A Photoacoustics based Continuous Non-Invasive Blood Glucose Monitoring System Praful P. Pai∗ , Pradyut K. Sanki, and Swapna Banerjee Department of Electronics & Electrical Communication Engineering Indian Institute of Technology Kharagpur Kharagpur, West Bengal, India - 721302 ∗ [email protected]

Abstract—The paper examines the use of photoacoustic spectroscopy (PAS) for making continuous non-invasive blood glucose measurements. An apparatus for performing photoacoustic (PA) measurements is constructed and the technique is verified in vitro and in vivo through measurements on glucose solutions and live tissue. The signal amplitude is observed to increase with the glucose concentration in both cases. A linear calibration method is applied on each individual to obtain a glucose concentration value from each PA measurement. The glucose values obtained are compared with reference glucose concentrations measured using a standard glucose meter, giving a mean absolute difference (MAD) of 23.75 mg/dl and a mean absolute relative difference (MARD) of 18.03%. A plot of 196 measurement pairs taken over 30 normal subjects on a Clarke Error Grid gives a point distribution of 67.86%, 31.12%, 0.0%, 1.02% and 0.0% over zones A to E of the grid. This performance is an improvement over those obtained previously using PAS and point to the potential of the technique for non-invasive glucose measurements. An FPGA based reconfigurable embedded architecture is proposed for high speed data acquisition, noise reduction and display of PA measurements. The architecture operates at 274.823 MHz on a Xilinx Virtex-II Pro FPGA providing an SNR improvement of 30 dB and enabling a portable blood glucose monitoring system. Index Terms—Biomedical Measurement, Biomedical Monitoring, Calibration, Diabetes, Noninvasive Measurement, Photoacoustic Effects

I. I NTRODUCTION Diabetes is a highly prevalent disease, with an estimated 366 million people suffering from it worldwide [1]. In diabetics, defects in insulin secretion or insulin action prevents cells from taking up glucose leading to hyperglycaemia, i.e. higher than normal levels of glucose in blood. Persistent hyperglycaemia leads to a host of micro- and macro- vascular complications such as impaired vision and renal function. Control of blood glucose levels in the normal physiological range with multiple glucose measurements and intensive insulin therapy delays onset and progression of complications [2]. Current state-of-the-art glucose monitors require diabetics to prick their fingers and test drops of blood using electrochemical testing strips at regular intervals throughout the day. The discomfort and high cost per measurement associated with the procedure leads to irregular testing, resulting in poor treatment outcomes. Non-invasive glucose monitoring offers painless and sample free measurements in a continuous

manner, encouraging regular monitoring and control. At present a variety of measurement methods are being investigated for non-invasive blood glucose monitoring [3], [4]. We have selected photoacoustic (PA) spectroscopy for the same due to its ability to measure small absorption coefficients produced by glucose without affecting tissue. It also requires little or no sample preparation before measurement, and is not affected by optical scattering by tissues. The PA effect involves the generation of acoustic waves in a sample following absorption of modulated optical excitation. Optical energy from an intense light source is absorbed by a sample in proportion to its optical properties, resulting in excitation of its constituent molecules. The absorbed energy is released non-radiatively as heat, causing localized temperature rise and thermoelastic expansion of the irradiated sample volume. This volumetric expansion produces a pressure wave, which is dependent on the optical and physical properties of the sample constituents, and can be used to determine their concentration. The phenomenon was first reported by Bell [5], but it was not until a century later that Rosencwaig demonstrated the similarity of optical and PA spectra of biological molecules [6]. The PA response generated in a sample can be used to estimate glucose concentration through selection of excitation wavelengths which are absorbed in particular by glucose molecules. Optical absorption by glucose is strongest in the mid infra red, but high absorption by water at these wavelengths leads to low penetration depths, making it impractical for glucose measurement. For the purpose of the current study, an excitation wavelength of 905 nm was selected as it is close to 939 nm near infra red overtone band of glucose [7], [8]. Moreover, it lies within the 600-1300 nm region of the spectrum, also known as the tissue optical window, which allows for maximum optical penetration in tissue giving PA measurements from greater depth [9]. In addition, pulsed laser diodes operating at 905 nm were readily available. The PA pressure, pP A , can be related to the sample properties by the following relation [10]:   αβυ 2 1 E0 (1) pP A = πR2 CP where, α = optical absorption coefficient of the sample

β = thermal expansion coefficient of the sample υ = sound velocity in the sample CP = specific heat of the sample E0 = optical energy incident on the sample R = radius of PA source generated in the sample Equation 1 shows the amplitude of the PA signal to be a function of the optical and physical properties of the sample. Also, the sample properties are proportional to the concentration of glucose in the sample, cg , as follows [11]: α ∝ cg β ∝ cg υ ∝ cg 1 C ∝ cg P

pP A ∝

αβυ 2 CP

(2)

∝ cg

The PA pressure waveform is described using the wave equation [12]. It has a bipolar nature, with a short positive peak followed by a broad negative peak. Keeping that in mind, and looking at 1 and 2, the peak-to-peak amplitude of the PA signal is taken as a measure of the glucose concentration. II. S YSTEM D EVELOPMENT A. Experimental Apparatus A block diagram of the experimental set up constructed for making PA measurements is shown in Figure 1. A pulsed laser diode (PLD) module operating at 905 nm (LS9-Series, LaserComponents GmbH) is used as an optical excitation source. The PLD output is regulated by a set of control circuits which provide a 100 Hz TTL trigger pulse and control voltages to fix the pulse width and peak power of the PLD to 100 ns and 100 W respectively. The sample is positioned in front of the PLD emission window and placed in contact with a lead zirconate titanate (PZT) piezoelectric transducer (V108, Olympus Panametrics) to measure the generated PA signal. Liquid samples are taken in a cuvette (Type G, OptiGlass Limited) which is placed in contact with the PZT and held in place using a holder mechanism. Solid samples and tissues are placed in a housing which fits on to the transducer body and restricts movement during measurement. Acoustic attenuation due to impedance mismatch at the sample-transducer interface is reduced by using a thin layer of ultrasound gel. The output of the PZT is amplified using a low-noise amplifier (LNA)(Model 351A, Analog Modules, Inc.) and acquired using a digital storage oscilloscope (DSO)(Model 54622A, Agilent Technologies). The signal is averaged over 1024 frames using the DSO to reduce random noise present in the signal, and the averaged signal is recorded to a computer.

Fig. 1. Block Diagram of Experimental PA Measurement Setup

B. Embedded System Development The back-end of the measurement apparatus, consisting of the PLD control circuits and the DSO, is replaced with an FPGA based embedded solution. The LNA output is digitized using an ADC (AD9256, Analog Devices, Inc.) and the samples are coherently averaged using FPGA (Virtex-II Pro, Xilinx, Inc.). The FPGA implements additional circuits for PLD trigger control along with a coherent averaging IP-core for noise reduction. A MicroBlaze soft core processor is also integrated on the FPGA to implement an LCD (MI0283KT, MI Technology Co. Ltd.) module and realize a portable system. The random noise present in the PA signal is reduced by ensemble averaging of multiple acquisitions of the signal. Averaging reduces the variance of random noise while preserving the amplitude of the PA signal coherent with the beginning of the sampling intervals. The improvement in SNR by coherent averaging of N successive signal frames is given by [13]: SN Rcoh =

√ σin SN Ravg = = N SN Rin σavg

SN Rcoh (dB) = 10.log10 (N )

(3) (4)

Figure 2 shows a block level diagram of the architecture proposed for coherent averaging of the PA signal by frame accumulation. A Trigger Generator Module (TGM) uses a 20bit up counter to generate a 100 Hz TTL trigger pulse for operation of the PLD and synchronizing ADC data acquisition. A Frame Counter Module (FCM) uses a 10-bit up counter to obtain 1024 signal frames from the ADC. The counter is incremented at each trigger pulse from the TGM and the counter output is used by the Controller Module (CM) to control the IP-core. An Address Generator Module (AGM) generates addresses to access the RAM. A two-stage 26-bit pipelined frame adder (FA) is used to add frames with a critical path delay of 2.11 ns as compared to a simple 26-bit adder which takes 3.24 ns. The CM controls the overall operation of the averaging core and generates initialization signals for the TGM, FCM, and AGM along with RAM read / write signals. Two architectures, I and II, are proposed for coherent averaging. Architecture I, shown in Figure 3a, uses two RAMs to accumulate signal frames in a ping-pong manner. Data samples are read frame by frame using the TGM output, with

Fig. 2. Block Level Diagram of the Coherent Averaging Module

data_in

Address Generator

Controller

rst

done

10

16 0

0 0

1

0

M1

Trigger Generator

Registers

(2)

(2)

0

(2) rw2

rSel

1

0

M10

1

0

M9

1

rd2 0 0

1

M8

0

0

addRd

done

10

1

data_in 16

16

Registers

Registers

(3)

(1)

26

Trigger Generator

wr1 1

Address Generator

Controller

M3

M6

adrWr

(2) rd1

0

1

M5

Reg

Reg

1 M2

Registers

0

M7

RAM1

RAM2

1000x26bit

1000x26bit

Registers

0 0

1 2

rSel

0

26

Addwr

0 1

0 M2

Addrd

26 1

1 M1

Registers (3)

26

Pipelined Adder

(3)

1 M3

1 3

RAM

0 M4

1000x26bit

rd wr

M4

Pipelined Adder(26bit)

16 data_out [25:10]

16 data_out [25:10]

(a) Averaging Architecture I

(b) Averaging Architecture II

Fig. 3. Architectures for Coherent Averaging of PA Signals for Noise Reduction

each incoming frame added with previously acquired frame data. The first frame is stored in RAM1, after which the second frame is added to the first and the result is stored in RAM2. Each incoming frame is added to the result and stored in the alternate RAM block. Alternatively, Architecture II, shown in Figure 3b, uses a single RAM block with a FA for coherent averaging. The initial data frame is stored directly to RAM and each incoming frame sample is added to the stored frame sample and written to the same RAM location. After accumulation of 1024 frames, the FCM prompts 10-bit right-shifting of the accumulated result to divide by 1024 and get a single averaged data frame. Coherent averaging of 1024 successive frames gives an SNR improvement of 30 dB. A MicroBlaze processor based embedded back-end solution was implemented by integration of the ADC, coherent averaging IP-core, LCD Module and other required digital design using Xilinx Embedded Development Kit (EDK). Figure 4 shows a block level diagram of the embedded system with a MicroBlaze soft core processor being used for overall system control. The coherent averaging IP-core is connected to the processor through a Fast Simplex Link (FSL) bus. The LCD Module, indicator LEDs, and push buttons are operated through the Xilinx GPIO IP-core which is connected to the processor using the processor local bus (PLB). Upon

Embedded Back-end on FPGA Interrupt Controller (Xilinx)

Interrupt FSL

Clock 100MHz

Microblaze (Xilinx) FSL GPIO Core (Xilinx)

PLB

Coherent Averaging with Trigger IP Core Trigger UART Core (Xilinx)

completion of averaging process, the coherent averaging IPcore interrupts the processor, which executes an interrupt service routine (ISR) to read the averaged output over the FSL bus and store it in an internal buffer. On receiving the entire data frame, the processor accesses the LCD module and displays the averaged PA signal. A Bluetooth module is also integrated with the system using a UART IP-core and Xilinx interrupt controller connected to the processor over the PLB. The averaged data frame is transferred to the Bluetooth module over RS-232, from where it can be transmitted to a phone for storage and analysis. C. Calibration The averaged PA signal is processed to extract features which are used to calibrate the system and obtain the sample glucose concentration. From 1 and 2, we know that the amplitude of the PA signal varies linearly with the sample glucose concentration. This linear relationship was used to calibrate PA measurements using the peak-to-peak amplitude of the PA signal, pP A , as follows: pP A = mcg + C

where, m and C are slope and intercept of the straight line representing this linear relationship. Their values are obtained using two reference PA measurements, pP A1 and pP A2 , taken at known glucose concentrations, cg1 and cg2 , as follows:

Analog Front-end AD9265 (16bit) PLD

PZT Mobile Handset

Push Buttons & LEDs Graphic LCD

pP A1 − pP A2 cg1 − cg2 C = pP A2 − mcg2 − mcg1

m=

LNA

Bluetooth Module

C = pP A1

(6a) (6b)

The values of m and C are used to calibrate PA measurements made thereafter to obtain an estimate of the glucose concentration, cgest , as follows: pP A − C (7) m This calibration is currently performed off-line on a computer and will be ported to the FPGA at a later stage. cgest =

Fig. 4. Block Level Diagram of the Embedded System for Photoacoustic based Continuous Non-Invasive Blood Glucose Monitoring

(5)

III. M ETHODOLOGY The system was initially used to make PA measurements in vitro on glucose solutions, followed by in vivo measurements on tissue. A. In Vitro Testing The measurement principle was initially verified in vitro on glucose solutions at different concentrations. Solutions at four different glucose concentrations (0, 75, 150, and 250 mg/dl) were made by combining varying amounts of dextrose (D-Glucose) with distilled water. A fixed quantity (4 ml) of each solution was taken in a cuvette and their PA response was measured at two excitation wavelengths, 905 nm (LS9 Series PLD Module, LaserComponents GmbH) and 1064 nm (Nd:YAG Laser, Continuum Lasers) using the developed system. The concentration of the solutions was varied and changes in the peak-to-peak amplitude of the PA response were observed at both excitation wavelengths. B. In Vivo Testing and Calibration Following in vitro testing, in vivo PA measurements were performed on healthy human volunteers at an excitation wavelength of 905 nm. The measurements were taken at the forefinger, while operating within safety limits as specified by laser safety standards [14], [15]. Glucose measurements were made simultaneously within capillary blood and venous plasma using a regular blood glucose meter (OneTouch Ultra 2, LifeScan, Inc.) and with a laboratory analyzer (Glucose GOD FS, DiaSys Diagnostic Systems International) at the local hospital laboratory respectively. The PA and blood glucose measurements were performed on volunteers over the course of glucose/meal tolerance tests. Volunteers fasted for 8 hours, following which a fasting glucose measurement was taken. Thereafter, they drank a 75 g oral dose of glucose or had meals to elevate their blood glucose levels. Postprandial measurements were taken 2 hours after consumption of the glucose load or meal. In vivo testing was performed over two phases: 1) Initial Trials: To see whether changes in peak-to-peak PA amplitude correspond to changes in blood glucose concentration, PA and blood glucose measurements were taken simultaneously during fasting and postprandial stages of a tolerance test and the correlation between the two was observed. 2) Extended Trials: To check whether the PA amplitude followed blood glucose concentration over time, PA and blood glucose measurements were taken at intervals of 15-20 minutes over the course of a tolerance test. The PA measurements were calibrated using 5 to get glucose concentration values. The initial two PA and reference glucose measurements made on each individual were used in 6 to compute the calibration slope and intercept. All future PA measurements made on the individual were calibrated by employing these values in 7 to obtain an estimate of the blood glucose concentration. The resultant glucose concentration estimates are compared to reference glucose measurements using a Clarke Error Grid (CEG) to check the accuracy of

the proposed method [16]. The CEG divides the correlation plot into zones A to E, depending upon the risk of using the calibrated glucose values for treatment. Results in zones A and B are considered clinically acceptable, while results in zones C, D, and E are potentially dangerous and clinically significant errors. In addition, the Mean Absolute Difference (MAD) and the Mean Absolute Relative Deviation (MARD) are also used to analyse device performance [17]. IV. R ESULTS AND D ISCUSSION A. Coherent Averaging The architectures for coherent averaging were implemented on XC2VP-30 Virtex-II Pro FPGA using Xilinx ISE 10.1i. Maximum post place and route (P&R) frequencies of 224.792 MHz and 274.823 MHz were obtained for Architecture I and Architecture II respectively. The synthesis reports of the two architectures are compared in Table I, which shows the lower memory utilization and greater speed offered by Architecture II. Table II shows a comparison of Architecture II with existing architectures for coherent averaging [18], [19]. The Architecture II IP-core was used in the embedded backend of the system which was implemented on XCVP30 VirtexII Pro FPGA. Figure 5 shows a comparison of 1024-frame averaged PA signal computed using MATLAB (R2009b) and obtained using FPGA. A snapshot of the PA signal averaged in real-time and displayed on the LCD module is also shown. B. In Vitro Testing The variation in PA measurements from glucose solutions at 905 nm and 1064 nm is shown in Figure 6. The peak-to-peak amplitude of the PA measurements increases linearly with the sample glucose concentration at both excitation wavelengths. Also, the PA response at 1064 nm is found to be greater than that at 905 nm. This is in agreement with (1-2), as the optical TABLE I S YNTHESIS R EPORTS OF C OHERENT AVERAGING A RCHITECTURES (D EVICE : XC2VP-30-7FF896 V IRTEX -II P RO ) Architecture I

Architecture II Number Utilized (% Utilization)

Resource Type

Available

Number Utilized (% Utilization)

Slices Slice Flip Flops 4 I/P LUTs Bonded IOBs BRAMs

13696 27392 27392 556 136

190 (1%) 279 (1%) 273 ( 0%) 314 (56%) 4 (2%)

113 ( 0%) 189 ( 0%) 118 ( 0%) 161 (28%) 2 (1%)

224.792 MHz

274.823 MHz

Maximum Operating Frequency

TABLE II P ERFORMANCE C OMPARISON OF C OHERENT AVERAGING A RCHITECTURES Author

Platform

Speed

Memory

[18] [19] Proposed

Xilinx System Generator Spartan-3E (XC3S250E-4FT256) Spartan-3E (XC3S250E-4FT256)

90 MHz 200 MHz 228 MHz

4 BRAMs 4 BRAMs 2 BRAMs

4

0.1

0.05

0

100

200

300

400

500 Sample

600

700

800

900

1000

0.05

(a) Using MATLAB

0

100

200

300

400

500 Sample

600

700

800

900

(b) Using FPGA

1000

150

4.058

100

(c) Displayed on LCD

Fig. 5. A comparison of the averaged PA signal computed from 1024 signal frames using MATLAB and FPGA before being displayed on an LCD.

absorption of glucose at 1064 nm is greater than that at 905 nm [7]. Thus, in vitro measurements show the PA amplitude to increase linearly with the glucose concentration of the solution. C. In Vivo Testing and Calibration Following in vitro measurements on glucose solutions, in vivo PA measurements were taken on human volunteers over the course of glucose tolerance tests. These trials were conducted over two phases: 1) Initial Trials: The initial phase of in vivo trials involved 116 volunteers(75 male, 41 female, Ages: 51±13 years) on whom PA and blood glucose measurements were performed only at fasting and postprandial stages of the tolerance test. The peak-to-peak amplitude of the PA measurements was found to follow the glucose concentration between fasting and postprandial stages for all 116 volunteers. However, the change in peak-to-peak PA amplitude with glucose concentration varied across individuals. Individuals with near equal glucose concentrations have different PA amplitude, pointing to changes in the optical absorption or other physical parameters between individuals. This variation is due to biomolecules, such as, melanin and haemoglobin, which vary in amount across individuals. The results obtained point to the need for individual calibration of the PA measurements.

50

0.4233 1

2

3

350

Predicted Glucose Concentration [mg/dl]

PA Peak-to-Peak Amplitude [A.U.]

17

13

12

11

905 nm 1064 nm 50

100

150

200

100

50

5

4.0225 1

2

250

Glucose Concentration [mg/dl]

Fig. 6. The peak-to-peak PA amplitude measured at excitation wavelengths of 905 nm and 1064 nm for solutions at different glucose concentrations.

3

4

5

Measurement Number

Clarke Error Grid Analysis Plot

C

300

A

E

250

B 200

150

D

D

100

B

50

E

C

10 0

4.1659

2) Extended Trials: Following initial trials, extended in vivo trials were carried out on a cohort of 30 healthy volunteers (29 Male and 1 Female, Ages: 29 ± 8 years). Herein, PA measurements were taken at the forefinger using the developed apparatus along with reference glucose measurements using a regular glucose meter (OneTouch Ultra 2, LifeScan, Inc.). measurements were made every 15-20 minutes over the 2hour duration of a tolerance test. Figure 7 show trends in the two quantities for two of the individuals tested. The PA measurements and blood glucose concentrations are observed to follow a similar trend over an extended duration of time. However, differences in PA amplitude across individuals in Figure 7 highlights the need for individual calibration of PA measurements. The calibration algorithm outlined in Section II-C is applied on measurements taken on each individual, and the estimated glucose values were compared to reference blood glucose measurements using a CEG as shown in Figure 8. Of the 196 measurement pairs taken over 30 individuals, 133 (67.86%) fell in zone A of the CEG, followed by 61 (31.12%) in zone B and 2 (1.02%) in zone D of the grid. No measurement pairs

400

14

150

Fig. 7. Variation in the blood glucose concentration and the peak-to-peak amplitude of PA measurements for two individuals taken at intervals of 15-20 minutes over the course of glucose tolerance testing.

In Vivo PA Response at Different Wavelengths

15

200

Measurement Number

18

16

4

PA Peak-to-Peak Amplitude [A.U.]

0.15

0.1

200

Glucose Concentration [mg/dl]

0.15

Blood Glucose PA Amplitude

PA Peak-to-Peak Amplitude [A.U.]

0.2

Glucose Concentration [mg/dl]

0.25

0.2

x 10 4.3092

250

Blood Glucose PA Amplitude

Amplitude (V)

0.25

Amplitude (V)

0.3

4

x 10 7.6927

250 0.3

0 0

50

100

150

200

250

300

350

400

Reference Glucose Concentration [mg/dl]

Fig. 8. A Clarke Error Grid Analysis showing a comparison of predicted and reference blood glucose measurements from thirty subjects.

TABLE III P ERFORMANCE C OMPARISON OF N ON - INVASIVE G LUCOSE M EASUREMENT M ETHODS

Raman [25] Sentris [26] Proposed Aprise GlucoTrack Solianis MGMS Pendra

[22] [23] [21] [24] [20]

Zone A

Zones A&B

Zones C-E

MARD (%)

86.66 83.00 67.86

99.00 98.98

1.00 1.02

11.50 18.03

66.50 60.00 56.00 39.20 32.40

94.60 96.00 93.00 89.00 78.40

5.40 4.00 7.00 11.00 21.60

22.40 40.80 -

fell in zones C and E of the grid. A total of 194 (98.98%) measurement pairs fell within acceptable zones A and B of the CEG. A MARD of 18.03% ± 23.16% and a MAD of 23.75 mg/dl (1.32 mmol/l) was obtained for the glucose values estimated using the proposed method. Table III shows a comparison of the performance of the system to existing literature. The performance of the proposed method using a single excitation wavelength is an improvement over results obtained by earlier researchers using PAS or other non-invasive measurement methods [20]–[24]. V. C ONCLUSIONS An apparatus for making PA measurements has been demonstrated for continuous non-invasive glucose monitoring and an FPGA based embedded system has been realized for making such a portable device. A fast, memory efficient architecture is proposed for de-noising of PA signals by coherent averaging. The suitability of PA measurements at 905 nm for non-invasive glucose measurement has been demonstrated in vitro on glucose solutions and in vivo over human subjects, and the PA amplitude was found to vary linearly with the glucose concentration. Individual calibration of PA measurements is required to account for variations occurring in the PA signal due of influence of molecules other than glucose. A linear calibration method applied over in vivo PA measurements showed improved accuracy in comparison to a majority of earlier studies using PAS or other non-invasive methods. This can be improved further through multi-wavelength PA measurements and use of multivariate calibration techniques. R EFERENCES [1] D. R. Whiting, L. Guariguata, C. Weil, and J. Shaw, “IDF Diabetes Atlas: Global Estimates of the Prevalence of Diabetes for 2011 and 2030,” Diabetes Research and Clinical Practice, vol. 94, no. 3, pp. 311–321, Dec. 2011. [2] The Diabetes Control and Complications Trial Research Group, “The Effect of Intensive Treatment of Diabetes on the Development and Progression of Long-Term Complications in Insulin-Dependent Diabetes Mellitus,” The New England Journal of Medicine, vol. 329, no. 14, pp. 977–986, 1993. [3] V. V. Tuchin, Handbook of Optical Sensing of Glucose in Biological Fluids and Tissues, V. V. Tuchin, Ed. CRC Press, 2008. [4] D. D. Cunningham and J. A. Stenken, In Vivo Glucose Sensing. John Wiley & Sons, Inc., 2010. [5] A. G. Bell, “The Production of Sound by Radiant Energy,” Science, vol. 2, no. 48, pp. 242–253, May 1881.

[6] A. Rosencwaig, “Photoacoustic Spectroscopy of Biological Materials,” Science, vol. 181, no. 4100, pp. 657–658, Aug. 1973. [7] J. Tenhunen, H. Kopola, and R. Myllyl¨a, “Non-invasive glucose measurement based on selective near infrared absorption; requirements on instrumentation and spectral range,” Measurement, vol. 24, no. 3, pp. 173–177, Oct. 1998. [8] O. S. Khalil, “Spectroscopic and Clinical Aspects of Noninvasive Glucose Measurements,” Clinical Chemistry, vol. 45, no. 2, pp. 165– 177, Feb. 1999. [9] A. N. Bashkatov, E. A. Genina, V. I. Kochubey, and V. V. Tuchin, “Optical properties of human skin, subcutaneous and mucous tissues in the wavelength range from 400 to 2000 nm,” Journal of Physics D: Applied Physics, vol. 38, no. 15, pp. 2543–2555, Aug. 2005. [10] A. C. Tam, “Applications of photoacoustic sensing techniques,” Reviews of Modern Physics, vol. 58, no. 2, pp. 381–431, 1986. [11] Z. Zhao, “Pulsed Photoacoustic Techniques and Glucose Determination in Human Blood and Tissue,” Ph.D. Thesis, University of Oulu, May 2002. [12] G. J. Diebold, “Photoacoustic Monopole Radiation: Waves from Objects with Symmetry in One, Two, and Three Dimensions,” in Photoacoustic Imaging and Spectroscopy, L. V. Wang, Ed. CRC Press, Dec. 2009, ch. Chapter 1, pp. 3–17. [13] R. G. Lyons, “Signal Averaging,” in Understanding Digital Signal Processing, 2nd ed. Pearson Education India, 2009, ch. 11, pp. 429– 456. [14] International Electrotechnical Commission, International Standard - IEC 60825-1 - Safety of Laser Products. International Electrotechnical Commission, 2001. [15] American National Standards Institute, American National Standard for Safe Use of Lasers in Health Care Facilities. Laser Institute of America, 2005. [16] W. L. Clarke, D. J. Cox, L. A. Gonder Frederick, W. Carter, and S. L. Pohl, “Evaluating Clinical Accuracy of Systems for Self-Monitoring of Blood Glucose,” Diabetes Care, vol. 10, no. 5, pp. 622–628, Sep. 1987. [17] W. L. Clarke and B. Kovatchev, “Continuous Glucose Sensors: Continuing Questions about Clinical Accuracy.” Journal of Diabetes Science and Technology, vol. 1, no. 5, pp. 669–675, Sep. 2007. [18] A. Kalashnikov, V. Ivchenko, and R. Challis, “VLSI Architecture for Repetitive Waveform Measurement with Zero Overhead Averaging,” in IEEE International Workshop on Intelligent Signal Processing. IEEE, 2005, pp. 334–339. [19] O. C. Kulkarni and S. Banerjee, “Implementation of High Speed Coherent Averaging Block for Portable Noninvasive Blood Glucose Measurement System,” in International Conference on Communication, Computers, and Devices, Kharagpur, Dec. 2010, pp. 1–5. [20] I. M. E. Wentholt, J. B. L. Hoekstra, A. Zwart, and J. H. DeVries, “Pendra goes Dutch: Lessons for the CE Mark in Europe,” Diabetologia, vol. 48, pp. 1055–1058, 2005. [21] A. Caduff, F. Dewarrat, M. Talary, G. Stalder, L. Heinemann, and Y. Feldman, “Non-invasive glucose monitoring in patients with diabetes: A novel system based on impedance spectroscopy,” Biosensors and Bioelectronics, vol. 22, no. 5, pp. 598–604, Dec. 2006. [22] R. Weiss, Y. Yegorchikov, A. Shusterman, and I. Raz, “Noninvasive Continuous Glucose Monitoring Using Photoacoustic TechnologyResults from the First 62 Subjects,” Diabetes Technology & Therapeutics, vol. 9, no. 1, pp. 68–74, Feb. 2007. [23] I. Harman Boehm, A. Gal, A. M. Raykhman, E. Naidis, and Y. Mayzel, “Noninvasive Glucose Monitoring: Increasing Accuracy by Combination of Multi-Technology and Multi-Sensors,” Journal of Diabetes Science and Technology, vol. 4, no. 3, pp. 583–595, 2010. [24] A. Caduff, M. Mueller, A. Megej, F. Dewarrat, R. E. Suri, J. Klisic, M. Donath, P. Zakharov, D. Schaub, W. A. Stahel, and M. S. Talary, “Characteristics of a multisensor system for non invasive glucose monitoring with external validation and prospective evaluation,” Biosensors and Bioelectronics, vol. 26, no. 9, pp. 3794–800, May 2011. [25] N. C. Dingari, I. Barman, J. W. Kang, C.-R. Kong, R. R. Dasari, and M. S. Feld, “Wavelength selection-based nonlinear calibration for transcutaneous blood glucose sensing using Raman spectroscopy,” Journal of Biomedical Optics, vol. 16, no. 8, pp. 087 009–1–10, Aug. 2011. [26] R. A. Gabbay and S. Sivarajah, “Optical Coherence Tomography-Based Continuous Noninvasive Glucose Monitoring in Patients with Diabetes,” Diabetes Technology & Therapeutics, vol. 10, no. 3, pp. 188–193, Jun. 2008.