Integrated machine-learning-assisted wearable platform for stress state assessment based on sweat cortisol and multimodal physiological biosensing.

Abstract

External stimuli can induce stress-related physiological states characterized by rapid autonomic and cardiovascular fluctuations and slower endocrine-related biochemical responses. To address this challenge, we developed an integrated wearable biosensor platform coupled with machine-learning-assisted interpretation for rest versus stress assessment, with labels defined by baseline and task phases. At the sensing level, a stepwise-modified electrochemical interface enabled sensitive detection of cortisol, thereby providing low-frequency endocrine-related biochemical context. Simultaneously, the flexible wearable circuit continuously monitored heart rate, blood volume pulse, galvanic skin response, and skin temperature, enabling continuous physiological monitoring of rapid autonomic and peripheral dynamics. To integrate these heterogeneous signals, we employed an Inception-MABFDNN model that fuses windowed physiological signals with low-frequency periodic cortisol measurements through modality-specific attention mechanisms. The single-use immunosensor exhibited a log-linear calibration range from 1 pg mL-1 to 1 μg mL-1 and a detection limit of 0.68 pg mL-1. In an independent test cohort of 10 participants, the machine-learning model achieved a classification accuracy of 94.75 %. Paired enzyme-linked immunosorbent assay measurements of human sweat showed good agreement with the immunosensor-derived cortisol concentrations. These results demonstrate the potential of next-generation wearable biosensing systems to bridge biochemical and physiological information across temporal scales through multimodal fusion for more robust assessment of experimentally induced stress-related states.