# Real-time estimations of blood glucose concentrations from sweat measurements using the local density random walk model

## TL;DR

This paper introduces a faster model for estimating blood glucose from sweat, enabling real-time monitoring without blood draws.

## Contribution

The Local Density Random Walk model enables efficient real-time glucose estimation from sweat, replacing slower biophysical simulations.

## Key findings

- The LDRW model achieves similar accuracy to the original biophysical model with a root mean square difference of 0.04 mmol/L.
- The LDRW model reduces computational time to 2.6 seconds per data point, 0.7% of the original method's time.
- The LDRW model has a better balance of fit and complexity, as indicated by a lower corrected Akaike Information Criterion value.

## Abstract

Sweat provides a non-invasive alternative to blood draws, enabling glucose-concentration monitoring for both healthy individuals and diabetic patients. In our previous work, we demonstrated a strategy that accurately estimates blood glucose concentrations from sweat measurements. However, this method involves time-consuming simulations using a biophysical model, limiting its application to offline use. The goal of this study is to propose an approach that increases computational efficiency, thereby facilitating real-time estimation of blood glucose concentrations using sweat-sensing technology. To this end, we propose replacing the original biophysical model with the Local Density Random Walk (LDRW) model. This is justified because both models describe the pharmacokinetics of glucose transport through a convective-diffusion process. The performance of the LDRW model and the original biophysical model are compared in terms of estimation accuracy, computational efficiency, and model complexity, using seven datasets from the literature. The estimation of blood glucose concentrations using the LDRW model closely approximates that of the original model, with a root mean square difference of just 0.04 mmol/L between the two models' estimates. Remarkably, the LDRW model significantly reduces the average computational time to 2.6 s per data point, representing only 0.7% of the time required by the original method. Furthermore, the LDRW model demonstrates a smaller corrected Akaike Information Criterion value than the original method, indicating an improved balance between goodness of fit and model complexity. The proposed novel approach paves the way for the clinical adoption of sweat-sensing technology for non-invasive, real-time monitoring of diabetes.