In this section the proposed microwave glucose sensor is presented alongside with a systematic design approach, featured design parameters, dependency of resonance characteristics, numerical analysis, glucose sensing mechanism and the various accomplished in-vitro and in-vivo experiments for glucose concentration measurements. A comprehensive discussion including the influence of different geometrical parameters on the resonance profile, the sensitivity analysis for glucose detection under different topologies, and the signal processing and machine learning algorithms used to intelligently analyze the sensor’s raw data are also provided.

CSRR-based sensor design

The proposed biosensor is numerically designed to operate around 2.45 GHz when used for glucose detection. This is chosen to match the Industrial, Scientific, and Medical (ISM) band 2.4–2.5 GHz when the sensor is integrated with the radar system45. In fact, operating in this range of frequency would also allow adequate penetration depth for the induced EM waves to reach the blood vessels in the fingertip. This penetration is to a certain extent maximized in our sensor given the concentrated energy of the sensing region where the fingertip is placed46. Additionally, at this frequency there is a higher possibility to identify different glucose concentrations of interest due to the small loss introduced (tanδ ~ 0.13). Therefore, sensing at this frequency would probably bring a considerable sensitivity despite the small changes in dielectric properties from one glucose level to another47. These variations in dielectric constant and loss tangent are studied for the desired concentrations across the centimeter-band 1–6 GHz using the first order Debye relaxation model proposed in48 for glucose aquatic solutions. A slight increase in ({{epsilon }_{r}}^{^{prime}}) and, conversely, a tenuous decrease in tanδ are observed with an increased glucose concentration. In this tendency, it is also observed that the percent variation in tanδ is much higher than its corresponding on ({{epsilon }_{r}}^{^{prime}}) (e.g. for a 10 mg/dL change at 2.4 GHz, the percent change in tanδ is about 0.4% while the variance in ({{epsilon }_{r}}^{^{prime}}) is approximately 0.005%).

The sensor structure is primarily inspired from the metamaterial technology by incorporating four similar cells of hexagonal-shaped complementary split-ring resonators (CSRRs) with localized elements in a novel configuration as shown in Fig. 1b. The four CSRRs are configured in a honey-cell pattern and engraved at depth 35 µm in the copper ground plane of an FR4 dielectric PCB (Fig. 1d) of dielectric permittivity ({epsilon }_{r}^{^{prime}}) = 4.4, loss tangent (tandelta) = 0.02, length L = 66 mm, width W = 20 mm, and thickness h = 0.8 mm. The magnetic walls of the passive CSRRs are oriented perpendicularly to a planar microstrip-line (MTL) etched on the upper face of the substrate as shown in Fig. 1e. This alignment would guarantee an electric excitation for the resonance with time-varying electric field between the MTL strip and the ground plane. The dimensions of the feed line are optimized to 66 × 1.5 mm2 in order to obtain a 50 Ω characteristic impedance that perfectly matches the SMA (SubMiniature version A) connectors of the driving power source. Two topologies for the honey-cell configuration are realized, compact and dispersed. Accordingly, two hexagonal cells are placed horizontally along the axis of the MTL strip and detached with distance CL = 7.6 mm (compact) and CL = 12.6 mm (dispersed) between their corresponding geometric centres. The honey-cell design is completed by setting the other two hexagonal cells in a vertical placement with CW = 12 mm between their corresponding centres. As sketched in Fig. 1b, the unit-cell of each CSRR is composed of two concentric rings (outer and inner) of hexagonal shape nested within each other with a coupling-gap t = 0.4 mm in between. The loop of each ring has a dielectric slit of side-width s = 0.4 mm and ends in a metallic slot of width g = 0.4 mm. The diagonal length of the outer ring is a = 7.6 mm, while that of the inner one is set to b = 6.0 mm. The split gaps for both rings are on diametrically opposite sides of each other.

For reliable blood glucose sensing, it is vital to maximize the resonance strength and to perfectly confine the resonating electromagnetic fields within the permittivity sensing region of the sensor. Therefore, the design and geometrical parameters of the planar transmission line as well as the engraved CSRR cells have been optimized to obtain a steep transmission resonance around f0 = 3.0 GHz for the unloaded sensor after studying the primary influence of each parameter using a Finite Element Method (FEM)-based numerical simulator as shown in Fig. 2. To elaborate, the resonance characteristics of the proposed sensor are basically determined by the substrate dielectric specifications and geometries (i.e. width and height). In addition, the design parameters of each hexagonal unit-cell play a crucial role in defining the complete resonance profile for the sensor. These parameters include the diagonal length a and b, coupling-gap t, side-width s, split-gap g, and coupling-controllers (CL and CW) of the integrated honey-cell. In contrast, before engraving the hexagonal CSRRs, these resonance characteristics are solely determined by only the geometrical width and length of the dielectric patch.

Figure 2

(a) Equivalent electrical model of the CSRR sensor under loading condition. (bf) Simulated transmission coefficient S21 of the unloaded honey-cell CSRR at different (b) diagonal length of a (c) split gap g (d) side width s (e) slots-coupling distance t, and (f) center-to-center coupling distance CL. The primary influence of each of these parameters on fR is numerically investigated. Since the diagonal length a is proportional to the length of the dielectric slot, increasing a would directly increase the resultant Le and therefore reducing fR as shown in (b) for a changing from 5.6 to 7.6 mm. (c) depicts how fR is shifted towards higher frequencies (i.e. increasing) as the split gap g is increasing from 0.1 to 0.5 mm. Similar functionality can be observed in (d) for the case of slots-coupling distance t. Moreover, fR is also slightly increasing with an increased side-width s as shown in (e). Both behaviors can be explained by the fact that the side width, slots-coupling distance, and split gap are all inversely proportional to the effective capacitance Ce, thereby increasing these parameters will probably increase fR. The effect of the center-to-center separation distance is studied in (f) that shows a higher magnitude for the transmission response (i.e. insertion loss), and hence higher Q-factor, for larger distance CL. This is reasonable as increasing CL would reduce the effect of the mutual coupling between the adjacent hexagonal CSRRs.

Glucose detection mechanism

A simple lumped element model is used to describe the physical behaviour of the integrated CSRR when used for glucose sensing49,50. This electrical model consists of three parts as shown in Fig. 2a: first, the signal line (MTL) (of 1.5 mm width) used for exciting the resonator is modelled by an inductance Lc to represent the magnetic coupling resulting from the MTL when crossing the CSRR structure. The second part is the dielectric substrate (intermediate junction of length 66 mm, width 20 mm, and thickness 0.8 mm) used to electrically couple the MTL with the resonators in the ground plane. This part can be approximately modelled by a shunt capacitance Cc in parallel with a resistance Rc that represents both the dielectric losses of the substrate and the conductive losses of the metal strip. The last part is the honey-cell resonator (of four cells) that is modelled by an RLC parallel resonant circuit where the winding of the dielectric rings (of width s) in each hexagonal cell acts as inductance LR, the metallic gaps g and spacings t between rings create a parallel capacitance CR, the conductive and dielectric losses are modelled by a resistance RR that takes into account the very small inductive magnetic losses associated with LR. When the SUT (i.e. glucose sample) is placed on top of the CSRR surface, the last resonator part in the model is modified by adding a parallel RC circuit that models the dielectric characteristics of the loaded glucose sample. In specific, CM is directly related to the relative permittivity of the sample, while RM is mainly dependent on its loss properties. This configuration of the lumped elements stores an oscillating electric and magnetic energy in the inductance and capacitance that are arisen by the induced charges and currents inside the patterned dielectric loops or slots upon exciting the CSRRs. When both energies are in balanced state, the microwave sensor resonates at a specific frequency as given in Eq. (1), directly observed on the minimum frequency response of the transmission coefficient S21 with a loaded Q-factor given by Eq. (2)

$${f}_{R}({{epsilon }_{r}}^{^{prime}})=frac{1}{2pi sqrt{{L}_{R}{(C}_{e}({{epsilon }_{r}}^{^{prime}})+{C}_{c})}},$$


$${Q}_{R}({{epsilon }_{r}}^{^{prime}},tandelta )=frac{{f}_{R}}{{Delta f}_{3dB}}={2pi {f}_{R}R}_{P}left(tandelta right)left({C}_{e}({{epsilon }_{r}}^{^{prime}})+{C}_{c}right),$$


where ({Delta f}_{3dB}) denotes the 3 dB bandwidth of the resonance peak, Ce ~ CR//CM denotes the effective capacitance of the integrated CSRRs when loaded by the glucose samples of various concentrations49. Rp is the equivalent resistance of paralleling the coupling-part Rc and CSRR-part Re that represents the dielectric losses of the CSRR and its loaded glucose sample. In this sense, while the magnetic field is confined between the slit traces, the stored electric energy widely spread and concentrated in the CSRRs region interacts with the glucose samples placed in vicinity of the CSRRs. Consequently, subtle changes in the dielectric permittivity of the loaded glucose samples would disturb the electric field distribution, and thereby reflected in fR through the perturbation of the CSRR effective capacitance Ce. This shift in the resonance frequency is therefore a measure for determining the sample glucose concentration for a constant volume. Similarly, Re is mainly dependent on the loss property of the glucose sample and therefore the contrast in tanδ is reflected as amplitude variations in the resonance profile. The resulting changes in resonance attributes are considered as a coded signature of the glucose sample dielectric properties, that could be correlated effectively to the corresponding glucose level through careful analysis of the modified resonance behaviour.

Numerical analysis

Both topologies, compact and dispersed, were numerically analyzed using an FEM-simulator to test the functionality and quantify the resonant frequencies for three different cases: when unloaded, loaded with an empty container, and loaded with pure distilled-water (DI). A cylindrical glass container was designed to load the glucose samples onto the CSRRs surface (this is replaced later by a suitable fixture to place the fingertip) as shown in Fig. 3a. The container has 8.9 mm outer diameter, 7.4 mm inner diameter, 1.5 mm wall thickness, and 10 mm height. In addition, a thin glass layer of hg = 0.15 mm was introduced between the container and CSRRs in order to evade short-circuiting the dielectric slits of the CSRR cells when loaded with the glucose specimens of moist nature. The intrinsic transmission resonant frequency prior-loading is 3.0 GHz for both topologies as shown in Fig. 3b. Loading the sensors with the empty cylindrical container will introduce a few MHz shifts in resonance as can be clearly observed in Fig. 3b, where fR for the compact and dispersed are tuned to 2.83 GHz and 2.75 GHz, respectively. The empty container response will be used as a reference state when the sensors are used for sensing the glucose concentrations of interest. Filling the container with the DI-water of volume 600 µL will disturb the electric field generated in the sensitive area of the dielectric substrate, thereby resulting into further shift in fR due to the change in effective permittivity of the media surrounding the resonators. A frequency shift of about 770 MHz (compact) and 710 MHz (dispersed) was noticed when the DI-water was simulated inside the container. Moreover, a damping effect was observed on these resonances with a significant drop in the resonance peaks due the lossy nature of the DI-water sample. Furthermore, an additional resonance is established around 1.56 GHz (compact) and 1.66 GHz (dispersed) as a result of this new configuration of the sensor structure that accommodates the DI-water sample.

Figure 3

(a) Geometrical model and parameters of the proposed honey-cell CSRR integrated with the glass container used for loading the glucose samples, (b) simulated transmission response (S21) of the sensor topologies (compact and dispersed) for three different states: unloaded, loaded with empty container, and filled with DI-water. Simulated transmission response of different glucose samples loaded on top of the compact sensor with (c) no skin, (e)–(g) skin thickness of 0.5, 1.0, and 1.5 mm, respectively. Cross section of the electric field distribution in the midst of the glucose tissue in the respective cases (d) no skin, (h)(j) skin thickness of 0.5, 1.0, and 1.5 mm, respectively.

Next, glucose aqueous samples of concentrations C1–C11 in the range 40–500 mg/dL were simulated on top of the sensing region inside the glass container. This range of concentrations covers the wide diabetes spectrum including hypoglycemia (< 70 mg/dL) and hyperglycemia (> 140 mg/dL). The single-pole (first order) Debye model given by Eq. (3) was used to construct the numerical models for the dispersive dielectric properties of the glucose-water tissues at different concentrations. It is noteworthy to mention, this model was developed in48 based on the spectroscopy measurements of watery solutions of 50, 250, 1,000, and 2,000 mg/dL as collected using a commercial coaxial probe kit connected to VNA. This is the most rational model to use for approximating the blood-glucose behaviour since no other mathematical model is yet developed for the dielectric properties of actual blood-glucose solutions.

$${epsilon }_{r}left(w,xi right)={epsilon }_{infty }left(xi right)+left(frac{{epsilon }_{stat}left(xi right) – {epsilon }_{infty }left(xi right)}{1 + jomega tau (xi )}right)+frac{{sigma }_{s}}{jomega {epsilon }_{o}},$$


where ({epsilon }_{r}left(w,xi right)) is the complex permittivity of the aquatic solution of glucose concentration (xi) (in mg/dL) at angular frequency (w), while ({epsilon }_{stat}), ({epsilon }_{infty }), and (tau) are the concentration-dependent Debye coefficients46, ({sigma }_{s}) is the static conductivity, and ({epsilon }_{o}) is the permittivity of free-space.

Figure 3c shows the simulated sensor’s transmission responses for the different glucose samples C1–C11 loaded at volume 600 µL each (corresponds to 2 mm thickness inside the container). Four resonance frequencies were induced in the frequency range 1.5–5.0 GHz. All these resonances show variations in both frequency and amplitude for glucose level changes. However, the fourth harmonic resonance around 4.8 GHz is more responsive to the dielectric contrast of the glucose samples. This is reflected not only in the slight shift in resonance frequency, but also more distinctly in the significant change in resonance amplitude, implying that the loss property of the dissolved glucose contributes more to the system frequency response. The simulated response also demonstrates the sensor capability for glucose detection over the whole diabetes range including the hypoglycemia concentrations 40–60 mg/dL (C1–C3, shown in blue), normal condition concentrations 70–140 mg/dL (C4–C7, shown in red) and hyperglycemia concentrations 200–500 mg/dL (C8–C11, shown in black).

To mimic the realistic scenario of placing a fingertip onto the sensing region, we performed another simulation where the glucose samples C4–C7 of thickness 2 mm each were placed on top of a skin layer (({{epsilon }_{r}}^{^{prime}}) = 38.1 and tanδ = 0.28)51,52 of varying thicknesses 0.5, 1.0, and 1.5 mm, and the corresponding transmission responses were analyzed over the frequency range 1–6 GHz as depicted in Fig. 3d–f, respectively. In fact, skin thickness is generally related to several factors such as body site, gender, skin type, age, pigmentation, etc., however, the simulated thicknesses are within the practical range of human fingertips53. It is observed that three distinct resonances were induced in these scattering responses around f = 1.9–1.95 GHz, f = 2.44–2.51 GHz and f = 4.77–6.5 GHz. The third resonance in each response conveys more information about the glucose concentrations of the simulated samples buried under the skin. This is demonstrated very clear in the larger amplitude variations compared to those existed at lower resonant frequencies as plotted in the enclosed graphs. However, skin layers of larger thickness would damp the corresponding resonance strength, and therefore weaken the coupled electric field that interacts with the glucose sample as shown in Fig. 3g–j that compare the electric field intensity coupled to the glucose tissue in the respective cases. Consequently, a lower sensitivity is reflected in the resonance readings with reduced amplitude resolution for sensing the glucose concentrations. For instance, by calculating the amplitude variation per unit change in ({epsilon }_{r}^{^{prime}}) and (tandelta) for the two cases of 0.5- and 1.0-mm skin thickness, we got ∆|S21|/∆|({epsilon }_{r}^{^{prime}})|of about 4 and 1, respectively. While that for ∆|S21|/∆|(tandelta)|is around 40 and 10, respectively.

In-vitro experiments

Various measurements have been conducted to verify the performance of the proposed glucose sensor using two different setups incorporating either (i) VNA or (ii) 2.45 GHz radar board. For the sake of simplicity, aquatic glucose solutions were used in these experiments to imitate the blood behaviour at different glucose concentrations (70–120 mg/dL) of clinically relevant to Type-2 diabetes. This would also help to secure the reproducibility of measured scattering data. Many other studies have also adopted the glucose aqueous solutions for preliminary experiments on non-invasive glucose detection using RF sensors54,55,56. This approximation is valid since water contributes about 50% of the entire human blood that contains other vital components at varying proportions (e.g. Na, Ca, Mg, K, Cl, etc.). These minerals are present at lower concentrations compared to the dominant glucose whereby the blood dielectric properties are dominantly affected46,57,58,59. In particular, compared to the broad range of glucose concentrations in blood; other minerals only vary in limited ranges, Na: 310–333 mg/dL, Cl: 337–372 mg/dL, Mg: 1.8–3.4 mg/dL, Ca: 8.5–10.5 mg/dL, K: 13.6–21.4 mg/dL)59, and other components in blood like the lactic acids, have also very small concentrations (normal range 4.5–19.8 mg/dL) when compared to the dominant glucose. Moreover, Na and Cl that exist in larger order of magnitudes compared to other minerals would mostly influence the conductivity of the blood and therefore their effect is negligible on the frequency shift in comparison with the effect of glucose variation. Since the frequency shift is considered as the main sensing parameter of the sensor, it is reasonable to assume that the small mineral concentration variations do not interfere with results from glucose related frequency shift.

Measurements using the VNA

First, the fabricated prototypes (compact and dispersed) were experimentally tested using the VNA setup shown in Fig. 4a to record the intrinsic transmission responses before loading any glucose sample. Figure 5a presents the measured transmission coefficients as a function of frequency for the unloaded compact and dispersed sensors while comparing against the conventional single-hexagonal CSRR (shown in dotted green). Under unloading conditions, the three CSRR structures experience transmission resonances around fR = 3.0 GHz with resonance minima of about − 22.85 dB, − 29.57 dB, and − 43 dB, for the single-hexagonal, compact, and dispersed, respectively, as depicted in Fig. 5a. Both compact and dispersed sensor have steeper resonance depth that would provide a better sensitivity for characterizing lossy glucose samples of high tanδ (i.e. detecting the tiny variations in the loss tangent for various glucose concentrations). These measurement results show an excellent agreement with the predicted numerical simulations. Minor differences are due to typical tolerances in the fabrication process.

Figure 4

Glucose sensing experiments (a) VNA experimental setup, (b) sensor loaded with an empty container, the micropipette and glucose aqueous solutions used in the experiments.

Figure 5

(a) Measurements and simulations of the transmission coefficients of the fabricated sensors before loading glucose samples (resonance around 3.0 GHz), (b) measured transmission coefficient S21 when the compact sensor is experimented for the 70 mg/dL glucose sample at different volumes 200, 400, and 600 µL to compare the corresponding sensor behaviors. Increasing the sample volume, and implicitly the thickness, would tune the S21 resonant frequencies towards lower frequencies.

In order to test the developed sensors for measuring glucose samples at the levels of interest, we fabricated a cylindrical glass container to hold the samples on top of the CSRRs surface. The empty cylindrical container was placed on the honey-cell CSRR structure, as shown in Fig. 4b. This will introduce a few MHz shifts from the reference resonance in S21 of the unloaded state. In each measurement trial, the micropipette device was used to measure a precise volume of V = 600 µL from each concentration, load it inside the container and the change in the transmission resonance frequency response was recorded. This volume was adopted for testing various glucose samples to minimize the measurement errors due to uncertainty in the sample volume. In this regard, the sensor response was studied at different volumes of the tested sample as shown in Fig. 5b. It is observed that the sensor shift in resonance frequency would converge to a lower limit as the sample volume amount to 600–700 µL which enables a homogeneous distribution of the sample inside the container whereby the sensing region is completely covered. To test the sensor response for another concentration, a clean tissue paper was used to completely remove the previously tested glucose sample with attention to retrieve the exact reference resonance at S21 prior-loading the sample for a fair comparison. Figure 6a,b show the transmission response (magnitude and phase) of the compact sensor when the concentration of glucose is changing in 70, 90, and 110 mg/dL. Figure 6c,d capture the same responses when the dispersed sensor was used for testing. In both sensors’ responses, it was observed that the resonant frequencies at which the transmission is minimized are shifted towards lower frequencies as the glucose concentration in the sample increases. Additionally, trackable amplitude changes are exhibited in some resonances due to the tiny variations in the loss tangent of the tested glucose sample. Minor variations were also noticed in the transmission phase responses as shown in Fig. 6b,d. Similar observations on the resonance frequency and amplitude were recorded when the compact sensor was tested for sensing higher glucose concentrations in the range 200–500 mg/dL that represent the hyperglycemia condition as shown in Fig. 7.

Figure 6

Measured transmission response S21 as function of frequency for tested glucose samples of various concentrations. The transmission coefficient and phase for the (a), (b) compact and (c), (d) dispersed sensors, respectively.

Figure 7

Measured transmission response |S21| as a function of frequency for tested glucose samples of higher concentrations of clinically relevant to the hyperglycemia condition in diabetes.

Sensitivity analysis

The VNA measurements of the scattering-parameters of the glucose-loaded CSRR sensors show that both reflection and transmission responses (magnitudes and phases) are varying according to the glucose level changes in the loaded sample. Particularly, considerable shifts in the resonance frequencies of the S11 and S21 are observed for varying between 70, 90, and 110 mg/dL. The data sheet of the VNA instrument used in this experiment poses for higher uncertainties in the readings of the reflection coefficient (magnitude and phase) compared to the transmission coefficient that is more accurate and stable. For this reason, we only considered the frequency shifts in the transmission coefficient S21 to assess and compare the sensitivity range of the proposed sensors. As seen from Fig. 8a,b, both sensors exhibit impressive frequency resolutions that would be beneficial to distinguish various glucose concentrations of relatively small contrast in dielectric properties. The resonance frequencies resultant from loading a specific glucose concentration can be estimated using the linear regression models derived for each sensor, compact and dispersed, as shown in Fig. 8a,b, respectively. Conversely, the inverse models could be used to estimate the unknown glucose level of a tested sample. In these plots, it is observed that the respective resonant frequencies decrease with increased levels of glucose concentrations. However, the frequency resolutions for glucose level changes at the respective resonances are not perfectly identical. The sensitivity of the compact sensor is estimated as − 1.25 × 10–3 GHz/(mg/dL) representing the gradient of f1 and f2 linear models. However, the best sensitivity slope for the dispersed topology is recorded as − 9.5 × 10–4 GHz/(mg/dL) at f2 which is a bit lower than the compact counterpart.

Figure 8

Linear correlation models for the resultant resonant frequencies of (a) compact and (b) dispersed sensors at different glucose concentrations. (a) For the compact sensor, three resonances, f1, f2, and f3, are established in the frequency range 1.5–2.8 GHz. Sensing the glucose samples on (b) The dispersed sensor exhibits four distinct resonances f1, f2, f3, and f4 in the frequency range 1.6–2.8 GHz. A reduction in the resonant frequencies is observed with growing glucose concentrations.

The VNA microwave system was strictly calibrated to limit the transmission loss at 0.02 dB over the operating frequency range. This is important to acquire reproducible measurements that are more accurate. It is noteworthy to mention that, all the glucose measurements were repeated three times and the average is reported for precision, repeatability, and reproducibility verifications. This averaging would help to eliminate random noise introduced by the source power (i.e. VNA) or any other uncorrelated noise and improve the SNR accordingly. In each repeatability trial, careful attention was paid towards retrieving the exact initial resonance response of the empty container-loaded sensor after removing the samples using a tissue paper. The robustness of the sensor could be seen from the stable and repeatable frequency responses that vary in a small range of ± 0.5 MHz as indicated by the error bars enclosed in these visuals. Given that the sensor frequency reading for 5 mg/dL is about 4.7 MHz (average sensitivity of 0.94 MHz/(mg/dL)), the glucose concentrations as small as 1 mg/dL could be identified using the proposed sensing platform with a reasonable accuracy. Since the sensor scattering response is quite dependent on the sample EM properties which are temperature dependent, therefore, we stored all the prepared glucose samples in the same room of temperature regulated at 25 ± 1 °C. Additionally, the heat and temperature were controlled and monitored at 25 ± 1 °C during the glucose experiments to minimize the instrumental and environmental impact on the collected measurements. However, small fluctuations in temperature will not bring a significant effect on the resonance measurements of the CSRR sensors (± 1 MHz shift in fR). This is explained by the fact that, the far field radiation of our small resonators of sub wavelengths could be effectively suppressed as depicted in Fig. 9 where the total electric field is quantified in the far field region of the sensors with maximum magnitude of about 1 V/m. Previously, it is shown that most of the electromagnetic energy is concentrated in the honey cell area in the near-field region. Therefore, any unwanted environmental reflections will not have notable effect on the sensor measurements.

Figure 9

Far-field radiation of the CSRR sensor.

Lastly, in what follows the sensitivity performance of the proposed honey-cell sensors is compared to other microwave sensors in the recent literature. To this end, a comparative statement is presented in Table 2 where state-of-the-art glucose sensors are sorted in the order of increasing sensitivity with their respective parameters. Herein, the sensitivity is defined as the frequency shift/variation ∆fR with respect to 1 mg/dL of glucose concentration change in SUT for a certain volume and specific test setup. Following this measure, the sensitivity achieved in earlier works based on diverse microwave sensing mechanism is much lower than the least resolution adopted by our proposed sensors. The sensitivity of the proposed sensors is also superior to other techniques based on tracing the slight variations of S11 and S21 resonant magnitudes for which a measuring tool of high-precision is needed. The limited sensitivity in the literature is due to the confined EM fields coupled to the resonator surface that limit the interaction with the tested glucose sample. Because of this phenomenon, the substrate in traditional resonators has a more important role in defining the resonance frequency rather than the SUT of glucose content. However, the improved design of the CSRR elements in the presented work, where this interaction is greatly expanded over the sensing region, allows the resonance frequency response of the sensor to be mainly defined by the SUT permittivity. To the best of our knowledge, the achieved sensitivity of this work, 0.94 MHz/[mg/dL] is far beyond the best results reported in literature regardless of shape and volume of SUT. This would render the sensor response less susceptible to environmental and instrumental noises than its conventional counterparts. Therefore, it can be emphasized that the proposed sensor can in principle be used to detect normal blood sugar range quite conveniently as well as those for hypoglycemia and hyperglycemia. It could also be used as a warning tool to diabetic patients when attempting to consume energy drinks and fruit juices which have inherently high glucose concentrations.

Table 2 Comparison of different microwave glucose sensors.

Machine learning processing

The sensor’s scattering responses collected from the VNA instrument are analyzed using the Principal Component Analysis (PCA) unsupervised machine learning algorithm to clearly distinguish various glucose concentrations tested by the sensor. This is a vital add-on to the integrated microwave sensor that could further enhance its detection sensitivity for the blood glucose concentrations of interest that vary in a narrow range representing the normal diabetes condition. It is shown in the previous experimental results that the sensor exhibits distinctive scattering responses in terms of the transmission coefficient (magnitude and phase) when loaded by various glucose concentrations, however these responses recorded by the VNA (i.e. |S21| or S21) would exhibit small changes/shifts in fR to reflect the varying glucose concentrations of the tested samples (e.g. Fig. 6a,c). For instance, the slight frequency shifts acquired by the VNA for the compact |S21| resonances that vary in a small resolution of 1.25 MHz/(mg/dL) as shown in Fig. 8a. These changes would be tricky to trace using exclusively common sense and hence the difficulty of precise identification of various glucose levels. The PCA as a robust classification algorithm can be used to further analyze the raw data induced by the sensor for different glucose samples. In particular, to enhance the small scattering differences between various glucose levels, we used to apply the PCA algorithm that maps the feature vectors (measured scattering data) for each glucose sample into a two-dimensional space where each glucose concentration is represented by only two indices called the principal components (mathcal{K}). This will help to bring a higher resolution when the sensor’ responses are correlated to different blood glucose levels. In other words, the PCA algorithm creates a mapping between 1) X  RN: the N-dimensional input variables’ vector (also referred to as the features’ vector) which corresponds to the measured |S21| or S21 obtained at the different frequencies over the operating frequency range and 2) y R: the outcome variable which corresponds to the estimated glucose level.


The theory of the PCA algorithm is based on the dimensionality reduction of the problem where a vector space transform is performed68. Herein, the PCA is exploited to extract important features from the data set of the sensor scattering responses and further to express this information as a set of orthogonal variables called principal components69. Generally, these principal components are derived from the eigen-decomposition of positive semi-definite matrices and the singular value decomposition of rectangular matrices69. In this sense, original datasets of high dimensionality could be reduced to smaller number of distinctive variables via mathematical projection without missing much information to analyze patterns, tendencies and anomalies.

The pseudocode in Algorithm 1 describes this classification routine when applied to the VNA measurements in Fig. 6a,b, the transmission coefficient parameters (magnitude and phase) are considered the two feature vectors to be extracted from the frequency response of each tested glucose sample. Since these magnitude and phase responses were recorded at N = 201 frequency points spanning the range from 1.5 to 2.8 GHz, then each glucose sample is described by two feature vectors magnitude and phase with 201 length each. It is shown that samples of various concentrations have these magnitude and phase values varying at some of the respective frequency points especially at resonance frequencies. As described in Algorithm 1, the PCA algorithm could use either the magnitude (|S21|) or phase (S21) vectors as input features to classify various glucose concentrations. Once the algorithm is provided with measured feature vector ({x}_{{C}_{m}}) for each glucose sample ({C}_{m}), it processes the data using the steps in Algorithm 1 to extract the principal components ({mathcal{K}}_{{C}_{m}}^{i}) for all the glucose samples and cluster them accordingly based on their concentration-dependent principal components as shown in Fig. 10a,b where samples of different glucose concentrations are notably separated in the PCA space when the magnitude and phase feature vectors are used, respectively. We also observe that both feature vectors (i.e. magnitude and phase) of the sensor are effective and efficient to bring higher discrimination or spatial separation between different glucose samples. This is considered as a training stage for the developed PCA model; therefore, any new tested sample could also be clustered accordingly following the same analytical recipe.

Figure 10

ML processing of the measured transmission responses of the compact sensor for various glucose concentrations (Fig. 6a,b), (a) Magnitude of S21 using PCA and (b) phase of S21 using PCA.

Testing using the radar system

A portable prototype of the compact sensor was developed by interconnecting a small low-cost and low-power radar module as a driving source instead of the bulky VNA instrument. In particular, the open source QM-RDKIT that supports the frequency modulated continuous wave (FMCW) functionality, was utilized to couple the CSRR sensor at the ISM frequency range (2.4–2.5 GHz)45. To do so, the transmitting “Tx” and receiving “Rx” ports of the radar board were connected to the feeding ports of the CSRR sensor using the SMA RF coaxial cables as shown in Fig. 11a. The major sections of the radar board are shown in Fig. 11b. It has an RF section to generate and output the transmitted signal and to downconvert the received signal to a frequency range that can be easily digitized using the onboard Analog-to-Digital converter (ADC). Specifically, the onboard Voltage Controlled Oscillator (VCO) and Phase Lock Loop (PLL) are used to generate the transmitted signal of defined frequency. The PLL serves to frequency lock the output of the VCO using the onboard reference to provide a stable and repeatable output frequency. The output of the VCO is amplified before being passed to the input port of the interconnected sensor. The corresponding signal from the sensor output port is first mixed with a sample of the transmitted signal to produce a frequency offset (beat frequency), then it is filtered to remove any unwanted signals developed from the mixing process. Afterwards, the signal is passed to the ADC for digitization, and either stored in memory or streamed over the USB/Bluetooth connection. In addition to the ADC, the digital section also contains the PIC microcontroller, USB and power interfaces. The PIC microcontroller coordinates all functions of the radar board, responds to all control commands and data requests received through the USB/Bluetooth connections.

Figure 11

Glucose sensing experiments in (a) A portable setup for the honey-cell sensor connected to the low-frequency radar board, (b) The major sections inside the radar board: digital, Bluetooth radio, RF, filter prototyping, and lightbar for audiovisual feedback45.

For glucose measurements, a micropipette was used to measure a precise volume of V = 600 µL from each blood sample and transfer the blood into a cylindrical container integrated on top of the sensor. Once the glucose sample is fully loaded, the RF power button was triggered from the PC interface to transmit a radar signal of one-way single sweep mode into the CSRR sensor. For repeatability verifications, each measurement trial for a glucose sample was repeated three times with a repeatability error of about ± 0.02 V. The corresponding raw data was collected accordingly and sent to the host PC over a USB connection. The data could also be sent wirelessly via a Bluetooth connection. The average of M = 3 repeatable voltage signals ({A}_{C}(n)=frac{1}{M}sum_{i=1}^{M}{x}_{C}^{i}(n)) for each glucose sample C is plotted in Fig. 12a. Next, the respective average signals ({A}_{C}(n)), C = 70, 90, and 110 mg/dL, were further processed to reveal the energy density corresponding to each glucose sample. Particularly, the finite energy for each average signal ({A}_{C}(n)) of a tested glucose sample was evaluated using ({varepsilon }_{s}=sum_{n=1}^{N}{left|{A}_{C}(n) right|}^{2}), where N = 600 is the total number of time samples in each average signal. The dielectric contrast of different glucose samples was distinctly captured by the sensor and demonstrated in the varying energy density 1,604, 1,382, and 1,153 Volts2 for the 70, 90, and 110 mg/dL, respectively, as depicted in Fig. 12b. The sensor measured data was also processed using the PCA classification algorithm that clusters the corresponding glucose samples based on their extracted principal components as depicted in Fig. 12c.

Figure 12

Glucose level detection using the radar setup (a) Raw data for tested glucose samples as collected on the receiving channel, (b) Comparison of energy density between different glucose levels and (c) The PCA processed results.

In-vivo experiments

In the light of previous positive results, the integrated portable sensing system was further tested for a simple in-vivo experiment as a proof-of-concept for this technology when revised for intermittent or continuous blood glucose monitoring. For this purpose, the CSRR compact sensor was attached to a fixture structure suitable for finger placement as shown in Fig. 13a. The geometrical dimensions of the fixture are shown in Fig. 13b. The new prototype was integrated with the radar board (QM-RDKIT) as captured in Fig. 13d for the entire system setup. The radar system was interconnected to a laptop where the radar operation was controlled through a GUI. Preliminary tests were performed on a healthy male volunteer at age 29 before and after having the lunch meal while comparing the non-invasive measurements against a standard glucometer used as a reference for comparison. This testing recipe was guided by the fact that, in healthy non-diabetic people, the blood glucose level should measure between 72–99 mg/dL before a meal and should be less than 140 mg/dL two hours after a meal32,33. Therefore, a pre-prandial test was first performed for the tested subject by placing his fingertip suitably in the sensing region inside the fixture as shown in Fig. 13c. The finger should be in contact with the sensing region (firmly attached to the fixture) to perturb the electromagnetic fields and induce noticeable changes in the sensor transmission response. The sensing process from the fingertip would take a short time of about one-minute max during which no changes in the temperature status of the subject finger is expected. The corresponding raw data in response of a one-way single sweep transmission was collected from the radar receiving channel using the featured GUI. The same test was repeated three times for repeatability verification and the average of the three readings (with ± 0.03 V error max) is plotted in Fig. 13e (black curve). Afterwards, the individual’s BGL was measured using the commercial invasive glucometer to get the actual pre-prandial BGL of about 4.4 mmol/L (≈ 80 mg/dL). Similarly, a second test was performed for the tested subject two hours after having a lunch meal for normal diet. The test on the non-invasive sensor was repeated for three consecutive times and the average voltage signal is plotted in Fig. 13e (blue curve). The post-prandial BGL was measured at 6.9 mmol/L (≈ 124 mg/dL) on the glucometer. Following these measurements, the transmission results of the CSRR sensor were observed to be reliably consistent and aligned with the glucometer readings for the individual BGL variations before and after the meal. Particularly, the sensor’ transmitted signal exhibits a change in amplitude and a shift in time domain in response to the varying BGL of the tested subject. The black curve is corresponding to 80 mg/dL BGL while the blue one is representing the 124 mg/dL reading that leveled up two hours after the meal intake.

Figure 13

Blood glucose level tracking from an individual’s fingertip (a) fabricated fixture for fingertip placement, (b) geometrical dimensions of the fixture, (c) fingertip placement in the fixture, (d) complete setup for the in-vivo experiment, (e) analysis of the BGL discrete readings from the CSRR sensor and glucometer. Continuous BGL monitoring over a course of 30 min for (f) pre-prandial, and (g) post-prandial testing. The peak amplitude of the sensor response versus time (blue curve). The invasive measurements for BGL variations over time (red curve).

To better understand the BGL detection, the measured sensor data was further analyzed and processed using the discrete Fourier transform (DFT) algorithm. The consequent energy density has shown to be varying for the two processed data corresponding to the two different BGL readings, 80 and 124 mg/dL, as depicted in the enclosed plot in Fig. 13e, that shows an energy density metric of about 1,207 and 1,122 Volts2, respectively. This would also imply that the sensitivity to glucose variation is slightly reduced when compared to that of the samples in glassy container. In fact, the coupled electric field in the sensing region has less interaction with the glucose molecules in this case given the lossy nature of the fingertip biological structure including the cornified skin layer. This is seen very clearly when a skin layer model was introduced in the numerical simulations showing the field intensity with decaying magnitudes halfway through the glucose-contained layer. However, the sensitivity could be enhanced by modifying the design specifications through incorporating a flexible substrate of smaller loss tangent or utilizing a more powerful driving circuit (> 1 Watt output power) instead of the RDKIT used in this preliminary prototype as a proof-of-concept.

To confirm the correlation of the sensor readings to that of the actual BGL in real-time setting, another experiment was performed while continuously monitor a volunteer’s BGL over a course of 30 min before and after a meal. First, the pre-prandial test was conducted, and the corresponding sensor data were collected every 10 min resulting into four distinct readings. At each trial instant, the measurement was repeated for three times while placing the fingertip and the average of was plotted in Fig. 13f in terms of peak amplitudes (blue curve). The invasive readings were collected accordingly using the Glucometer and plotted in Fig. 13f (red curve). The sensor measurements follow the trend of the reference BGL that increases slightly in the range 93.7–101 mg/dL. In this narrow range of BGL variation, the sensor readings exhibited a repeatability error of about ± 0.0198 V max.

The post-prandial test was performed similarly right after the meal (~ 10 min) by collecting four distinct readings over a period of 30 min. The average of three repeatable sensing trials was plotted in Fig. 13g in terms of the peak amplitudes with a repeatability error of about ± 0.019 V. The invasive measurements revealed a significant jump to 165.7 mg/dL 10 min after the meal intake, then dropped slightly to 155, 146, then 10 mg/dL by the fourth check performed 40 min after the meal. The sensor readings follow this descending pattern as depicted in Fig. 13g. The sensor results for both pre- and post-prandial measurements shows no delay compared to the reference BGL, thus indicating the direct BGL monitoring from blood. The in-vivo measurements attained in this study have not investigated the effect of physical activities or the physiological differences between different subjects on the sensor responses; and that will be further explored in future studies on many patients with different diabetes conditions.

Source link

Leave a Reply

Your email address will not be published.