Noninvasive Blood Glucose Measurement Using RF Spectroscopy and a LightGBM AI Model
Diabetes mellitus affects over 530 million people globally, leading to severe complications such as cardiovascular disease, blindness, and nerve and kidney damage if not managed. Traditional blood glucose (BG) monitoring methods like finger pricking are invasive, costly, and wasteful, while continuous glucose monitors (CGMs) still involve discomfort and recurring costs.
RF (Radio Frequency) or microwave detection uses changes in blood's dielectric properties caused by BG levels. However, achieving clinical precision remains challenging because of difficulties correlating dielectric properties (such as effective permittivity) with glucose levels due to tissue property variations and the need for noise-reducing signal processing.
The research proposes a novel RF sensor with broadband spectroscopy and a LightGBM (gradient boosting machine) AI model to measure BG across multiple frequencies. It differs from the Cole-Cole model, which struggles with human measurement complexities or traditional methods focusing on resonant frequencies and reflection parameters.
The approach, operating across a frequency range of 100 MHz to over 4000 MHz, uses a linear array of loosely coupled antenna elements to detect BG-related variations in power and resonant frequencies.
The sensor was evaluated using data from five healthy participants (ages 29-61, two females, three males) who consumed 37.5 g of liquid D-glucose. Dexcom G6 CGM readings serve as a BG proxy for sensor characterization.
Participants rested their forearms on a chair integrated with the RF dielectric sensor, wearing a Dexcom G6 CGM after fasting for 90 minutes. BG values were recorded every 5 minutes, with the first 30 minutes establishing baseline levels. After consumption, the levels rose, and testing continued for 3.5 hours.
The RF sensor collected data through 22-second sweeps across the 500–1500 MHz range at 0.1-MHz intervals. Data was preprocessed into 5-minute averages and reduced to 25-MHz intervals, resulting in 1555 observations for model training and testing.
The LightGBM model, trained with MARD (mean absolute relative difference) as the loss function and overfitting penalties, achieved a significantly low 12.9% MARD and MAE (mean absolute error), outperforming empirical chance and accurately predicting BG levels non-invasively as opposed to traditional CGM sensors.
Error grid analysis by Clarke and SEG methods also confirmed the model's predictions were clinically acceptable, with 79.1% of the predicted blood glucose values falling within Zone A.
The sensor's broad frequency range enhanced predictions, surpassing models like linear regression and convolutional neural networks, with the best performance in the normoglycemic range and improved hyperglycemic predictions.
Despite limitations such as the small sample size and insufficient hypoglycemic data, the hardware and software techniques developed by the researchers demonstrate the potential of RF spectroscopy and the LightGBM model in future clinical applications.
Further research should focus on improving accuracy, expanding to diverse populations, and validating the model's performance with broader glucose data and reference device comparisons. Beyond diabetes management, this could enable real-time monitoring and predictive analysis for early detection of glycemic abnormalities.
These efforts will drive model development and work toward FDA (Food and Drug Administration) clearance, making glucose monitoring more accessible, affordable, and comfortable for patients.