Non-invasive Blood Glucose Monitoring Using a Radiofrequency Sensor and a Machine Learning Model

Abstract

Diabetes Mellitus (DM) affects over 530 million people globally and is a condition characterized by high blood glucose (BG) that can result in severe long-term health consequences if poorly managed. To mitigate the risks associated with DM, it is crucial to regulate BG levels through regular monitoring. Current methods of monitoring BG come with drawbacks in the form of pain and expense. With the rapid increase in prevalence of DM, there is an increasing need for economical, accurate, and non-invasive continuous measures of BG. Here we present a validation for a novel sensor that rapidly scans through a wide range of radio frequencies (RF) and uses machine learning (ML) techniques for the purpose of non-invasive BG measurement. After model training, the RF sensor and ML techniques employed in this study can predict BG at a level significantly better than chance, as shown below. Data were collected from 13 healthy participants over a series of tests lasting 2 – 3 hours. During each test, participants placed their forearm on an RF sensor that measured their BG levels using sweeps across the 500 MHz – 1500 MHz range at 0.1 MHz intervals. Participants ingested 37.5 grams of glucose solution to generate BG readings across normoglycemic and hyperglycemic ranges. Each sweep took approximately 22 seconds, including a one second pause between sweeps. Across 110 tests, we collected 3,311 BG observations. Concurrent measurements from a Continuous Glucose Monitor (CGM) were taken as reference. Data were divided using a 60-20-20 (training-validation-test) split and trained on a Light Gradient-Boosting Machine (lightGBM) model to predict BG values. A ‘blind’ evaluation of model performance was determined using the held-out test dataset. We predicted BG values in the held-out test dataset with a Mean Absolute Relative Difference (MARD) of 10.8% in the normoglycemic range and 15.9% in the hyperglycemic range (11.27% overall MARD). These results were significantly better than a chance model (two-sample t-test, p <0.01). In a Surveillance Error Grid analysis of model accuracy, 89.0% of predictions fell in Risk Level 0, 10.1% in Risk Level 1, and 0.9% in Risk Level 2. These results demonstrate that the novel RF sensor and ML techniques described can predict a reference BG value in the population under study. More research is underway to refine and expand these methods using a gold-standard blood glucose reference device and among a broader participant population with an expanded glucose range. DK and VS are consultants for and own stock in Know Labs. JA and KC are employed by and have stock options in Know Labs. CW and KP are consultants for Know Labs. NA. This is the full abstract presented at the American Physiology Summit 2024 meeting and is only available in HTML format. There are no additional versions or additional content available for this abstract. Physiology was not involved in the peer review process.