Machine learning guided electrochemical sensor for passive sweat cortisol detection

Abstract

A noninvasive sensor was developed on a flexible, nano-porous substrate with the capability to detect cortisol passively from human sweat. A Machine Learning (ML) algorithm was also developed to analyze sensor data to detect rising and falling trends of cortisol with time, which can potentially inform the wearer of rising or falling cortisol levels. Sensor response was measured by Electrochemical Impedance Spectroscopy (EIS) assay with spiked synthetic sweat within the physiological concentration range of cortisol, for which responses to low, medium, and high concentrations were significant. Calibration studies established a dynamic range of 8–140 ng/ml and EIS demonstrated a dose-dependent response and operational frequency range of 100–500 Hz. Inter-assay and intra-assay variations were below 20%. One-way ANOVA and post hoc Tukey test established significant response (p < 0.05) to different concentrations of cortisol. EIS was performed with dosing ranging from high-low and low-high concentrations of cortisol, to simulate rising and falling cortisol levels over time. Rate of change of response to shifts in cortisol concentration was used to train a weighted K nearest neighbor (KNN) machine learning algorithm to detect and classify increasing and decreasing cortisol in sweat. Accuracy was validated to be 100% by k means cross validation. Sensor analysis of sweat samples from human subjects showed a 100% accuracy in testing with the algorithm along with a 100% True positive and 0% False negative rate, thus showing that a wearable sweat sensor can be used with ML to detect rising and falling trends in cortisol for on demand, noninvasive sensing.

Keywords

• Cortisol
• Electrochemical sensing
• Machine learning
• Electrochemical impedance spectroscopy