Towards Photoplethysmogram Based Non-Invasive Blood Pressure Classification

Full text

Towards Photoplethysmogram based Non-Invasive
Blood Pressure Classification
Rajdeep Kumar Nath1, Himanshu Thapliyal1and Allison Caban-Holt2
1Department of Electrical and Computer Engineering
2Sanders-Brown Center on Aging
University of Kentucky, Lexington, USA
hthapliyal@uky.edu
Abstract—A novel blood pressure classification model using
Phototplethysmogram (PPG) is proposed in this work. The
proposed model uses signal processing and machine learning
algorithms to classify blood pressure in four stages: normal,
elevated, stage 1 and stage 2. A total of 83 features were extracted
from the PPG signal which includes 71 statistical features and
12 characteristic features. We have used random forest classifier
to train and test our predictive model. The proposed method is
evaluated on publicly available MIMIC database for 20 different
individuals. The database contains raw PPG data for different
users and Arterial Blood Pressure (ABP) to calculate the systolic
and diastolic blood pressure to be used as the ground truth for
training and validation purposes. We have achieved an overall
accuracy of 90.8% over the four classes of blood pressure levels.
The results indicate that the proposed model will be ideal for
integration into a non-invasive blood pressure monitoring system
with significant accuracy.
Index Terms—Photoplethysmogram (PPG), Blood Pressure
(BP), Systolic Blood Pressure (SBP), Diastolic Blood Pressure
(DBP), Frequency Spectrum, Arterial Blood Pressure.
I. INTRODUCTION
Hypertension is a condition where blood exerts higher than
normal pressure in the blood vessels. High blood pressure
or hypertension is a common health problem that can affect
individuals of any age groups. With no initial symptoms,
hypertension can be difficult to detect and develops over years
causing serious health conditions like heart failure, vision loss,
stroke, and complications related to kidney function. American
Heart Association (AHA) estimated that about 85 million
people, one in every three adults in the US have high blood
pressure related complications [1]. However, monitoring blood
pressure on regular basis can prevent situation from getting
worse. The conventional method to measure blood pressure is
through Sphygmomanometer which is invasive and most of the
general population are not familiar with the interpretation of
blood pressure reading from Sphygmomanometer along with
the issue of portability and the complexity associated in using
the device. Although, there are portable devices that are used
for blood pressure estimation, these devices still requires the
use of cuff in the arm for monitoring blood pressure which
does not facilitate continuous monitoring. Hence, a cuffless
non-invasive blood pressure monitoring system is required
to meet the health needs of the growing population and to
improve the quality of life of the people.
In this work, we have presented a non-invasive cuffless
blood pressure classification model using PPG signal. The rest
of the paper is organized as follows: Section II presents the
related work, Section III presents the proposed model, Section
IV presents the results and Section V presents the conclusion
and future work.
II. RE LATE D WORK
Various works has been done in the literature to estimate
systolic and diastolic blood pressure. However, most of the
existing works focus on blood pressure estimation rather than
blood pressure classification. To the best of our knowledge,
in [2], an attempt has been made to classify blood pressure
and is the only work attempting the classification problem so
far. Also, in [2], the researchers has classified BP values using
both ECG and PPG and over three classes of BP values. In
the current work, we propose a model to classify BP values
over four classes [3] using only PPG signal. The purpose
of implementing a blood pressure classification model for
integration into continuous blood pressure monitoring is be-
cause the interpretation of the readings is more understandable
to the general population when blood pressure readings are
presented as classes rather than continuous values. Also, the
time complexity associated with classification is lower than
that of continuous prediction hence making it ideal for real
time monitoring.
III. PROP OS ED MO DE L
In this work, the raw PPG signal and the corresponding
ABP signal collected from the MIMIC database [4] is used to
build and validate our model. The dataset contains 72 complete
records of varying length. Out of the 72 records, 20 records
were selected to build our model. The selection criteria is
based on two main factors: (i) only those subjects whose
records included both PPG signal streams and ABP signals,
(ii) of those subjects whose PPG data are available, subjects
were filtered out where the missing data consisted of more
than 10% of the length of the data. The ABP data is used
for calculating the systolic blood pressure and diastolic blood
pressure for training and validating the model.

Fig. 1. Overview of the proposed model
The proposed model broadly consists of three stages (Figure
1): (i) Data Preprocessing, (ii) Feature Extraction, (iii) BP
Classification.
A. Data Preprocessing
The data preprocessing step takes as input raw PPG signal
stream and generates processed PPG signal stream. The raw
PPG data is first normalized so that all the values remain in
the range [0,1]. The normalized data is represented in such a
way that all the data points can be represented as a factor the
mean of the dataset as: , x(t) = n.mean(x), where x is the
signal stream and x(t) is the value of the signal at any instant
of time t and n is an integer. This step is done by calculating
the residue, when the sample point is divided by the mean of
the signal by using the modulus operation and then subtracting
the residue from the signal stream. This step performs some
preliminary noise removal. Subsequently, the processed signal
stream is filtered for noise removal by applying a butterworth
band pass filter with cut off frequency 40 hz. The filtered data
is then used for feature extraction.
B. Feature Extraction
The PPG signal is characterized by two main points systolic
and diastolic points. Figure 2 shows a typical PPG signal.
The diastolic point marks the lowest point of the diastole
where the pressure exerted by the blood flow is minimum
and the beginning of systole where the heart begins pumping
blood to the lungs and the blood pressure reaches a maximum
value at the systolic point. The period from systolic point
to the next diastolic point is known as the diastole period
where blood begins filling into the heart for replenishing
oxygen content. PPG signal is also characterized by Dicrotic
notch and point of second wave. However these points of
inflection are not easily identifiable from the original PPG
signal. The information obtained from the first derivative and
second derivative of the PPG signal when combined with the
information obtained from original PPG signal can provide
reliable information about the critical points of a PPG signal.
For example the diastolic point is the point of the PPG signal
where the first derivative of the signal is closest to zero. The
second derivative of the PPG signal also called Acceleration
PlethysmoGram (APG), has five critical points which are
helpful in providing important information required to monitor
Fig. 2. A typical PPG signal and its components
Fig. 3. Original PPG signal and its first and second derivative
cardiac functionality. This concept is illustrated in figure 3 [5].
The frequency spectrum of the PPG signal along with the
derivatives can also provide critical information about the
cardiac function that are otherwise not readily extracted from
PPG signal alone. Hence, features are extracted from six signal
streams of PPG signal: (i) first derivative of PPG signal, (ii)
second derivative of PPG signal, (iii) frequency spectrum of
PPG signal, (iv) first derivative of the frequency spectrum and
(v) second derivative of frequency spectrum.
Table I presents the feature type and their characteristics.
The statistical feature consists of estimation of important
statistical features like mean, variance, kurtosis, skewness
etc. The characteristic features of the frequency spectrum
consists of estimation of the peak amplitude and the frequency
associated with the signal in a given window.
C. Blood Pressure Classification
The Blood pressure classification functionality takes the
original PPG signal stream and predicts class of the blood

TABLE I
FEATU RE SE T DE SCR IP TIO N
Signal Stream Feature Type No. of Features Window Size Overlap
PPG first derivative Statistical 14 2 1
PPG second derivative Statistical 14 2 1
PPG signal Statistical 15 2 1
PPG frequency spectrum Characteristic 12 1 0.5
Frequency Spectrum first derivative Statistical 14 2 1
Frequency Spectrum second derivative Statistical 14 2 1
pressure based on the computed features. The blood pressure
ranges are classified in four classes:
(i) Normal: SBP <120 mmHg and DBP <80 mmHg.
(ii) Elevated: SBP (120 −139) mmHg and DBP (80 −90)
mmHg.
(iii) Stage 1: SBP (140−159) mmHg and DBP (90-99)mmHg.
(iv) Stage 2: SBP >=160 mmHg and DBP >= 100 mmHg.
The training class of the classifier assigns labels to each sample
in the feature set according to the ground truth of observation.
The blood pressure classification is implemented using the
Random Forest classifier. The Random Forest classifier is
an ensemble for various base estimators (decision trees in
this context) to fit different sub samples of the data. In our
model, we have used the number of estimators as 1000 and
the criterion gini is used to measure the quality of a split and
the maximum depth to which a solution is explored is chosen
at 30.
IV. RES ULT S
The classification model achieved an overall accuracy of
90.8% across all classes. The confusion matrix is shown in
Table II.
TABLE II
CONFUSION MATRIX FOR BLOOD PRESSURE CLASSIFICATION
Predicted/True Normal Elevated Stage 1 Stage 2 Accuracy
Normal 962 78 154 2 79.83%
Elevated 92 4269 10 0 96.86%
Stage 1 145 47 1277 139 80.9%
Stage 2 6 13 137 1669 92.2%
.
The performance of the proposed classification model is
further evaluated using different metrics namely specificity,
sensitivity and precision, which are calculated from the con-
fusion matrix . These metrics are evaluated for all classes and
are listed in Table III.
TABLE III
PERFORMANCE MEASURE FOR CLASSIFICATION MODEL
Class Specificity Sensitivity Precision
Normal 0.96 0.79 0.80
Elevated 0.97 0.96 0.97
Stage 1 0.95 0.80 0.79
Stage 2 0.97 0.92 0.91
From Table III, it can be seen that the specificity (true neg-
ative rate) is almost equal to 1 (perfect specificity). Although
the sensitivity for normal and stage 1 classes are a little lower
than that of elevated and stage 2 classes, the number obtained
is sufficiently good for a real time model.
V. FUTURE WO RK A ND CONCLUSION
As an extension of this work, we propose to build a
continuous blood pressure prediction model alongside blood
pressure classification model. Combining the results from the
two models will increase the efficiency and robustness of the
proposed model. Although, the blood pressure classification
can also be done by predicting the blood pressure and then
classifying the stages using a look up table, the process is not
reliable due to errors in predicting the blood pressure with high
deviation in any instances, the model might give inaccurate
classification of the blood pressure range. However, if different
models are employed to perform the tasks separately, it is
highly unlikely that both the models will provide erroneous
readings at the same instance.
In this paper, we have presented a blood pressure classifi-
cation model which can predict over four classes of BP values
with an overall accuracy of 90.8%. The proposed model can
predict normal, elevated, stage 1 and stage 2 classes with an
accuracy of 79%, 96%, 80% and 92% respectively and with
a precision of 80%, 97%, 79% and 91% respectively.
REFERENCES
[1] Christian Nordqvist, “What’s to know about high blood pressure?. Avail-
able from: https://www.medicalnewstoday.com/articles/159283.php,
November 2017.
[2] M. Kachuee, M. M. Kiani, H. Mohammadzade and M. Shabany,
”Cuffless Blood Pressure Estimation Algorithms for Continuous Health-
Care Monitoring,” in IEEE Transactions on Biomedical Engineering, vol.
64, no. 4, pp. 859-869, April 2017. doi: 10.1109/TBME.2016.2580904
[3] National High Blood Pressure Education Program. The Seventh Re-
port of the Joint National Committee on Prevention, Detection, Eval-
uation, and Treatment of High Blood Pressure. Bethesda (MD):
National Heart, Lung, and Blood Institute (US); 2004 Aug. Ta-
ble 3, Classification of blood pressure for adults. Available from:
https://www.ncbi.nlm.nih.gov/books/NBK9633/table/A32/
[4] G. B. Moody and R. G. Mark, ”A database to support development
and evaluation of intelligent intensive care monitoring,” Computers
in Cardiology 1996, Indianapolis, IN, USA, 1996, pp. 657-660. doi:
10.1109/CIC.1996.542622
[5] Millasseau, Sandrine C. et al. Determination of age-related increases in
large artery stiffness by digital pulse contour analysis. Clinical science
103 4 (2002): 371-7.