Algorithm Refinement in the Non-Invasive Detection of Blood Glucose via Bio-RFID™ Technology¹

Abstract

Diabetes Mellitus (DM) is a highly prevalent and increasingly common disorder that can have dire health consequences if not properly managed. Management of DM involves monitoring of blood glucose levels which can be both cumbersome and invasive, limiting adherence. We present a validation for a novel sensor designed to measure blood glucose (BG) non-invasively using Radio Frequency (RF) waves. In this n=5 study, we trained a Light Gradient-Boosting Machine (lightGBM) model to predict BG values using 1,555 observations from over 130 hours of data collection from 5 participants. An observation is defined as data collected from 13 Bio-RFID sensor sweeps paired with a single Dexcom G6 value. Using this model, we were able to predict BG in the test set with a Mean Absolute Relative Difference (MARD) of 12.7% in the normoglycemic range and 14.0% in the hyperglycemic range. Overall, 70.7% of the estimates fell within 15% of the reference value, and 79.1% fell within 20% of the reference value. While this is a relatively small participant sample, these strong initial results indicate the efficacy of this technique, and that with further refinement and more data, there is promise to achieve a clinically relevant level of accuracy.

Novel data preprocessing techniques in an expanded dataset improve machine learning model accuracy for a non-invasive blood glucose monitor

“…We describe the development of a method that employs novel aspects of data collection utilizing the Bio-RFID sensor and data preprocessing techniques to determine stability and reproducibility of a machine learning model trained on an expanded dataset. With a MARD of 11.27%, we have demonstrated stability of our model and an improvement in model accuracy when compared to previous test results 3 . At this stage, these developments support the use of such techniques as appropriate and feasible for larger scale testing of the validity of RF devices for BG management.…”

“…Similar studies have been conducted in the past with 5 unique individuals of a similar demographic 3 . In the current study, data was collected from 13 healthy adults (4 female, 9 male) aged 24 -61.…”

“…The dataset used in this analysis contained 3,311 observations from over 330 hours of data collected from 13 healthy participants across 110 two-to three-hour tests. In order to minimize noise in the data and to reduce the number of variables passed to the machine learning model, we maintained downsampling techniques in the temporal domain outlined in an earlier analysis on a smaller dataset 3 . In this analysis on an expanded dataset, we represented the frequency domain with more granularity by applying less aggressive feature reduction techniques of the RF spectrum.…”

“…Here we reduced 10,001 spectral features to 167 by dividing each sweep into blocks of 60 frequency values and taking their mean, so that our data was in 6MHz intervals. In the previous analysis we downsampled to 40 spectral features 3 .…”