1003-P: A New Machine-Learning Model and Expanded Dataset for a Noninvasive BGM

Introduction: Continuous glucose monitoring (CGM) is essential in diabetes management, but global adoption is hindered due to economic costs and discomfort. A non-invasive, cost-effective, and accurate CGM would support the patient population and increase adoption. This study evaluates the accuracy of a multi-frequency RF sensor for non-invasive blood glucose (BG) monitoring in people with prediabetes and Type 2 diabetes using venous blood as a reference. Methods: Using a sensor that records data from several thousand radio frequencies (RF), participants' forearms were scanned during an Oral Glucose Tolerance Test. From 22 participants, 1,430 venous blood samples were collected using a peripheral intravenous catheter. Using the RF data, a CatBoost machine learning (ML) model was built on 80% of these values to estimate BG as a dependent variable. This model was applied to a held-out test dataset and a Mean Absolute Relative Difference (MARD) was calculated. Results: The CatBoost model returned an overall MARD of 11.8% ± 1.5% on the test dataset. It performed similarly on normoglycemic (12.1% ± 1.8%) and hyperglycemic (11.0% ± 2.3%) ranges. Notably, 100% of predictions fell in Risk Grade A or B in a SEG analysis. Conclusion: The ML techniques applied to data collected by this RF sensor hold promise for the non-invasive measurement of BG. Ongoing studies will include expanding the participant population and continuing model refinement.