Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.

Digital Object Identifier 10.1109/ACCESS.2017.DOI

A Novel Technique for Non-Invasive

Measurement of Human Blood
Component Levels from Fingertip Video
Using DNN Based Models
MD. REZWANUL HAQUE, S. M. TASLIM UDDIN RAJU, MD. ASAF-UDDOWLA GOLAP, AND
M. M. A. HASHEM
Department of Computer Science and Engineering, Khulna University of Engineering & Technology, Khulna 9203, Bangladesh
Corresponding author: S. M. Taslim Uddin Raju (e-mail: [email protected]).

ABSTRACT
Blood components such as hemoglobin, glucose, creatinine measuring are essential for monitoring one’s
health condition. The current blood component measurement approaches still depend on invasive techniques
that are painful, uncomfortable for the patients. To facilitate measurement at home, we proposed a novel non-
invasive technique to measure blood hemoglobin, glucose, and creatinine level based on PPG signal using
Deep Neural Networks (DNN). Fingertip videos from 93 subjects have been collected using a smartphone.
The PPG signal is generated from each video, and 46 characteristic features are then extracted from the
PPG signal, its derivatives (1st and 2nd ) and from Fourier analysis. Additionally, age and gender are also
included to feature because of the significant effects on hemoglobin, glucose, creatinine. A correlation-
based feature selection (CFS) using genetic algorithms (GA) has been used to select the optimal features
to avoid redundancy and over-fitting. Finally, DNN based models have been developed to estimate the
blood Hemoglobin (Hb), Glucose (Gl), and Creatinine (Cr) levels from the selected features. The approach
provides the best-estimated accuracy of R2 = 0.922 for Hb, R2 = 0.902 for Gl, and R2 = 0.969 for
Cr. Experimental aftermaths show that the proposed method is a suitable technique to be used clinically
to measure human blood component levels without taking blood samples. This paper also reveals that
smartphone-based PPG signal has a great potential to measure the different blood components.

INDEX TERMS Smartphone, Non-Invasive Technique, Hemoglobin, Glucose, Creatinine, NIR LED,
Photoplethysmogram, Features Extraction, Features Selection, Deep Neural Networks.

I. INTRODUCTION blood glucose level. Similarly, hemoglobin is the protein

molecule in red blood cells, and low hemoglobin count is
Blood is the essential life-maintaining fluid that strews over
the reason behind anemia [3]. So, continuous measurement
the whole body and is responsible for carrying heat, hor-
of blood hemoglobin level is essential for the treatment of
mones, antibodies, immune cells, etc. necessary for every
anemic patients [4] and premature babies [5], and dengue
cell. Everything related to the life cycle depends on blood
fever [6]. Likewise, the elevated level of creatinine intends
directly or indirectly. As a result, the measurement and anal-
weakened kidney [7]. So, for evaluating one’s medical con-
ysis of blood component levels can indicate the possibility
dition, it is imperative to measure the blood component levels
of numerous diseases. Too much glucose level in the blood
regularly.
indicates the possibility of diabetes. Diabetes is the most
common chronic, metabolic disorder and the leading cause of Blood component levels measurement techniques can be
death in the worldwide [1]. Patients with long-time diabetes grouped into invasive, minimally-invasive and non-invasive
can be affected by several diseases like heart disease, kidneys [8]. Conventional measurement techniques require a blood
damage, and lead to blindness [2]. Complications that occur sample of patients which is collected using finger-prick or
due to diabetes can be minimized by regular monitoring of venepuncture [9]. These methods are painful, inconvenient

VOLUME x, xxxx 1
 Author et al.: Preparation of Papers for IEEE ACCESS

and costly for the patient due to frequent blood collection, for the extraction of the PPG signal.
and do not allow real-time monitoring [10], [11]. However, In this paper, a novel non-invasive technique has been
invasive techniques are more precise and reliable, but it needs proposed to estimate the blood component levels using PPG
a well-resourced laboratory with trained personnel, most of signal characteristics features extracted from fingertip video.
whom are inaccessible in a remote area [12]. On the contrary, The finger is illuminated using near-infrared light-emitting
the non-invasive technique needs only a bio-signal (image diodes (NIR LED), and a smartphone is used to capture the
or spectrum) to estimate the blood component levels. Non- fingertip video. Fingertip video is preprocessed and gener-
invasive technology can overcome the above shortcomings ated the PPG signal by applying several filtering. Analyzing
and has become a more popular topic in smart healthcare re- the PPG signal and its derivatives signals and using Fourier
search [13]. Several techniques already exist for this purpose, analysis on the signal, a total of 46 features are extracted.
but those have some limitations like portability, relatively Additionally, age and gender are also included in the fea-
high cost, and poor penetration of light. So, we have been tures set. A correlation-based feature selection (CFS) method
motivated to develop a system that can measure the blood using genetic algorithms (GA) has been used to select the
component levels conveniently and painlessly. appropriate features for developing DNN models. Finally,
Now, near-infrared (NIR) spectroscopy and photoplethys- the DNN based models have been developed to estimate
mogram (PPG) are most widely used to measure blood hemoglobin, glucose and creatinine level in blood. The major
component levels non-invasively [14]. PPG is an optical contributions of this paper are summarized as follows:
measurement technique that has been applied to measure the • Invoking a non-invasive system for monitoring blood
changes in blood volume in certain parts of the body [15], glucose, hemoglobin and creatinine using smartphone
[16]. It reflects the movement of blood from the heart to the which is easy to use and comfortable to patients.
fingertip. The PPG system is designed with a light source, • Constructing a wearable device using NIR LED for
and a photodetector, where the light source illuminates the collecting fingertip video.
tissues region (e.g., finger) and the detector senses the re- • Generating PPG signal from the fingertip video and
flected light. The amount of light absorbed varies periodically extracting the PPG characteristics features from the
according to the variations of blood volume in the circulation generated PPG signal.
system and can be used to acquire the PPG signal. A lot • Selecting appropriate features using a correlation based
of researches have been done based on the PPG signal to feature selection method to avoid redundancy in features
monitor different physiological parameters because of its and over-fitting problem.
simple, low cost, and comfortable setup [17]. For example, • Developing DNN based models for estimating of blood
heart rate monitoring [18], anemia detection [19], [20], heart- component levels using both 48 features and the selected
rate validation [21], blood pressure estimation [22] and blood features.
glucose level [23]. Conventionally, PPG signals are acquired The rest of this paper is organized as follows: Section
using optical techniques like sensor-based device, chip, or II briefly described the existing works of this field. The
pulse oximeters [19], [21]–[24]. proposed methodology, data collection, features extraction
Recently, several smartphones which have built-in sensor and model construction are explained in Section III. Perfor-
systems for instantaneous measurement of heart rate, oxygen mance analysis of our work and the comparison of results
saturation based on PPG signals [25]. These non-invasive with previous works are demonstrated in Section IV. The
techniques are not only useful for patients those required conclusion and future direction of this paper are drawn in
regular health monitoring but also needed for healthcare Section V.
professionals. Several studies are described briefly in the
literature concerning the measurement of the different blood II. RELATED WORKS
component levels non-invasively based on smartphones and The smartphone is a portable, affordable and convenient
RGB analysing. Among these, Wang et al. developed a platform for developing point-of-care health tools. Non-
smartphone-based application called HemaApp using Nexus- invasive techniques are essential for patients who require to
5p to predict the hemoglobin concentration in blood in monitor blood tests regularly. In this section, we focus on the
2016 [26]. In 2017, the same authors used Nexus-6p and work related to measuring blood component based on PPG
updated the hardware configuration for the same purpose signals and smartphones, and summarize the representative
[3]. In 2018, Hasan et al. developed a smartphone appli- researches as follows.
cation called SmartHeLP using Nexus-4p to measure the Kavsaoğlu et al. [19] introduced a non-invasive method
hemoglobin value [27]. In 2019, Zhang et al. proposed a to predict blood Hb levels using the characteristics of the
non-invasive technique to estimate the blood glucose level PPG signals with eight regression machine learning algo-
based on smartphone PPG signal [28]. A few PPG signal rithms. Al-Baradie and Bose [29] developed a portable non-
acquisition methods have been described in the literature, but invasive Hb measurement sensor using LED for acquiring
most of them used professional equipments. Therefore, we the PPG signal at the wavelength of 670 nm and 810 nm.
aim at using the smartphone camera and external lighting Ramasahayam et al. [30] developed a system consists PPG
source to capture a video of the fingertip that can be used signal acquisition module (including LEDs with wavelength
2 VOLUME x, xxxx
Author et al.: Preparation of Papers for IEEE ACCESS

of 935 nm, 950 nm, and 1070 nm and a photodetector) for Least Square methods and extracted features from the PPG
the assessment of Gl levels. signal. Finally, Principal Component Regression (PCR) had
Wang et al. [26] developed a smartphone-based application been applied for the prediction of Gl level from the extracted
as HemaApp, using the smartphone camera and multiple features. The proposed model could predict the glucose level
lighting sources, including infrared LEDs that illuminate the with a Standard Error of Prediction (SEP) of 18.31 mg/dL.
patient’s fingertip. A smartphone Nexus-5p was used for Zhang et al. [28] introduced a non-invasive blood glucose
recording a series of videos with white, 880 nm, and 970 estimation technique using smartphone (smartphone camera
nm LED array. They focused on three different hardware was utilized for data acquisition) PPG signals and generated
embodiments, where the first embodiments included white an output using a machine learning algorithm (Subspace
flash + infrared emitter, the second one consisted of incan- KNN classifier). A smartphone camera with a frame rate of
descent lamp + white flash + infrared emitter, and the final 28 fps (sampling rate 28 Hz) was used to record a 30-40
one was made white flash + series infrared LED array. A seconds long video of the left-hand index finger. Features
high-band pass filter was used to calculate average intensity were extracted from a single PPG signal and its derivatives.
for each channel. Then Fast Fourier Transform (FFT) and Finally, the blood glucose level was estimated from the valid
Support Vector Machine (SVM) regression were applied for signals by applying a subspace KNN classifier. The overall
each combination of datasets. However, this is not sufficient accuracy was 86.2%.
for the people whose Hb level below 8 g/dL (heavily anemic). Considering the existing literature, we observed that dig-
The limitation of HemaApp is that they collected data by ital image or video data of eye conjunctiva and fingertip
using a Nexus-5 device and only one brand of the intense can measure different blood components such hemoglobin,
light bulb. Results could be varied according to the different glucose, creatinine. It also shows that the PPG signal contains
brand devices. They achieved a correlation R result between potential information regarding different blood components.
0.69 and 0.82, and RMSE value between 1.26 and 1.56 g/dL. Most of the existing techniques use specific hardware, e.g.,
Edward et al. [3], improved the configuration of hardware laser light and finger chip, to acquire the PPG signal, whereas
for amplifying the weaker signal of Blue and Green. They our system uses only a smartphone and an external lighting
estimated the blood Hb level without using an IR LEDs. In source to acquire the PPG signal. To sum up, there is up to
this case, they obtained Pearson correlation of 0.62 and an now no non-invasive method for measuring blood component
RMSE of 1.27. levels based on PPG signal from a fingertip video, which
M. Anggraeni and A. Fatoni [31] introduced a non- has the advantages of time efficiency and no calibration
invasive anemia detection system based on a digital image requirement.
of palpebral conjunctiva captured by a smartphone camera.
Digital image of the inferior palpebral conjunctiva was cap- III. MATERIALS AND METHODS
tured with an Asus Zenphone-2 Laser smartphone, ambient In this section, the materials and methods are illustrated
lighting without flash, and then color-corrected with white concisely. Collection of video data, generation of PPG signal,
paper. The color intensity (R, G, B) was extracted from raw features extraction, features selection, and finally, model
data using Colorgrab software (Loomatix), then evaluated construction process have been explained briefly. The overall
using regression analysis. Among the three-color (R, G, architecture of the system is shown in Figure 1.
B) intensity levels, red color intensity resulted in a high
correlation with clinically measured Hb levels and gained A. EXTERNAL HARDWARE DESIGN
R2 = 0.8139. A wearable device has been constructed as accessories for
Hasan et al. [27] developed SmartHeLP, a smartphone- illuminate the finger while capturing the fingertip video
based Hb level estimation technique using ANN and fingertip through smartphone. The device consists of six NIR LED
videos. The authors collected 10-second (300 frames) finger- with wavelength of 850 nm and, a white LED. NIR LEDs
tip video each from 75 participants of 20-56 year of ages and are placed in a circle with the white LED in the centre of
the levels of Hb were from 7.6 to 13.5 gm/dL. Red, Green, the circle shown as Figure 2 (b). The function of the white
and Blue pixel intensities from each frame were separated for LED is to amplify the intensity of NIR LEDs. The external
feature extraction, and the ANN-based model was developed surface of the device is made to black so that the reflectance
using these features to predict the Hb level. They observed factor can hardly affect the analysis. Figure 2 illustrates the
a correlation of R2 0.93. To reduce the necessary feature prototype of the external lighting source device.
space, they identified a specific Region of Interest (ROI) in Conventionally, PPG signals are acquired using optical
the image frame. techniques like sensor-based device or chip. In these cases,
Chowdhury et al. [32] proposed non-invasive Gl esti- visible or NIR LEDs are used for illuminating the finger
mation method from smartphone-based video analysis. A or different parts of body tissues and a photodetector for
smartphone camera at 30 fps was used for collection the measuring the amount of light either transmitted or reflected.
fingertip videos from 18 subjects and subsequently converted Therefore, we aim to obtain PPG signals through the use of
the frames of the video into PPG waveform. PPG signal was the smartphone camera to capture fingertip video and then
preprocessed using Gaussian filter along with Asymmetric analyzing the variation of light intensity reflected from a
VOLUME x, xxxx 3
 Author et al.: Preparation of Papers for IEEE ACCESS

Red, Blue and

Green Channel

Subjects Video Time Series Red Signal

Smartphone Frames

Systolic Peak
Second Wave Preprocessing
Dicrotic Notch

Diastolic Peak Original PPG

Hemoglobin (g/dL)
1st Derivative
Blood
Glucose (mmol/L) Components Model Genetic
Level (DNN) Algorithm
2nd Derivative
Creatinine (mg/dL)
Measurement Applied Model Features Selection Features Extraction PPG Wave

FIGURE 1: The overall system architecture of blood Hemoglobin, Glucose, and Creatinine level measurement.

most of the works used 475 − 2500 nm wavelength light

for acquiring the PPG signal. For example, Ramasahayam
et al. [30] were used 935 nm, 950 nm, and 1070 nm NIR
LED to acquire the PPG signal for estimating Gl level. In
HemaApp [3], 880 nm and 970 nm, as well as 500-700 nm
and 1300 nm NIR LED, were used to measure the blood
Hb concentration in two different studies. Al-Baradie and
Bose [29] were developed a LED-based Hb sensor system
for acquiring the PPG signals at the wavelengths of 670 nm
and 810 nm. Light absorbed of NIR wavelength range from
(a) (b) 700−2500 nm by tissues [17] is considerable to get the strong
PPG signal as well as it can penetrate through the tissues of
FIGURE 2: Illustration of external lighting source device: (a) the finger between 1 − 2 cm effectively [33]. As it is easier to
Circuit diagram (b) External wearable device
get NIR LED of 850 nm wavelength, so this LED has been
used in our device.

finger caused by the change of blood volume in systolic B. DATA COLLECTION

and diastolic cycles. Although some new smartphones are A 15-second video of the right index finger was recorded
starting to be equipped with sensor or IR LEDs, the most using a Nexus-6p camera (30 fps), while the finger was
smartphone have only white LED and do not have sensors illuminated using an external wearable device. Meanwhile,
for detecting the reflected light. External lights, near-infrared gold standard data of hemoglobin, glucose, and creatinine
lights, are required when a smartphone has no support to level of these subjects were also collected using the clinical
sense blood components non-invasively in living tissues. For method. These two procedures were performed consecutively
fingertip-based data collection, we pointed out that a covered but separately (with an interval of less than one minute), so
external NIR light source can provide the best PPG signal that the blood component levels did not fluctuate in such a
from a smartphone video. But the selection of near-infrared short period. The gold standard value for hemoglobin was
wavelength is the first step of hardware design and a critical measured with Sysmex XS-800i Haematology Analyzer, and
issue to acquire the strong and clean PPG signal because we glucose and creatinine were measured with Thermo Scien-
have to take wavelength as consideration for the absorption tific Konelab 60i, Chemistry Analyzer respectively in the
of light by blood, muscle, and skin tissues. From the study clinical laboratory. The blood sample of each subject for
of several previously existing techniques, it is observed that clinical measurement was collected just before taking the
4 VOLUME x, xxxx
Author et al.: Preparation of Papers for IEEE ACCESS

(a) (b) (c) (d) (e)

FIGURE 3: Data Collection procedure: (a) External wearable device (b) Power ON of the external device (c) Index finger placed on
the device when the power is OFF (d) Power ON with index finger on the device (e) Capturing 15-second video using smartphone
(Nexus-6p) camera and store it in this device.

fingertip video. 93 subjects (59 males and 34 females; age: phone camera (30 fps).
33.54 ± 17.12 years) were participated in the whole proce- After reviewing all the facts, the right-hand index finger
dure. Prior to the collection of data, information regarding was placed on the wearable device and illuminating the finger
the work and training on how to use the wearable device with an wearable device. Finally, a 15-second video was
and smartphone to capture fingertip video were given to the captured, placing a smartphone back camera on the finger,
healthcare workers. Table 1 shows the statistical information according to Figure 3. The first 3-second and the last 2-
of these collection. second are discarded from each fingertip video to avoid
Data collection is one of the vital stages of conducting unstable frames.
research. Data can be corrupted within a moment for a simple
mistake and affect all the next processes. The following C. GENERATION OF PPG SIGNAL AND FEATURES
guidelines were followed while recording the fingertip video. EXTRACTION
• The subject’s right hand and fingers were clean and dry Photoplethysmography (PPG) is a signal which is optically
before capturing the fingertip video. obtained through a plethysmograph used for detecting the
• The index finger was preferred, but other fingers were volumetric variation through blood circulation [15], [16]. It
used according to the condition of the tissue if the index reflects the movement of blood from heart to fingertip. For
finger was injured. generating the clean and good PPG signal from the image
• The fingertip video was recorded after taking clinical frame, the selection of optimal Region of Interest (ROI) is
blood sample. indispensable [44], [45]. Another vital issue of the generation
• External wearable device was constructed in a user- of PPG signals from the image frame is the selection of
friendly manner so that participants can easily place the channel (Red, Green, and Blue). Because it is shown that
finger on the device. if the range of image pixel intensity is below 200, then
• Same conditions (room temperature and light) were it is not possible to extract all characteristic features from
maintained during the acquisition of data from any the generated PPG signal. Among three channels, the Red
participant. channel is the highest intensity (225−245) compare to Green
• Fingertip video was recorded using Nexus-6p smart- (0 − 3) and Blue (15 − 25). To identify the ROI, HemaApp
[26] used a centre segment of the frame of an image and
calculated the average intensity for each channel of the centre
TABLE 1: Patient demographics and clinical laboratory data segment of that image, Scully et al. [46] picked 50 × 50
(gold standard value/reference value). array of pixels on the Green channel and Jonathan et al. [47]
extracted 10 × 10 block of the mean intensity value of the
Physical Index Statistical Data
pixels from the central region. SmartHeLP [27] separated a
Age (years) 33.54 ± 17.12 (0 − 79)
Gender 59 male (63.5%); 34 female (36.5%)
10 × 10 pixels block from the image frame for identifying
Hemoglobin (g/dL) 12.84 ± 2.15 (7.90 − 21.49) the best generated PPG signal’s most adequate position. In
Glucose (mmol/L) 6.65 ± 2.94 (3.33 − 21.11) this study, the average intensity for each channel is calculated
Creatinine (mg/dL) 1.03 ± 0.87 (0.37 − 9.02) by cropping image frame of 500 × 500 pixels from right to
*
Note: Results are stated as Mean ± S.D. (range) for quanti- left section of the image. This is done because the image from
tative variables and frequency distribution (%) for categorical this section is the most consistent and stable when the video
variables. of the fingertip is captured with the smartphone.
VOLUME x, xxxx 5
 Author et al.: Preparation of Papers for IEEE ACCESS

Video Frame Sequence

Algorithm 1: Generation of PPG signal from
videos data and features extraction
290 Pixels
Input: N -number of videos
Output: ListF N -extract features for N videos
/* For NIR-0850 lighting condition, 500 x 500 Pixels
capture 10sec(30fps) videos with
Nexus-6p smartphone */
/* Loop up to terminal status */ 290 Pixels
1 for i ← 1 to N do
2 Extract only 300 frames for each video;
(a)
3 From right to left section of frame, measuring
500 px as width and 500 px as height is 238
Red Channel
cropped;
/* Initialization list: ListH */ 236

Average Intensity
4 List of select highest intensity channel ListH;
5 for j ← 1 to 300 do
234
/* Calculate M eanR, M eanG,
M eanB */
6 Average value of red channel M eanR; 232
7 Average value of green channel M eanG;
8 Average value of blue channel M eanB;
0 50 100 150 200 250 300
9 for k ← 1 to 3 do Number of frames
10 if channelk is red then (b)
11 M eanR ← average intensity for
3
red channel from F ramesj ; Acquired PPG Wave
2.5 Systolic Peak
12 else if channelk is green then 2
13 M eanG ← average intensity for
1.5
Amplitude

green channel from F ramesj ;

1
14 else 0.5
15 M eanB ← average intensity for 0
blue channel from F ramesj ; -0.5

/* Calculate maximum value of 0 1.7 3.4 5.1 6.8 8.5 10

channels Time (s)
*/
16 M axvalC ← (c)
3
max(M eanR, M eanG, M eanB); Single PPG Wave
2.5
/* Append maximum value from
three channels to list: 2
Amplitude

ListH */ 1.5
17 ListHj ← M axvalC; 1

/* Generate PPG signals: P P GSigi 0.5

*/ 0
18 P P GSigi ← generateP P G(ListH); -0.5
/* extract features from a
1.1 1.3 1.5 1.7 1.9
single PPG signal and append Time (s)
features to ListF N */ (d)
19 ListF Ni ← extractF eatures(P P GSig);
FIGURE 4: Generation of PPG signal from fingertip video: (a)
/* Returning the extracted Crop 500 × 500 pixels from the middle of right side from each
features */ frame (b) Red channel from video’s frames (c) Continuous PPG
waveform after applying Butter-worth Filter (d) One single
20 return ListF N
PPG waveform with highest positive systolic peak.

6 VOLUME x, xxxx
Author et al.: Preparation of Papers for IEEE ACCESS

PPG signal
1st Derivative
2nd Derivative
Systolic peak
+ Dicrotic notch
Diastolic peak
Amplitude

x/2 Δ𝐓 Hole
x
+ + +
width
𝑽𝟐
𝑨𝟏 𝑨𝟐
z y 𝑽𝟏

𝒕𝟏
𝒕𝟐
𝒕𝟑
𝒂𝟏 𝒕𝒑𝒊
𝒂𝟐
Amplitude

𝒆𝟐

𝒆𝟏
𝒍𝟐
𝒍𝟏
𝒕𝒂𝟏
𝒃𝟏 𝒕𝒂𝟐
𝒕𝒃𝟏
𝒕𝒆𝟏
𝒕𝒃𝟐
𝒕𝒍𝟏 𝒃𝟐
𝒕𝟐
𝒕𝒑𝒊
𝒕𝟑
Time
FIGURE 5: The characteristic features acquired from PPG signal and its corresponding first and second derivatives.

As the PPG signal reflects the movement of blood from the lar in terms of characteristics [28]. One single PPG signal
heart to the fingertip, the characteristics of the PPG signal with the highest positive systolic peak is extracted from the
can provide information on the levels of blood constituent continuous PPG waveform because of its highest intensity
[19], [48]. Therefore, it is necessary to identify appropriate changes. This single PPG signal is processed for the features
preprocessing and feature extraction methods to analyze the extraction. The overall process of PPG signal generation
PPG signal precisely [49]. A video data comprises a series from the fingertip video is demonstrated in Figure 4.
of digital images called frames. 300 frames are extracted Before extraction of features, PPG signal is preprocessed
from a 10-second (30 fps) video. Then, the time series of to minimize noise as raw PPG signals are prone to the
average intensity is plotted for three channels (Red, Green, noise and motion artifacts. A finite impulse response (FIR)
and Blue). The Red channel is the highest intensity channel. filter is applied to remove noise. Then 46 characteristics
Bandpass Butterworth Filter [50] is compatible with the red features are extracted from the PPG signal and it’s deriva-
channel and manages to generate a good PPG signal. So, the tives, shown in Figure 6. The pipeline of the PPG signal
green and blue channels are discarded, and the red channel is generation from video data and features extraction process
adopted. PPG signal is continuous and repeated waveform is described algorithmically in Algorithm 1. These features
and every individual PPG signal contains approximately have been classified as follows: 21 (f3 to f22 and f48 )
same information. A peak detection algorithm is used to features are extracted from the PPG signal, 19 (f23 to f41 )
locate the signal peaks [51]. The signal is then divided into features extracted from 1st and 2nd derivatives [52], and 6
single periods. The waveform of single-period PPG signals (f42 to f47 ) features extracted using fourier analysis (Fast
might vary slightly for different people, but they are simi- Fourier Transformation). In addition, Age (f1 ) and Gender
VOLUME x, xxxx 7
 Author et al.: Preparation of Papers for IEEE ACCESS

TABLE 2: Extracted characteristics features from PPG signal.

Features Definition Features Definition

f3 : x Systolic peak [34] f26 : tl1 Interval time from point l1 to next point l1
f4 : y Diastolic peak [34] f27 : ta1 /tpi Ratio between time interval of a1 (ta1 ) and pulse
interval (tpi )
f5 : z Dicrotic notch f28 : tb1 /tpi Ratio between time interval of b1 (tb1 ) and pulse
interval (tpi )
f6 : tpi Pulse interval [35] f29 : te1 /tpi Ratio between time interval of e1 (te1 ) and pulse
interval (tpi )
f7 : y/x Augmentation index [36] f30 : tl1 /tpi Ratio between time interval of l1 (tl1 ) and pulse
interval (tpi )
f8 : (x − y)/x Alternative augmentation index [37] f31 : b2 /a2 Ratio of b2 and a2 [36], [38]
f9 : z/x Ratio of dicrotic notch and systolic peak [39] f32 : e2 /a2 Ratio of e2 and a2 [36], [40]
f10 : (y − x)/x Negative relative augmentation index f33 : (b2 + e2 )/a2 Ratio of (b2 + e2 ) and a2 [40]
f11 : t1 Systolic peak time f34 : ta2 Interval time from point l2 to next point a2 for
second derivative of PPG
f12 : t2 Dicrotic notch time f35 : tb2 Interval time from point l2 to next point b2
f13 : t3 Diastolic peak time f36 : ta2 /tpi Ratio between time interval of a2 (ta2 ) and pulse
interval (tpi )
f14 : ∆T = t3 - t1 Time between systolic and diastolic peak [35] f37 : tb2 /tpi Ratio between time interval of b2 (tb2 ) and pulse
interval (tpi )
f15 : t1 /2 Time between half systolic peak points f38 : (ta1 + ta2 )/tpi Ratio of (ta1 + ta2 ) and pulse interval (tpi )
f16 : A2 /A1 Inflection point area ratio [41], [42] f39 : (tb1 + tb2 )/tpi Ratio of (tb1 + tb2 ) and pulse interval (tpi )
f17 : t1 /x Systolic peak rising slope f40 : (te1 + t2 )/tpi Ratio of (te1 + t2 ) and pulse interval (tpi )
f18 : y/(tpi − t3 ) Diastolic peak falling slope f41 : (tl1 + t3 )/tpi Ratio of (tl1 + t3 ) and pulse interval (tpi )
f19 : t1 /tpi Ratio of systolic peak time (t1 ) and pulse interval f42 : fbase Fundamental component frequency obtained from
time (tpi ) Fast Fourier Transformation (FFT)
f20 : t2 /tpi Ratio of dicrotic notch time (t2 ) and pulse interval f43 : |sbase | Fundamental component magnitude acquired from
time (tpi ) FFT
f21 : t3 /tpi Ratio of diastolic peak time (t3 ) and pulse interval f44 : f2nd Second component frequency obtained from FFT.
time (tpi ) Such that, fbase < f2nd
f22 : ∆T /tpi Ratio of time between systolic and diastolic peak f45 : |s2nd | Second component magnitude acquired from FFT
(∆T ) and pulse interval time (tpi )
f23 : ta1 Interval time from point l1 to point a1 for 1st f46 : f3rd Third component frequency obtained from FFT.
derivative of PPG Such that, fbase < f2nd < f3rd
f24 : tb1 Interval time from point l1 to point b1 f47 : |s3rd | Third component magnitude acquired from FFT
f25 : te1 Interval time from point l1 to point e1 f48 : sVRI = V2 /V1 Stress-induced vascular response index [43]

(f2 ) of each subject are included as features. Figure 5 is D. FEATURES SELECTION

shown features with characteristics point of PPG signal. The
extracted features have been listed in Table 2. Feature selection is the most important step before model
construction since the prediction power of the model depends
on the features. Feature selection has lots of advantages
before developing the model like it avoids the opportunities
of model overfitting, improves the model accuracy by dis-
carding the redundant features, and reduces the complexity
of the algorithm and hence model trains faster. To discard
redundant and irrelevant features before construction of the
model and to determine the optimum subset of features, a
selection process should be applied [22]. There are a number
of feature selection methods to select the optimal features
set. In this work, a correlation-based feature selection (CFS)
using a genetic algorithm (GA) is applied [53]. It works
based on the genetic algorithm. Genetic algorithm is a search-
FIGURE 6: Block diagram of feature extraction. based optimization technique derived from the biological
8 VOLUME x, xxxx
Author et al.: Preparation of Papers for IEEE ACCESS

phenomenon of natural selection [54]. It is started with a If Eo ≥ 0,

population of individuals. Every individual is capable of q Pn
2
k=1 (fi,k −fj,k )
reaching the goal. The initial population is produced by Ef = efi ,fj = n (3)
randomly choosing of leaf nodes (variables) and functions.
If Eo < 0,
According to the problem, the variable and function are q Pn
changed. By applying crossover and mutation techniques, Ef = efi ,fj = − k=1 (fi,k −fj,k )
2
(4)
n
the old population is replaced with a new generation. The
crossover brings together two individuals into one individ- Finally, the correlation Rcc (as objective function) between
ual. The mutation strategy switches the selected individual’s Ef and Eo can be calculated as (5).
function node and generates a new individual. The features
SE E
selection process has been described algorithmically in Al- Rcc = p f o (5)
gorithm 2. The extracted features set randomly generate a SEo SEo
subset of features as initial population. The fitness of each where,
individual in the population is evaluated by an objective P ¯ ¯
i (Efi −Ef )(Eoi −Eo )
function (5) that qualifies how good a set of features is. The SEf Eo = n−1 ,
closer-fit features are selected and altered to create a new P ¯ 2 P ¯ 2 (6)
i (Efi −Ef ) i (Eoi −Eo )
SEf = n−1 , SEo = n−1
generation. Fitness of the new generation is calculated in the
same way as the initial population and repeat the process. The
search technique is applied to select the features (individuals)
of each generation based on fitness. The individual with a Algorithm 2: Genetic algorithm for features selec-
higher fitness value has a better chance of being selected. The tion.
process terminates, when a maximum number of generations m
Input: dataset D = {fi , oi }i=1 (fi ∈ F, oi ∈ O),
has been attained, or the population has reached a satisfactory α = population size, δ = crossover rate,
fitness level. Overall feature selection process is illustrated in γ = mutation rate, T = maximum iterations
Figure 7. The probability of selecting j th individual from the Output: Optimal feature subset
population is indicated in (1).
/* Initialization */
vj 1 Initialize population: P ← Generate randomly α
p j = PM , (1)
k=1 vk feasible solutions (subset of features);
/* Evaluate the fitness of each
where, vj is the value of fitness of an individual j in the
solution,S in P */
population, M is the size of population.
2 Evaluate the initial population according to the
Therefore, an appropriate fitness function is needed to
fitness function (5);
evaluate fitness. Here, a correlation-based fitness function is
/* Initialize i ← 0 */
used. Let, o1 , o2 , · · · , oN be the response values and f1 , f2 ,
3 while i ≤ T or
· · · , fN be the corresponding n-dimensional feature set. The
[argmaxs f itness(S) < f itness_threshold] do
distance for each pair {oi , oj } of the response values can be
/* Selection */
calculated as follows:
4 Select two parents from the population using
roulette wheel selection ;
Eo = eoi ,oj = oi − oj (2)
5 The probability to select each individual in the
At the same time, the distance between corresponding population using (1) ;
features sets can be referred as (3) and (4). /* Crossover */
6 Generate two offspring by crossover operation
between two parents ;
Extracted
/* Mutation */
Features Set 7 if rand(0.0, 1.0) ≤ γ then
GA Based Search
8 Mutate the offspring solutions ;
9 Evaluate each offspring’s according to the
Features Fitness fitness function (5);
Subset Value /* Updating */
10 Update P ← PS ;
/* PS ← new offspring */
Feature Set
Evaluation /* The best solution return */
Selected
Features 11 return the best solution Sbest from P that has the
highest fitness;
FIGURE 7: Block diagram of feature selection.

VOLUME x, xxxx 9
 Author et al.: Preparation of Papers for IEEE ACCESS

TABLE 3: The selected features by means of CFS using genetic algorithm for different blood component levels.

Dataset Selected Features Count

f1 f2 f3 f4 f5 f6 f11 f15 f17 f18 f19 f21 f24 f25 f30 f31 f32
Hemoglobin 27
f35 f36 f37 f40 f41 f44 f45 f46 f47 f48
f1 f3 f5 f7 f11 f12 f14 f15 f16 f17 f18 f20 f22 f23 f28 f32 f34
Glucose 25
f36 f38 f40 f41 f45 f46 f47 f48
f3 f4 f5 f9 f10 f15 f16 f17 f19 f24 f26 f28 f29 f33 f34 f35 f37
Creatinine 24
f39 f42 f44 f45 f46 f47 f48

TABLE 4: Control parameters and their values used in genetic

Feature Set Fitness algorithms for the feature selection process.
Best Fitness
Mean Fitness
Parameters Value
0.8
Population Size 50
0.7 Number of Generation 50
Selection Method Roulette-wheel
Crossover Type Two-point
Fitness value

0.6
Probability of Crossover 0.1
0.5 Probability of Mutation 0.2

0.4

As fitness quality improves, Rcc becomes bigger. Hence, in

0.3
this study, the optimization issue is described as maximizing
0 10 20
Generation
30 40 50 the fitness function Rcc .
(a)
Feature Set Fitness
E. FEATURES SELECTION RESULTS
Best Fitness
Mean Fitness
In this work, a CFS-based feature selection method has been
1.6 used to discard the irrelevant and redundant features from
the feature set. To select the best features subset using GA,
1.4 the following control parameters are used: population size =
50, probability of crossover = 0.1, probability of mutation =
Fitness value

1.2 0.2, selection criteria = roulette wheel and total generation

= 50 (Table 4). The improvement in features set fitness
1.0
over time with the best fitness value is graphically plotted
in Figure 8. The best feature subset obtained using GA for
0.8
blood component levels are tabulated in Table 3. We have
obtained 27 features for hemoglobin, 25 for glucose, and
0 10 20 30 40 50
Generation 24 for creatinine, and their corresponding best fitness values
(b) are 0.854, 1.642, and 2.421, respectively. From Table 3, it
Feature Set Fitness is shown that the hemoglobin values are varied according
Best Fitness to age and gender. On the other hand, creatinine levels are
2.4 Mean Fitness
not dependent on age and gender because abnormality of
2.2
creatinine value in the blood indicates weakened kidney.
2.0

F. MODEL CONSTRUCTION
Fitness value

1.8
A Deep Neural Network is a feed-forward, artificial neural
1.6 networks comprising an input layer, several hidden layers,
1.4
and an output layer. It is equipped with biases, weights, and
activation functions such as a rectified linear unit (ReLU)
1.2 [55]. The input layer consists of neurons equal to the number
of features in the dataset, and the output layer is composed of
0 10 20 30 40 50
Generation a single neuron [56]. In this work, we explored a multilayer
(c) feed-forward network, where nodes of each layer receive the
FIGURE 8: Feature selection results: (a) Hemoglobin (Best inputs from the previous layer. The output of nodes in one
Fitness = 0.854) (b) Glucose (Best Fitness = 1.642) (c) Creatinine layer will be the input of the next layer. The architecture
(Best Fitness = 2.421) and training procedure of DNN is illustrated in Figure 9. As
shown in the figure, four hidden layers are used. There are
10 VOLUME x, xxxx
Author et al.: Preparation of Papers for IEEE ACCESS

𝑓1

𝑓2

𝑦ෝ (Estimated Blood
𝑓3
Component Levels)

𝑓𝑛

Input layer: First hidden Second hidden Third hidden Fourth hidden
Output
all features layer with layer with 200 layer with layer with 300
Layers,
or selected 150 neurons. neurons. 250 neurons. neurons.
Activation
features with Activation Activation Activation Activation
Function
GA Function Function “ReLU”. Function Function “ReLU”.
Linear
“ReLU” Dropout 0.25. “ReLU”. Dropout 0.5.

FIGURE 9: The architecture of proposed Deep Neural Networks (DNN) based models for blood component levels estimation. Not all
connections and units are displayed, can be marked as representative of DNN based model.

150 neurons in the first hidden layer and 250 neurons in the G. TEN-FOLD CROSS-VALIDATION
third hidden layer. There are 200 neurons with a 0.25 dropout In this study, a 10-fold cross-validation method was used to
unit in the second hidden layer and 300 neurons with a 0.50 construct and evaluate the performances of DNN models. At
dropout unit in the fourth hidden layer. The dropout method is first, reference Hb values and PPG characteristics features
an alternative and more efficient option for addressing DNN (age and gender were also included as features) of 93 subjects
overfitting [57]. The output of each processing unit in the were divided into 10 almost equal subgroups or folds to train
hidden layer can be expressed as in . and test the model. In each iteration, 9 subgroups were used
X for training the model, and the rest one was used for testing
qj = ωj fj + β (7) the model. This process went on until 10 iterations were
j completed. After 10 times of training and testing, the reliable
Hb estimation model was established. The same procedure
where fj , ωj and β indicate input feature vector, weights was followed for the remaining DNN models to estimate Gl
and bias to the neuron respectively. The hidden layer will and Cr levels respectively.
modify the input above using a nonlinear function by

ϕre (q) = max(0, q) (8) IV. EXPERIMENTAL RESULTS AND DISCUSSIONS

In this section, the experimental results and comparison of
where, ϕre is the ReLU activation function. Finally, a our work with previous exiting methods have been discussed
linear activation function is used in the output layer.

ϕli (q) = q 0 (9) TABLE 5: Hyperparameters and their values used in DNN
based models
Where q 0 = (−∞, +∞)
Parameters Status
Thus, the dense layer returns the sum of the activation
Batch size 32
function. DNN model was trained on 100 epochs with respec- Learning rate α 0.01
tive learning rate and batch sizes as 0.01 and 32. To facilitate The number of hidden layers 4
the training processing and update the parameter of param- The number of nodes at 4 hidden layers (150, 200, 250, 300)
eters of DNN, Adam was used as optimizer function. The Dropout at 2nd and 4th hidden layer (0.25, 0.5)
hyperparameters used in the proposed DNN based models The number of nodes at input layer 48 or selected features
are shown in Table 5. The DNN models were trained and The number of node at output layer 1
tested with both all features and CFS-based selected features Activation function ReLU, Linear
for each blood component level. Optimizer Adam

VOLUME x, xxxx 11
 Author et al.: Preparation of Papers for IEEE ACCESS

TABLE 6: Performance measurement of blood component levels using DNN based models with all features.

With 48 features (All features)

Blood Component Algorithm Performance measure criteria
R2 R MAE MSE RMSE MAPE AI
DNN
Hemoglobin Mean 0.877 0.936 0.329 0.563 0.750 2.531 0.966
Std. Dev. 0.031 0.018 0.045 0.140 0.059 0.319 0.013
DNN
Glucose Mean 0.770 0.880 0.516 1.975 1.405 6.521 0.929
Std. Dev. 0.042 0.032 0.083 0.486 0.113 0.986 0.020
DNN
Creatinine Mean 0.865 0.993 0.072 0.100 0.316 4.834 0.975
Std. Dev. 0.055 0.029 0.010 0.028 0.028 0.398 0.017

TABLE 7: Performance measurement of blood component levels using DNN based models with selected features by means of CFS-
based feature selection.

With selected features via CFS ( 27 features for Hb, 25 features for Gl and 24 features for Cr)
Blood Component Algorithm Performance measure criteria
R2 R MAE MSE RMSE MAPE AI
DNN
Hemoglobin Mean 0.922 0.962 0.318 0.355 0.596 2.548 0.979
Std. Dev. 0.052 0.022 0.036 0.073 0.042 0.271 0.015
DNN
Glucose Mean 0.902 0.951 0.375 0.840 0.917 5.035 0.973
Std. Dev. 0.019 0.011 0.057 0.210 0.074 0.693 0.007
DNN
Creatinine Mean 0.969 0.995 0.052 0.023 0.152 4.508 0.993
Std. Dev. 0.037 0.026 0.005 0.005 0.012 0.298 0.020

briefly.
n
1X
M SE = (yi − ŷi )2 (13)
A. PERFORMANCE EVALUATING CRITERIA n i=1
The models were implemented in Python programming lan- v
guage with computing environment named Spyder available u n
u1 X
in Anaconda. The experiments were carried out on a single RM SE = t (yi − ŷi )2 (14)
computer (Asus X556U , Intelr Core(T M ) i5−72000U n i=1
CPU @ 2.50GHz, 8.0 GB RAM and Nvidia GeForce n  
940M X) with W indows − 10 operating system. 1 X  yi − ŷi 
M AP E = (15)
The performance evaluating criteria to test the perfor- n i=1  yi 
mance of developed DNN based models is R2 (Coefficient of Pn
Determination), R (Pearson Correlation Coefficient), Mean (yi − ŷi )2
IA = 1 − Pn i=1 0 0 2
(16)
Absolute Error (MAE), Mean Squared Error (MSE), Root
i=1 [|yi | + |ŷi |]
Mean Square Error (RMSE), Mean Absolute Percentage
Error (MAPE), and IA (Index of Agreement) values. where,
If y1 , y2 , · · · , yn are n reference values and ŷ1 , ŷ2 , · · · , yi0 = yiP− ȳ and ŷi0 = ŷi − ȳ,
1 n
ŷn are the corresponding estimated values, then equations for ȳ = n Pi=1 yi ,
n
these metrics are as follows [19]: f¯ = n1 i=1 fi ,
th
Pn fi is the i input feature.
2 (yi − ŷi )2
R = 1 − Pi=1 n 2
(10)
i=1 (yi − ȳ) B. ROBUSTNESS PERFORMANCE OF THE MODELS
Pn The performance of the proposed method has been evaluated
(fi − f )(yi − y)
R = P i=1
q (11) with seven performance measurement indices: R, R2 , MAE,
n 2 2
i=1 (fi − f ) (yi − y)
MSE, RMSE, MAPE, and IA. It is noteworthy that the
n
proximity of R2 to 1 indicates the strength of the relationship
1X between model outputs and reference values. The RMSE and
M AE = |yi − ŷi | (12)
n i=1 MSE show relative errors and MAE represent the absolute
12 VOLUME x, xxxx
Author et al.: Preparation of Papers for IEEE ACCESS

Bland-Altman Plot
22 MAE = 0.329±0.045
R 2 = 0.877±0.031 md + 1.96*sd
4 md
20
md - 1.96*sd

Difference Hemoglobin(g/dL)
3
Estimated Hemoglobin(g/dL)

18
2
16
1
14 0
12 1
10 2

8 3
8 10 12 14 16 18 20 22 8 10 12 14 16 18 20
Reference Hemoglobin(g/dL) Average Hemoglobin(g/dL)
(a) (b)
Bland-Altman Plot
MAE = 0.516±0.083 md + 1.96*sd
20.0 R 2 = 0.770±0.042 8
md
17.5 6 md - 1.96*sd

Difference Glucose(mmol/L)
Estimated Glucose(mmol/L)

15.0 4
12.5 2
10.0 0
7.5 2
5.0
4
2.5
2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 4 6 8 10 12 14 16 18 20
Reference Glucose(mmol/L) Average Glucose(mmol/L)
(c) (d)
Bland-Altman Plot
12 0.5

10 0.0
Difference Creatinine(mg/dL)
Estimated Glucose(mg/dL)

MAE = 0.072±0.010
R 2 = 0.865±0.055 0.5
8
1.0
6
1.5
4 2.0
2 md + 1.96*sd
2.5 md
md - 1.96*sd
0 3.0
0 2 4 6 8 0 2 4 6 8 10
Reference Creatinine(mg/dL) Average Creatinine(mg/dL)
(e) (f)
FIGURE 10: Relationship and agreement (Bland-Altman plot) between estimated values and reference values at testing for DNN
models without feature selection.

error. R and R2 for both give a correlation of blood compo- mmol/L. Similarly, for the clinically measured individual
nent (Hb, Gl, and Cr) levels on datasets. The higher values of sample, the gold standard Cr levels ranged between 0.37
R2 indicates that the model is performed well with the input mg/dL and 9.02 mg/dL, with a mean of 1.03 mg/dL and
datasets. standard deviation of Cr is 0.87 mg/dL.
In this study, 93 subjects were studied aged between 0 In the first stage, the DNN based models were trained
and 79 years, while 59 males and 34 females. The range and validated with all features for each blood component
of gold standard clinically measured Hb levels (Reference level. The 10-fold cross-validation was applied to validate the
Hemoglobin) for this sample of individuals ranged from 7.90 DNN models. After the 10-fold cross-validation, each folds
g/dL to 21.49 g/dL, with Hb mean values of 12.84 g/dL and contained reference Hb values and estimated Hb values. Then
standard deviation of Hb is 2.15 g/dL. The range of clinically the mean performances of the model were calculated. The
measured gold standard Gl levels ranged from 3.33 mmol/L same procedure was followed for models of estimated the
to 21.11 mmol/L for this individual sample, with mean Gl Gl and Cr levels. Table 6 demonstrates the results of blood
values of 6.65 mmol/L and standard deviation of Gl is 2.94 component levels measurement using DNN models with all
VOLUME x, xxxx 13
 Author et al.: Preparation of Papers for IEEE ACCESS

Bland-Altman Plot
22 MAE = 0.318±0.036
R 2 = 0.922±0.052 md + 1.96*sd
20 2 md
md - 1.96*sd

Difference Hemoglobin(g/dL)
Estimated Hemoglobin(g/dL)

18
1
16
14 0
12
10 1
8
8 10 12 14 16 18 20 22 8 10 12 14 16 18 20
Reference Hemoglobin(g/dL) Average Hemoglobin(g/dL)
(a) (b)
Bland-Altman Plot
MAE = 0.375±0.057 5 md + 1.96*sd
20.0 R 2 = 0.902±0.019
4 md
17.5 md - 1.96*sd

Difference Glucose(mmol/L)
Estimated Glucose(mmol/L)

3
15.0 2
12.5 1
10.0 0
7.5 1
2
5.0
3
2.5
2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 5.0 7.5 10.0 12.5 15.0 17.5 20.0
Reference Glucose(mmol/L) Average Glucose(mmol/L)
(c) (d)

10 0.50
MAE = 0.052±0.005
Estimated Creatinine(mg/dL)

0.25
R 2 = 0.969±0.037
Difference Creatinine(mg/dL)

8
0.00

6 0.25

0.50
4
0.75

2 1.00 md + 1.96*sd
md
1.25 md - 1.96*sd
0
0 2 4 6 8 0 2 4 6 8 10
Reference Creatinine(mg/dL) Average Creatinine(mg/dL)

(e) (f)
FIGURE 11: Relationship and agreement (Bland-Altman plot) between estimated values and reference values at testing for DNN
models with CFS based feature selection.

features. The estimated accuracies of proposed DNN models level, 25 for Gl level, and 24 for Cr level respectively. From
are R2 = 0.877 and MAE = 0.329 for Hb, 0.770 and 0.516 Table 3, it is shown that the optimal features are varied with
for Gl and 0.865 and 0.072 for Cr level, respectively. regard to measurement level. It is occurred due to different
Furthermore, a correlation-based features selection (CFS) features are correlated to different blood component levels.
method was used to select the best optimal features. Features In afterwards, the optimal features were applied to the DNN
selection is necessary in order to reduce the probability of models to estimate the blood component levels. Models were
models over-fitting. Correlation is a measure to determine validated using 10-fold cross-validation for each reference
whether or not a feature is highly correlated with class label blood component level.
but not highly correlated with any of the other relevant From Table 7, it can be observed that the CFS based feature
features [22], [59]. After applied the CFS method to the selection method provided the best results when combining
features dataset with Hb, Gl and Cr estimation levels, the with DNN models for estimation of Hb, Gl, and Cr levels.
number of features have been decreased from 48 to 27 for Hb According to the obtained results in Table 7, this approach
14 VOLUME x, xxxx
Author et al.: Preparation of Papers for IEEE ACCESS

TABLE 8: Comparison of our proposed DNN based models with several exiting smartphone-based non-invasive methods offered in
physiological parameter monitoring processes.

Authors and Year Purpose Smartphone Participants Captured Algorithm(s) Performance

Dantu et al. (2014) [58] HTC One X 68 − p − value = 0.005

Wang et al. (2016) [26] Nexus-5p 31 SVR R = 0.82

Wang et al. (2017) [3] Nexus-6p 32 LR R = 0.62

Anggraeni and Fatoni (2017) [31] Asus Zen-Fone 2 Laser 20 LR R2 = 0.81

Hasan et al. (2018) [27] Nexus-4p 75 ANN R2 = 0.93

Zhang et al. (2019) [28] iPhone 6s Plus 14 DT, BT and KNN Acc = 86.2%

Chowdhury et al. (2019) [32] iPhone 7 Plus 18 PCR SEP = 18.31mg/dL

 
2

  R = 0.922

Proposed (In this work) Nexus-6p 93 DNN R2 = 0.902

 

R2 = 0.969

*
= Hemoglobin, = Glucose, = Creatinine, = Video, = Image, Acc = Accuracy, SVR = Support Vector Regression,
LR = Linear Regression, ANN = Artificial Neural Network, DT = Decision Tree, BT = Bagged Trees, KNN = K-nearest neighbor, PCR
= Principal Component Regression, SEP = Standard Error of Prediction, DNN = Deep Neural Network.

provides the highest estimated accuracy of R2 = 0.922 for veloped a smartphone-based application named as HemaApp
Hb, R2 = 0.902 for Gl, and R2 = 0.969 for Cr. From the (FDA-approved device) to measure Hb level non-invasively.
Table 6 and Table 7, it is noticed that the accuracy (R2 ) of The authors captured fingertip video data from 31 subjects
DNN models with CFS based selected features is increased using Nexus-5p, used the SVR model, and achieved highest
by 5.13%, 17.14%, and 12.02% for Hb, Gl and Cr levels relationship with correlation coefficient (R) of 0.82. In [3],
comparing with DNN models with all features. the same authors improved the configuration of hardware and
A relationship between the reference values (gold standard used the LR model, and achieved a Pearson correlation of
values) and estimated values for blood component levels has 0.62. Anggraeni and Fatoni [31] developed LR model using
been established to understate the above results better. A conjunctiva image of 20 participants captured by Asus Zen-
correlation-based comparison between the estimated value Fone 2 and estimated the Hb level that highly correlated with
and reference value has been shown in Figure 10 ((a), (c), clinical Hb value and gained a relationship with R2 of 0.81.
and (e)) and Figure 11 ((a), (c), and (e)) for Hb, Gl, and In [27], the authors developed a smartphone-based applica-
Cr respectively. Furthermore, the Bland-Altman plot is il- tion named as SmartHeLP to measure the Hb concentration
lustrated to evaluate how much the estimated value differs in blood. In this case, they collected 75 fingertip video data
from the reference value. Bland-Altman [60] plot shows the via Nexus-4p smartphone and applied in the ANN model,
agreement between the reference value and the estimated and achieved relationship with R2 of 0.93. Zhang et al. [28]
value by constructing limits of agreement (LOA). It is found developed a system for the estimation of blood Gl level
that most of the estimated values for hemoglobin, glucose, based on smartphone fingertip video. The authors acquired
and creatinine are within the LOA (bias ±1.96 × SD) in the PPG signal from video, used subspace KNN classifier
Figure 10 ((b), (d), and (f)) and Figure 11 ((b), (d), and (f)) and obtained the accuracy Acc = 86.2%. Chowdhury et al.
[32] proposed a smartphone-based approach to estimate the
C. COMPARISON OF RESULTS blood Gl level in a non-invasive. The authors recorded the
Compassion of our work is shown in Table 8 for estimating fingertip video, converted the frames into PPG signal, applied
the blood component levels to validate our contribution with PCR algorithm on the extracted features, and achieved a SEP
previous existing smartphone-based non-invasive techniques. as low as 18.31 mg/dL. However, it is difficult to compare
existing works related to this filed due to different datasets,
Looking at individual related works in Table 8, Dantu
devices, and sensors to collect the data from various parts
et al. [58] introduced a method for measuring Gl level via
of the subject’s body and different evaluation criteria. In our
smartphone camera by capturing the transmitted laser light
proposed method, we have used a smartphone device (Nexus-
through a fingertip. The blue and green intensities of modi-
6p) to capture the fingertip video data, generated PPG signal
fied Bear-Lambert law was used to determine the GL level
from fingertip videos, and developed DNN based models to
and gained a (p − value) = 0.005. Wang et al. [26] de-
VOLUME x, xxxx 15
 Author et al.: Preparation of Papers for IEEE ACCESS

estimate the hemoglobin, glucose, and creatinine levels and [3] E. J. Wang, W. Li, J. Zhu, R. Rana, and S. N. Patel, “Noninvasive
achieved a good performance. hemoglobin measurement using unmodified smartphone camera and white
flash,” in 2017 39th Annual International Conference of the IEEE Engi-
neering in Medicine and Biology Society (EMBC). IEEE, 2017, pp.
V. CONCLUSION 2333–2336.
This paper introduced a non-invasive hemoglobin, glucose, [4] S. S. Morris, M. T. Ruel, R. J. Cohen, K. G. Dewey, B. de la Brière,
and M. N. Hassan, “Precision, accuracy, and reliability of hemoglobin
and creatinine measurement system based on PPG signal assessment with use of capillary blood,” The American journal of clinical
obtained from fingertip video using the DNN model. It nutrition, vol. 69, no. 6, pp. 1243–1248, 1999.
provides a good basis for observing hemoglobin, glucose, [5] M. K. Hasan, N. Sakib, R. R. Love, and S. I. Ahamed, “Analyzing the
existing noninvasive hemoglobin measurement techniques,” in 2017 IEEE
and creatinine in real-time from home. Firstly, fingertip video 8th Annual Ubiquitous Computing, Electronics and Mobile Communica-
is recorded using a smartphone, illuminating the finger with tion Conference (UEMCON). IEEE, 2017, pp. 442–448.
850 nm NIR LED. Then, PPG signals are generated from the [6] J. L. A. Nirupa and V. J. Kumar, “Non-invasive measurement of
hemoglobin content in blood,” in 2014 IEEE International Symposium on
video. From the PPG signal, its derivatives (1st and 2nd ) and Medical Measurements and Applications (MeMeA). IEEE, 2014, pp.
Fourier analysis, 46 characteristic features are extracted. A 1–5.
correlation-based features selection method using a genetic [7] L. Scarfe, A. Rak-Raszewska, S. Geraci, D. Darssan, J. Sharkey, J. Huang,
N. C. Burton, D. Mason, P. Ranjzad, S. Kenny et al., “Measures of kidney
algorithm is used to select appropriate features. Finally, DNN function by minimally invasive techniques correlate with histological
based models are developed, and 10-fold cross-validation glomerular damage in scid mice with adriamycin-induced nephropathy,”
method is applied to validated the models. The obtained Scientific reports, vol. 5, p. 13601, 2015.
results show that the combination of CFS and DNN models [8] S. A. Siddiqui, Y. Zhang, J. Lloret, H. Song, and Z. Obradovic, “Pain-free
blood glucose monitoring using wearable sensors: Recent advancements
are very effective for estimating blood component levels. and future prospects,” IEEE reviews in biomedical engineering, vol. 11,
The performance of our proposed technique has also been pp. 21–35, 2018.
compared against existing methods, which indicates the ap- [9] H. Von Schenck, M. Falkensson, and B. Lundberg, “Evaluation of"
hemocue," a new device for determining hemoglobin.” Clinical chemistry,
plicability of our proposed method. vol. 32, no. 3, pp. 526–529, 1986.
In the future, we would like to prefer the following ap- [10] P. P. Pai, P. K. Sanki, S. K. Sahoo, A. De, S. Bhattacharya, and S. Banerjee,
proaches for further enhancement of our work: (1) the dataset “Cloud computing-based non-invasive glucose monitoring for diabetic
care,” IEEE Transactions on Circuits and Systems I: Regular Papers,
will be increased from a heterogeneous people to make vol. 65, no. 2, pp. 663–676, 2017.
the dataset more balanced, (2) the video will be recorded [11] S. Lekha and M. Suchetha, “Real-time non-invasive detection and clas-
using different smartphone to increase the reliability of the sification of diabetes using modified convolution neural network,” IEEE
journal of biomedical and health informatics, vol. 22, no. 5, pp. 1630–
proposed technique, (3) the fingertip video collection pro- 1636, 2017.
cess will be handled automatically, (4) the whole estimation [12] M. Parker, Z. Han, E. Abu-Haydar, E. Matsiko, D. Iyakaremye,
process will be cloud-based, (5) the data will be captured L. Tuyisenge, A. Magaret, and A. Lyambabaje, “An evaluation of
hemoglobin measurement tools and their accuracy and reliability when
from an end-user and sent the data to a cloud-server using
screening for child anemia in rwanda: A randomized study,” PloS one,
a smartphone-based application as well as the result of cap- vol. 13, no. 1, p. e0187663, 2018.
tured data will be displayed to the end-user by analyzing the [13] J. Harvey, S. M. Salehizadeh, Y. Mendelson, and K. H. Chon, “Oxima: A
data with the models. frequency-domain approach to address motion artifacts in photoplethys-
mograms for improved estimation of arterial oxygen saturation and pulse
rate,” IEEE Transactions on Biomedical Engineering, vol. 66, no. 2, pp.
ACKNOWLEDGMENT 311–318, 2018.
The authors would like to thanks all the voluntary par- [14] J. Kraitl, D. Klinger, D. Fricke, U. Timm, and H. Ewald, “Non-invasive
measurement of blood components,” in Advancement in Sensing Technol-
ticipants of this study and Dr. Moniruzzaman, Managing ogy. Springer, 2013, pp. 237–262.
Director, physicians, nurses, and supporting members from [15] F. Rundo, S. Conoci, A. Ortis, and S. Battiato, “An advanced bio-inspired
Medical Centre Hospital, Chittagong, Bangladesh, for their photoplethysmography (ppg) and ecg pattern recognition system for med-
ical assessment,” Sensors, vol. 18, no. 2, p. 405, 2018.
constant support during the study. This work was financially [16] J. Allen, “Photoplethysmography and its application in clinical physiologi-
supported in part by the Khulna University of Engineering cal measurement,” Physiological measurement, vol. 28, no. 3, p. R1, 2007.
& Technology (Khulna-9203, Bangladesh). The authors also [17] J. Yadav, A. Rani, V. Singh, and B. M. Murari, “Prospects and lim-
itations of non-invasive blood glucose monitoring using near-infrared
thank anonymous reviewers for suggestions on improving spectroscopy,” Biomedical signal processing and control, vol. 18, pp. 214–
this manuscript. 227, 2015.
[18] S. Kwon, H. Kim, and K. S. Park, “Validation of heart rate extraction
using video imaging on a built-in camera system of a smartphone,” in 2012
CONFLICTS OF INTEREST
Annual International Conference of the IEEE Engineering in Medicine and
Thee authors declare that there are no conflicts of interest Biology Society. IEEE, 2012, pp. 2174–2177.
regarding the publication. [19] A. R. Kavsaoğlu, K. Polat, and M. Hariharan, “Non-invasive prediction of
hemoglobin level using machine learning techniques with the ppg signal’s
characteristics features,” Applied Soft Computing, vol. 37, pp. 983–991,
REFERENCES 2015.
[1] J. Li and C. Fernando, “Smartphone-based personalized blood glucose [20] M. A.-u. Golap and M. Hashem, “Non-invasive hemoglobin concentration
prediction,” ICT Express, vol. 2, no. 4, pp. 150–154, 2016. measurement using mggp-based model,” in 2019 5th International Confer-
[2] Z. Li, G. Li, W.-J. Yan, and L. Lin, “Classification of diabetes and ence on Advances in Electrical Engineering (ICAEE). IEEE, 2019, pp.
measurement of blood glucose concentration noninvasively using near 1–6.
infrared spectroscopy,” Infrared Physics & Technology, vol. 67, pp. 574– [21] J. Moraes, M. Rocha, G. Vasconcelos, J. Vasconcelos Filho, V. de Albu-
582, 2014. querque, and A. Alexandria, “Advances in photopletysmography signal

16 VOLUME x, xxxx
Author et al.: Preparation of Papers for IEEE ACCESS

analysis for biomedical applications,” Sensors, vol. 18, no. 6, p. 1894, [41] L. Wang, E. Pickwell-MacPherson, Y. Liang, and Y. T. Zhang, “Noninva-
2018. sive cardiac output estimation using a novel photoplethysmogram index,”
[22] F. Miao, N. Fu, Y.-T. Zhang, X.-R. Ding, X. Hong, Q. He, and Y. Li, Proc. of 31st IEEE EMBS, pp. 1746–1749, 2009.
“A novel continuous blood pressure estimation approach based on data [42] E. Seitsonen, I. Korhonen, M. Van Gils, M. Huiku, J. Lötjönen, K. Korttila,
mining techniques,” IEEE J. Biomed. Health Inform, vol. 21, pp. 1730– and A. Yli-Hankala, “Eeg spectral entropy, heart rate, photoplethysmogra-
1740, 2017. phy and motor responses to skin incision during sevoflurane anaesthesia,”
[23] N. Nirala, R. Periyasamy, B. K. Singh, and A. Kumar, “Detection of type- Acta Anaesthesiologica Scandinavica, vol. 49, no. 3, pp. 284–292, 2005.
2 diabetes using characteristics of toe photoplethysmogram by applying [43] Y. Lyu, X. Luo, J. Zhou, C. Yu, C. Miao, T. Wang, Y. Shi, and K.-i.
support vector machine,” Biocybernetics and Biomedical Engineering, Kameyama, “Measuring photoplethysmogram-based stress-induced vas-
vol. 39, no. 1, pp. 38–51, 2019. cular response index to assess cognitive load and stress,” in Proceedings
[24] C. Choi, B.-H. Ko, J. Lee, S. K. Yoon, U. Kwon, S. J. Kim, and Y. Kim, of the 33rd Annual ACM Conference on Human Factors in Computing
“Ppg pulse direction determination algorithm for ppg waveform inversion Systems. ACM, 2015, pp. 857–866.
by wrist rotation,” in 2017 39th Annual International Conference of the [44] T.-H. Fu, S.-H. Liu, and K.-T. Tang, “Heart rate extraction from photo-
IEEE Engineering in Medicine and Biology Society (EMBC). IEEE, plethysmogram waveform using wavelet multi-resolution analysis,” Jour-
2017, pp. 4090–4093. nal of medical and biological engineering, vol. 28, no. 4, pp. 229–232,
[25] P. Schoettker, J. Degott, G. Hofmann, M. Proença, G. Bonnier, A. Lemkad- 2008.
dem, M. Lemay, R. Schorer, U. Christen, J.-F. Knebel et al., “Blood [45] L. Chen, A. T. Reisner, and J. Reifman, “Automated beat onset and peak
pressure measurements with the optibp smartphone app validated against detection algorithm for field-collected photoplethysmograms,” in 2009
reference auscultatory measurements,” Scientific Reports, vol. 10, no. 1, Annual International Conference of the IEEE Engineering in Medicine and
pp. 1–12, 2020. Biology Society. IEEE, 2009, pp. 5689–5692.
[26] E. J. Wang, W. Li, D. Hawkins, T. Gernsheimer, C. Norby-Slycord, and [46] C. G. Scully, J. Lee, J. Meyer, A. M. Gorbach, D. Granquist-Fraser,
S. N. Patel, “Hemaapp: noninvasive blood screening of hemoglobin using Y. Mendelson, and K. H. Chon, “Physiological parameter monitoring
smartphone cameras,” in Proceedings of the 2016 ACM International Joint from optical recordings with a mobile phone,” IEEE Transactions on
Conference on Pervasive and Ubiquitous Computing. ACM, 2016, pp. Biomedical Engineering, vol. 59, no. 2, pp. 303–306, 2011.
593–604. [47] E. Jonathan and M. Leahy, “Investigating a smartphone imaging unit for
[27] M. K. Hasan, M. M. Haque, R. Adib, J. F. Tumpa, A. Begum, R. R. Love, photoplethysmography,” Physiological measurement, vol. 31, no. 11, p.
Y. L. Kim, and I. A. Sheikh, “Smarthelp: Smartphone-based hemoglobin N79, 2010.
level prediction using an artificial neural network,” in AMIA Annual [48] H. Yuan, S. F. Memon, T. Newe, E. Lewis, and G. Leen, “Motion artefact
Symposium Proceedings, vol. 2018. American Medical Informatics minimization from photoplethysmography based non-invasive hemoglobin
Association, 2018, p. 535. sensor based on an envelope filtering algorithm,” Measurement, vol. 115,
pp. 288–298, 2018.
[28] Y. Zhang, Y. Zhang, S. A. Siddiqui, and A. Kos, “Non-invasive blood-
[49] C. Wei, L. Sheng, G. Lihua, C. Yuquan, and P. Min, “Study on condition-
glucose estimation using smartphone ppg signals and subspace knn classi-
ing and feature extraction algorithm of photoplethysmography signal for
fier,” Elektrotehniski Vestnik, vol. 86, no. 1/2, pp. 68–74, 2019.
physiological parameters detection,” in 2011 4th International Congress on
[29] R. S. Al-Baradie and A. S. C. Bose, “Portable smart non-invasive
Image and Signal Processing, vol. 4. IEEE, 2011, pp. 2194–2197.
hemoglobin measurement system,” in 10th International Multi-
[50] A. Chatterjee and A. Prinz, “Image analysis on fingertip video to obtain
Conferences on Systems, Signals & Devices 2013 (SSD13). IEEE, 2013,
ppg,” Biomedical and Pharmacology Journal, vol. 11, no. 4, pp. 1811–
pp. 1–4.
1827, 2018.
[30] S. Ramasahayam, S. H. Koppuravuri, L. Arora, and S. R. Chowdhury, [51] A. R. KAVSAOĞLU, K. Polat, and M. R. Bozkurt, “An innovative peak
“Noninvasive blood glucose sensing using near infra-red spectroscopy detection algorithm for photoplethysmography signals: an adaptive seg-
and artificial neural networks based on inverse delayed function model of mentation method,” Turkish Journal of Electrical Engineering & Computer
neuron,” Journal of medical systems, vol. 39, no. 1, p. 166, 2015. Sciences, vol. 24, no. 3, pp. 1782–1796, 2016.
[31] M. Anggraeni and A. Fatoni, “Non-invasive self-care anemia detection [52] A. R. Kavsaoğlu, K. Polat, and M. R. Bozkurt, “A novel feature ranking
during pregnancy using a smartphone camera,” in IOP Conference Series: algorithm for biometric recognition with ppg signals,” Computers in
Materials Science and Engineering, vol. 172, no. 1. IOP Publishing, 2017, biology and medicine, vol. 49, pp. 1–14, 2014.
p. 012030. [53] R. Tiwari and M. P. Singh, “Correlation-based attribute selection using
[32] T. T. Chowdhury, T. Mishma, M. Osman, T. Rahman et al., “Estimation of genetic algorithm,” International Journal of Computer Applications, vol. 4,
blood glucose level of type-2 diabetes patients using smartphone video,” no. 8, pp. 28–34, 2010.
arXiv preprint arXiv:1911.12619, 2019. [54] J. R. Koza, “Genetic programming as a means for programming computers
[33] L. Carroll and T. R. Humphreys, “Laser-tissue interactions,” Clinics in by natural selection,” Statistics and computing, vol. 4, no. 2, pp. 87–112,
dermatology, vol. 24, no. 1, pp. 2–7, 2006. 1994.
[34] D. McDuff, S. Gontarek, and R. W. Picard, “Remote detection of pho- [55] C. C. Aggarwal, “Training deep neural networks,” in Neural Networks and
toplethysmographic systolic and diastolic peaks using a digital camera,” Deep Learning. Springer, 2018, pp. 105–167.
IEEE Transactions on Biomedical Engineering, vol. 61, no. 12, pp. 2948– [56] I. Goodfellow, Y. Bengio, and A. Courville, “Deep learning. book in prepa-
2954, 2014. ration for mit press,” URL¡ http://www. deeplearningbook. org, vol. 1,
[35] M. Elgendi, “On the analysis of fingertip photoplethysmogram signals,” 2016.
Current cardiology reviews, vol. 8, no. 1, pp. 14–25, 2012. [57] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdi-
[36] K. Takazawa, N. Tanaka, M. Fujita, O. Matsuoka, T. Saiki, M. Aikawa, nov, “Dropout: a simple way to prevent neural networks from overfitting,”
S. Tamura, and C. Ibukiyama, “Assessment of vasoactive agents and vas- The journal of machine learning research, vol. 15, no. 1, pp. 1929–1958,
cular aging by the second derivative of photoplethysmogram waveform,” 2014.
Hypertension, vol. 32, no. 2, pp. 365–370, 1998. [58] V. Dantu, J. Vempati, and S. Srivilliputhur, “Non-invasive blood glucose
[37] U. Rubins, A. Grabovskis, J. Grube, and I. Kukulis, “Photoplethysmogra- monitor based on spectroscopy using a smartphone,” in 2014 36th Annual
phy analysis of artery properties in patients with cardiovascular diseases,” International Conference of the IEEE Engineering in Medicine and Biol-
in 14th Nordic-Baltic Conference on Biomedical Engineering and Medical ogy Society. IEEE, 2014, pp. 3695–3698.
Physics. Springer, 2008, pp. 319–322. [59] S. S. Durduran, “A decision making system to automatic recognize of
[38] I. Imanaga, H. Hara, S. Koyanagi, and K. Tanaka, “Correlation between traffic accidents on the basis of a gis platform,” Expert Systems with
wave components of the second derivative of plethysmogram and arterial Applications, vol. 37, no. 12, pp. 7729–7736, 2010.
distensibility,” Japanese heart journal, vol. 39, no. 6, pp. 775–784, 1998. [60] J. M. Bland and D. G. Altman, “Statistical methods for assessing agree-
[39] S. A. Esper and M. R. Pinsky, “Arterial waveform analysis,” Best Practice ment between two methods of clinical measurement,” International journal
& Research Clinical Anaesthesiology, vol. 28, no. 4, pp. 363–380, 2014. of nursing studies, vol. 47, no. 8, pp. 931–936, 2010.
[40] H. J. Baek, J. S. Kim, Y. S. Kim, H. B. Lee, and K. S. Park, “Second
derivative of photoplethysmography for estimating vascular aging,” in
Information Technology Applications in Biomedicine, 2007. ITAB 2007.
6th International Special Topic Conference on. IEEE, 2007, pp. 70–72.

VOLUME x, xxxx 17
 Author et al.: Preparation of Papers for IEEE ACCESS

MD. REZWANUL HAQUE was born on July

25, 1996. He has persuaded his B.Sc. engineer-
ing degree in Computer Science and Engineer-
ing at Khulna University of Engineering & Tech-
nology, Khulna, Bangladesh. His research inter-
ests include Bioinformatics, Biomedical Engineer-
ing, Health Informatics, Biomedical Instrumenta-
tion, Computer Vision, Image Processing, Natural
Language Processing, Machine Learning, Deep
Learning, GANs, Deep Reinforcement Learning
and Artificial Intelligence.

S. M. TASLIM UDDIN RAJU was born on

February 19, 1996 in Feni, Bangladesh. He has
received his B.Sc. engineering degree in Computer
Science and Engineering (CSE) from the Khulna
University of Engineering & Technology (KUET),
Khulna, Bangladesh in 2019, where he is currently
pursuing the M.Sc. engineering degree. He is now
working as Lecturer in Department of CSE at
KUET, Khulna, Bangladesh. His research inter-
ests include Bioinformatics, Health Informatics,
Biomedical Instrumentation, Deep Learning, Machine Learning, Natural
Language Processing, and Image processing.

MD. ASAF-UDDOWLA GOLAP received the

bachelor’s and master’s degrees in computer sci-
ence and engineering from the Khulna Uni-
versity of Engineering & Technology, Khulna,
Bangladesh, in 2017 and 2019, respectively.
His research interests include Bioinformatics,
Biomedical Signal Processing, Deep Learning,
Machine Learning, Image processing, Genetic
Programming and Bio-inspired Computing Tech-
niques.

M. M. A. HASHEM received the Bachelors de-

gree in Electrical and Electronic Engineering from
Khulna University of Engineering & Technology
(KUET), Khulna, Bangladesh in 1988, Masters
degree in Computer Science from Asian Institute
of Technology (AIT), Bangkok, Thailand in 1993
and PhD degree in Artificial Intelligence Systems
from the Saga University, Japan in 1999. He is
currently a Professor in the Dept. of Computer
Science and Engineering, Khulna University of
Engineering & Technology (KUET), Bangladesh. His research interests
include Artificial Intelligence, Machine Learning, Bioinformatics, Biomed-
ical Engineering, Health Informatics, Biomedical Instrumentation, Evolu-
tionary Computations, Soft Computing, etc. He has published more than
100 referred articles in international Journals/Conferences. He is a co-
author of a book titled Evolutionary Computations: New Algorithms and
their Applications to Evolutionary Robots, Series: Studies in Fuzziness and
Soft Computing, Vol. 147, Springer-Verlag, Berlin/New York, ISBN: 3-540-
20901-8, (2004).

18 VOLUME x, xxxx