Design of a microwave sensor for non-invasive monitoring of blood glucose level with high sensitivity using electromagnetic properties

Abstract

This paper introduces a novel, portable microwave sensor for rapid, non-invasive blood glucose monitoring. The design features an octagonal array of complementary split ring resonators (CSRRs) on a dielectric substrate, operating safely in the industrial, scientific, and medical (ISM) frequency band. Its key innovation, an engineered 180^∘ phase difference between adjacent unit cells, generates a highly concentrated electromagnetic (EM) field at the sample interface. This focused interaction significantly enhances measurement sensitivity and overall detection capability. The sensor accurately detects glucose concentrations across the 50–500 mg/dL clinical range, demonstrating a remarkable sensitivity of 2.3 MHz/(mg/dL) in laboratory settings and 1.78 MHz/(mg/dL) in realistic scenarios, surpassing existing microwave sensors. This superior performance is attributed to the CSRR architecture, which maximizes the sample’s EM field interaction, enabling the precise quantification of subtle dielectric changes corresponding to varying glucose levels. Laboratory verification using a vector network analyzer (VNA) confirmed significant frequency shifts with glucose samples from 80 to 340 mg/dL. Beyond its high sensitivity, the sensor’s compact size, simple fabrication, affordability, and non-ionizing operation establish it as a promising candidate for developing practical, real-time, non-invasive glucose monitoring systems to advance diabetes management.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-41378-6.

Introduction

Diabetes is a chronic disorder in which the body cannot correctly produce or use insulin to regulate blood glucose levels¹. Maintaining blood glucose within the normal range (70–130 mg/dL before meals and under 180 mg/dL after meals) is crucial, as failing to do so can lead to hyperglycemia or hypoglycemia, resulting in severe complications, including cardiovascular disease, nerve damage, and death². Consequently, frequent glucose monitoring is vital for management. Traditional finger-prick testing is often painful and inconvenient. An alternative, Continuous Glucose Monitoring (CGM), uses a sensor inserted under the skin to measure glucose in the interstitial fluid. However, CGM systems are often limited by high costs, discomfort, low stability, and inaccuracies arising from the difference between interstitial fluid and blood glucose levels^3–6. These challenges create a strong demand for non-invasive monitoring techniques to improve patient comfort and adherence. The urgency is emphasized by data from the International Diabetes Federation (IDF), which shows that 589 million adults (1 in 9) have diabetes as of 2025, with over 40% being undiagnosed. The IDF projects this number will rise to 853 million by 2050. This growing prevalence, combined with the flaws of current methods, underscores the critical need for accurate, non-invasive blood glucose sensors⁷.

Non-invasive blood glucose monitoring: recent advances and challenges

Significant research has been conducted to develop innovative non-invasive (NI) glucose detection systems, aiming to provide more comfortable and continuous monitoring. Most of these systems are based on a variety of optical, spectroscopic, and electromagnetic methods, including near-infrared (NIR) technologies like IoT-enabled R-NIR sensors⁸, mbNIR sensors paired with neural networks⁹, photoplethysmography (PPG)¹⁰, custom diode arrays¹¹, and commercial NIRS systems¹². Other approaches utilize mid-infrared (MIR) spectroscopy through passive imaging¹³, tuneable quantum cascade lasers (QCLs)^14,15, and IR sensors with fuzzy logic¹⁶. Foundational work has also been conducted to create high-power MIR frequency combs for spectroscopy¹⁷. Further techniques include photoacoustic multispectral systems¹⁸ and microscopic photoacoustic spectroscopy¹⁹, Raman spectroscopy with individualized calibration models²⁰, polarization-based sensing with machine learning²¹, and low-power polarization-switching²². Additionally, AI-driven terahertz (THz) metasurface biosensors²³, and multisensor systems measuring impedance and temperature²⁴. Additionally, readily accessible biological fluids, such as tears²⁵, saliva^26–29, sweat^30–34, urine^35,36, and respiratory moisture, have been employed in numerous enzyme-based electrochemical techniques to establish a correlation between their glucose content and glycemic levels. However, these methods often exhibit a delayed correlation with actual blood glucose variations and are susceptible to metabolic fluctuations.

Radio frequency (RF)/microwave techniques are a promising avenue for non-invasive glucose monitoring. By measuring how glucose affects the dielectric properties of blood, these methods circumvent issues that plague optical sensors, such as interference from skin pigmentation or temperature fluctuations. RF waves also penetrate tissue more deeply than light, allowing for more reliable and real-time measurements. Among these technologies, resonant-based sensors are highly effective. They operate by detecting shifts in resonance characteristics (like frequency, amplitude, or quality factor) caused by changes in a sample’s permittivity. A significant advantage of this approach is its potential for simple, low-cost physical structures that can be miniaturized for wearable devices^37,38. However, the primary challenge remains achieving sufficient sensitivity to detect the minimal variations in blood glucose found in the human body³⁹. For a sensor to be viable, it must be validated under realistic conditions. Testing in simple water-based solutions is insufficient, as it doesn’t replicate the complex, “lossy” nature of biological fluids. Therefore, performance must be evaluated with tissue-mimicking phantoms⁴⁰ and, most critically, with actual human blood samples to ensure dependable results^41,42. Practical factors, such as operating within the physiological glucose range and considering the impact of sample volume and sensor placement are also crucial for demonstrating real-world performance⁴³. Overcoming these challenges necessitates sophisticated engineering based on the principles of electromagnetic wave interaction with biological tissues⁴⁴. A cornerstone of this approach is the strategic use of specialized resonant structures, notably Split-Ring Resonators (SRR) and Complementary Split-Ring Resonators (CSRR). These meticulously engineered structures exhibit a sharp resonant response and can confine a highly intense electric field within a tiny region. By placing the sample in this “hotspot,” the wave-sample interaction is maximized, producing a larger and more easily detectable resonance shift even for tiny changes in glucose, thereby dramatically boosting sensitivity. Successful applications of this principle are evident in high-sensitivity chipless sensors for wearables and advanced CSRR structures, which achieve exceptional performance through impedance matching⁴⁵. The next generation of these sensors integrates this core principle with other technologies. This includes using active electronic components to compensate for signal loss in blood⁴⁶ and applying advanced data processing, such as Convolutional Neural Networks (CNNs), to interpret complex sensor data accurately⁴¹. The ultimate goal is to develop a highly sensitive and reliable sensor platform that is structurally simple, cost-effective, and suitable for integration into wearable devices.

Sensor design and operating principle

The proposed microwave glucose sensor is based on a dual-octagonal complementary split-ring resonator (CSRR) configuration designed to enhance electromagnetic field confinement and sensitivity to dielectric variations in blood. As shown in Fig. 1a, the sensor consists of two identical octagonal CSRR unit cells etched on the copper ground plane and excited through a microstrip feeding network. The two resonators are driven with a controlled 180° phase difference, which is achieved by implementing a dual-branch microstrip feed in which one branch is deliberately extended to introduce the required phase delay. This engineered phase opposition enforces an antisymmetric coupling mode between the paired CSRRs, leading to strong electric-field confinement within the inter-cell region where the blood sample is positioned. As a result, the electromagnetic energy is preferentially concentrated in the sensing region, maximizing the interaction between the resonant field and the dielectric properties of the sample. The operating frequency was selected within the ISM band near 5 GHz to balance sufficient penetration depth in biological tissue with compatibility for wearable medical devices. The geometry of the proposed sensor was developed through a systematic analytical-to-numerical workflow. Initial estimates of the CSRR dimensions were obtained using standard resonance relations for CSRR-based structures in the 5 GHz range. These analytical dimensions were subsequently refined through iterative full-wave CST simulations. During the optimization process, key parameters—including the CSRR ring radius, split and gap widths, spacing between the two resonant cells, and the length difference between the dual feed branches responsible for the 180° phase shift—were systematically varied. Each iteration was evaluated based on its impact on the resonance frequency, electric-field confinement, and glucose-induced perturbations. This procedure resulted in convergence at a dominant resonance around 5.5 GHz. Driven frequency-domain simulations performed under the actual excitation conditions confirmed that the fourth resonance (f₄ ≈ 5.5 GHz) exhibits the strongest electric-field confinement and a clear phase reversal between the two resonant cells (Fig. 1b). The corresponding electric-field distribution (Fig. 1c) reveals a highly localized and intense field within the CSRR unit cells and the inter-cell sensing region, which is essential for achieving high sensitivity to small dielectric changes. Additionally, Fig. 1d provides a detailed analysis of the field distribution at the resonance frequency, highlighting the variation in sensitivity across the sensing region.

Fig. 1.

Structural configuration and electromagnetic characteristics of the proposed dual-octagonal CSRR glucose sensor.

The sensitivity enhancement mechanism can be further interpreted using a perturbation-based formulation. Under antisymmetric excitation (Δφ = 180°), the strong coupling between the two CSRR unit cells increases the effective electric-field energy stored in the sensing region, thereby enhancing the effective coupling capacitance. According to first-order perturbation theory, the resonance frequency shift of a dielectric-loaded resonator can be approximated as

1

where f_₀ is the unperturbed resonance frequency and denotes the effective permittivity experienced by the resonant mode. The engineered phase opposition amplifies the field–sample interaction, increasing induced by glucose-dependent dielectric variations in blood. Consequently, a larger resonance frequency shift is produced per unit change in glucose concentration, resulting in enhanced sensitivity.

(a) Photograph of the assembled microwave sensor, including the standalone prototype, aluminum enclosure, SMA connectors, and a glass container filled with blood. (b) Schematic configuration of the CSRR sensor, indicating the copper ground plane, microstrip feed line, octagonal unit-cell topology, and key geometric parameters used for optimization. (c) Electric field distribution on the resonator surface at 5.5 GHz, obtained from CST simulations. The figure includes the carpet plot, isoline representation, and the electric field lines in a plane perpendicular to the x-axis at the location of the resonator unit cells. In both the carpet and isoline plots, a strong concentration of the electric field is observed in the unit cell regions, indicating high field intensity. The electric field magnitude scale in all three visualizations ranges from 0 to 80 dB(V/m). (d) Side-view visualization of the electric field lines around the sensor and the glass vial containing the blood sample. The field-line plot is shown qualitatively to indicate the propagation paths and coupling mechanism, and does not represent the electric field magnitude.

The table lists the key structural dimensions used in the CST model, including resonator radius, split gap, substrate thickness, and feed-line width, all optimized for resonance near 5.5 GHz within the ISM band.

Results and discussion

This section provides a detailed description of the proposed microwave glucose sensor, covering its design, key parameters, numerical simulations, and experimental validation. The sensor is designed to operate at approximately 5.5 GHz within the ISM band, a frequency chosen to ensure that electromagnetic waves penetrate sufficiently to reach the tissue. This choice optimizes penetration depth while minimizing tissue losses, thereby enhancing the sensor’s sensitivity to subtle changes in blood’s dielectric properties associated with varying glucose concentrations⁴⁷. These variations in the dielectric constant and loss tangent were studied from 1 to 7 GHz using the first-order Debye relaxation model, which shows a slight increase in dielectric constant and a corresponding decrease in loss tangent as glucose concentration rises⁴⁸. As depicted in Fig. 1c, the electric field lines at the resonant frequency of 5.5 GHz form a closed loop between two unit cells. This configuration, where the lines originate from one unit cell, traverse the glass and blood, and subsequently return to another unit cell, creates a highly concentrated electric field and a strong coupling between the unit cells. Consequently, the microstrip structure’s frequency response reaches its maximum potential effectiveness. The sensor’s physical architecture features two identical octagonal Complementary Split-Ring Resonators (CSRRs). As illustrated in Fig. 1a, these resonators are etched onto the 35 μm thick copper ground plane of a Rogers 4003 PCB, which has a dielectric permittivity (ɛ_r′​) of 3.55 and a loss tangent (tanδ) of 0.0027. For practical application and test repeatability, the sensor is housed within an aluminum enclosure. Excitation of the CSRRs is achieved through efficient coupling to a microstrip transmission line (MTL), which was optimized with a width of 1.1 mm, a thickness to 0.035 mm, and an impedance of 50 Ω for optimal power transmission. Two distinct topologies of this design were implemented and optimized: the first for blood samples contained in glass vessels (Fig. 1a) and the second for a simplified wrist model (Fig. 2a). In this configuration, the two octagonal cells are positioned vertically along the MTL axis, separated by a distance of D = 5.2 mm. Each CSRR unit cell comprises two concentric octagonal rings with precisely defined geometric parameters: an outer ring diagonal length (a) of 5 mm, an inner ring length (b) of 4.12 mm, a coupling-split (d) of 0.2 mm, a dielectric split (c) of 0.2 mm, and a metal split-gap (g) of 0.2 mm, with the splits for each ring on opposite diagonal sides. A comprehensive schematic of the sensor configuration, detailing all geometric parameters, is shown in Fig. 1b, with specific values provided in Table 1. To significantly boost sensor sensitivity, a novel approach incorporating a 180° phase shift between the two unit cells was introduced. This phase difference creates the closed-loop electric field configuration, effectively trapping electromagnetic energy within the sensor’s immediate vicinity. As a result, the dielectric properties of loaded samples, such as glucose-laden blood tissue, can be measured with high accuracy. The design and geometric parameters of the planar transmission line and the etched unit cells were therefore meticulously optimized to achieve a sharp transmission resonance around f₀ = 5.5 GHz when the sensor is loaded with a sample.

Fig. 2.

Multilayer wrist-tissue simulation model and corresponding electric-field distribution of the proposed CSRR sensor. (a) CST structural model incorporating skin, fat, and blood layers to emulate near-physiological sensing conditions. (b) Side-view electric-field lines at the 5.5-GHz resonance showing field penetration through tissue layers and the 180° phase difference between the two resonant cells, which enables strong coupling within the CSRR pair.

Glucose detection mechanism

To simulate the frequency behavior of the integrated CSRR sensor, a lumped-element model was developed, as illustrated in Fig. 3^49,50. In this model, each of the two identical octagonal cell resonators is represented by a parallel RLC resonant circuit (L_u1, C_u1, R_u1 and L_u2, C_u2, R_u2). The inductance (L_u) arises from the structure’s dielectric rings, the capacitance (C_u) is formed by the metal splits and spacers, and the resistance (R_u) accounts for conductive and dielectric losses. A coupling inductance L_c1 models the microstrip transmission line (MTL) that excites the resonators. In contrast, the coupling between the MTL and the CSRR structure is represented by a shunt capacitor, C_c1, in parallel with a resistor, R_c1, to account for substrate and conduction losses. A key aspect of the model is the method for creating the 180° phase shift: while the first cell is excited via L_c1, the second cell is coupled to the feed line through an additional, distinct inductor (L_c2), which effectively models the two different excitation paths. Finally, when a sample is loaded onto the sensor, its dielectric properties are incorporated by adding a parallel RC circuit (C_M, R_M) coupled to each resonator. The capacitance (C_M1, C_M2) is directly related to the sample’s relative permittivity, and the resistance (R_M1, R_M2) corresponds to its loss characteristics. Due to the symmetrical design, the values of these sample-related components are identical for both cells.

Fig. 3.

Equivalent lumped-element circuit model and corresponding microstrip layout of the proposed dual-octagonal CSRR sensor under sample-loading conditions. Each resonator is modeled as a parallel RLC branch, while inductors (Lc₁ and Lc₂) represent the feed-line coupling paths that generate the designed 180° phase shift between the two unit cells. The loading effect of the blood sample is modeled by parallel RC networks (Cm || Rm) connected to each resonator, representing the dielectric permittivity and loss characteristics of the sensing medium. The accompanying microstrip view illustrates the physical implementation of these circuit elements on the substrate.

Changes in the dielectric permittivity of the blood samples affect the electric field distribution, which can be observed in the resonance frequency f_R through changes in the effective capacitor C_e CSRR (C_e =C_m||C_u). Therefore, changes in resonance frequency can be used to determine the glucose concentration of the sample. The resistor R_e (R_e =R_m||R_u), which represents the combined resistance of R_m and the CSRR-part R_u, is mainly affected by the loss characteristics of the blood sample. Changes in tanδ are reflected as changes in the amplitude of the resonance profile. These changes in resonance properties are a signature of the dielectric properties of the blood sample, which can be related to the glucose level through analysis of the modified resonance behavior. The arrangement of the lumped elements stores oscillating electric and magnetic energy in the inductance and capacitance. These are caused by the induced charges and currents within the patterned dielectric loops or slots when the CSRRs are excited. When the electric and magnetic energies are balanced, the microwave sensor resonates at a specific frequency, as shown in Eq. (2). This resonance is directly seen as the lowest point in the transmission coefficient S₂₁.

2

To evaluate the performance and resonance frequencies of the proposed sensor under loaded conditions, numerical simulations were conducted using CST Microwave Studio. The simulation modeled a cylindrical glass container, designed to hold 0.5 mL of blood samples on the CSRR surface, as illustrated in Fig. 1a. The container had an outer diameter of 11 mm, an inner diameter of 9 mm, a wall thickness of 1 mm, and a height of 25 mm. The unipolar Debye model (first order), presented by Eq. (3), was used to create numerical models for the dielectric properties of dispersed glucose-blood samples at different concentrations. This model was developed in⁴⁸ based on spectroscopic measurements of 50, 250, 1000, and 2000 mg/dL aqueous solutions collected using a commercial coaxial probe kit connected to a VNA. This model is the most reasonable approximation for the behavior of blood glucose.

3

Equation (3) defines the complex permittivity of the blood solution of glucose concentration (in mg/dL) at the angular frequency . The parameters , , and τ are concentration-dependent Debye coefficients⁵¹, is static conductivity, and is the permittivity of free space. Blood samples with glucose concentrations ranging from S₁-S₁₄ (50–500 mg/dL) were simulated above the sensing area within the glass container. This concentration range encompasses a broad spectrum of diabetic conditions, including hypoglycemia (< 70 mg/dL) and hyperglycemia (> 130 mg/dL). As illustrated in Fig. 4(a), the sensor was first evaluated both with and without the aluminum enclosure to examine the effect of electromagnetic shielding on its transmission response. The comparison shows a negligible resonance shift while slightly improving the amplitude stability due to reduced radiation losses and external interference.

Fig. 4.

Simulated transmission (S₂₁) frequency responses of the proposed CSRR glucose sensor under different glucose concentrations. (a) Transmission frequency response of the proposed CSRR sensor measured with and without the aluminum enclosure, showing improved signal stability due to the shielded configuration. (b) Simulated frequency responses for glucose concentrations ranging from 50 to 140 mg/dL. (c) Simulated frequency responses for higher glucose concentrations from 200 to 500 mg/dL, showing progressive resonance shifts with increasing permittivity of the blood sample.

Subsequently, five distinct resonances were identified in the transmission scattering parameters, centered approximately at f₁ = 2.05 GHz, f₂ = 3.4 GHz, f₃ = 4.15 GHz, f₄ = 5.5 GHz, and f₅ = 6.7 GHz. Resonance frequency shifts were simulated for selected glucose concentrations (Table 1), and the corresponding transmission responses are depicted in Figs. 4(b) and 3(c). The most pronounced frequency variations were observed at the fourth resonance (≈ 5.5 GHz), which therefore became the focus of detailed sensitivity analysis. In addition, all resonances exhibit damping behavior, characterized by a decrease in the amplitude of resonance peaks, primarily due to the absorptive nature of the blood samples.To highlight the linear correlation between glucose concentration and the frequencies of the second to fifth resonances, Fig. 5a and b present linear regression models for each blood glucose range (the first resonance shows almost no change). These results demonstrate a strong linear relationship, suggesting that the sensor can be calibrated for individual patients to accurately measure blood glucose levels continuously.

Table 1.

Optimized geometric parameters of the proposed dual-octagonal CSRR sensor.

Parameter	Value(mm)	Parameter	Value(mm)	Parameter	Value(mm)
a	5	D	5.2	L1	7.45
b	4.12	X	30	L2	2.6
c	0.2	Y	30	L3	11.1
d	0.2	h	0.8	L4	9.1
g	0.2	t	1.1	L5	26

Fig. 5.

Linear correlation between simulated resonant frequency and glucose concentration for the proposed CSRR sensor. (a) Linear fitting of resonance frequency versus glucose concentration for the low-range samples (50–140 mg/dL). (b) Linear fitting for the higher glucose concentrations (200–500 mg/dL), showing consistent frequency downshift with increasing glucose level.

To assess the effective sensing depth and exclude volumetric artifacts, the sensitivity of the proposed sensor was evaluated as a function of sample volume by varying the fill height of the blood sample inside the cylindrical vial from 0.2 to 1.0 mL. As summarized in Table 2, the sensitivity increases progressively for smaller volumes up to approximately 0.5 mL, indicating enhanced interaction as the near-field sensing region becomes filled. For larger volumes, the rate of sensitivity increase gradually diminishes and approaches saturation beyond approximately 0.8–1.0 mL. This behavior indicates that the sensing response is governed by the localized near-field region above the CSRR unit cells, and additional sample volume outside this effective sensing region contributes negligibly to the measured response. These results confirm that the extracted sensitivity primarily reflects localized electromagnetic interaction rather than bulk volumetric effects.

Table 2.

Simulated glucose concentration cases (S1–S14) and corresponding dielectric properties of blood at 5.5 GHz. The table lists the relative permittivity (ε′) and conductivity (σ) values used in CST simulations for modeling blood samples with glucose levels ranging from 50 mg/dL to 500 mg/dL.

Parameter	Glucose (mg/dL)	εr′_blood	Parameter	Glucose (mg/dL)	εr′_blood	σ_blood (S/m)
S1	50	53	S8	120	53.14	6
S2	60	53.02	S9	130	53.16	6
S3	70	53.04	S10	140	53.18	6
S4	80	53.06	S11	200	53.3	6
S5	90	53.08	S12	300	53.5	6
S6	100	53.1	S13	400	53.7	6
S7	110	53.12	S14	500	53.9	6

To simulate a more realistic scenario, a second analysis was performed using a simplified wrist model placed in the sensing area, as shown in Fig. 2a. This model consisted of a = 1 mm thick skin layer, a = 1 mm thick fat layer, and a = 1 mm radius cylinder filled with blood. The detailed structural and electromagnetic parameters of the multilayer model used in the simulation are summarized in Table 3. The model includes skin, fat, and blood layers with the corresponding dielectric constants and thicknesses adopted from the IT’IS tissue property database⁵². The center frequency of the simulation (5.5 GHz) and the glucose concentrations used for each blood layer are also listed for reference. The resulting electric field lines at the resonant frequency of 5.2 GHz (Fig. 2b) show a diminished interaction with the blood compared to the previous lab-based scenario. This reduction in intensity is an expected consequence of the high-permittivity skin (ɛ_r=40) and fat (ɛ_r=10) layers, which absorb and disperse the field. Nevertheless, a significant coupling between the two unit cells persists, ensuring that the electric field intensity interacting with the blood is maximized under these more challenging, realistic conditions. The sensitivity of many prior sensors has been characterized in vitro using phantoms, where performance relies on the stark dielectric contrast between the blood sample and the surrounding air. This high contrast naturally concentrates the electric field within the sample, yielding a strong response. However, this mechanism is less effective for realistic in-vivo applications. The proximity of skin and fat tissue, which have high dielectric constants, disperses the electric field and diminishes its intensity in the target blood vessels, rendering the simple contrast-based approach unreliable. An effective in-vivo sensor must therefore achieve field confinement through deliberate design, creating the focus of the electric field that is robust to such environmental loading effects. Our work introduces a novel approach that accomplishes this by leveraging a 180-degree phase differential between two unit cells. This configuration induces strong electromagnetic coupling, which becomes the primary mechanism for focusing the electric field into the target region. Consequently, the sensor’s high sensitivity is an engineered feature arising from this internal coupling, rather than a passive reliance on the dielectric properties of its surroundings. This methodology represents a significant and necessary departure from previous designs.

Table 3.

Sensitivity of the proposed sensor versus sample volume.

Sample volume (mL)	Sensitivity (MHz/(mg/dL))	Sample volume (mL)	Sensitivity (MHz/(mg/dL))	Sample volume (mL)	Sensitivity (MHz/(mg/dL))
0.2	1.6	0.5	2.3	0.8	2.65
0.3	1.85	0.6	2.45	0.9	2.7
0.4	2.1	0.7	2.6	1	2.7

To investigate the influence of skin thickness on sensor sensitivity, simulations were conducted using a wrist model with varying skin thicknesses of 0.5 mm, 1 mm, and 1.5 mm. The corresponding transmission responses were analyzed. While skin thickness varies due to factors like body position, gender, skin type, age, race, and geographic location, the chosen thicknesses represent a practical range for the human wrist. As depicted in Fig. 6a, three distinct resonances were observed in the scattering responses, centered at specific frequencies: f₁ = 3.1 GHz, f₂ = 5.2 GHz, and f₃ = 6.3 GHz. As illustrated for the 1 mm case in Fig. 6b and c, shifts in glucose concentration correlate directly with shifts in the resonant frequency. Notably, as depicted in Fig. 6b and c, the second resonance in each response provided valuable information about the glucose concentration of the simulated samples beneath the skin and fat layers. However, the analysis confirmed that signal attenuation is directly proportional to skin thickness. This phenomenon weakens the electric field’s interaction with the glucose sample, leading to a decline in sensor sensitivity. The average sensitivity was measured to be 1.85 MHz/(mg/dL) for 0.5 mm skin, 1.78 MHz/(mg/dL) for 1 mm skin, and 1.6 MHz/(mg/dL) for 1.5 mm skin. As further quantified in Fig. 7, a strong linear relationship is observed between the resonant frequency shift and glucose concentration across both the physiological range (50–140 mg/dL) and higher concentrations (200–500 mg/dL).

Fig. 6.

Simulated transmission (|S₂₁|) frequency responses of the multilayer wrist-tissue model incorporating skin, fat, and blood layers. (a) Baseline transmission response of the sensor without sample loading. (b) Simulated frequency responses for glucose concentrations from 50 to 140 mg/dL. (c) Simulated frequency responses for higher concentrations from 200 to 500 mg/dL, showing a gradual downward shift in resonance frequency with increasing blood permittivity.

Fig. 7.

Linear correlation between resonant frequency and glucose concentration for (a) 50–140 mg/dL and (b) 200–500 mg/dL.

Crucially, while this decline in sensitivity is relatively small, it highlights a significant source of inter-individual variability. Such variations, if unaddressed, could lead to inaccurate measurements between different users. Therefore, to ensure clinical accuracy in practical applications, a participant-specific calibration protocol is essential.

In-vitro experiments.

Although the proposed sensor is fundamentally designed for non-invasive operation on the skin surface, extracted blood samples placed inside sealed borosilicate vials were used in this study solely to enable controlled laboratory validation. This approach was chosen to incorporate the natural dielectric contributions of salts, proteins, and metabolites present in human blood, which cannot be accurately reproduced using glucose–water mixtures commonly used in previous studies. Importantly, the blood never came into direct contact with the sensor; the electromagnetic interaction occurred entirely through the glass layer separating the sample from the CSRR structure. To validate the proposed glucose sensor, rigorous laboratory testing was conducted using a Vector Network Analyzer (VNA), as shown in Fig. 8a. Two identical sensor prototypes were used in the experimental setup to assess manufacturing tolerances and record transmission responses for multiple glucose concentrations. Real blood samples from three individuals aged 24, 27, and 32 years were used to mimic clinically relevant glucose levels within the range of 80–320 mg/dL. Each blood sample was divided into identical 0.5-mL vials, and different glucose concentrations were prepared by adding controlled amounts of pharmaceutical-grade dextrose powder. Performing the measurements on samples obtained from multiple volunteers with different blood characteristics (two O⁺ and one B⁺) ensured reproducibility and minimized individual-specific bias in the scattering response. Thus, the invasive aspect pertains only to the sample-collection step required for accurate laboratory validation, not to the sensing principle or intended mode of wearable operation. In future work, the CSRR sensor will be integrated into a wrist-mounted prototype to enable truly non-invasive glucose monitoring.

Fig. 8.

Experimental setup and sample loading for glucose sensing measurements. (a) Photograph of the measurement setup using a vector network analyzer (VNA) connected to the fabricated CSRR sensor. (b) Sensor loaded with a real human-blood sample contained within a glass vial placed directly above the resonator region.

Effect of sample volume on sensitivity

The sensitivity increases with sample volume up to approximately 0.5 mL, beyond which it saturates due to full occupation of the sensor’s near-field region. Further increases in volume result in sensitivity changes below 0.2 MHz/(mg/dL), indicating that volumetric artifacts are negligible beyond this point.

Figure 9 presents the measured transmission coefficients (S₂₁) for the fabricated sensors in both unloaded and loaded states within the 1–6 GHz frequency range. The measured resonant frequencies for both prototypes closely matched the values predicted by simulations. The two sensors exhibited nearly identical performance, with only minor variations in resonance depth and frequency, which are attributed to normal manufacturing tolerances.

Fig. 9.

Measured transmission (S₂₁) frequency responses of the fabricated CSRR sensor under unloaded and loaded conditions. The plots compare the baseline response with those obtained for blood samples having glucose concentrations between 80 mg/dL and 350 mg/dL, showing a clear resonance shift corresponding to glucose variation.

For loaded measurements, cylindrical glass containers were used to hold the prepared blood samples on the sensor’s surface, as shown in Fig. 8b. Placing the empty container on the sensor introduced a baseline frequency shift of a few megahertz from the unloaded resonance frequency. In each experiment, a micropipette was used to dispense a precise 0.5 mL volume of each sample to minimize errors arising from volume uncertainty, and the corresponding shifts in the transmission resonance were recorded. The actual glucose concentration of each prepared sample was independently verified using a commercial blood glucose level (BGL) monitoring device.

The experimental transmission response of the sensor was measured across a range of glucose concentrations (80–340 mg/dL), as shown in Fig. 9. These results revealed two key phenomena: a consistent downward shift in resonant frequency with increasing glucose concentration, and detectable changes in resonance amplitude due to slight variations in the sample’s loss tangent.

It should be noted that the use of extracted blood samples in this study was limited to laboratory validation and does not represent a non-invasive clinical measurement.

The quality factor (Q) of the dominant sensing resonance (f₄) was evaluated from the transmission response using the standard 3-dB bandwidth method. Table 4 summarizes the extracted Q values for the unloaded reference condition, the vial-based laboratory measurements with real blood samples, and the simplified wrist tissue model. In the unloaded reference state, a high Q is observed, indicating low intrinsic loss and effective electromagnetic confinement. When a blood sample is introduced inside a glass vial, the resonance shifts to lower frequencies due to dielectric loading and Q decreases moderately with increasing glucose concentration, consistent with the increased effective dielectric loss of blood. For the wrist tissue model, additional losses from skin and fat layers modify the resonance profile and reduce Q compared with the unloaded case; however, the dominant resonance remains clearly identifiable across the investigated glucose range. Importantly, although Q varies across the different scenarios, the glucose-induced resonance shift remains significantly larger than the corresponding resonance linewidth, confirming that Q degradation does not limit frequency-tracking resolution and that the sensing response remains robust in wrist-representative conditions.

Table 4.

Structural and dielectric parameters of the multilayer tissue model used in simulations, including skin, fat, and blood layers, along with the corresponding thicknesses, dielectric constants, and glucose concentrations at the center frequency of 5.5 GHz (values adapted from the IT’IS database).

Layer	Thickness (mm)	εr′	σ (S/m)	Glucose (mg/dL)
Skin	1.0	40.0	1.46	–
Fat	1.0	10.0	0.23	–
Blood (S1)	1.0	53.00	6.0	50
Blood (S2)	1.0	53.02	6.0	60
Blood (S3)	1.0	53.04	6.0	70
Blood (S4)	1.0	53.06	6.0	80
Blood (S5)	1.0	53.08	6.0	90
Blood (S6)	1.0	53.10	6.0	100
Blood (S7)	1.0	53.12	6.0	110
Blood (S8)	1.0	53.14	6.0	120
Blood (S9)	1.0	53.16	6.0	130
Blood (S10)	1.0	53.18	6.0	140
Blood (S11)	1.0	53.30	6.0	200
Blood (S12)	1.0	53.50	6.0	300
Blood (S13)	1.0	53.70	6.0	400
Blood (S14)	1.0	53.90	6.0	500

Linear regression and statistical validation

To ensure measurement accuracy and repeatability⁵³, each glucose concentration was tested three times, and the mean resonant frequency was used for analysis. This averaging procedure effectively minimized random noise originating from the vector network analyzer (VNA) and other external sources, thereby enhancing the signal-to-noise ratio (SNR). All experiments were performed in a temperature-controlled laboratory environment (25 ± 1 °C) to eliminate temperature-induced variations in the dielectric properties of blood and ensure stable resonant responses. The detailed experimental results for the three volunteer participants are summarized in Table 5 (a–c), showing the measured resonance frequencies at various glucose concentrations. As predicted by the simulations, the fourth resonance (f₄ ≈ 5.5 GHz) exhibited the highest sensitivity to changes in glucose concentration. A linear regression analysis of these results, illustrated in Fig. 10 (a–c), reveals a strong and consistent linear relationship between glucose concentration and resonant frequency. The slopes of the fitted lines correspond to an average experimental sensitivity of approximately 2.3 MHz per mg/dL, in excellent agreement with the simulation-based prediction. The nearly identical slopes among the three participants confirm the uniform sensitivity and reproducibility of the proposed CSRR-based sensor. With this level of precision, the measurement platform can reliably detect glucose variations as small as 1 mg/dL. The frequency deviation among repeated measurements was less than 0.5 MHz, corresponding to a glucose uncertainty below 0.2 mg/dL—demonstrating the outstanding accuracy, linearity, and stability of the proposed microwave sensing system.

Table 5.

Summary of Q-factor at the dominant sensing resonance (f₄) under different loading scenarios.

Configuration	Condition	Resonance freq, f₄ (GHz)	3-dB BW (MHz)	Q-factor
Vial-based laboratory test	Unloaded	5.8	19.6	296
Vial-based laboratory test	Blood, 50 mg/dL	5.512	63.3	87
Vial-based laboratory test	Blood, 140 mg/dL	5.392	68.1	79
Wrist tissue model	Unloaded	5.494	11.8	465
Wrist tissue model	Blood, 50 mg/dL	4.996	14	357
Wrist tissue model	Blood, 140 mg/dL	4.75	16	322

Fig. 10.

Linear correlation between measured resonant frequency and glucose concentration for three volunteer participants. (a) participant 1 (aged 23 years), (b) participant 2 (aged 27 years), and (c) participant 3 (aged 32 years). Each data point represents the mean of three repeated measurements, with error bars corresponding to one standard deviation. Dashed gray lines indicate linear regression fits, and insets show the fitted equations with their coefficients of determination (R²).

A comprehensive comparison of the sensitivity performance of the proposed sensor with other recent microwave sensors is presented in Table 6. This comparison ranks advanced glucose sensors based on their sensitivity to relevant parameters. Sensitivity is defined as the change in frequency (Δf_R) per unit change in glucose concentration (1 mg/dL) for a given volume and specific test setup. As a result, the sensitivity obtained in previous works based on different microwave sensing mechanisms was significantly lower than the minimum resolution adopted by the proposed sensor of this research. Additionally, most of the previously proposed sensors have not been investigated under the more complex and realistic conditions considered in this study. Therefore, the sensitivity obtained for the more realistic conditions investigated in this study is another distinguishing feature of this research. The sensitivity of the proposed sensor surpasses that of other techniques, which rely on tracking slight changes in the S₁₁ and S₂₁ resonance magnitudes, requiring high-precision measuring instruments. The improved design of the CSRR elements in this work enhances the interaction between the sample and the sensor in the sensor region, allowing the resonant frequency response of the sensor to be defined mainly by the passage of the sample under test (SUT).

Table 6.

Experimental frequency measurements for blood samples obtained from three volunteer participants at different glucose concentrations. Each value represents the average of three independent measurements taken under identical conditions (room temperature 25 ± 1 °C). (a) Participant 1. (b) Participant 2. (c) Participant 3.

(a) Participant 1
Glucose (mg/dL)	Avg. Freq (GHz)	1 st Test (GHz)	2nd Test (GHz)	3rd Test (GHz)
89	5.4721	5.47204	5.47218	5.47208
111	5.4235	5.42356	5.42344	5.4235
134	5.3705	5.3701	5.371	5.3704
183	5.2584	5.258	5.259	5.2582

(b) Participant 2
Glucose(mg/dL)	Avg. Freq (GHz)	1st Test (GHz)	2nd Test (GHz)	3rd Test (GHz)
82	5.493.	5.4936	5.4928	5.4926
138	5.361.	5.3604	5.3618	5.3608
261	5.079.	5.0785	5.0803	5.0782
332	4.914.	4.9149	4.9132	4.9139

(c) Participant 3
Glucose(mg/dL)	Avg. Freq (GHz)	1st Test (GHz)	2nd Test (GHz)	3rd Test (GHz)
87	5.402.	5.413	5.4026	5.4021
129	5.312.	5.3111	5.3124	5.3125
193	5.166.	5.1663	5.1654	5.1663
348	4.813.	4.8137	4.8126	4.8127

The sensitivity achieved in this study, at 2.3 MHz/[mg/dL] under standard conditions and 1.78 MHz/[mg/dL] under more practical circumstances, surpasses the highest reported values to date, as determined by our investigations. Moreover, the sample volume utilized in this research aligns with realistic conditions and the volume of blood in the wrist that engages with the sensor. The sensor’s performance under these conditions demonstrates enhanced reliability compared to conventional counterparts, suggesting its suitability for real-world applications.

The proposed sensor can be effectively used to detect the normal blood glucose range as well as cases of hypoglycemia and hyperglycemia.

Table 7.

Comparison of the proposed CSRR-based glucose sensor with previously reported non-invasive glucose-sensing approaches. All listed methods, including the present work, are based on non-invasive measurement principles; however, the type of testing sample differs among them—ranging from simulated models and tissue phantoms to real blood samples.

Ref	Sensing Technique	Operation Freq. (GHz)	Test solution	Concentration (mg/dL)	Sensing parameter	S (MHz/[mg/dL])
54	split ring resonator	4.18	Aqueous solution	0–5000	fR (S21)	0.0026
55	Band-stop filter based on SIW cavity	5–5.5.5	The fingertip	100–500	fR (S21)	0.24
56	three-loop microstrip patch antenna	3	Aqueous solution	50–500	fR (S11)	0.25
57	meta-structured antenna	4	Aqueous solution	50–250	fR (S11)	0.352
58	Two-port Rectangular Dielectric Resonator (RDR)	2.47	The fingertip	90–403	fR (S21)	0.39
59	Omega-coupled split-ring resonator	1.41	The fingertip	0–200	fR (S11)	0.56
60	closed-loop split ring resonator	2–5	On the forearm	80–155	fR (S11)	0.82
61	patch antenna	5.7	Aqueous solution	-	fR (S11)	0.089
62	Open-loop resonator with electric coupling	2.35	Aqueous solution	89–456	fR (S21)	0.95
63	Substrate Integrated Waveguide (SIW)	5.74	Aqueous solution	10–200	fR (S11)	1.218
64	Complementary split-ring resonator	2.45	Aqueous solution	40–140	fR (S21)	1.25
65	Fingertip placed on a planar resonator	1.5–2.3	The fingertip	98–188	fR (S11)	1.34
66	square-shaped spiral ring resonator	1–2	Aqueous solution	0.01–0.05	fR (S11)	1.99
67	Dual-band bandpass filter	2.45–5.2	Aqueous solution	0–400	fR (S11)	2.026
This work	Octagonal-shaped complementary split ring resonator	4.3–5.1	The Wrist	80–340	fR (S21)	1.78
This work	Octagonal-shaped complementary split ring resonator	4.4–5.5	Blood sample	50–500	fR (S21)	2.3

Conclusion

This paper presents the design, simulation, and experimental validation of a novel microwave sensor for non-invasive blood glucose monitoring. By employing an innovative octagonal CSRR configuration with an engineered 180-degree phase difference between unit cells, we achieved a highly concentrated and intense electric field, maximizing its interaction with the sample under test. This advanced design principle yielded unprecedented sensitivity, which was rigorously verified through both realistic simulations and in vitro experiments using real human blood samples. The sensor demonstrated an exceptional experimental sensitivity of 2.3 MHz/(mg/dL) under laboratory conditions and a simulated sensitivity of 1.78 MHz/(mg/dL) in a more practical wrist model, values that surpass those of previously reported microwave-based sensors. The strong linear correlation observed between resonance frequency shifts and glucose concentrations, ranging from 50 to 500 mg/dL, confirms the sensor’s potential for reliable and accurate glucose tracking across the full clinical range.

Future work and limitations

While the present study demonstrates a highly sensitive microwave-based glucose sensor, several challenges must be addressed before translation toward practical and clinical use. Future research will focus on implementing the proposed design on a flexible substrate that can conform to the skin surface, enabling stable, continuous, and wearable operation similar to a medical patch. The integration of machine learning (ML) and artificial intelligence (AI) algorithms is envisioned to facilitate automated calibration and intelligent processing of the high-resolution sensing data. Such data-driven approaches can adaptively compensate for inter-individual and environmental variability, including differences in skin thickness, body temperature, perspiration, and dielectric properties across users. Embedding the sensor within wearable platforms such as smartwatches or wristbands represents a promising pathway toward compact and user-friendly non-invasive glucose monitoring systems. In the longer term, coupling the proposed sensor with AI-based decision-support frameworks could enable predictive alerts for hypo- and hyperglycemia, real-time dietary feedback, and closed-loop glucose management, contributing to personalized and automated diabetes care. The present work is primarily focused on electromagnetic validation under controlled laboratory conditions and through a simplified wrist tissue model. Experimental evaluation under dynamic wrist motion and varying contact pressure inherently requires a fully wearable prototype and on-body testing, which will be addressed in future studies. Owing to the frequency-based sensing mechanism and the strong confinement of the electric field within the CSRR unit cells, the proposed sensor is expected to exhibit improved robustness against motion-induced artifacts compared with amplitude-based sensing approaches.

Finally, since the sensor operates under low-power excitation solely for resonance tracking, a detailed specific absorption rate (SAR) analysis has been deferred to future work involving a fully wearable implementation and anatomically representative tissue models.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Alireza Jamili conceived the initial idea, designed the sensor, and developed the theoretical framework. He was responsible for performing the CST simulations and analyzing the data. Majid Tayarani supervised the project, provided critical guidance on the design and methodology, and reviewed and revised the manuscript. All authors have read and approved the final version of the manuscript.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Data availability

The data supporting the conclusions of this research can be obtained from Majid Tayarani and Alireza Jamili; however, there are limitations to accessing these data. The data used in this study were obtained under a license and are therefore not publicly available. Nevertheless, the authors can provide the data upon reasonable request, provided that permission is obtained from Majid Tayarani.

Declarations

Competing interests

The authors declare no competing interests.

Ethical considerations

All experimental protocols were approved by the Research Ethics Committee of Iran University of Science and Technology and were conducted in accordance with relevant guidelines and regulations.

Informed consent

was obtained from all volunteer participants. All participants were volunteers and were fully informed about the purpose of the study and the nature of the procedures before providing their consent.