A machine learning-based on-demand sweat glucose reporting platform

Abstract

Diabetes is a chronic endocrine disease that occurs due to an imbalance in glucose levels and altering carbohydrate metabolism. It is a leading cause of morbidity, resulting in a reduced quality of life even in developed societies, primarily affected by a sedentary lifestyle and often leading to mortality. Keeping track of blood glucose levels noninvasively has been made possible due to diverse breakthroughs in wearable sensor technology coupled with holistic digital healthcare. Efficient glucose management has been revolutionized by the development of continuous glucose monitoring sensors and wearable, non/minimally invasive devices that measure glucose concentration by exploiting different physical principles, e.g., glucose oxidase, fluorescence, or skin dielectric properties, and provide real-time measurements every 1-5 min. This paper presents a highly novel and completely non-invasive sweat sensor platform technology that can measure and report glucose concentrations from passively expressed human eccrine sweat using electrochemical impedance spectroscopy and affinity capture probe functionalized sensor surfaces. The sensor samples 1-5 µL of sweat from the wearer every 1-5 min and reports sweat glucose from a machine learning algorithm that samples the analytical reference values from the electrochemical sweat sensor. These values are then converted to continuous time-varying signals using the interpolation methodology. Supervised machine learning, the decision tree regression algorithm, shows the goodness of fit R² of 0.94 was achieved with an RMSE value of 0.1 mg/dL. The output of the model was tested on three human subject datasets. The results were able to capture the glucose progression trend correctly. Sweet sensor platform technology demonstrates a dynamic response over the physiological sweat glucose range of 1-4 mg/dL measured from 3 human subjects. The technology described in the manuscript shows promise for real-time biomarkers such as glucose reporting from passively expressed human eccrine sweat.

Figure 1

(A) Glucose monitoring system overview and its components. (B) Examples of some of the activities that can be performed by wearing a glucose monitoring system.

Figure 2

(A) Exploratory data analysis of input parameters obtained from the sensor Zmod, Zphase, Temperature and %RH. (B) Correlation matrix visualization between the input parameters for the machine learning model.

Figure 3

(A) Concept diagram of interpolation signal generation using reference points, (B) bar plot comparison of R² value obtained from k-fold cross-validation for regression algorithms, (C) bar plot comparison for RMSE value obtained from k-fold cross-validation for regression algorithms.

Figure 4

Effect of Gaussian white noise addition on the distribution of Ytrain with statistics and respective loss functions for the decision tree regression algorithm.

Figure 5

Results obtained on the test dataset for the three human subject datasets plotted concerning actual progression from reference values.

Figure 6

Test and train loss function plotted for every iteration. The last point on the iterations shows the optimized combination of hyperparameters.

Figure 7

System-level block diagram of sensor and technology with cloud integration.

Figure 8

Sensing architecture block diagram of sensor and technology.