0
Research Papers

# Investigations on Multisensor-Based Noninvasive Blood Glucose Measurement SystemOPEN ACCESS

[+] Author and Article Information

Research Lab,
Instrumentation and Control
Engineering Division,
NSIT,
Block-6, Dwarka,
New Delhi 110078, India
e-mail: bmjyoti@gmail.com

Asha Rani

Research Lab,
Instrumentation and Control
Engineering Division,
NSIT,
Block-6, Dwarka,
New Delhi 110078, India
e-mail: ashansit@gmail.com

Vijander Singh

Research Lab,
Instrumentation and Control
Engineering Division,
NSIT,
Block-6, Dwarka,
New Delhi 110078, India
e-mail: vijaydee@gmail.com

Department of Sensors and
Biomedical Technology,
VIT University,
Vellore 632014, India

Manuscript received July 20, 2016; final manuscript received March 13, 2017; published online June 27, 2017. Assoc. Editor: Rafael V. Davalos.

J. Med. Devices 11(3), 031006 (Jun 27, 2017) (7 pages) Paper No: MED-16-1271; doi: 10.1115/1.4036580 History: Received July 20, 2016; Revised March 13, 2017

## Abstract

Noninvasive blood glucose (NIBG) measurement technique has been explored for the last three decades to facilitate diabetes management. Photoplethysmogram (PPG) signal may be used to measure the variations in blood glucose concentration. However, the literature reveals that physiological perturbations such as temperature, skin moisture, and sweat lead to less accurate NIBG measurements. The task of minimizing the effect of these perturbations for accurate measurements is an important research area. Therefore, in the present work, galvanic skin response (GSR) and temperature measurements along with PPG were used to measure blood glucose noninvasively. The data extracted from the sensors were used to estimate blood glucose concentration with the help of two machine learning (ML) techniques, i.e., multiple linear regression (MLR) and artificial neural network (ANN). The accuracy of proposed multisensor system was evaluated by pairing and comparing noninvasive measurements with invasively measured readings. The study was performed on 50 nondiabetic subjects with body mass index (BMI) 27.3 ± 3 kg/m2. The results revealed that multisensor NIBG measurement system significantly improves mean absolute prediction error and correlation coefficient in comparison to the techniques reported in the literature.

<>

## Introduction

Diabetes is a chronic disease in which the body fails to regulate blood glucose concentration in the normal range (90–140 mg/dl) [1,2]. The reason can be inadequate production of insulin or blood cells do not respond to insulin. Poorly managed diabetes can lead to severe health complications. According to a latest survey in 2016, 415 million people have diabetes and this will rise to 592 million by 2035 [3]. Diabetes caused at least 673 billion USD on health expenditure in 2015 which is 12% of total spending on adults across the world [3]. World Health Organization (WHO), Geneva, Switzerland prognosticates that diabetes will be the seventh leading cause of death in 2030 [4].

The diabetes complications can be averted by regular blood glucose monitoring. At present, most of the available glucose measurement devices are invasive or minimally invasive in nature which are painful, costly, and have limited lifetime. Therefore, realization of a painless and clinically accurate, noninvasive blood glucose (NIBG) system is needed to reform diabetes management [5,6]. This NIBG system will also be useful for diabetes-II patients and intensive care patients where continuous monitoring of glucose is required without the risk of infection. NIBG measurement is an intriguing task for researchers for the last three decades [6]. The researchers have been exploring noninvasive technologies for many years; but it is in primal stages of development. The basic principles of NIBG measurement are based on physical phenomena such as changes in optical properties, electrical impedance, and skin temperature due to variations in blood glucose. There are various NIBG measurement technologies available in the literature, but none of the methods provides same accuracy as that of invasive methods [7,8]. Due to low absorption and high penetration, near-infrared spectroscopy (NIRS) is a conceivable optical technique for NIBG measurement [9,10]. Recently, the problems and prospects of NIRS-based NIBG measurements [11] are reviewed which reveal that physical parameters such as variation in pressure, temperature, and chemical parameters cause interferences with glucose measurement. In addition to this, environmental variations such as changes in temperature, humidity, and skin hydration also interfere with the measurement. Further the data sets using only near infrared (NIR) optodes were found less reliable which affect the accuracy of NIBG measurement [12]. Therefore, there is a need of a multisensor based system in order to improve the overall accuracy of the system. Many researchers have attempted NIBG measurement using multiple sensors [1315].

The major difficulty associated with NIBG measurement is the dynamic background noise due to physiological factors which adversely affect NIBG measurement [11]. As investigated by Liu and coworkers [16], temperature is one of the physiological parameters which affects NIBG measurement and alters the optical and electrical properties of skin. Zhang and Yeo [17] showed in in vitro studies that the increase in water temperature causes a change in water absorbance. Therefore, consideration of skin temperature may greatly improve the prediction results. Further, electrical properties of skin also change due to high glucose level in diabetic subjects. Such variations may be compensated by simultaneous consideration of temperature as well as galvanic skin response (GSR). Therefore, in the present work, multiple physiological parameters, i.e., PPG, heart rate (HR), GSR, and skin temperature were considered to minimize the effect of perturbations and thus accurate NIBG measurements were obtained. The relationship between measured data and blood glucose was built with the help of ML techniques. The features extracted from sensor data after suitable processing were taken as inputs and invasively measured blood glucose from glucometer was considered as target output of the neural network. The correlation coefficient R2 was considered as the performance index for selection of ML technique. Two ML algorithms, i.e., MLR and artificial neural network (ANN), were used to estimate blood glucose [14].

###### The Physiological Relevance of Each Signal With Blood Glucose.

The physiological relevance of signals considered for NIBG measurement with blood glucose is discussed in this subsection.

###### Relation Between PPG Signal and Diabetes.

Several studies demonstrate the use of PPG signal for NIBG measurement [1820]. The literature reveals that blood glucose is directly related to blood viscosity, i.e., elevated blood glucose level results in an increase in blood viscosity. Blood viscosity varies with the flux of blood capillaries which manifests as an alteration in the shape of PPG waveform [14,21]. The differential PPG signal was also used to detect blood glucose in some studies. However, in the present work, physiological variations in the shape of PPG signal were measured through spectral analysis of the waveform for NIBG measurement.

###### Relation Between Heart Rate Variability and Diabetes.

Heart rate (HR) was measured from PPG signal to consider the effect of HR variability. The increase in blood glucose level of diabetic subject results in microvascular injury to small blood vessels which results in quite low HR variability [22,23]. The relationship between HR and blood glucose may be captured through the power spectrum and statistics of HR.

###### Relation Between Galvanic Skin Response and Diabetes.

Petrofsky and McLellan [24] have shown in their study that diabetic subjects have higher GSR due to low blood flow and sweat which causes microdamage to the skin. Some studies have shown that skin resistance of a diabetic patient is almost double in comparison to normal subjects. Thus, variation in blood glucose changes the electrical properties of skin which may be measured using GSR.

###### Relation Between Temperature and Diabetes.

Electrical and optical properties of skin are also affected by temperature perturbations. Skin temperature plays a major role in functioning of sweat glands which affects the mechanism for generation of GSR. In addition, Liu and coworkers [16] investigated temperature as one of the physiological parameters which affect NIBG measurement as it changes optical properties of skin. Such variations were compensated by simultaneous temperature measurement using thermister pod. Moreover, Zhang and Yeo [17] have observed that there is around 2–8 times change in absorbance with 1 deg change in water temperature. The authors suggested that including temperature for NIBG measurement improves the prediction considerably. Further Irace et al. [25] suggested that the increase in temperature causes the increase in blood viscosity. According to Poiseuille's law, there is an inverse relationship between viscosity and blood flow [26]. An increase in viscosity results in decreased blood flow, which in turn alters the shape of PPG signal. The skin temperature was therefore taken into consideration in the present work.

The aim of this work was to analyze the global prediction accuracy of multisensor-based system using ML algorithm. The performance of the device was tested on a different database to verify robustness of the model. The subsequent sections are organized as follows: system description and data analysis for the estimation of blood glucose are discussed in Sec. 2. The experimental results are given in Sec. 3. Finally, discussion and conclusion are presented in Secs. 4 and 5, respectively.

## Materials and Methods

The NIBG measurement system was broadly divided into three stages:

1. (1)Data acquisition was carried out using ADInstruments PL3516 Power Lab 16/35 data acquisition system. The ADInstruments system has 16 bit (313 μV resolution in the ±10 V range) resolution. The Lab Chart Version 7.2.1 (ADInstruments) software was used.
2. (2)Feature extraction was carried out using matlab signal processing toolbox.
3. (3)Finally, machine learning based calibration model was developed in matlab to obtain the glucose concentration from multisensor data.

The block diagram of multisensor system is shown in Fig. 1(a) and the sensor attachment is shown in Fig. 1(b). It consists of a sensor unit, signal conditioning circuit, and a calibration model. The sensor unit further consists of a PPG probe, GSR electrodes, and temperature detector along with driving circuit. It was assumed that the combination of these sensor signals allows a more reliable NIBG monitoring.

###### Measurement of PPG Signal.

The present work utilizes PPG signal to detect blood glucose. MLT1010F pulse transducer was used to measure the PPG signal. The skin was illuminated by NIR LED and changes in light reflectance were obtained as PPG signal using photodiode. The NIR–LED considered has 940 nm wavelength at which glucose has a fairly good absorption as compared to other blood chromophores. Bioamp (FF136) was used for the amplification of PPG signal. It also provided isolation of human body from electrical circuitry. The PPG signals were recorded at a rate of 100 samples/s; after necessary amplification, a low pass filter having 10 Hz cutoff frequency was used to filter high-frequency components and power line interference. High pass filter of 0.5 Hz cut off frequency was used to remove lower frequency signals due to various physiological noises such as respiration, baseline wander, and thermoregulation.

###### Measurement of GSR.

The measurement of GSR or electrodermal response (EDR) is a harmless and noninfectious method as it involves two stainless steel dry electrodes wrapped around the middle finger and ring finger of subjects. A small amount of voltage was applied to the electrodes in order to measure skin conductance. GSR was obtained by measuring the amount of current flowing through the electrodes. The unit of conductance is micro Siemens. Generally, GSR is measured from body parts which have more sweat glands. Only specific sites are considered for GSR measurement where eccrine glands have maximum sensitivity [27] to get the maximum changes in output response. Two electrodes were used to measure the GSR signal. One active electrode was placed on ventral side of the finger, as this side has more eccrine glands. Second electrode is reference electrode, placed on dorsal side of finger, and has very less eccrine glands.

GSR response of the skin was monitored using MLT116F module having two finger electrodes which uses 75 Hz AC excitation. AC excitation minimizes polarization of electrodes caused by the DC source. The manufacturer advised zeroing of the apparatus before taking each measurement [28] so as to get the precise and accurate measurements and it takes a few seconds. The electrical output was amplified with a gain of 1000 using FE116 GSR amplifier. GSR signal was recorded at the rate of 100 samples/s. As GSR varies slowly, a 4250 ms epoch was extracted. The frequency of GSR signal lies in the range 0–8 Hz which requires a low pass filter of 8 Hz to remove noise.

###### Measurement of Skin Temperature.

ML309 temperature module having thermistor pod was placed on the skin to monitor temperature. The change in resistance is converted into a voltage signal which is further amplified to produce the output at a rate of 50 mV/ °C. The signal was sampled at a rate of 200 samples/s. The recalibration of temperature transducer is carried out twice a month for more accurate measurements as specified by manufacturer. The recalibration option is provided in the apparatus by manufacturer.

###### PPG Signal.

The optical sensor probe attached to index finger was used to observe a diffused reflectance spectrum. The measured signal has both AC and DC components. The DC component is mainly due to bones and other tissues. The AC component of PPG carries absorption information due to glucose changes. Optical density is calculated using the below equation: Display Formula

(1)$ΔODλ=log[1+(ΔIλ(ti)Iλ(ti+1))]$

where $ΔODλ$ is the difference between optical densities at different intervals of time, $Iλ(ti)$ is the optical density at ti, and $Iλ(ti+1)$ is the optical density at ti+1.

A set of 5000 consecutive samples Xwindow(t) was extracted from PPG signals which were further divided into 500 frame sizes. As 10 s signal was sufficient to calculate heart rate, an array of X(τ, n) was extracted consisting ten frames (FFrame) of 10 s duration each. τ is the time index of frame and n is the frame index. A set of features YM was extracted from each frame as follows.

###### Kaiser–Teager Energy (KTE).

Kaiser–Teager energy of PPG signal was calculated and used to find the energy profile of periodic signals. It was calculated for each frame of signal and then added to obtain the sum of energy. A high KTE means signal amplitude is high in that frame, whereas low KTE is due to background noises and transients. Average value of features, that is mean $KTEnμ$, variance $KTEnσ$, interquartile range $KTEniqr$, and skewness $KTEnskew$ were also calculated.

###### Spectral Entropy.

$Hns$ of PPG signal was also calculated to classify each frame. For noises or transients, spectral entropy was high, whereas for clean signal it had low value. Therefore, the signal can easily be separated from a noisy signal. This feature gives information about spectral shape, damping, and noise-related information. FFT of the frame was calculated with the help of zero padding in order to increase the length of the signal.

The mean value of entropy $Hμ$, the variance $Hσ$ interquartile range $Hiqr$, and skewness $Hskew$ was also calculated for PPG signal.

###### Heart Rate.

Heart rate was calculated as one feature from $Xwindow(t)$. Mean heart rate $HRμ$, variance $HRσ$ interquartile range $HRiqr$, and skewness $HRskew$ were also calculated for PPG signal.

###### Person-Specific Information.

Information about the subjects, such as age (AG), weight (Wt), and BMI, was also considered.

###### Galvanic Skin Response.

GSR amplitude was measured as the difference of peaks of response to the baseline amplitude.

###### Skin Temperature.

Skin temperature was measured using thermocouple probe.

The overall feature vector $Yn$ becomes Display Formula

(2)$Yn=[KTEnμ,KTEnσ,KTEniqr,KTEnskew,Hσ,Hμ,Hiqr,Hskew,HRμ,HRσ,HRiqr,HRskew, GSR, temp, AG, BMI,Wt]$

Fifty normal subjects participated in this pilot study, out of which 35 subjects were male and 15 were females. The average age was 25 ± 5 yr and the average BMI was 27.3 ± 3 kg/m2. The pilot study was approved by the Institute Ethical Committee (IEC). Informed consent was obtained from all participants before experimentation. The blood glucose of 50 subjects was measured using multiple sensors in fasting condition. Subjects were asked to sit on a chair in relaxed condition with hand maintained at heart level. Physical activity during the measurement was prohibited to minimize the motion artifacts. Later on, the subjects were asked to take 75 g of glucose dissolved in 300 ml of water and four measurements were taken at an interval of 30 min [15]. Thus, five measurements were taken for each subject. The results were validated by simultaneously measuring the blood glucose with the help of commercially available invasive fingertip glucometer. The complete measurement was repeated for all the subjects, which generated a data of 50 × 5 × 2 = 500 patterns. All the measurements were carried out in the duration of one month. The dataset used for testing and validation purpose was not used in training of the neural network.

###### Data Analysis for Estimation of Blood Glucose.

The relationship between feature matrix and blood glucose was formulated using ML techniques. The relationship built by ML should be generalized so that it does not give inaccurate results for new data points [29]. The technique should tackle peculiarities of data and deal with noisy data as well. The correlation coefficient (R2) was considered as performance index for selecting the ML technique. Blood glucose concentration was estimated from extracted feature vector $Yn$ with the help of multiple linear regression and ANN techniques. PPG signal is highly sensitive to motion artifacts and varies from person to person due to the difference in fatty tissues as well as the path length (thickness of a finger). In addition, there were various factors which affect the amount of blood glucose level, such as the amount of meal, physical work, and level of stress [30]. Thus, a nonlinear relationship exists between the input data (feature vector) and the target data (actual glucose level). ANN is well suited for the characterization in complex nonlinearities [18]. The main advantage of ANN is its learning capability for development of new solution to problems which is not well defined. Therefore, it was used to model the relationship between blood glucose level and sensors output. There are various types of ANN model based on architecture, learning algorithm, and activation function [30]. The estimation of blood glucose level was investigated by multilayer perceptrons (MLP), Levenberg–Marquardt (LM) [3135] method. It is a more powerful learning technique than conventional gradient descent method. It generally provides faster convergence, better estimation, and has ability to train online in comparison to other training algorithms. It is also known to deal with computationally complex problems and functional relationships between sets of data.

The input data to the neural network were a feature vector obtained from various sensors and target output was blood glucose concentration which was obtained from Glucometer. The neural network weights were adjusted to get minimum mean squared error (MSE). The trained ANN was used to estimate the glucose values from multisensor data.

###### Multiple Linear Regression (MLR).

Initially, MLR was used to build the relationship between multiple sensor data and blood glucose. Linear regression was used to estimate the outcome $Y$. The general form of regression is

$Y=β0+β1x1+β2x2⋯βnxn+ε$

where $Y$ is the variable to be estimated, x1 to xn are the independent variables, and $ε$ is the error.

###### Levenberg–Marquardt Based Calibration Model to Estimate Blood Glucose.

The Levenberg–Marquardt (LM)-based ANN is widely used in the medical field for analysis purpose [32,33]. LM is a combination of steepest descent and Gauss–Newton method. If the present solution is distant from the accurate one, the algorithm acts as a steepest descent method; it is sluggish, but assures convergence. In case the present solution is near the accurate solution, then the algorithm acts as Gauss–Newton method.

LM method is very simple and robust method for nonlinear least square problems [34,35]. The weights of ANN are updated according to LM algorithm while minimizing the MSE between the target and network output. Initially, solution matrix at step k + 1 can be written as Display Formula

(3)$xk+1=xk−Ak−1gk$

where $xk$ is the solution vector at step kDisplay Formula

(4)$Ak=∇2F(x)|x=xk$
Display Formula
(5)$gk=∇F(x)|x=xk$

Ak is the Hessian matrix evaluated at $xk$ and gk is the gradient vector evaluated corresponding to $xk$.

Let us assume F(x) as sum of squares of function then Display Formula

(6)$F(x)=∑i=1nv2i(x)=vT(x)v(x)$

where vi(x) is the error function and v(x) the error vector.

The gradient vector can be written in matrix form as Display Formula

(7)$∇F(x)=2JT(x)v(x)$

The Hessian matrix can be approximated as Display Formula

(8)$∇2F(x)≅2JT(x)J(x)$

where J(x) is the Jacobian matrix.

From Eqs. (3)(5), the subsequent value of xk is obtained as Display Formula

(9)$xk+1=xk−[JT(xk)J(xk)]−1JT(xk)v(xk)$

Equation (7) is known as Gauss Newton method. After modification, the estimated Hessian matrix becomes the improved Hessian matrix, G, and evaluated as Display Formula

(10)$G=H+μI$

where H is the approximate Hessian matrix, μ is the eigenvalue of approximated Hessian matrix (momentum parameter), I is the unity matrix.

Equations (7) and (8) lead to LM algorithm Display Formula

(11)$xk+1=xk−[JT(xk)J(xk)+μkI]−1JT(xk)v(xk)$

To calculate the Jacobian matrix, back propagation algorithm was used which is the key step in LM algorithm. The structure of LM neural network under consideration consisted of input layer, one hidden layer, and output layer. Tangent sigmoid and log sigmoid activation functions were used for input layer and hidden layer, respectively, whereas purelin was used in output layer.

###### Sensitivity Analysis to Choose the Number of Neurons in Hidden Layer.

The number of neurons in hidden layer was chosen using Eq. (12) [36]. N is the number of neurons in the hidden layer and S is the total number of samples. I and O are the number of input and output variables, respectively, Display Formula

(12)$N=I+O2+S$

The sensitivity analysis was carried out to validate the number of neurons in hidden layer by measuring the change in correlation coefficient R2. The number of neurons considered in the experimentations is ±5 of neurons calculated as per Eq. (12).

###### Measures for Performance Evaluation.

The performance of in vivo system was evaluated in terms of widely used indices, i.e., mean squared error (MSE), mean absolute percentage error (MAPE), and correlation coefficient (R2).

The value of R2 signifies the relationship between input and output of neural network and was obtained between 0 and 1. The plot of predicted values against target values having 45 deg slope indicates strong correlation. The performance of the system was also evaluated using MAPE. 10% ≤ MAPE ≤ 20% shows good estimation, 20% ≤ MAPE ≤ 50% indicates poor estimation, and MAPE ≥ 50% shows an inaccurate estimation [33,34]. Training of the network is stopped when MSE is minimized. The low values of MSE and MAPE associated with a high R2 are an indication of efficient estimation of blood glucose.

## Results

Seventeen features were extracted (Yn) and considered as input to MLR and ANN. The desired output was glucose concentration. The dataset of 500 input–output patterns was divided into three parts: 70% of the dataset was used for training, 15% for validation, and remaining 15% for testing the neural network. ML technique was implemented using matlab on Intel core TM i3-370M, 2 GB RAM, 2.40 GHz processor with real patient experimental data set. The characteristics of reference blood glucose values are given in Table 1 and the characteristics of the subjects under study are given in Table 2.

The correlation coefficient using MLR technique was observed to be 0.612. The nonlinearity between data and blood glucose was observed due to the presence of several components with overlapping spectral features. Further, the value of R2 for MLR technique was very low, therefore, ANN was considered in the present work for the estimation of blood glucose. An MLP network was trained using LM algorithm. This network was selected because it approximates nonlinear relationships efficiently. The structure (Fig. 2) consisted of 17, 32, and 1 neurons in input layer, hidden layer, and output layer, respectively. The hidden layer and output layer had biasing to avoid saturation. The outputs shared the information from hidden units in order to make use of common traits of the functions to be estimated.

Regression plot for validation is given in Fig. 3. In order to evaluate the prediction accuracy of ANN, different numbers of neuron in hidden layer were considered for sensitivity analysis. The number of samples in the experiment is 500 and the number of input and output variables was 17 and 1, respectively. The number of neurons in the hidden layer estimated from Eq. (12) was 31. Therefore, hidden neurons were considered in the range of 26–36. Maximum value of R2 was observed using 32 neurons in the hidden layer. Therefore, further analysis was made by considering 32 neurons in the hidden layer. The sensitivity analysis for the number of neurons in the hidden layer is given in Table 3. R2 achieved for validation was 0.94 as shown in Fig. 3 which implies that the estimated blood glucose was very close to invasively measured values.

The MAPE observed in the present work was 9.21%, which indicates the good accuracy of proposed approach.

The performance plot shown in Fig. 4 reveals that MSE is reduced for more number of epochs. The test and validation samples had analogous characteristics. No major overfitting was caused near the epoch 14 at which both test and validation performance was best. The error histogram demonstrated in Fig. 5 provides additional verification of the proposed approach.

Clark error grid analysis shown in Fig. 6 was performed to analyze the clinical accuracy. Clark error grid is generally used for clinical authentication of medical devices. It was observed from results that 86.01% points lie in region A and 13.99% points lie in region B, none of the points lies in C, D, and E regions. The predicted result reveals that the maximum inaccuracy in measurement was only 20%; however, it will not lead to inappropriate treatment.

###### Tenfold Cross-Validation.

To check over fitting of data by ML technique, tenfold validation was performed. Where 90% of data were used for training purpose and rest 10% was used for validation. This was performed ten times and every time data were rotated. The result of tenfold validation is shown in Table 4.

## Discussion

The results obtained from experimentation were compared with the results reported in the literature as shown in Table 5. It is revealed from the results that the values of performance measures were significantly improved as compared to the previous works. Thus multisensor-based approach provides a significant improvement in the noninvasive blood glucose measurements. The results were further validated using the commonly used methods for global validation of experimental results. Tenfold validation and Clark error grid analysis [8] used for clinical authentication of medical devices were considered for the purpose.

In this work, data collection was carried out in low movement condition to avoid movement artifacts and to establish good baseline conduction. PPG signals can be easily influenced by motion artifacts, which may lead to inaccurate interpretation of the waveform [37]. The signal measured during normal activities can cause motion artifacts which can be removed with the help of suitable filters. The artifacts can be reduced in real-time devices by least mean square based active noise-cancelation method [38]. Moreover in the literature, research groups proposed various noise-cancelation algorithms to improve the accuracy of PPG sensor (e.g., moving average filter, Fourier analysis, adaptive filters, Kalman filter, and wavelet transformation) [39].

The present work is a preliminary study based on the dataset of only 50 nondiabetic healthy subjects,; however, it will be extended to include the diabetic subjects for further improvement of the proposed NIBG system. The multisensor approach was tested only for short-term measurements. The results obtained were encouraging and may be tested for long-term monitoring also. The long-term blood glucose measurements are erroneous due to unpredictable and variable physiological parameters [8]. The effects of these parameters were compensated by considering them as inputs to the proposed system. Therefore, the NIBG measurement technique may serve as a viable tool for accurate long-term diagnostics.

## Conclusion

It is evident from studies that physiological perturbations severely affect the accuracy of NIBG measurement. This research work presents an effective approach to minimize the effect of these perturbations with the help of multisensor measurements. ANN-based calibration model was used to estimate the blood glucose from multiple sensor data. The pilot study was carried out on 50 nondiabetic subjects with BMI 27.3 ± 3 kg/m2. MLR and ANN were used to estimate the glucose concentration from extracted data. A significantly low MAPE (9.21%) and high R2 (0.94) demonstrated the capability of the multisensor approach to improve the accuracy of NIBG system. All the measurements were lying in clinically acceptable A and B zones of Clarke error grid. In future, larger dataset including diabetic subjects will be tested for long-term monitoring.

## References

Diabetes.co.uk, 2017, “ Blood Sugar Level Ranges,” Diabetes Digital Media Ltd., Coventry, UK, accessed May 4, 2017,
American Diabetes Association, 2009, “ Diagnosis and Classification of Diabetes Mellitus,” Diabetes Care, 32(Suppl. 1), pp. S62–S67. [PubMed]
Sadikot, S. , 2016, “ IDFs Global Voice in the Diabetes Landscape,” Diabetes. Res. Clin. Pract., 114, pp. 180–181.
WHO, 2014, “ Global Status Report on Non Communicable Diseases,” World Health Organization, Geneva, Switzerland, accessed May 5, 2017,
Yadav, J. , Rani, A. , Singh, V. , and Murari, B. M. , 2014, “ Near-Infrared LED Based Non-Invasive Blood Glucose Sensor,” International Conference on Signal Processing and Integrated Networks (SPIN), Noida, India, Feb. 20–21, pp. 591–594.
Nilsson, G. E. , Tenland, T. , and Öberg, P. A. , 1982, “ Laser-Doppler Methods for the Assessment of Microcirculatory Blood Flow,” Trans. Inst. Meas. Control, 4(2), pp. 109–112.
Liakat, S. , Bors, K. A. , Xu, L. , Woods, C. M. , Doyle, J. , and Gmachl, C. F. , 2014, “ Noninvasive In Vivo Glucose Sensing on Human Subjects Using Mid-Infrared Light,” Biomed. Opt. Express, 5(7), pp. 2397–2404. [PubMed]
Tamada, J. A. , Garg, S. , Jovanovic, L. , Pitzer, K. R. , Fermi, S. , Potts, R. O. , and Cygnus Research Team, 1999, “ Noninvasive Glucose Monitoring: Comprehensive Clinical Results,” JAMA, 282(19), pp. 1839–1844. [PubMed]
Luo, Y. , An, L. , Ma, Z. , Liu, R. , and Xu, K. , 2006, “ Discussion on Floating-Reference Method for Noninvasive Measurement of Blood Glucose With Near-Infrared Spectroscopy,” Proc. SPIE, 6094, p. 60940K.
Yadav, J. , Rani, A. , Singh, V. , and Murari, B. M. , 2015, “ Comparative Study of Different Measurement Sites Using NIR Based Non-Invasive Glucose Measurement System,” Procedia Comput. Sci., 70, pp. 469–475.
Yadav, J. , Rani, A. , Singh, V. , and Murari, B. M. , 2015, “ Prospects and Limitations of Non-Invasive Blood Glucose Monitoring Using Near-Infrared Spectroscopy,” Biomed. Signal Process. Control, 18, pp. 214–227.
Zanon, M. , Sparacino, G. , Facchinetti, A. , Talary, M. S. , Mueller, M. , Caduff, A. , and Cobelli, C. , 2013, “ Non-Invasive Continuous Glucose Monitoring With Multi-Sensor Systems: A Monte Carlo-Based Methodology for Assessing Calibration Robustness,” Sensors, 13(6), pp. 7279–7295. [PubMed]
Caduff, A. , Talary, M. S. , Mueller, M. , Dewarrat, F. , Klisic, J. , Donath, M. , Heinemann, L. , and Stahel, W. A. , 2009, “ Non-Invasive Glucose Monitoring in Patients With Type 1 Diabetes: A Multisensor System Combining Sensors for Dielectric and Optical Characterisation of Skin,” Biosens. Bioelectron., 24(9), pp. 2778–2784. [PubMed]
Monte-Moreno, E. , 2011, “ Non-Invasive Estimate of Blood Glucose and Blood Pressure From a Photoplethysmograph by Means of Machine Learning Techniques,” Artif. Intell. Med., 53(2), pp. 127–138. [PubMed]
Chowdhury, M. K. , Srivastava, A. , Sharma, N. , and Sharma, S. , 2015, “ Noninvasive Blood Glucose Measurement Utilizing a Newly Designed System Based on Modulated Ultrasound and Infrared Light,” Int. J. Diabetes Dev. Countries, 36(4), pp. 439–448.
Zhang, W. , Liu, R. , Zhang, W. , Jia, H. , and Xu, K. , 2013, “ Discussion on the Validity of NIR Spectral Data in Non-Invasive Blood Glucose Sensing,” Biomed. Opt. Express, 4(6), pp. 789–802. [PubMed]
Zhang, X. , and Yeo, J. H. , 2009, “ Temperature Influence on Non-Invasive Blood Glucose Measurement,” Proc. SPIE, 7186, p. 71860R.
Ramasahayam, S. , Haindavi, K. S. , and Chowdhury, S. R. , 2015, “ Noninvasive Estimation of Blood Glucose Concentration Using Near Infrared Optodes,” Sensing Technology: Current Status and Future Trends IV, Springer International Publishing, Cham, Switzerland, pp. 67–82.,
Ramasahayam, S., Koppuravuri, S. H., Arora, L., and Chowdhury, S. R., 2015, “ Noninvasive Blood Glucose Sensing Using Near Infra-Red Spectroscopy and Artificial Neural Networks Based on Inverse Delayed Function Model of Neuron,” J. Medical Systems, 39(1), p. 166.
Yamakoshi, K. , and Yamakoshi, Y. , 2006, “ Pulse Glucometry: A New Approach for Noninvasive Blood Glucose Measurement Using Instantaneous Differential Near-Infrared Spectrophotometry,” J. Biomed. Opt., 11(5), p. 054028. [PubMed]
Çinar, Y. , Şenyol, A. M. , and Duman, K. , 2001, “ Blood Viscosity and Blood Pressure: Role of Temperature and Hyperglycemia,” Am. J. Hypertens., 14(5), pp. 433–438. [PubMed]
Ducher, M. , Cerutti, C. , Gustin, M. P. , Abou-Amara, S. , Thivolet, C. , Laville, M. , Paultre, C. Z. , and Fauvel, J. P. , 1999, “ Noninvasive Exploration of Cardiac Autonomic Neuropathy. Four Reliable Methods for Diabetes?,” Diabetes Care, 22(3), pp. 388–393. [PubMed]
Van Ravenswaaij-Arts, C. M. , Kollee, L. A. , Hopman, J. C. , Stoelinga, G. B. , and van Geijn, H. P. , 1993, “ Heart Rate Variability,” Ann. Intern. Med., 118(6), pp. 436–447. [PubMed]
Petrofsky, J. S. , and McLellan, K. , 2009, “ Galvanic Skin Resistance—A Marker for Endothelial Damage in Diabetes,” Diabetes Technol. Ther., 11(7), pp. 461–467. [PubMed]
Irace, C. , Carallo, C. , Scavelli, F. , Esposito, T. , De Franceschi, M. S. , Tripolino, C. , and Gnasso, A. , 2014, “ Influence of Blood Lipids on Plasma and Blood Viscosity,” Clin. Hemorheol. Microcirc., 57(3), pp. 267–274. [PubMed]
Palanisamy, K. , Murugappan, M. , and Yaacob, S. , 2013, “ Multiple Physiological Signal-Based Human Stress Identification Using Non-Linear Classifiers,” Elektron. Elektrotech., 19(7), pp. 80–85.
Conesa, J. , 1995, “ Electrodermal Palmar Asymmetry and Nostril Dominance,” Perceptual Mot. Skills, 80(1), pp. 211–216.
ADINSTRUMENTS, 2000, “ Front-End Signal Conditioners,” ADInstruments Pty Ltd., Sydney, Australia, accessed May 5, 2017,
Ramasahayam, S. , and Chowdhury, S. R. , 2016, “ Non Invasive Estimation of Blood Urea Concentration Using Near Infrared Spectroscopy,” Int. J. Smart Sens. Intell. Syst., 9(2), pp. 449–467.
Ghobadian, B. , Rahimi, H. , Nikbakht, A. M. , Najafi, G. , and Yusaf, T. F. , 2009, “ Diesel Engine Performance and Exhaust Emission Analysis Using Waste Cooking Biodiesel Fuel With an Artificial Neural Network,” Renewable Energy, 34(4), pp. 976–982.
Gaidhane, V. H. , Hote, Y. V. , and Singh, V. , 2012, “ Nonrigid Image Registration Using Efficient Similarity Measure and Levenberg-Marquardt Optimization,” Biomed. Eng. Lett., 2(2), pp. 118–123.
Gaidhane, V. H. , Hote, Y. V. , and Singh, V. , 2016, “ Emotion Recognition Using Eigenvalues and Levenberg–Marquardt Algorithm-Based Classifier,” Sādhanā, 41(4), pp. 1–9.
Rani, A. , Singh, V. , and Gupta, J. R. P. , 2013, “ Development of Soft Sensor for Neural Network Based Control of Distillation Column,” ISA Trans., 52(3), pp. 438–449. [PubMed]
Rani, A. , Singh, V. , and Gupta, J. R. P. , 2011, “ Soft Sensor Based on Adaptive Linear Network for Distillation Process,” Int. J. Comput. Appl., 36(1), pp. 39–45.
Troy, T. L. , and Thennadil, S. N. , 2001, “ Optical Properties of Human Skin in the Near Infrared Wavelength Range of 1000 to 2200 Nm,” J. Biomed. Opt., 6(2), pp. 167–176. [PubMed]
Yadav, A. K. , Malik, H. , and Chandel, S. S. , 2014, “ Selection of Most Relevant Input Parameters Using WEKA for Artificial Neural Network Based Solar Radiation Prediction Models,” Renewable Sustainable Energy Rev., 31, pp. 509–519.
Couceiro, R. , Carvalho, P. , Paiva, R. P. , Henriques, J. , and Muehlsteff, J. , 2012, “ Detection of Motion Artifacts in Photoplethysmographic Signals Based on Time and Period Domain Analysis,” Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), San Diego, CA, Aug. 28–Sept. 1, pp. 2603–2606.
Han, H. , and Kim, J. , 2012, “ Artifacts in Wearable Photoplethysmographs During Daily Life Motions and Their Reduction With Least Mean Square Based Active Noise Cancellation Method,” Comput. Biol. Med., 42(4), pp. 387–393. [PubMed]
Hwang, S. , Seo, J. , Jebelli, H. , and Lee, S. , 2016, “ Feasibility Analysis of Heart Rate Monitoring of Construction Workers Using a Photoplethysmography (PPG) Sensor Embedded in a Wristband-Type Activity Tracker,” Autom. Constr., 71(Pt. 2), pp. 372–381.
View article in PDF format.

## References

Diabetes.co.uk, 2017, “ Blood Sugar Level Ranges,” Diabetes Digital Media Ltd., Coventry, UK, accessed May 4, 2017,
American Diabetes Association, 2009, “ Diagnosis and Classification of Diabetes Mellitus,” Diabetes Care, 32(Suppl. 1), pp. S62–S67. [PubMed]
Sadikot, S. , 2016, “ IDFs Global Voice in the Diabetes Landscape,” Diabetes. Res. Clin. Pract., 114, pp. 180–181.
WHO, 2014, “ Global Status Report on Non Communicable Diseases,” World Health Organization, Geneva, Switzerland, accessed May 5, 2017,
Yadav, J. , Rani, A. , Singh, V. , and Murari, B. M. , 2014, “ Near-Infrared LED Based Non-Invasive Blood Glucose Sensor,” International Conference on Signal Processing and Integrated Networks (SPIN), Noida, India, Feb. 20–21, pp. 591–594.
Nilsson, G. E. , Tenland, T. , and Öberg, P. A. , 1982, “ Laser-Doppler Methods for the Assessment of Microcirculatory Blood Flow,” Trans. Inst. Meas. Control, 4(2), pp. 109–112.
Liakat, S. , Bors, K. A. , Xu, L. , Woods, C. M. , Doyle, J. , and Gmachl, C. F. , 2014, “ Noninvasive In Vivo Glucose Sensing on Human Subjects Using Mid-Infrared Light,” Biomed. Opt. Express, 5(7), pp. 2397–2404. [PubMed]
Tamada, J. A. , Garg, S. , Jovanovic, L. , Pitzer, K. R. , Fermi, S. , Potts, R. O. , and Cygnus Research Team, 1999, “ Noninvasive Glucose Monitoring: Comprehensive Clinical Results,” JAMA, 282(19), pp. 1839–1844. [PubMed]
Luo, Y. , An, L. , Ma, Z. , Liu, R. , and Xu, K. , 2006, “ Discussion on Floating-Reference Method for Noninvasive Measurement of Blood Glucose With Near-Infrared Spectroscopy,” Proc. SPIE, 6094, p. 60940K.
Yadav, J. , Rani, A. , Singh, V. , and Murari, B. M. , 2015, “ Comparative Study of Different Measurement Sites Using NIR Based Non-Invasive Glucose Measurement System,” Procedia Comput. Sci., 70, pp. 469–475.
Yadav, J. , Rani, A. , Singh, V. , and Murari, B. M. , 2015, “ Prospects and Limitations of Non-Invasive Blood Glucose Monitoring Using Near-Infrared Spectroscopy,” Biomed. Signal Process. Control, 18, pp. 214–227.
Zanon, M. , Sparacino, G. , Facchinetti, A. , Talary, M. S. , Mueller, M. , Caduff, A. , and Cobelli, C. , 2013, “ Non-Invasive Continuous Glucose Monitoring With Multi-Sensor Systems: A Monte Carlo-Based Methodology for Assessing Calibration Robustness,” Sensors, 13(6), pp. 7279–7295. [PubMed]
Caduff, A. , Talary, M. S. , Mueller, M. , Dewarrat, F. , Klisic, J. , Donath, M. , Heinemann, L. , and Stahel, W. A. , 2009, “ Non-Invasive Glucose Monitoring in Patients With Type 1 Diabetes: A Multisensor System Combining Sensors for Dielectric and Optical Characterisation of Skin,” Biosens. Bioelectron., 24(9), pp. 2778–2784. [PubMed]
Monte-Moreno, E. , 2011, “ Non-Invasive Estimate of Blood Glucose and Blood Pressure From a Photoplethysmograph by Means of Machine Learning Techniques,” Artif. Intell. Med., 53(2), pp. 127–138. [PubMed]
Chowdhury, M. K. , Srivastava, A. , Sharma, N. , and Sharma, S. , 2015, “ Noninvasive Blood Glucose Measurement Utilizing a Newly Designed System Based on Modulated Ultrasound and Infrared Light,” Int. J. Diabetes Dev. Countries, 36(4), pp. 439–448.
Zhang, W. , Liu, R. , Zhang, W. , Jia, H. , and Xu, K. , 2013, “ Discussion on the Validity of NIR Spectral Data in Non-Invasive Blood Glucose Sensing,” Biomed. Opt. Express, 4(6), pp. 789–802. [PubMed]
Zhang, X. , and Yeo, J. H. , 2009, “ Temperature Influence on Non-Invasive Blood Glucose Measurement,” Proc. SPIE, 7186, p. 71860R.
Ramasahayam, S. , Haindavi, K. S. , and Chowdhury, S. R. , 2015, “ Noninvasive Estimation of Blood Glucose Concentration Using Near Infrared Optodes,” Sensing Technology: Current Status and Future Trends IV, Springer International Publishing, Cham, Switzerland, pp. 67–82.,
Ramasahayam, S., Koppuravuri, S. H., Arora, L., and Chowdhury, S. R., 2015, “ Noninvasive Blood Glucose Sensing Using Near Infra-Red Spectroscopy and Artificial Neural Networks Based on Inverse Delayed Function Model of Neuron,” J. Medical Systems, 39(1), p. 166.
Yamakoshi, K. , and Yamakoshi, Y. , 2006, “ Pulse Glucometry: A New Approach for Noninvasive Blood Glucose Measurement Using Instantaneous Differential Near-Infrared Spectrophotometry,” J. Biomed. Opt., 11(5), p. 054028. [PubMed]
Çinar, Y. , Şenyol, A. M. , and Duman, K. , 2001, “ Blood Viscosity and Blood Pressure: Role of Temperature and Hyperglycemia,” Am. J. Hypertens., 14(5), pp. 433–438. [PubMed]
Ducher, M. , Cerutti, C. , Gustin, M. P. , Abou-Amara, S. , Thivolet, C. , Laville, M. , Paultre, C. Z. , and Fauvel, J. P. , 1999, “ Noninvasive Exploration of Cardiac Autonomic Neuropathy. Four Reliable Methods for Diabetes?,” Diabetes Care, 22(3), pp. 388–393. [PubMed]
Van Ravenswaaij-Arts, C. M. , Kollee, L. A. , Hopman, J. C. , Stoelinga, G. B. , and van Geijn, H. P. , 1993, “ Heart Rate Variability,” Ann. Intern. Med., 118(6), pp. 436–447. [PubMed]
Petrofsky, J. S. , and McLellan, K. , 2009, “ Galvanic Skin Resistance—A Marker for Endothelial Damage in Diabetes,” Diabetes Technol. Ther., 11(7), pp. 461–467. [PubMed]
Irace, C. , Carallo, C. , Scavelli, F. , Esposito, T. , De Franceschi, M. S. , Tripolino, C. , and Gnasso, A. , 2014, “ Influence of Blood Lipids on Plasma and Blood Viscosity,” Clin. Hemorheol. Microcirc., 57(3), pp. 267–274. [PubMed]
Palanisamy, K. , Murugappan, M. , and Yaacob, S. , 2013, “ Multiple Physiological Signal-Based Human Stress Identification Using Non-Linear Classifiers,” Elektron. Elektrotech., 19(7), pp. 80–85.
Conesa, J. , 1995, “ Electrodermal Palmar Asymmetry and Nostril Dominance,” Perceptual Mot. Skills, 80(1), pp. 211–216.
ADINSTRUMENTS, 2000, “ Front-End Signal Conditioners,” ADInstruments Pty Ltd., Sydney, Australia, accessed May 5, 2017,
Ramasahayam, S. , and Chowdhury, S. R. , 2016, “ Non Invasive Estimation of Blood Urea Concentration Using Near Infrared Spectroscopy,” Int. J. Smart Sens. Intell. Syst., 9(2), pp. 449–467.
Ghobadian, B. , Rahimi, H. , Nikbakht, A. M. , Najafi, G. , and Yusaf, T. F. , 2009, “ Diesel Engine Performance and Exhaust Emission Analysis Using Waste Cooking Biodiesel Fuel With an Artificial Neural Network,” Renewable Energy, 34(4), pp. 976–982.
Gaidhane, V. H. , Hote, Y. V. , and Singh, V. , 2012, “ Nonrigid Image Registration Using Efficient Similarity Measure and Levenberg-Marquardt Optimization,” Biomed. Eng. Lett., 2(2), pp. 118–123.
Gaidhane, V. H. , Hote, Y. V. , and Singh, V. , 2016, “ Emotion Recognition Using Eigenvalues and Levenberg–Marquardt Algorithm-Based Classifier,” Sādhanā, 41(4), pp. 1–9.
Rani, A. , Singh, V. , and Gupta, J. R. P. , 2013, “ Development of Soft Sensor for Neural Network Based Control of Distillation Column,” ISA Trans., 52(3), pp. 438–449. [PubMed]
Rani, A. , Singh, V. , and Gupta, J. R. P. , 2011, “ Soft Sensor Based on Adaptive Linear Network for Distillation Process,” Int. J. Comput. Appl., 36(1), pp. 39–45.
Troy, T. L. , and Thennadil, S. N. , 2001, “ Optical Properties of Human Skin in the Near Infrared Wavelength Range of 1000 to 2200 Nm,” J. Biomed. Opt., 6(2), pp. 167–176. [PubMed]
Yadav, A. K. , Malik, H. , and Chandel, S. S. , 2014, “ Selection of Most Relevant Input Parameters Using WEKA for Artificial Neural Network Based Solar Radiation Prediction Models,” Renewable Sustainable Energy Rev., 31, pp. 509–519.
Couceiro, R. , Carvalho, P. , Paiva, R. P. , Henriques, J. , and Muehlsteff, J. , 2012, “ Detection of Motion Artifacts in Photoplethysmographic Signals Based on Time and Period Domain Analysis,” Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), San Diego, CA, Aug. 28–Sept. 1, pp. 2603–2606.
Han, H. , and Kim, J. , 2012, “ Artifacts in Wearable Photoplethysmographs During Daily Life Motions and Their Reduction With Least Mean Square Based Active Noise Cancellation Method,” Comput. Biol. Med., 42(4), pp. 387–393. [PubMed]
Hwang, S. , Seo, J. , Jebelli, H. , and Lee, S. , 2016, “ Feasibility Analysis of Heart Rate Monitoring of Construction Workers Using a Photoplethysmography (PPG) Sensor Embedded in a Wristband-Type Activity Tracker,” Autom. Constr., 71(Pt. 2), pp. 372–381.

## Figures

Fig. 1

(a) Block diagram of NIBG measurement and (b) multiple sensor based NIBG measurement system

Fig. 2

Detailed ANN structure

Fig. 3

Regression analysis of neural network based on LM-based ANN model

Fig. 4

Best validation performance

Fig. 5

Error histogram

Fig. 6

Clark error grid analysis (EGA) for LM

## Tables

Table 1 Characteristics of reference blood glucose values
Table 2 Characteristics of the study population
Table 3 Sensitivity analysis for number of neurons in hidden layer
Note: With 32 numbers of neurons, maximum value of correlation coefficient is achieved.
Table 4 Results of conventional validation and tenfold validation
Table 5 Comparison of result with the literature

## Discussions

Some tools below are only available to our subscribers or users with an online account.

### Related Content

Customize your page view by dragging and repositioning the boxes below.

Related Journal Articles
Related Proceedings Articles
Related eBook Content
Topic Collections