Noninvasive Blood Glucose Measurement Using RF Spectroscopy and a LightGBM AI Model

We present a validation for a novel sensor and data processing pipeline designed to measure blood glucose (BG) noninvasively using the rapid collection of a broad range of radio frequency (RF) waves via a decoupled antenna array. Five healthy human subjects ingested 37.5 g of glucose solution to generate BG readings across two glycemic ranges: normoglycemic and hyperglycemic. Concurrent measurements from a continuous glucose monitor (CGM) and the RF sensor were collected for comparative analysis. A light gradient-boosting machine (LightGBM) model was trained to predict BG values using 1555 observations, where an observation is defined as data collected from 13 RF sensor sweeps paired with a single Dexcom G6 CGM value. Using this model, we predicted BG in the held-out test dataset with a mean absolute relative difference (MARD) of 12.7% in the normoglycemic range and 14.0% in the hyperglycemic range. While in early-stage validation, these results demonstrate the promise of this hardware and software technique for the noninvasive measurement of BG for practical application.


I. INTRODUCTION
D IABETES mellitus (DM) affects over 530 million indi- viduals globally [1].DM is a condition characterized by high blood glucose (BG) and, if unmanaged, can result in severe long-term health consequences, such as cardiovascular diseases [2], nerve damage (neuropathy), kidney damage (nephropathy), and eye disease resulting in visual loss or even blindness [3].To mitigate the risks associated with DM, it is crucial to regulate BG levels through regular monitoring, lifestyle modifications, oral medications, and/or insulin injections [4].To monitor BG, patients often rely on invasive portable measurement devices that require frequent finger pricking and disposable test strips, which are painful, and generate both ongoing expenses and biomedical waste.In some cases, patients may need up to ten BG measurements per day, which can be exceedingly uncomfortable and interfere with manual tasks, such as typing, craftsmanship, or artistry.
Modern less invasive and continuous glucose monitors (CGMs) exist [5], [6], [7], although these devices are not without limitations and come with the cost of regular replacement and the discomfort of probe insertion.Nevertheless, monitoring BG regularly is fundamental to effective treatment.To date, there is no solution providing noninvasive, inexpensive, and reliable point-of-care measurement for BG regulation.
One technique for the noninvasive monitoring of BG that has shown promise involves radio frequency (RF)/microwave detection [8], [9].This technique relies on the fact that BG affects the dielectric properties of the blood, which in turn changes the way that microwave electromagnetic fields behave as they interact with it [10].Most studies applying such techniques have used devices that operate at a single frequency in the 1-10 GHz range [11], [12], [13], [14], [15], [16], [17], though some researchers have reported using a range of different frequencies [18].Results from those studies have demonstrated that there is a merit in this approach, though clinical precision has been elusive [19].
The primary difficulty of using microwaves to sense BG is establishing the correlation between effective permittivity and BG [20].The hurdles in understanding this correlation are 1) sensitivity and 2) selectivity [19].The changes in the dielectric properties of tissues because of changes in BG are extremely small, so creating a sensor capable of measuring these minute changes in living subjects has proven to be a challenge, because alongside sensitivity comes susceptibility to both noise and interference [11].Similarly, selectivity is a problem, because the RF waves must penetrate skin, fat, and vascular layers-all of which have different dielectric properties-before reaching blood, and therefore, signal postprocessing is necessary to reduce interference from the waves being influenced by these other tissues.
Most of the studies designed to predict BG seek to identify one or a small number of "resonant" frequencies and values of S11 parameters.This study investigates the efficacy of a novel RF sensing device that applies broadband spectroscopyrapidly scanning a large range of RF frequencies and recording voltage values detected at each frequency-to quantify BG continuously.A proof of principle study demonstrated the ability of a similar RF antenna architecture to use S21 parameters for the in vitro quantification of several molecular solutes in water [21].In this work, we collect data from the RF sensor while also collecting CGM data from five healthy participants, whose BG levels are affected by the ingestion of a glucose solution.To support our goal of rapidly characterizing our sensor with data that can be collected quickly and with minimal burden on participants, we use readings from a Dexcom G6, 1 a popular CGM, as a proxy for BG.The sensor's ability to predict BG is measured using the mean absolute relative difference (MARD)the standard accuracy metric for glucose measurement studies.
The final way in which this work differs from other studies concerns the model used.One approach found in the literature attempts to model the complex dielectric constant of a solution from Cole-Cole model [22].This method works particularly well for a spiked electrolyte with an otherwise unchanging composition except for the analyte concentration but faces significant difficulties in the context of the complex set of variables inherent in measuring humans.To this end, we employ machine learning techniques.Using a light gradient-boosting machine (LightGBM) model, we demonstrate an overall MARD of 12.9% relative to a CGM, providing an early indication that the unique hardware and machine learning model used in this study are the promising techniques for testing on a larger, more diverse population to ultimately determine viability for clinical application.

A. Overview of Participants
The study was approved by Core Human Factors IRB [IORG0007854; IRB Registration # IRB00009432], and all participants provided verbal and written informed consent to participate in the study.Five healthy adults (two females and three males) aged 29-61 participated in the study.Participants had no clinical history or diagnosis of diabetes or other 1 Registered trademark.significant medical conditions that could interfere with data collection or BG more broadly.

B. Data Collection
Data were collected from each participant once each day in a research laboratory located in Seattle, WA, USA.The research laboratory used for this study contains two testing rooms, each with a testing armchair equipped with a patented Know Labs, Inc., RF dielectric sensor built into each arm of the chair, allowing participants to simply rest their arms on the chair for the duration of the test.
1) RF Sensor: The sensor employed in this study was the patented Know Labs, Inc., RF dielectric sensor, which consists of a printed circuit board assembly (PCBA) that generates RF signals and measures received power after passing those signals through an antenna array.Fig. 1 shows a block diagram of the sensor and antenna array.Although not all available frequencies were used in this study, the sensor can generate RF signals that can range from just over 100 MHz to just over 4000 MHz.There is a transmit (Tx) amplifier to boost the signal, and then, the RF signal is routed through a switch matrix that allows for it to be sent to any one of the four supported antenna elements, or optionally through an onboard fixed-attenuation path called the calibration path that allows the system to test itself and provide a known benchmark.
The same switch matrix also establishes the receive (Rx) path, where one of the four supported antenna elements (or the calibration path) is chosen to receive the transmitted signal.Once through the switch matrix, the signal is amplified with a low-noise amplifier (LNA) to set it in the appropriate range of the power measurement circuitry, which translates the received RF signal's power into a voltage output that can be sampled by the onboard analog to digital converter (ADC).
These digitized values are accumulated in the firmware of the microcontroller unit (MCU), which is responsible not only for data collection but also controlling the whole system.The MCU accumulates samples during a dwell, which is defined as the period of time that the signal generator is active at a specific frequency, and then averages those samples to provide a single power measurement value per dwell over the device's USB connector.
2) Characteristics of the Sensor: The initial sensor antenna design used a single antenna element or a pair of transmit and receive elements, with design topologies derived from common narrow-and broadband radiating structures (loops, monopoles, patches, and spirals).Through experimentation, the system's best results were associated with "loosely coupled" structures.These are structures that, in air, had a nominal average value of transmission coefficient that was not "too strong" due to large fringing or near fields, which made them highly susceptible to dielectric loading effects when exposed to a possible analyte.More designs were created based on this principle with the aid of finite element modeling (FEM) simulation tools and subjected to fast design/test/verify cycles with the use of rapid prototyping equipment.
This work led to the current instantiation of the Know Labs, Inc., RF dielectric sensor, which uses an array of antenna elements.These elements are not designed to radiate efficiently into free space, nor are they designed for any specific resonant frequency, though the system does display several resonant modes.Instead, they are intended to provide a mechanism for RF fields to be capacitively coupled from a selected transmit element to a selected receive element.The elements are spaced in a linear array, and the selection of transmit and receive elements depends on the relative element positioning that works best for the target analyte under test.
The RF PCBA measures the power received by the selected element of the antenna array, which is a proxy for the antenna array's transmission coefficient, or S21 in scattering parameter terms, over its frequencies of operation.For the antenna array, the measured pseudo-S21 response varies widely depending on the specific analyte's dielectric properties.This results in a pair of effects that we can observe in the pseudo-S21 measurement over frequency; namely, the aforementioned resonant frequencies move to new frequencies, and the overall received power also varies.The sample plots of the pseudo-S21 signal received for differing BG levels in a single user are shown in Fig. 2.

C. Study Protocol
Roughly 24 h before the start of testing, participants were fitted with a Dexcom G6 on the posterior of the upper left arm.Participants were asked to fast for at least 90 min before testing.During data collection, participants sat in the testing chair with the left forearm placed on the antenna of the sensor.Participants made efforts to minimize body movement for the duration of the test and were seated 10 min prior to the start of the test.Glucose values were recorded from the Dexcom G6 every 5 min (the sampling rate of the reference device) for the duration of the test.The Dexcom G6 values collected during the first 30 min of the test were considered to be the individual's "baseline" BG.At 30 min, the participant consumed 37.5 g of liquid D-Glucose (Azer Scientific Glucose Drink #10-LL-075), which caused their BG to rise.Testing continued until the participant's BG returned to baseline for 30 min or until a maximum test time of 3.5 h had passed.
Data were collected continuously from the RF dielectric sensor using sweeps across the 500-1500 MHz range at 0.1-MHz intervals, so each sweep collected data on 10 001 frequencies.Each sweep took approximately 22 s, including a 1-s pause between sweeps.

D. Data Preprocessing
The dataset used in this analysis contained 1555 Dexcom G6 values collected at 5-min intervals and 22 615 RF dielectric sensor sweeps collected at 22-s intervals.In order to minimize noise in the data and to reduce the number of variables passed to the machine learning model, the RF dielectric sensor data were grouped in two ways.We averaged values in the temporal domain by taking the mean of the 5 min of data (consisting of 13 frequency sweeps) leading up to a Dexcom G6 measurement.We also reduced features in the frequency domain by taking the mean of each set of 250 consecutive frequencies, so that the model received data in 25-MHz intervals rather than the 0.1-MHz intervals of the original data.
After preprocessing, the final dataset used in model development contained 1555 observations, each comprising data from 13 RF dielectric sensor sweeps paired with a single Dexcom G6 value.According to the reference device, participants' BG during the tests ranged from 65 to 278 mg/dL (3.61-15.4mmol/L), with 88.1% of values in the normoglycemic range (defined as 70-180 mg/dL or 3.89-10 mmol/L), 11.6% in the hyperglycemic range (over 180 mg/dL or 10 mmol/L), and 0.3% in the hypoglycemic range (under 70 mg/dL or 3.89 mmol/L).

E. Model Architecture and Training
We employed a LightGBM model to the predict values of the Dexcom G6 using the RF dielectric sensor data.The model was implemented in the LightGBM package [23] (version 3.3.5) in Python (version 3.10.11).Importantly, LightGBM models make very few assumptions about the structure of the data.This lack of assumptions avoids trying to model the dielectric properties of human tissue from first principles (an unsolved problem) and instead allows significant flexibility to learn the relationship between frequency response and BG.The LightGBM model was trained on the training dataset with MARD [see (1) and ( 2)] as the loss function.L1, L2, and feature fraction penalties were applied to limit overfitting.Hyperparameter tuning was conducted on the penalty terms, and the lowest MARD achieved was taken for the final model, resulting at L1 = 0.4, L2 = 0.4, and feature fraction = 0.5.A sample tree taken from the resulting LightGBM model is shown in Fig. 3.
All hyperparameter tuning was conducted on the training dataset created using an 80-20 training-test split.All reported performance metrics were calculated on the test dataset performance.The training and test datasets were stratified by participant (n = 5), RF dielectric sensor (two devices), and glycemic status using scikit-learn's "train_test_split" function.That is, the 80% of data were randomly selected for the training dataset, with the constraint that it contained equal representation from each participant, device, and glycemic status.

F. Performance Metrics
We evaluated the model's performance using the MARD.The absolute relative difference (ARD) is calculated as where x ref is the reference value-the value reported by the Dexcom G6, and x pred is the value predicted by the model.The MARD is the mean of these ARDs across the entire dataset in question While the model was selected to minimize MARD, other metrics were calculated after model development was complete as a hedge against overfitting a particular outcome.These metrics included mean absolute error (MAE), the proportion of predicted values that fell within 15% of the reference values for BG (±15%), the proportion of predicted values that fell within 20% of the reference values for BG (±20%), a Clarke error grid analysis, and a surveillance error grid (SEG).The "percent within threshold" metrics are based on those given by the Food and Drug Administration (FDA), which imposes approval requirements on CGMs [24].To contextualize these accuracy metrics in a way that is specific to this unique dataset, we calculated empirical chance for each metric.Empirical chance was calculated by randomly shuffling the test dataset's reference values and comparing the model's predictions for each observation compared against these values.This approach allows us to keep the true distribution of our test dataset's reference values when calculating chance.Results were also broken down by each factor used for stratifying the test dataset, i.e., glycemic level, RF dielectric sensor, and participant.

LightGBM Model Compared Against Empirical Chance
the model was trained, inference was conducted on the test dataset, which contained 311 observations.Each BG level predicted by RF dielectric sensor was then compared against the Dexcom G6 reference device's BG measurement, with results shown in Table I.In each metric, our model performed significantly better than the empirical chance model.The MARD in the test dataset was significantly lower than empirical chance [t(310) = 11.978,p < 0.001].Although the model was trained using MARD as the loss function, we also used other metrics to evaluate the model.The MAE in the test dataset was also significantly lower than empirical chance [t(310) = 10.860,p < 0.001].Via a proportion test, there were significantly more values that fell within 15% of the target value (z = 8.995, p < 0.001) and significantly more values that fell within 20% of the target value (z = 8.533, p < 0.001) than empirical chance.

B. Comparing the Training and Test Datasets
Despite precautions taken for overfitting described in Section II-E, there is some evidence that overfitting occurred in the training dataset.In particular, the results were better in the training dataset than in the test dataset, as seen in Table II.

C. Comparing the Results Across Glycemic Ranges
The MAE in the normoglycemic range was significantly better than empirical chance (t = 10.354,p < 0.001).In the hyperglycemic range, there was a tendency (t = 1.822, p = 0.078) to a better MAE than empirical chance.There were insufficient observations in the hypoglycemic range in the test dataset, and thus, only normoglycemic and hyperglycemic ranges were compared.For a detailed comparison, see Table III.

D. Clarke Error Grid Analysis
We also performed a Clarke error grid analysis of our results.Developed in 1987 by Mondal and Mondal [25] and Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

E. SEG Analysis
A more modern error analysis grid that uses a finer resolution to assess the potential clinical impact of errors made by a predictive glucose model is the SEG.The SEG draws on the experience of 206 diabetes clinicians to assess the potential  risk of varying error [27].The SEG analysis of our results is shown in Fig. 5.

IV. DISCUSSION
We demonstrated that the novel RF dielectric sensor and LightGBM prediction algorithm were able to predict BG continuously and noninvasively with a measurement of accuracy (calculated as MARD) of 12.9% against the commercial standard Dexcom G6 CGM as a reference.The sensor captured a spectrum of thousands of frequencies during every 22-s sweep over the course of a 2-3.5 h period.While other studies exploit the effect that glucose has on the dielectric properties of blood, they have generally focused on detecting this change in one or a narrowband of frequencies.The varying dielectric responses of glucose at different frequencies allowed us to build a model using information from all these frequencies to make accurate predictions.
The CGM was used as a reference device to enable rapid characterization of the sensor and iteration of the study protocol while generating a data pipeline, in which data were collected without significant discomfort to participants.In constructing a model, the LightGBM model performed better for our data than other machine learning algorithms employed in the literature, such as linear regression models [28] and global convolutional neural network models [29].The flexibility of the assumptions of this particular type of random forest model proved to be most effective for the number of features (frequencies) in the dataset.
In the test dataset, containing 311 observations, the model predicted BG measured by the Dexcom G6 CGM with an MARD of 12.9%.This was significantly lower than empirical chance on the same test dataset, which had an MARD of 29.8%.The model performed best in the normoglycemic range (between 70 and 180 mg/dL), due to the fact that the large majority (88.1%) of the training and test datasets contained normoglycemic values.BG estimations in the hyperglycemic range (>180 mg/dL) were still better than chance but were worse than the normoglycemic range.We were not able to assess values in the hypoglycemic range in this study due to the lack of data in this range.Overall, these results suggest that the RF dielectric sensor (hardware) and the LightGBM model (software) can accurately predict BG as reported by the CGM.Building on these results, future work will use the RF sensor and exploit the machine learning methods in this study to measure BG among a population of people with diabetes (PWD) and utilize a gold-standard venous blood reference device to further investigate accuracy as measured by MARD to ultimately validate clinical applicability.

A. Limitations
There are several limitations to this study.The most significant of these may be the number of participants in the study (n = 5), which limited our ability to gather data from a biologically diverse population.Because of this small sample size, we took an approach to stratify our training and test datasets by participant.While this stratification allows us to control variance in our test dataset to allow it to accurately represent the variance in our training dataset, it does limit our ability to understand to what degree these results would generalize to other participants.Additionally, because all participants were healthy and none was a person with diabetes, we had a limited amount of data in the hyperglycemic range and insufficient data to assess the hypoglycemic range at all.More data collection and analysis is required to accurately report glucose values on new users.
As described above, our model was designed to predict the values of a Dexcom G6 as a proxy for BG.While our ultimate goal is to quantify BG, the reference device provides an imperfect estimate of this, as an independent validation suggests its MARD (to a gold-standard reference) is 12.8% [30].In future work, it will be beneficial to explore the capability of the RF sensor to measure an entire cross section of tissue, which could include arterial glucose, venous glucose, capillary glucose, interstitial glucose, and intracellular glucose.Further studies could compare the RF dielectric sensor against a more precise BG reference device, with venous blood as the gold standard.

V. CONCLUSION
Overall, these results suggest that this hardware and software techniques can be applied to the noninvasive measurement of BG.Further investigations are merited to assess whether the performance can be extended to a larger participant population and a wider range of BG values, namely, in the hypoglycemic and hyperglycemic ranges to determine accuracy in intended use for DM management among PWD.Future clinical studies should aim to generate larger volumes of high-resolution RF dielectric sensor data compared with industry-leading reference device data to enable further data science and model development, and ultimately achieve the goal of developing an FDA-cleared noninvasive BG monitoring device.

Fig. 3 .
Fig. 3. Visualization of a sample tree from the LightGBM architecture.
Clarke al. a Clarke error grid is a graphical representation used to assess the clinical accuracy of BG measurement systems.It is a 2-D representation of the accuracy of the predictions, in which x-axis represents the reference values and the y-axis represents the values measured by the BG meter under evaluation (in this case, the RF dielectric sensor predictions).The lines on the grid demarcate five BG zones (A-E).The Clarke error grid analysis resulted in 246 of 311 (79.1%) of the BG values falling into Zone A, 63 of 311 (20.25%) of the values in Zone B, 0% in Zone C, 2 of 311 (0.6%) of the values in Zone D, and 0 in Zone E, as shown in Fig. 4, where Zones A and B are considered clinically acceptable.

Fig. 4 .
Fig. 4. Clarke error grid analysis is depicted here to demonstrate which values in the test dataset fell into each error zone based on the glycemic range.

TABLE I ACCURACY
OF PREDICTIONS IN TEST DATASET

TABLE III RESULTS
BY GLYCEMIC STATUS