demonstrates the model’s ability to provide personalized diagnosis. Thus, the proposed framework based on a physics-driven calibration technique holds the potential to enhance the reliability of wearable sweat analyte monitoring devices.