High-accuracy noninvasive continuous glucose monitoring using OCT angiography-purified blood scattering signals in human skin

The accuracy of noninvasive continuous glucose monitoring (CGM) through near-infrared scattering is challenged by mixed scattering signals from different compartments, where glucose has a positive correlation with a blood scattering coefficient but a negative correlation with a tissue scattering coefficient. In this study, we developed a high-accuracy noninvasive CGM based on OCT angiography (OCTA)-purified blood scattering signals. The blood optical scattering coefficient (BOC) was initially extracted from the depth attenuation of backscattered light in OCT and then purified by eliminating the scattering signals from the surrounding tissues under the guidance of a 3D OCTA vascular map in human skin. The purified BOC was used to estimate the optical blood glucose concentration (BGC) through a linear calibration. The optical and reference BGC measurements were highly correlated (R = 0.94) without apparent time delay. The mean absolute relative difference was 6.09%. All optical BGC measurements were within the clinically acceptable Zones A + B, with 96.69% falling in Zone A on Parke's error grids. The blood glucose response during OGTT was mapped with a high spatiotemporal resolution of the single vessel and 5 seconds. This noninvasive OCTA-based CGM shows promising accuracy for clinical use. Future research will involve larger sample sizes and diabetic participants to confirm these preliminary findings.


Introduction
Diabetes is a chronic metabolic disease that can lead to severe complications such as heart disease, kidney disease, and stroke [1,2].Regular monitoring of blood glucose levels is crucial for improving patients' quality of life [3].Glucose levels are typically measured in two compartments: blood and interstitial fluid (ISF) [4].Blood glucose concentration (BGC) is usually measured through enzymatic-based electrochemical devices that involve taking blood samples from finger or forearm pricks, which can be painful and carry a risk of infection [3].This leads to low compliance and poor glycemic control [5].Alternatively, ISF glucose concentration (IGC) can be continuously monitored by inserting needle sensors in subcutaneous adipose tissue [6].However, this method is less accurate than blood glucose monitoring since ISF glucose levels lag behind blood glucose levels and can cause discomfort and skin irritation [6].A noninvasive continuous glucose monitoring (CGM) with accuracy equal to or better than the current invasive methods would be highly beneficial.
Various noninvasive optical techniques have been developed to monitor blood glucose levels over the past few decades, including near-infrared and mid-infrared absorption spectroscopy, near-infrared scattering measurements, polarimetry, Raman spectroscopy, and photoacoustic spectroscopy [3,7].However, current accuracy needs to be improved for measuring glucose levels at clinically relevant levels.Optical methods based on the scattering of near-infrared light are considered one of the most promising approaches for noninvasive glucose monitoring [8,9], but detecting glucose-induced weak scattering changes in a strong scattering background is a significant obstacle.Optical coherence tomography (OCT) has been proposed for detecting the glucose-induced scattering changes in specific tissue layers, e.g., the dermis in the skin, which eliminates the influence of other tissue layers on the detected scattering signals [10,11].However, distinguishing between blood and interstitial fluid compartments remains a challenge, which may lead to less accurate measurements since blood and interstitial glucose have opposite effects on the optical scattering coefficients of surrounding tissues [12,13].
OCT angiography (OCTA) is an extension of OCT that allows for 3D mapping of blood perfusion within tissue down to the capillary level [14][15][16][17].OCTA vascular mapping has enabled the measurement of glucose-induced scattering changes in the specific compartment of blood or ISF with high sensitivity [18], which showed a high correlation between blood optical scattering coefficient (BOC) and BGC in mouse retinas in vivo.However, due to individual variations in BOC, calculating accurate BGC values for different individuals using a general equation is challenging.To overcome this issue, we propose using the percentage change of BOC for continuous glucose level monitoring with calibration for each individual.Our work involved verifying the correlation between OCTA-purified blood scattering signals and blood glucose levels in human skin in vivo, developing a calibration method to calculate BGC based on the OCTA-purified blood scattering signals, and testing the accuracy and reproducibility of the OCTA-based CGM (OCTA-G).

OCTA setup and scanning protocol
The OCTA setup was built based on a swept source configuration.Briefly, the system operated at a central wavelength of 1300 nm with a spectral bandwidth of 100 nm at a line-scan rate of 100 kHz and had a measured resolution of ∼14 and 13 µm in the axial and lateral directions in air, respectively.A stepwise raster scanning protocol (z-x-y) was adopted to acquire the human skin volumetric dataset.Each OCTA volume contained 450 A-lines per B-scan (fast-scan, x direction) and 900 B-scans repeated 3 times at 300 tomographic positions per volume (slow-scan, y direction), corresponding to a total acquisition time of ∼5 s.The OCTA raw data sequences were taken at the junction area of a finger with a field of view (FOV) of 4.5 mm × 3.0 mm (x-y).

Subjects
All 10 volunteers who participated in our study were of Asian descent, comprising of 6 males and 4 females, with an average age of 25 years.All volunteers were in good health, and none of them were taking any medications.This study was approved by the Ethics Committee of Zhejiang University, and signed informed consent was obtained from all volunteers before the experiment.

Experimental protocol
Standard oral glucose tolerance tests (OGTT) were conducted after an overnight fast at 9:00 A.M. Each test involved a 10-minute baseline period, followed by the ingestion of a glucose solution (75 g glucose dissolved in 300 ml water) for 5 minutes and an 85-minute recovery period to return to baseline levels.OCTA imaging was performed at the same region of interest (ROI) on the finger every 5 minutes during the trial.A fingertip blood sample was taken every 10 minutes, using a portable blood glucose meter (OneTouch Ultra, Johnson & Johnson, USA) to measure the BGC, which served as the reference value (G r ).Participants were allowed to take a break during the 5-minute intervals.After a rest of more than 2 days, similar experiments were conducted with the glucose solution replaced by water to serve as the control group.During the study, volunteers were not permitted to eat or drink, and the room temperature was maintained at 27°C to avoid temperature fluctuations.

Data processing
The data processing involved several steps, as shown in Fig. 1(a).(i) The raw spectral signal was Fourier transformed to create 3D structural images.(ii) Using a depth-resolved method proposed by Vermeer et al. [19], the local scattering coefficient was calculated under the assumption that optical scattering dominates the total attenuation of light at a wavelength of 1300 nm.(iii) The inverse signal-to-noise ratio and decorrelation OCTA (ID-OCTA) algorithm was used to create 3D angiograms of human skin [16,20].To speed up OCTA data processing, a graphics processing unit (GPU) was used to display enface angiograms in real-time, allowing instant feedback on angiogram quality and a higher yield rate of data acquisition with fewer motion artifacts [21].(iv) The OCT angiograms were binarized using the Otsu's method to generate the 3D vascular mask [22].(v) The vascular mask was overlapped onto the corresponding 3D matrix of the scattering coefficient, which was divided into two sub-parts: BOC and surrounding tissue scattering coefficient (TOC).The glucose levels of blood and ISF were then estimated from the scattering coefficient of blood and surrounding tissue, respectively.The TOC estimation only included the dermis layer of the skin.In addition, the BGC map sequences were spatially aligned to facilitate high-resolution visualization of the dynamic changes in BGC during OGTT.Figures 1(b)-1(f) show the representative cross sections of OCT structure, optical scattering coefficient, OCT angiogram and vascular mask, and enface OCT angiogram.
Our computational unit was equipped with an Intel Core i7-11700k 3.60 GHz CPU, 64 GB of RAM, and an NVIDIA GeForce RTX 3080Ti GPU.This configuration allowed us to process and display the skin OCTA map in real-time [21].Currently, the only calculation that requires post-processing is the determination of the blood scattering coefficient, which takes approximately 1.1 seconds.However, this calculation will be integrated into the GPU real-time processing algorithm in the future study.

Linear relation between BGC and BOC
Blood is a turbid substance comprising 55% plasma and 45% blood cells, with erythrocytes making up 99% of the cell count [23,24].Thus, the erythrocytes and blood plasma determine the optical properties of whole blood.In the near-infrared spectral range, scattering is substantially greater than absorption in most biological tissues [12].To simplify the light scattering analysis, we assumed that erythrocytes are uniform spheres with a volume equivalent to the average erythrocytes volume [25].According to the Mie theory, the reduced scattering coefficient µ ′ sb , the scattering coefficient µ s , and the anisotropy factor g of whole blood were calculated using the following equations: where the size of the scattering particles was simplified as six classes, denoted by M=6.The size distribution of erythrocytes in blood was obtained from Table 1 of the literature [26].For each size class j, the number of scatterers per unit volume of the medium N j was calculated as follows: where C j is the volume fraction occupied by the scatterers of class j, and V j is the volume of the erythrocyte of class j.In addition, each class is assigned a σ sj and g j to represent the scattering cross section and anisotropy factor of the individual scatterer, respectively.The impact of interparticle correlations W s is expressed in the following manner [27]: where H, r e represent the blood hematocrit and mean radius of erythrocytes, respectively.H 0 and r e0 represent the hematocrit of blood and mean radius of erythrocytes under isotonic condition, approximately 0.45 and 2.78 µm, respectively [25].
According to the Mie theory [25], the scattering cross section σ sj and the anisotropy factor g j of individual scatters can be described as follows: where λ stands for the central wavelength, and n p refers to the refractive index of plasma at the central wavelength.The asterisk * denotes complex conjugation.The Mie coefficients a l and b l can be determined using the following equations: where the prime ′ denotes differentiation with respect to the argument in parentheses.Ψ l (ρ) = ρJ l (ρ) and ξ l (ρ) = ρH (1)  l (ρ) are the Riccati-Bessel functions, where J l (ρ) is the l th order of the first kind of spherical Bessel function and H (1)  l (ρ) is the l th order of the spherical Hankel function of the first kind.The relative refractive index m and the size parameter x are: where n e and n p are the refractive index of erythrocyte and plasma, respectively.A change in blood glucose concentration BGC (G b , unit: mg/dL) impacts the plasma refractive index n p , and they are linearly related as stated in [25]: where n p0 is the refractive index of the blood plasma under the condition of G b = 0 mg/dL, which is approximately 1.327 at the 1300 nm wavelength [28].
A change in BGC also impacts the plasma osmolarity osm [29], resulting in a deformation of the erythrocyte's shape and volume (V e , unit: µm 3 ): ) where [Na + ] represents the concentration of sodium (unit: mEq/L), and [BUN] denotes the concentration of blood urea nitrogen (unit: mg/dL).osm 0 and G bo are the plasma osmolarity and BGC at isotonic condition, assumed to be 300 mosm/L and 90 mg/dL respectively [25].V e0 is the mean volume of the erythrocytes in blood under the isotonic condition and its approximate value is 90 µm 3 [26].Accordingly, the erythrocyte radius r e is written as follows: The erythrocyte volume, which depends on plasma osmolarity, has an impact on the concentration of hemoglobin (C Hb , unit: g/ml) within the erythrocytes (Eq.( 17)).This, in turn, affects the refractive index of erythrocytes n e (Eq.( 18)): ) where n e0 is the refractive index of the cell fluid without hemoglobin in erythrocytes, and is assumed to be 1.331 at the wavelength of 1300 nm [25,30].
We conducted a numerical simulation using Eq. ( 1)-( 18) and created a graph of the blood optical scattering coefficient BOC (µ ′ sb ) against the BGC (G b ).In Fig. 2(a), the red line shows a linear correlation between BOC and BGC at a wavelength of 1300 nm within the BGC range of 0-1000 mg/dL.This range includes the normal range of 70-160 mg/dL as well as hypoglycemia (< 65 mg/dL) and hyperglycemia (> 230 mg/dL) in humans [3].Within this BGC range, we can simplify Mie theory Eq. ( 1)-( 18) as follows [31]: As blood glucose concentration G b increases, it leads to a decrease in erythrocyte radius r e (see Eq. ( 16)) and a reduction in blood hematocrit H (see Eq. ( 6)).This change also causes an increase in the refractive indexes of plasma n p (see Eq. ( 13)) and erythrocytes n e (see Eq. ( 18)), ultimately resulting in an increase in blood scattering coefficient µ ′ sb within the range of 0-1000 mg/dL.This linear relationship is demonstrated in Fig. 2(a), where the BOC-BGC curve (blue line) of the simplified Eq. ( 19) closely resembles the curve obtained from Mie theory (red line) for blood glucose levels in the range of 0-1000 mg/dL.The theoretical linearity was achieved by assuming erythrocytes as isotropic spheres, but it's important to note that the shape and orientation of the erythrocytes vary in different types of blood vessels [32].For example, in capillaries, the erythrocytes are squeezed to a bullet-like shape, while in large vessels, they have a disk-like shape [32].Because of this, the blood has a lower scattering coefficient in capillaries (∼1.8 mm −1 ) compared to big vessels (∼2.6 mm −1 ).Accordingly, the blood scattering coefficients of big vessels with a diameter larger than 20 µm were used in the experimental validation of the derived BGC-BOC relation and the subsequent BGC measurements [33].As shown in Figs.2(b) and 2(c), this linear relationship between BOC and BGC was experimentally confirmed in human skin in vivo using OCTA.During the OGTT, the changes in BOC and BGC were closely correlated without significant delay.

Linear relation between IGC and TOC
The human skin consists of three primary layers: stratum corneum, epidermis, and dermis, with a well-developed blood microvascular network.Glucose is transferred from the capillary endothelium to the ISF in the dermis, leading to changes in the refractive index of the ISF and subsequently affecting the scattering characteristics of the skin [12].The reduced scattering coefficient µ ′ si (unit: mm −1 ) of the tissue can be calculated using the simplified Mie theory [31]: where the mean radius of the tissue scatterers r s is assumed to be 2.80 µm.The density of the scatters, denoted by ρ s , is approximately 6.53 × 10 6 mm −3 [34].n s and n ISF represent the refractive indices of scatterers and surrounding medium ISF, with approximate values of 1.401 and 1.360, respectively [35,36].Glucose level in the ISF is denoted by G i (unit: mg/dL).Using Eq. ( 20), we performed a numerical simulation and presented a plot of tissue optical scattering coefficient TOC (µ ′ si ) against interstitial glucose concentration IGC (G i ) in Fig. 3(a).Distinct from the BOC-BGC relation, TOC exhibits a negative linear correlation with increasing IGC in the range of 0-1000 mg/dL.This can be attributed to the reduction in the refractive index mismatch between the ISF and scatters in the tissue as IGC increases, leading to an overall decrease in the scattering coefficient of the tissue.And the correlation between TOC and BGC was further verified in human skin in vivo, as depicted in Fig. 3(b) and 3(c).During OGTT, there was a time lag of approximately 15 minutes between BGC and TOC changes.

Calibration
Based on the linear relation derived above, the BGC (G b ) and IGC (G i ) were readily estimated from the extracted BOC (µ ′ sb ) and TOC (µ ′ si ), respectively, using the following equation: where the parameters α b/i and β b/i were determined through a calibration procedure.Considering individual differences, the calibration procedure was performed independently for each subject, and two pairs of (G r , µ ′ sb/si ) acquired at fasting and 30-minute postprandial were used for calibration.

Accuracy analysis
The data from all samples were presented as mean ± standard deviation (SD).The correlation between measured BGC (or IGC) and reference BGC was linearly fitted based on the least squares criterion.To evaluate the accuracy of OCTA-G, we adopted Parke's error grid, ISO 15197 standards and mean absolute relative difference (MARD) [37][38][39].

Results
Our proposed OCTA-G technology allows for dynamic mapping of the BGC response during OGTT with single vessel resolution, as shown in the BGC-encoded en-face OCTA video (Visualization 1).The frames in Fig. 4(a) represent the response of glucose levels at different time points: baseline (-10 minutes), immediately after glucose intake (t = 0 min), and 20-and 50-minutes post-intake.It is noticeable that the BGC level in the capillaries (shown in green in Fig. 2(a)) is lower than that in the bigger vessels (shown in red in Fig. 2(a)).This is because the blood has a lower scattering coefficient in capillaries as compared to bigger vessels [32].To calibrate and measure the BGC, the scattering coefficients of blood obtained from bigger vessels with a diameter larger than 20 µm are used for the BGC-BOC linear calibration (shown in Fig. 2(b)), as well as for the BGC measurement (shown in Fig. 4(b) and 4(c)).
Figure 4(b) shows the plotted time courses of blood glucose, with our optical BGC response following a similar curve as the reference method without any apparent time lag.The optical and reference BGC measurements were highly correlated, with an R-value of 0.94 and a P-value less than 0.01, as shown in Fig. 4(c).On the other hand, the IGC measurement had a time lag of approximately 15 minutes after the reference BGC, as shown in Fig. 4(d).Due to this delay, the initial correlation between IGC and reference BGC was low (R-value 0.15), but it improved significantly to 0.89 with a P-value less than 0.01 after correcting for the time lag of IGC.
We verified the effectiveness of our optical BGC measurement in monitoring rapid changes in plasma glucose levels by examining the peak times of BGC response during OGTT in a cohort of 10 healthy subjects in Table 1.Our optical BGC measurements had a peak time of 25.5 ± 8.3 minutes, consistent with the reference measurements of 26.0 ± 8.4 minutes and the literature [40].In contrast, the peak time of our optical IGC was 36.5 ± 10.6 minutes, with a time lag of 10.5 minutes after the reference BGC.The peak time of BGC during OGTT is closely linked to the functioning of pancreatic beta cells and the body's ability to tolerate glucose [40,41].
We collected 121 pairs of optical and reference BGC values from 10 healthy subjects and evaluated the accuracy of our optical BGC measurements using Parke's error grid, ISO 15197 standards, and MARD.All optical BGC values were within the clinically acceptable Zones A + B, with 96.69% falling in Zone A on Parke's error grids (refer to Fig. 5(a)).Moreover, 94.21% of the optical BGC measurements met the ISO accuracy criteria (refer to Fig. 5(b)), and their MARD was 6.09%.

Discussion
Effective management of diabetes requires accurate noninvasive CGM.Near-infrared light scattering optical methods are considered one of the most promising approaches [8,9].However, these methods face challenges due to the high scattering background in bio-tissues and the complex effect of glucose on optical scattering coefficients in the compartments of blood and ISF.To ensure accurate glucose measurements, it is essential to differentiate the glucose-induced scattering changes in the specific compartments of blood and ISF.Our proposed OCTA-G technique has several advantages over previous methods.Firstly, it allows for compartmentspecific measurements of BGC and IGC under the guidance of OCTA.Secondly, it eliminates the requirement of remaining still to human subjects to minimize motion artifacts because of the homogeneous blood scattering in the vessels of the similar types and offers high practicability during long-term monitoring.Thirdly, using real-time processing, display, and spatial registration techniques in OCTA makes it possible to achieve dynamic 2D or 3D mapping of the changes in BGC with single vessel resolution.
Our OCTA-G technology has many potential applications.Firstly, it can help prediabetics identify their risk of developing diabetes by accurately measuring beta cell function and insulin resistance [42,43].Current invasive sampling methods have limited time resolution, making it difficult to quantify the OGTT glucose response curve accurately.However, OCTA-G provides noninvasive continuous blood glucose monitoring with a 5-second time resolution, allowing for more precise and detailed OGTT response curve quantification.This helps identify early metabolic risk factors.OCTA-G can also measure the time delay between BGC and IGC, which reflects the glucose diffusion time from the capillary to the tissue and is related to the metabolic rate of nearby cells [4].A higher-speed swept source can further enhance the time resolution of glucose monitoring.Secondly, OCTA-G technology can help diabetic patients by directly achieving noninvasive continuous BGC monitoring, overcoming the lag between IGC and BGC, a common challenge in current CGM [6].This enables better glucose tracking of glucose changes and enhances precision in glucose monitoring.Thirdly, OCTA-G's real-time processing, display, and spatial registration capabilities allow for the visualization and analysis of real-time changes in glucose distribution within the vascular network.This can significantly contribute to understanding glucose-related physiological processes and provide insights into conditions such as diabetes, vascular diseases, or metabolic disorders that impact glucose regulation in the vasculature [44].
Although our OCTA-G technique has shown promising results, there are still limitations that need improvement.One such area is glucose specificity.Our BGC measurements rely on changes in plasma osmolarity induced by glucose.However, other osmolytes like urea and sodium can also influence plasma osmolarity.Fortunately, glucose shows higher variability levels (40.4%) than other osmolytes (less than 5%) under normal conditions (refer to Fig. 6).Additionally, other osmolytes have a negligible effect on BOC compared to glucose (refer to Table 2).Although other osmolytes have negligible diurnal variations, they might have large variations between different individuals, resulting in a significant change of calibration coefficients α and β.Therefore, individual calibration procedures were conducted in our BGC measurements to calculate the coefficients α and β as described in Eq. ( 21).And the sample size used for calibration is highly related to the final accuracy of the measurements.Although a further lowering of MARD < 10% has little additional benefit for insulin dosing [45], increasing the calibration points from 2 to 10 can reduce the MARD of the system from 6.09% to 4.7%.Although our study has demonstrated accurate results with two calibration points, sophisticated calibration is necessary for improving accuracy with limited calibration frequency.Although our OCTA-G technology cannot replace invasive methods entirely, it can significantly reduce the frequency of invasive sampling and improve the accuracy of continuous glucose monitoring.In our BGC measurements, we neglected the influence of blood oxygen saturation because the difference in attenuation coefficients between oxyhemoglobin and deoxyhemoglobin was only 0.06 mm −1 in the operating wavelength range from 1250 to 1350 nm [27].Lastly, while our current implementation of OCTA-G relies on a desktop OCT system, the development of microelectronics and the miniaturization of optical components have led to a significant reduction in the size of OCT systems.Several studies have reported miniaturized OCT systems [46], making the future development of a portable OCTA-G system possible.This would enhance its usability and portability.It is known that the shape and orientation of erythrocytes can vary in different types of blood vessels [32], and this can significantly affect the optical scattering properties of blood.However, it is difficult to create a specific analytical expression that accounts for all these factors due to the complex nature of blood hemodynamics and multiple-scattering effects.To overcome this challenge, we developed a linear calibration method to measure the BGC values.Fortunately, the linear relationship between BOC and BGC predicted by Mie theory aligns well with the experimental outcomes in human skin in vivo using OCTA (see Fig. 2), and our optical BGC measurements show promising accuracy for clinical use.

Fig. 1 .
Fig. 1.(a) Flow diagram of glucose monitoring using OCT angiography in human skin.(i) generating OCT structural image by taking Fourier transform of raw spectral sequence, (ii) extracting depth-resolved scattering coefficient, (iii) computing OCT angiogram, (iv) binarizing using Otsu's method, (v) calculating BGC and IGC.Representative cross-sectional OCT structure (b), optical scattering coefficient (OSC) map (c), OCT angiogram (d) and vascular mask (e).The insert in (b) is a representative depth profile at the location indicated by the white dashed line in (b).(f) Enface OCTA angiogram.The white dashed line in (f) indicates the location of the cross-sections in (b-e).Scale bar in (b) is 500 µm and applies to (b-f).E: epidermis, D: dermis, NM: nail matrix, NR: nail root, NB: nail bed.

Fig. 3 .
Fig. 3. Linear relation between TOC and IGC.(a) Using numerical simulation, plots of TOC against IGC within 0-1000 mg/dL.(b) Correlation between TOC and IGC through vivo experiment.The solid squares and the solid line represent the correlation between the TOC and IGC after correcting for the delay time.(c) Times course of percentage changes in TOC during OGTT.TOC: tissue optical scattering coefficient (µ ′ si ), IGC: ISF glucose concentration (G i ), Ref: reference BGC.

Fig. 4 .
Fig. 4. OCTA-based BGC and IGC measurements during OGTT.(a) Representative enface BGC mappings at instants: -10 min, 0 min, 20 min, 50 min.(b) Time courses of optical BGC and reference BGC.(c) Correlation between optical and reference BGC.(d) Time courses of optical IGC and reference BGC.(e) The correlation between the optical IGC and reference BGC.The solid squares and the solid line represent the correlation between the IGC and BGC after correcting for delay time.Opt.: optical BGC (or IGC); Ref.: reference BGC.

Fig. 5 .
Fig. 5. Accuracy of optical BGC measurement.(a) Parke's error grid.Red and blue circles indicate points within zones A and B, respectively.The five risk zones are: Zone A, clinically accurate; Zone B, benign; Zone C, excessive; Zone D, undetectable; Zone E, faulty.Both zones A and B are clinically acceptable.(b) ISO 15197:2013.The solid line is the relative difference of ±15 mg/dL when BGC < 100 mg/dL and ±15% when BGC ≥ 100 mg/dL.Red and blue circles represent values within or outside the accuracy criteria.A total of 121 pairs of optical and reference BGC are used.