Frequency selective metasurface based radio frequency glucose sensor with a periodic array of meta cells

Abstract

Nowadays, the operating principle of most of the experimental electromagnetic glucose sensors is based on the effect of an anomalous dispersion caused by a direct contact between an object under test and a two-dimensional metasurface. Due to the repeated uses, the metasurface will be subjected to a chemical attack over time. To avoid this problem, this work proposes a novel glucose sensor that is equipped with an interfacial dielectric layer of 0.254 mm thickness to separate the object under test and the metasurface. The results of our analysis suggest that, if the interfacial dielectric layer is sufficiently thin, there will be no shortage of sensitivity in the proposed sensor. Consistent with our theoretical prediction, the proposed sensor was found to resonate at 9–10 GHz, with its resonant frequency responding to the glucose concentration in a dose dependent manner. The correlation the resonant frequency shift and the glucose concentration was found to be highly linear over the clinical diabetic range with a sensitivity of 4 MHz/(mg dL⁻¹). However, we could not obtain the same or similar sensitivity when the interfacial layer was substituted with a similar dielectric layer of 1 mm thick. Overall, the presence of the interfacial dielectric layer has not negatively impaired the sensing sensitivity of the glucose sensor if and only if its thickness was sufficiently small. This implication of this work can be advantageously used to tailor the future invivo glucose methodology or other health-care electronic devices.

Introduction

Glucose sensing based on EM waves is an active area of research, particularity in the field of health care electronics. In contrast with what was perceived by some corners of the research community, there has been no shortage of research interests in this area. During the Covid pandemic, some research teams have even extended the existing RF/microwave glucose sensing methodologies to detect coronavirus in a sample^1–3.

The permittivity of a liquid loaded with glucose is a function of the glucose concentration. A change in the glucose concentration will accordingly induce a change in the permittivity. In theory, this fact can be advantageously used for measuring the glucose concentration in a test sample. However, the problem of a glucose measurement is not as simple as such. There are many known challenges which render the measurement process difficult, if not impossible. One of the challenges is the sensitivity. The permittivity of a glucose-load sample is simply insensitive to the change in the glucose concentration⁴. Within the clinical diabetic range, the actual change in the permittivity in response to a large change of the glucose concentration is often too small to be observable. The correlation between the permittivity of a glucose-load sample and the concentration is simply too weak to tell what the glucose concentration is in a test sample.

Although the permittivity of a test sample is often weakly dependent on the glucose concentration, the glucose concentration is highly dependent on its anomalous dispersion. Dispersion is the dependence of the permittivity on the frequency of an applied electric field. Anomalous dispersion happens when the real part of the permittivity decreases from a positive value to a negative value as increasing frequency⁵. Anomalous dispersion does not occur unless the operating frequency is right in the neighbourhood of a resonant frequency. Nowadays, most of the experimental glucose sensors are based on the effect of an anomalous dispersion because of a direct contact between a test sample and a sensing area^6–11. Here, the sensing area is usually a two-dimensional metasurface that contributes to a dispersion during a resonance. The effect of dispersion enables the glucose concentration to be determined in terms of a resonant frequency shift, either manually with a vector network analyser or automatically with the help of a gain/phase detector^12,13.

At RF/microwave frequencies, the metasurface is usually based on a matrix of unit cells made with a combination of metal and slots. According to our tests, however, the chemicals and ions in the test sample can attack the metasurface overtime, rendering it deteriorated through rusting or oxidation. Over time, the sensor will lose its originally intended performance, and the sensing performance will unlikely remain unchanged. At the very least, the Q-factor of the sensor will be negatively affected by the metallic contaminant accumulated on the metasurface. Contact-induced contamination not only degrades the sensing performance, but also induces some other seemingly unrelated problems such as cross-infection^2,3. This is one of the reasons why, during the time of measurements, the sensing head needs to be constantly cleaned with a phosphate-buffered saline solution. Many other research groups have produced a reproducible and promising results^14–16.

On the other hand, the sensing area of most of the glucose sensors of this date has been fabricated directly on a thick dielectric substrate. The other side of the sensing area is normally put in contact with the object under test during a glucose measurement. The degree of dispersion depends on the effective permittivity, which, according to Maxwell-Garnett approximation¹⁷, can be determined from the volume ratio, the plasmonic permittivity of the resonant elements on the sensing area, the substrate’s permittivity and the permittivity of the object under test.

To overcome the above-mentioned problems, we propose another glucose sensor based on the similar operating principle in this work, but the proposed sensor eliminates the need for any direct contact between the metasurface and the object under test. The metasurface is physically protected by a thin layer of interfacial dielectric material. In one of our previous investigations⁹, we found that the fat layer underneath the human skin is an insulating layer that renders the glucose sensor insensitive to blood glucose concentration in our invivo measurements. However, the proposed sensor has exhibited an almost equally good sensitivity in this investigation. As shown in Table 1, which summarizes the performance of the cited glucose sensors, the sensitivity of the sensor proposed in this work is no less than those of other counterparts published in recent years^6–11,18,19.

Table 1.

Comparative assessment of the cited glucose sensors based on the resonant frequency shift approach.

Ref.	Sensor	Frequency	Measured parameter	Sensitivity	Detection limit
6	RF resonators	~ 2 GHz	S11	1.99 MHz/ (mg dL−1)	0.033 µM
7	Rectangular meandered line resonator	9.2 GHz	S21	1.08 MHz/(mg dL−1)	8.90 mg dl−1
8	Double Split-ring resonators	1.4–1.8 GHz	S21	0.3287 MHz per mM for resonant frequency shift method; 0.4375 MHz per mM for the bandwidth change method	0.75 mM for the bandwidth change method
19	LC (inductor–capacitor) tank resonator	0–1 GHz	S11	0.15 MHz/(mg/dL)	Not specified
9	Bifilar Spiral Resonator	~ 650 MHz	S11	1.3 MHz/(mg dL−1) & 0.0251 /(mg dL−1)	< 0.5 mg/ml
10	Planar Split-ring resonator	2 –2.12 GHz	S11	3.72 MHz/(mg dL−1) for ex-vivo; 1.25 MHz/(mg dL−1) for in-vivo	< 0.05 wt%
18	Stepped-impedance resonators	6.53 GHz	S11	9.787 MHz/(mg dL−1)	0.01928 µM
11	Multilayered Split-ring Split-ring	1.5–2 GHz	S11	2.25 MHz/(mg dL−1)	< 250 mg/dL
14	Split ring resonator	1.156 GHz	S11	5 KHz	18 mg/dL
15	Coupled resonators	3 GHz	S11	4 dB / (mg/dL−1)	--
16	Split ring	3 GHz	S21	0.3 dB /(mg mL−1)	--
This work	Resonant cavity fabricated on a coated metasurface	9–11 GHz	S11	4 MHz/(mg dL−1)	< 25 mg/dL

Methodology

The proposed glucose sensor was designed for measurements of the glucose concentration in an unknown liquid. While the sensor can theoretically be used for measuring the concentration of other chemicals, our focus in this work was placed on in-vitro measurements of the glucose concentration only.

The details of the proposed glucose sensor are shown in Fig. 1. The proposed glucose sensor was a (20 mm × 10 mm) printed circuit fabricated in 0.254 mm thick flexible Duroid RT5880 substrate. As shown in Fig. 1a, the metallization of the top of the substrate was a microstrip line connecting port 1 and port 2. The microstrip line was closely coupled with two complementary rectangular resonators. Figure 1c shows the dimensions of the proposed design (the top view and the bottom unit cell of metasurface). As can been seen in the photo of Fig. 1d, the metallization of the complementary resonators has been intentionally thickened by electroplating with copper, forming two complementary rectangular containers of a height of 1 mm. The longest length of each of the complementary rectangular resonators was exactly 13 mm. The bottom metasurface is shown in Fig. 1e. During the time of measurement, ports 1 and 2 of the proposed sensor were connected to the vector analyser connected in the manner as depicted in Fig. 1f, and both complementary resonators were fully filled with a test glucose solution using a precision micropipette.

Fig. 1.

Proposed glucose sensor. (a) The layout of the metalization on the top of a PCB (brown is copper); (b) the layout of the metalization on the back side of the PCB (blue is removed/etched part); (c) dimensions of the proposed unit cell; (d) the prototype showing the top of the proposed glucose sensor; (e) the prototype showing the back side of the proposed glucose sensor; (f) measurement setup and the cross-sectional view of the proposed sensor.

As shown in Fig. 1b and e, the back side of the substrate was a frequency selective surface with a matrix of dipole unit cells, which is referred hereafter as a metasurface. The purpose of this reflective metasurface was to preferentially improve the Q-factor of the resonance. Each unit cell was a I-shaped slot which was further surrounded by a rectangular split slot ring, thereby yielding a negative permittivity and a negative permeability during a resonance. The metasurface is by default reflective when it is in non-resonant state. The resonant frequency of the unit cell was largely dependent on the circumference of the rectangular split slot ring.

The single unit cell dispersion analysis shows the operating frequency range around 11 GHz as obtained from the Eigen mode solver in the 3D simulation software. The single unit cell was optimized carefully to obtain the desired frequency. The single unit cell then periodically extended to (7 × 15) element array that provides the best radiation beam characteristics along with desired operating frequency. This optimized array supports the beam propagation in the upward direction as shown in Fig. 2 of the manuscript.

Fig. 2.

Radiation pattern of the proposed sensor (Obtained from 3D EM simulation software CST Studio Suite: https://www.3ds.com/products/simulia/cst-studio-suite).

According to our electromagnetic simulation, the resonant frequency of each unit cell was roughly 11 GHz. At frequencies very far from the resonant frequency, the unit cell was reflective in the Z-direction, thereby blocking off the electromagnetic energy from the microstrip line in the Z-direction. At frequencies very close to the resonant frequency, the split slot ring together with the I-shaped slot exhibited a negative permittivity and a negative permeability, behaving as a left-hand metamaterial that is transparent in Z-direction. In other words, the frequency selective surface has acted like a perfect ground plane at frequencies far from 11 GHz, thereby enabling the microstrip line on the top of the substrate to operate in TEM modes as was originally intended. At frequencies very close to 11 GHz, however, the frequency selective surface was completely transparent and reflectionless in the z-direction, with the microstrip line acting like a Goubau line operating in TM modes. During the resonance, the complementary resonators in vicinity of the microstrip line have acted like a broadside antenna radiating electromagnetic energy in the positive z-direction.

The backside array consists of periodic unit cells optimized to obtain resonance around 10 GHz frequency. The single unit cell dispersion analysis shows the operating frequency range around 10 GHz as obtained from the Eigen mode solver in the 3D simulation software. The single unit cell was optimized carefully to obtain the desired frequency. This optimized array supports the beam propagation in the upward direction as shown in Fig. 2 of the manuscript. The thickness of the substrate is inversely proportional to the frequency of operation. The thickness is kept at 0.254 mm in this sensor design so as to match the 50-ohm impedance and to design a compact thin platform. The loss tangent separately has not been studied since this area leads to material study which is not the main target of or design. We chose the Rogers Duroid 5880 substrate which is considered as one of the best for antenna and sensor designing at GHz frequencies with low dielectric constant. Increasing or decreasing the thickness of the substrate will result in change in desired operating frequency from 10 GHz to some another frequency. The two rectangular slots on the top have been optimized well to keep the radiating beam in the upward direction supporting the backside array. The rectangular slots have been made bit thicker so as to contain the glucose solution for measurements and to stop the flow out of the desired sensing region. The distance between the rectangular slots and the microstrip line was kept at 0.1 mm so that these get excited by the coupling effects easily. If the distance increase, then the coupling effect will reduce and the rectangular slots containing the glucose solution will not get excited properly resulting in low sensitivity.

The proposed glucose sensor was experimentally found to resonate at 11.5 GHz under an unloaded condition. In the neighbourhood of 11.5 GHz, the microstrip line on the top of the sensor has behaved as a Goubau line operating at TM modes, with a propagation constant very close to the propagation constant of a free space. On the other hand, the thickness of the substrate was only 0.254 mm. Regardless of the permittivity of the substrate, the speed of light at around the resonant frequency was expected to be very close to the speed of light in free space. On the other hand, the longest physical length of the complementary resonators was both 13 mm, corresponding to a free space wavelength of exactly 11 GHz.

During a process of measurement, the solution under test was injected using a pipette into the rectangular openings of the complementary resonators until both complementary resonators are fully filled with the test solution. At the about same time, Ports 1 and 2 of the sensors will be connected to a vector network analyser. The resonant frequency at which a minimum reflection coefficient S11 occurred was manually sought using the vector network analyser so as to determine the glucose concentration in the test solution.

Glucose solutions are a highly dispersive and viscous substance that dissipates energy in the form of heat. Without the frequency selective surface, the complementary resonators loaded with a test glucose solution would have behaved like a low-Q resonator. In conjunction with the frequency selective surface, however, the complementary resonators exhibited a high-Q characteristic, with a Q-factor in the neighbourhood of 96. In the positive z-direction, the resonator acted like a broadside antenna that radiated out radiations in manner as depicted by the simulated result in Fig. 2 obtained from 3D electromagnetic (EM) simulation software CST Studio Suite. As shown in Fig. 2, the outgoing radiations were much confined in the complementary resonators.

The permittivity of the test solution can be conveniently modelled using Cole-Cole model¹ as:

1

where is the static dielectric permittivity, τ is the relaxation time of the dipole, is the frequency in radian and is the high frequency permittivity. α is an adjustment factor with a value between 0 and 1. In most cases, α is zero. σ is the static ionic conductivity of the glucose-loaded human blood. At microwave frequencies or above, the last term of Eq. (1) should be negligible. The main contributor to the effect of dispersion is the relaxation time of the dipole, τ.

One of the main objectives in this work was to make sure the sensor was highly sensitive to a glucose concentration change in the test solution in the absence of any form of direct contact between the metasurface of a sensor and the test sample. In the proposed sensor, the interface between the metasurface and the test solution was obviously non-existent. Instead, the metasurface and the test solution were separated by the Duroid RT5880 substrate of 0.256 mm thick. At this stage, what we need is to determine the effective permittivity of the bulk volume formed by the combination of the RT5880 substrate and the glucose-loaded test solution using Maxwell-Garnett approximation¹⁷.

Figure 1e shows the cross-sectional view of the proposed section as well as the connected ports setup. The thickness of the RT5880 substrate, , and the height of the complementary resonator were respectively 0.254 mm and 1 mm. The relative permittivity of the RT5880 substrate, , was fixed at 2.2, to which the corresponding wavelength at 11 GHz, , is 17.5 mm. The relative permittivity of the glucose loaded solution was a function of frequency, which is denoted as . On the other hand, the height of each of the rectangular complementary resonators, , was 1 mm. The Since , Maxwell-Garnett approximation is validly applicable²⁰. The volume ratio of the RT5880 substrate to the glucose-loaded test solution is given by . According to Maxwell-Garnett approximation, the permittivity of the contact interface in the X-Y plane can be obtained using:

2a

Since is very small compared to H, is approximately equal to 1 + . The cascade of two dielectric layers is equivalent to two capacitors connected in series. With this in mind, the effective permittivity in the Z-direction can be obtained using:

2b

According to Eq. (2b), the correlation between and is limited due to the finite value of h. Hence, for the dependency of and to be strong, the thickness of the RT5880 substrate must be much thinner than the value of .

This contact interface between the Duroid substrate and the frequency selective surface can be further modelled as a metasurface with an array of aligned unit cells having a negative plasmonic permittivity, , at frequencies very close to the resonant frequency. Here, the metasurface serves as a inclusion layer. The interface of this metasurface and the bulk volume yield another effective permittivity . To determine , we need to find the solution of quasi-static equations in the x-y plane as well as the boundary conditions which ensure the continuity of the electric field in the z-direction. Whilst the analytical solution of this problem is in general complicated, we can still use the Maxwell-Garnett approximation to obtain if the thickness of the metasurface is sufficiently thin:

2c

2d

where t is the effective thickness of the metasurface, is the excitation electric field and is the electric field inside the plasmonic inclusion given by .

Equation (2c) and (2d) clearly suggest that and are still a strong function of the permittivity of the glucose-loaded test solution if t is sufficiently small.

At frequencies very close to the resonant frequency, , the effective permittivity of the metasurface can be more easily modelled using the classical concept of anomalous dispersion⁵ in the following manner:

3

Here, is the plasma frequency of the sensing area formed by the metasurface. γ_m is the loss factor of the proposed glucose sensor at the m-th resonance. Typically, the loss factor is the reciprocal of the quality factor, i.e. . m is the index of the resonance, S_m is the oscillator strength of the m-th resonance, and S_o is the oscillator strength of the highest resonance.

Due to the presence of glucose in the test solution, the change in the glucose concentration will perturb the permittivity by perturbing the resonant frequency, . The derivative of (3) against the effective permittivity can be written as:

4

Equation (4) can be rearranged as:

5

Hence, if the glucose concentration is C, then the resonant frequency shift as a result of a change of the glucose concentration, , which is also known as sensitivity, can be empirically written as:

6

Or equivalently, can be substituted with , where and are respectively the real and imaginary parts of :

7

Since the sensitivity is a real number, Eq. (7) should be rewritten as:

8

has been proven to be negative, the sensitivity should be positive. It is very clear from Eq. (8) that the resonant frequency shift due to a change in the glucose concentration is proportional to the loss factor, , or inversely proportional to Q-factor of the metasurface (or the sensing surface). The Q-factor can be determined from the S11 parameter using the formula , where is the resonant frequency and is the bandwidth at 3 dB above the resonant magnitude. Obviously, the sensitivity of the glucose sensor can be further improved by increasing the loss factor, , lowering the Q-factor, and/or simply by lowering the resonant frequency, . The Q-factor can also be further engineered by changing the design of the frequency selective surface.

Results and discussion

Figure 3a shows the prepared glucose concentrations for experiments and Fig. 3b shows the measured reflection coefficient as a function of frequency under two conditions: (1) when the sensor was unloaded; and (2) when the sensor was loaded with a pure deionized water. In the absence of any liquid in the sensing area, which was the rectangular cavities in vicinity of the stripline connecting port 1 and port 2 (i.e. condition 1), the fabricated sensor was found to resonate at around 11 GHz with a minimum magnitude of S11 down to −60 dB. With an introduction of a deionized water onto the rectangular cavities next to stripline (i.e. condition 2), the measured resonant frequency has downshifted to 9.5 GHz.

Fig. 3.

(a) Prepared glucose concentrations (25 mg/dL–500 mg/dl). (b) Measured reflection coefficient as a function of frequency.

The measured insertion loss (as denoted as S12) is given as a function of the frequency in Fig. 4. Here, it can be seen that the proposed sensor is a broadside antenna with a insertion loss of slightly less than 2 dB over the stopband. The proposed sensor is not used as an antenna, the insertion loss at the stopband is not relevant.

Fig. 4.

Measured insertion loss as a function of frequency.

As shown in Fig. 5, filling the cavities with a glucose solution has noticeably increased the resonant frequency in a dose dependent manner. Whilst the predicted resonant frequencies were easily obtained by simulation, we were not able to simulate this glucose induced increment in the resonant frequency using any commercially available electromagnetic simulator. Consistent with our theoretical prediction as described in the previous section, the correlation between the measured resonant frequency and the glucose concentration turned out to be highly linear.

Fig. 5.

Measured resonant frequency of S11 parameter with varying glucose concentrations (50 mg/dl, 75 mg/dl, 100 mg/dl, 125 mg/dl, 150 mg/dl, 175 mg/dl).

The sensitivity of this sensor can be defined as how much change the resonant frequency undergoes in response to a unit change of glucose concentration. This parameter can be readily obtained from the slope of the graph in Fig. 5. The S11 resonant frequency shifts when the glucose concentration varies i.e. at 50 mg/dl, 75 mg/dl, 100 mg/dl, 125 mg/dl, 150 mg/dl, 175 mg/dl. By direct inspection on Fig. 5, the sensitivity of the proposed sensor was approximately 4 MHz/ (mg/dl). Though the tests were performed for prepared concentrations from 25 mg/dl to 500 mg/dl but only the clinical diabetic range has been chosen to depict useful results. Therefore, Fig. 6 further shows results for multiple measurements performed (10-times) i.e. the shifts in resonant frequencies for various glucose concentrations from 50 mg/dl to 175 mg/dl with error bar. The repeated experiments showed that the detection limit obtained is less than 25 mg/dL of glucose solution concentration which is good in terms of sensing capability.

Fig. 6.

Multiple measurements performed for same concentrations with error bar and measured resonant frequency of S11 parameter with varying glucose concentrations.

On the other hand, the same design of metallization has been fabricated with another similar but much thicker dielectric substrate in another experiment. The thickness of this dielectric layer was 1 mm. The S11 parameter have remained at around 10 GHz but the sensor with a thicker dielectric substrate has failed to yield a significant and highly observable resonant frequency shift even though the glucose concentration has significantly changed. This observation suggests that the sensing sensitivity is critically dependent on the thickness of the interfacial dielectric layer that separates the metasurface and the object under test.

In one of our previous work⁹, the resonant frequencies of our invivo measurements were ways less than those of our invitro measurements simply because of the presence of a thick fat layer underneath the skin. In this work, however, the sensitivity was found to be 4 MHz/ (mg dL⁻¹) even though the glucose-loaded water and the metasurface were separated by the interfacial dielectric layer, RT5880 simply because the interfacial dielectric layer is sufficiently thin. Equation (2a) and (2b), in conjunction with the measured results, have clearly proven the fact that the interfacial dielectric layer that separates the glucose loaded solution and the metasurface does not overly impair the sensing sensitivity of the glucose sensor if its thickness is sufficiently small.

Unlike other contact-based glucose sensors, the measured frequency response was found be relatively more stable against the ambient temperature over time although the ambient temperature has fluctuated at night. We believe that the thickness of the metallic wall of the cavity’s resonators have to some extent helped stabilize the temperature of the glucose solution.

Obviously, this implication of this work can be advantageously used to tailor the invivo glucose methodology. For example, instead of measuring the blood glucose levels on a human arm, which has a thick layer of fat that has contributed a very poor sensitivity in one of our previous studies, we can conduct our invivo blood glucose measurement on a human tongue, which has much less fat. This implication may well be subjected to other limitations that has not been identified in this work. Therefore, more research needs to be done. Table 1 further shows the comparative assessment of other glucose sensors reported.

In summary, the proposed design is quite useful for the measurements of glucose solutions. The sensor is mainly supported by a metasurface that has provided it a strong base to obtain desired frequency of operation and beam propagation to a particular upward direction. This design though proved to a good for glucose-based solution testing but it can be used for blood serum solution testing also as future scope of this work. Further the research work can be extended by modifying the design to work directly on the body as a wearable system. System level integration is possible in case the design can be made with more compact dimension by following the proposed design idea. The design can also incorporate/combine some ideas proposed in^16,21 in future works to enhance the sensitivity. Since this type of RF microwave designs are made on substrates that have high durability and tested for electronics research, these provide robust design option with low or negligible power consumption. Moreover, a power recycling sensor can be designed to avoid any kind of power loss as a future concept.

Conclusion

The proposed glucose sensor was successfully realized to enable the glucose concentration of a solution to be measured in the absence of any direct contact with the sensing area, with a sensitivity of 4 MHz/ (mg/dl). The proposed glucose sensor had a interfacial dielectric layer that separated the glucose loaded solution and the sensing surface, which was a reflective metasurface. Consistent with our theoretical prediction, the proposed sensor resonated at 9–10 GHz, with its resonant frequency responding to the glucose concentration in a dose dependent manner. The correlation the resonant frequency shift and the glucose concentration was found to be highly linear over the clinical diabetic range. In conjunction with the theoretical prediction, the experimental outcome suggests that the presence of the interfacial dielectric layer has not overly impaired the sensing sensitivity of the glucose sensor if its thickness was sufficiently small. This implication of this work can be advantageously used to tailor the invivo glucose methodology.

Author contributions

A. K: Conceptualization, Methodology, Simulations; S.D: Methodology, Simulations; L.WY.L: Simulations, Data curation, Experiments, Writing; Z.N: Supervision, Methodology, Experiments; R.J: Review, Writing; C.C.K: Modelling, Editing; A.M.A: Writing, Review, Editing; H.V: Editing, Reviewing. All authors reviewed the manuscript.

Data availability

Data sets generated during the current study are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.