PreprintPDF Available

Reliable Respiratory Rate Extraction using PPG

Authors:
Preprints and early-stage research may not have been peer reviewed yet.

Abstract and Figures

Wearable electronics enable a new look into the health of individuals in a fashion that was never possible before. Respiratory Rate (RR) is one of the parameters which is of interest for various health studies. However, many reliable methods for measuring RR require wearing gadgets that are impractical in a normal daily life setup. On the other hand, more practical methods which are less intrusive are often less reliable. Extracting RR using Photoplethysmo-gram (PPG) signals is one of the methods in the latter group. A major challenge for this method is the movement artifact which leads to wrong estimation of RR or failure in its calculation. In this work, we propose a new algorithm, Smart Fusion of Frequency Domain Peak (SFFDP), which outperforms existing algorithm by at least 37% improvement in terms of reliability; i.e., average error, Standard Deviation (STD), and Figure of Merit (FoM). This method does not require any signal other than PPG, and therefore can be used in a wide range of wearable devices such as smart watches without any hardware additions.
Content may be subject to copyright.
Pre-Publication
Reliable Respiratory Rate Extraction using PPG
David Pollreisz and Nima TaheriNejad
November 27, 2019
Abstract
Wearable electronics enable a new look into the
health of individuals in a fashion that was never pos-
sible before. Respiratory Rate (RR) is one of the pa-
rameters which is of interest for various health stud-
ies. However, many reliable methods for measuring
RR require wearing gadgets that are impractical in
a normal daily life setup. On the other hand, more
practical methods which are less intrusive are often
less reliable. Extracting RR using Photoplethysmo-
gram (PPG) signals is one of the methods in the
latter group. A major challenge for this method is
the movement artifact which leads to wrong estima-
tion of RR or failure in its calculation. In this work,
we propose a new algorithm, Smart Fusion of Fre-
quency Domain Peak (SFFDP), which outperforms
existing algorithm by at least 37% improvement in
terms of reliability; i.e., average error, Standard De-
viation (STD), and Figure of Merit (FoM). This
method does not require any signal other than PPG,
and therefore can be used in a wide range of wearable
devices such as smart watches without any hardware
additions.
1 Introduction
Wearable Health-care Systems (WHS) spread their
coverage to a wide range of applications [1, 2], be-
ing critical physical health domain [3, 4, 5, 6], men-
tal health [1, 7, 8], or well-being and sport activi-
ties [1, 2]. However, their development is not free
of challenges, in particular, constraints on resources
drives the engineers to try to extract as much in-
formation possible with as little hardware as possi-
ble [9]. Moreover, the uncontrolled environment in
which they operate poses challenges on their reliabil-
ity [5, 9]. Some of these challenges could be addressed
by using more complex processing methods, how-
ever, those require more computational and energy
resources which are often not sufficiently available
on these devices [4, 9]. Therefore, solutions which
improve the reliability of the system with minimum
overhead on the required resources are extremely
valuable. In this paper, we propose such a solution
for Respiratory Rate (RR) extraction.
RR is an important physiological measure used
in various medical studies [10]. However, common
methods for direct measurement of RR, such as
mounting a mouth piece, are rather uncomfortable
for the subjects. One of the least intrusive methods
of measuring RR is inferring it from Photoplethys-
mogram (PPG) signals. The basic principle of RR
extraction using PPG is to use the fact that breath-
ing influences the cardiac system [11]. These influ-
ences, depicted in Figure 1, include amplitude and
frequency modulation as well as wandering of base-
line. One of the main challenges that many WHS
face, which affects the extraction of RR from PPG
too, is the movement artifact [9]. In [12], we have
provided an in-depth insight into the existing meth-
ods for detection and removal of motion artifacts in
PPG signals. However, these methods require either
extra sensors, or extra computational and energy re-
sources to perform extensive complex calculations. In
this work, we present a method that does not need
any extra sensor and computationally is very similar
to existing algorithms which extract RR from PPG.
Nevertheless, the proposed algorithm, Smart Fusion
of Frequency Domain Peak (SFFDP), is considerably
1
Accepted for publication at 11th IEEE Latin American Symposium on Circuits and Systems – LASCAS 2020
D. Pollreisz and N. Taherinejad, “Reliable Respiratory Rate Extraction using PPG”, 11th IEEE Latin American
Symposium on Circuits and Systems – LASCAS 2020, pp. 1-4, 2020.
@InProceedings{Pollreisz2020_LASCAS,
author = { Davd Pollreisz and Nima TaheriNejad},
title = {Reliable Respiratory Rate Extraction using {PPG}},
booktitle = {11th IEEE Latin American Symposium on Circuits and Systems – {LASCAS}},
pages = {1-4},
year = {2020},
month = {February},
address = {San José, Costa Rica},}
978-1-7281-3427-7/20/$31.00 © 2020 IEEE
Pre-Publication
more robust against movement artifacts, i.e., at least
37% better in all three factors representing reliability.
2 Proposed Method
Figure 2 shows all the steps of the proposed RR ex-
traction algorithm. In the rest of this section, we
describe the details of each step.
2.1 Pre-Processing
To be able to extract the necessary features of the sig-
nal, first the raw data needs to be pre-processed and
prepared for the feature extraction step. This prepa-
ration includes band-pass filtering to remove the off-
set and any noise that lays outside the field of interest
and extracting the location of local maxima and min-
ima of the Blood Volume Pulse (BVP) signal. Using
a finite impulse response filter [13] the peaks can be
easily detected since the maxima of the BVP signal
cause significantly large spikes. The order of the fil-
ter, N, is defined by
N=2fs
25 ,(1)
Figure 1: Respiration caused modulations: (a) none
(b) Baseline Wanderer (BW) (c) Amplitude Modula-
tion (AM) (d) Frequency Modulation (FM).
Figure 2: Flow chart of the implemented RR extrac-
tion algorithm.
where fsis the sampling rate. The filter coeffi-
cients [13] are
bk=1for k = 0, ... N
21
1for k =N
2, ..., N 1(2)
Figure 3 shows the visible improvement on a sam-
ple PPG signal before and after application of the
filter. Out of the filtered signal, the local extrema
are searched with three criteria: (i) Extrema are rec-
ognized as such only if they are bigger than the mean
value of the signal, (ii) only detect extrema that are
0.4fsapart, (iii) a peak must be surrounded by two
troughs and vice-versa. After that, only the relevant
peaks and troughs remain and the extraction of the
features can begin.
2.2 Feature Extraction
First the Amplitude Modulation (AM) feature is ex-
tracted by calculating the difference of amplitude be-
tween the peaks and troughs, which at the end is
normalized to the mean of the signal. The Baseline
Wanderer (BW) is calculated by algebraic addition of
the amplitude of a peak to its following trough and
Figure 3: (a) Original PPG signal and (b) filtered
PPG signal.
2
Pre-Publication
dividing the value by two, which is then normalized to
the mean of the signal. The last feature, Frequency
Modulation (FM), is calculated by subtracting the
temporal location of each peak and the one after it.
At the end it is normalized to the mean of the signal.
2.3 Respiratory Rate Extraction
For RR extraction, we propose a new estimation
method and then use smart fusion which combines
the estimation and values extracted from each fea-
ture to find a point of agreement between all those
values.
2.3.1 Proposed Estimation
For estimation of RR, existing methods [14, 15, 16]
process the extracted feature and respective proper-
ties in the time domain. In our proposed algorithm,
Smart Fusion of Frequency Domain Peak (SFFDP),
we do not use the features in the time domain like the
other ones but work in the frequency domain. Our
algorithm searches for the Dominant Frequency (DF)
in the extracted signal. First, the signal is detrended
and after that, the dominant peak in the range of
0.033 2Hz (which corresponds to a breath rate of
2 to 120 per minute) is searched and found. The
breathrate corresponding to the DF is then consid-
ered as the estimated RR.
2.3.2 Smart Fusion (SFU)
The last step is the fusion of the estimated values.
To this end, first, the Standard Deviation (STD)
of the estimated values from each feature (BW, AM
and FM) is calculated for each window. If the STD is
below 4, the mean value of these values is calculated
and is taken as the RR value of the fusion method.
On the other hand, if only two estimations have a
STD below 4 and this value is lower than the STD
of all three estimations, the mean value of these two
estimations is calculated and set as the final RR value
of the fusion method. If all four STDs exceed 4 then
the value of the SFU is set to NaN .
Table 1: Distribution of the recorded data
No movement Movement
Normal breathing 10 12
Fast breathing 4 4
Slow breathing 4 7
3 Experimental Results
3.1 Collected Data and Setup
The data set consists of 41 samples from four male
healthy volunteers, aged between 26 and 29 years,
performing three different kinds of breathing and
movements. The first task was normal breathing in
the range of 10 to 15 breaths per minute. The sec-
ond, fast breathing with a breath rate over 15 and
the last, slow breathing with a breath rate below 10.
During the 60 seconds of measurement, the arm was
moved from the table straight into the air and this
was repeated three times. Table 1 shows the distri-
bution of the collected data, and Figure 4 shows an
example of a BVP signal with the three movements
and their respective artifacts.
The proposed algorithm is implemented in Mat-
lab. Sampling frequency of the BVP is 64 Hz and for
the filter we have used a 4th order high-pass Infinite
Impulse Response (IIR) filter, namely Butterworth,
with a 0.05 Hz cut-off frequency and a low-pass one
with a 5 Hz cut-off frequency.
Figure 4: A BVP with three movement artefacts.
3
Pre-Publication
Table 2: Statistical results of the proposed method,
SFFDP.
Window Mean|Error|STD Samples Window FoM
Overlap [BPM] [%] length [s]
No 3.148 3.214 75.61 28 10.644
Yes 3.351 3.387 75.61 28 11.022
3.2 Window Size Selection
The performance of the system depends on various
parameters such as processing window size. To find
the perfect window range, a parameter sweep was
performed where the window range was swept from 4
to 30 seconds in a step size of 2. The maximum was
set to 30 because the test data are 60 seconds long.
If the range exceeds 30 only half of the signal could
be analyzed since no second window could be fully
formed. That is the reason why the range was not
increased further. In addition, the same range was
tested with 50% overlapping windows.
3.3 Figure of Merit (FoM)
To concretely evaluate the performance and quality
of the proposed algorithm as well as other algorithms,
we define a FoM, which is calculated by
F oM =Mean(|Error|) + S T D + 10 ×(1 CS R2),(3)
where error is measured in Breath per Minute (BPM)
and CSR stands for the Computed Samples Ratio,
that is the number of samples with successful esti-
mation of RR divided by the overall number of sam-
ples. Since CSR is smaller than one and the other two
numbers are usually between one and ten, we multi-
ply (1CSR2) with a constant of 10, so that it can be
in the order of the other two numbers. Otherwise, the
last term would be too small and would have practi-
cally no effect on FoM. Thus, the proposed FoM com-
bines the amount of error and reliability (represented
by STD) of RR estimation, as well as the number
of samples it can successfully estimate. The smaller
FoM, the better the algorithm.
3.4 Results
The proposed algorithm, showed to be able to suc-
cessfully estimate 75% of the samples in our data
set, with a mean error and standard deviation of
approximately 3 BPM. This is an acceptable per-
formance for many applications (e.g., Early Warning
Score (EWS) [3, 5, 6]), especially given the fact that
the data was contaminated with movement artifacts.
Moreover, as we can see in Section 4, it is better than
other existing algorithms. A summary of the results
that we obtained for the proposed algorithm is in-
serted in Table 2. As we see in this table, the perfor-
mance of the system with and without window over-
lap is very similar. However, the algorithm without
overlapping windows has a slightly better FoM and
requires slightly less processing power which makes it
overall favorable. Moreover, we observe that a Win-
dow Length of 28s was selected for both cases since in
our parameter sweep it showed the best performance
in both cases.
4 Comparisons
4.1 Existing Algorithms
To be able to have a fair comparison, we implemented
three other principle algorithms in literature, namely
Time Domain Peak Detection (TDPD), Time Do-
main - Count Origin (TDCO), and Count Origin -
Smart and Time Fusion (COSTF). The first extrac-
tion algorithm, TDPD, [14] uses the Peak Detection
(PD) for estimation of RR. The second and third al-
gorithms, TDCO and COSTF [15], use Count Ori-
gin (CO) method to detect peaks and troughs. In
this method, a threshold as 0.2 times the 75th per-
centile of peak values is defined, and any peaks with
an amplitude smaller than this threshold is dismissed.
A breath is detected as two consecutive peaks sep-
arated by only one trough (with an amplitude less
than zero). Moreover, COSTF uses an additional fu-
sion method called Temporal Fusion (TFU) [16]. In
Table 3, a summary of all the four tested algorithms,
including the proposed method (SFFDP), and their
features are shown. The key new feature of the pro-
posed algorithm is using DF for its estimation and
4
Pre-Publication
Table 3: A summary of all RR extraction algorithms.
Feature Extraction RR Estimation Fusion
AM BW FM CO PD DF SFU TFU
TDPD D D D D D
TDCO D D D D D
COSTF D D D D D D
SFFDP D D D D D
the new enhanced SFU for the fusion.
It should be noted that compared to the origi-
nal implementation, in our implementation of other
works, we enhanced TDPD, TDCO, and COSTF by
changing the PPG peak detection from the detection
of the maximum in the raw signal to finding it after
applying a filter, as seen in Figure 3, to get a more
robust detection. In addition, the fusion algorithm
was changed so that it can fuse two estimated values
instead of only all three, as it was the case originally.
This increases the percentages of calculated samples.
A summary of the results we obtained for each algo-
rithm is inserted in Table 4. In Table 4, the STDs for
TDPD are 0 because only 2.44% of the samples got
calculated. That is only one sample and therefore no
STD could be calculated.
4.2 Comparison
We have summarized the comparison for the best re-
sult of each algorithm and their respective FoM im-
provements in Table 5. In this table, the improve-
ments of the proposed algorithm, p, compared to al-
gorithm iare calculated using Imp. =F oMiF oMp
F oMi.
We observe that the proposed system has the small-
est mean error (3.1 BPM) and STD (only 3.2). We
remind the readers that the STD of TDPD is not con-
sidered for comparison. Mainly because its value of
0 does not reflect its reliability, but rather its lack of
success in estimating more than one sample. With
regard to success in estimating RR, the proposed
method has a good performance of 75% which is sig-
nificantly larger than TDPD. This value however, is
slightly lower than TDCO and COSTF, which have
a corresponding ratio of 97% and 100%. Neverthe-
less, this slight degradation in the ratio of success-
fully estimated samples is compensated by a much
Table 4: Statistical results of other algorithms in
the literature.
Algorithm Mean|Error|STD Samples Window Window FoM
[BPM] [%] length [s] Overlap
TDPD 9.226 0 2.44 12 19.220
TDPD 6.877 0 2.44 12 D16.871
TDCO 15.418 6.353 97.6 22 22.253
TDCO 15.323 6.156 95.1 22 D22.431
COSTF 16.257 6.753 100 28 23.009
COSTF 15.676 6.877 100 30 D22.552
Table 5: Comparison of the best performance of all
algorithms.
Algorithm Mean|Error|STD Samples FoM Improv-
[BPM] [%] ement
TDPD 6.877 0 2.44 16.871 37%
TDCO 15.418 6.353 97.6 22.253 52%
COSTF 15.676 6.877 100 22.552 53%
Proposed 3.148 3.214 75.61 10.644 -
larger improvement in the mean error (approximately
five times smaller error) and standard deviation (ap-
proximately two times smaller). In particular, we
note that a mean error of 15, associated with TDCO,
and COSTF, is extremely large and being compa-
rable to the actual number of breath per minutes,
in most cases, renders it unacceptable. We observe
that FoMs reflects these factors as well. In sum-
mary, as we can see in Table 5, the proposed method
(SFFDP), compared to other three existing methods
(TDPD, TDCO, and COSTF), has a better perfor-
mance (smallest FoM) and improves them by 37-53%.
This shows the superiority of the proposed method
compared to other existing ones.
5 Conclusion
In this paper, we proposed a new RR extraction al-
gorithm, SFFDP, which uses DF and SFU to extract
the RR. The proposed algorithm proved to be signifi-
cantly more reliable than existing algorithms despite
introduction of movement artifacts. The proposed al-
gorithm has an average error of only 3.1 BPM and a
STD of 3.2, while successfully calculating 75.6% sam-
ples. In summary, compared to others, it shows more
than 37% improvement in the FoM, which combines
the mean error, STD, and the percentage of success-
fully estimated samples.
5
Pre-Publication
References
[1] P. Parego, A. M. Rahmani, and N. Taherinejad.
Wireless Communication and Mobile Healthcare.
ecture Notes of the Institute for Computer Sci-
ences, Social Informatics and Telecommunica-
tions Engineering. Springer International Pub-
lishing, 2017.
[2] H. Yin et al. Smart healthcare. Foundations
and Trends R
in Electronic Design Automation,
2(14):401–466, 2018.
[3] Maximilian G¨otzinger et al. Enhancing the
early warning score system using data confi-
dence. In International Conference on Wireless
Mobile Communication and Healthcare, pages
91–99. Springer, 2016.
[4] Farnaz Forooghifar, Amir Aminifar, and David
Atienza Alonso. Self-aware wearable systems in
epileptic seizure detection. In Euromicro Con-
ference on Digital System Design (DSD), page 7,
2018.
[5] Maximilian G¨otzinger et al. Enhancing the self-
aware early warning score system through fuzzi-
fied data reliability assessment. In International
Conference on Wireless Mobile Communication
and Healthcare. Springer, 2017.
[6] Arman Anzanpour et al. Self-awareness in re-
mote health monitoring systems using wearable
electronics. In Proceedings of Design and Test
Europe Conference (DATE), Lausanne, Switzer-
land, March 2017.
[7] Nima TaheriNejad and David Pollreisz. Assess-
ment of physiological signals during happiness,
sadness, pain or anger. In 6th MobiHealth, 2016.
[8] D. Pollreisz and N. TaheriNejad. A simple algo-
rithm for emotion recognition, using physiologi-
cal signals of a smart watch. In 2017 39th Annual
International Conference of the IEEE Engineer-
ing in Medicine and Biology Society (EMBC),
pages 2353–2356, July 2017.
[9] N. TaheriNejad. Wearable medical devices:
Challenges and self-aware solutions. In IEEE
Life Sciences Newsletter, pages 5–6, 2019.
[10] Peter Charlton, Timothy Bonnici, Lionel
Tarassenko, Jordi Alastruey, and Clifton. Ex-
traction of respiratory signals from the electro-
cardiogram and photoplethysmogram: Techni-
cal and physiological determinants. Physiological
Measurement, 38(5):669–690, 3 2017. Creative
Commons Attribution license.
[11] Roxana Alexandra Cernat, Constantin Ungure-
anu, Georgeta Mihaela Ungureanu, Ronald
Aarts, and Johan Arends. Real-time extrac-
tion of the respiratory rate from photoplethys-
mographic signals using wearable devices. Euro-
pean Conference on Ambient Intelligence, 2014.
[12] David Pollreisz and Nima Taherinejad. Detec-
tion and removal of motion artifacts in PPG sig-
nals. Mobile Networks and Applications, Aug
2019.
[13] C. C Mayer. Pulse contour analysis model for
non-invasive determination of hemodynamical
parameters. Phd Thesis, Vienna University of
Technology, 2007.
[14] Axel Schaefer and Karl W. Kratky. Estimation
of breathing rate from respiratory sinus arrhyth-
mia: Comparison of various methods. Annals of
Biomedical Engineering, 36, 03 2008.
[15] Syed Ahmar Shah. Vital sign monitoring and
data fusion for paediatric triage. Phd Thesis,
Oxford University, UK., 2012.
[16] Jesus Lazaro et al. Deriving respiration from
photoplethysmographic pulse width. Medical
and Biological Engineering and Computing, vol.
51:pp.233–242, 02 2013.
6
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
With the rise of wearable devices, which integrate myriad of health-care and fitness procedures into daily life, a reliable method for measuring various bio-signals in a daily setup is more desired than ever. Many of these physiological parameters, such as Heart rate (HR) and Respiratory Rate (RR), are extracted indirectly and using other signals such as Photoplethysmograph (PPG). Part of the reason is that in some cases, such as RR measurements, the devices which directly measure them are cumbersome to wear and thus, rather impractical. On the other hand, signals, such as PPG from which the RR can be extracted, are not very clean. This poses a challenge on reliable extraction of these metrics. The most important problem is that they are corrupted by motion artifacts. In this paper, we review the state of the art algorithms which are used to detect and filter motion artifacts in PPG signals and compare them in terms of their performance. The insight provided by this paper can help the scientists and engineers to obtain a better understanding of the field and be able to use the most suitable technique for their work, or come up with innovative solutions based on existing ones.
Article
Full-text available
Objective: Breathing rate (BR) can be estimated by extracting respiratory signals from the electrocardiogram (ECG) or photoplethysmogram (PPG). The extracted respiratory signals may be influenced by several technical and physiological factors. In this study, our aim was to determine how technical and physiological factors influence the quality of respiratory signals. Approach: Using a variety of techniques 15 respiratory signals were extracted from the ECG, and 11 from PPG signals collected from 57 healthy subjects. The quality of each respiratory signal was assessed by calculating its correlation with a reference oral-nasal pressure respiratory signal using Pearson's correlation coefficient. Main results: Relevant results informing device design and clinical application were obtained. The results informing device design were: (i) seven out of 11 respiratory signals were of higher quality when extracted from finger PPG compared to ear PPG; (ii) laboratory equipment did not provide higher quality of respiratory signals than a clinical monitor; (iii) the ECG provided higher quality respiratory signals than the PPG; (iv) during downsampling of the ECG and PPG significant reductions in quality were first observed at sampling frequencies of <250 Hz and <16 Hz respectively. The results informing clinical application were: (i) frequency modulation-based respiratory signals were generally of lower quality in elderly subjects compared to young subjects; (ii) the qualities of 23 out of 26 respiratory signals were reduced at elevated BRs; (iii) there were no differences associated with gender. Significance: Recommendations based on the results are provided regarding device designs for BR estimation, and clinical applications. The dataset and code used in this study are publicly available.
Article
Full-text available
A method for deriving respiration from the pulse photoplethysmographic (PPG) signal is presented. This method is based on the pulse width variability (PWV), and it exploits the respiratory information present in the pulse wave velocity and dispersion. It allows to estimate respiration signal from only a pulse oximeter which is a cheap and comfortable sensor. Evaluation is performed over a database containing electrocardiogram (ECG), blood pressure (BP), PPG, and respiratory signals simultaneously recorded in 17 subjects during a tilt table test. Respiratory rate estimation error is computed obtaining of 1.27 ± 7.81 % (0.14 ± 14.78 mHz). For comparison purposes, we have also obtained a respiratory rate estimation from other known methods which involve ECG, BP, or also PPG signals. In addition, we have also combined respiratory information derived from different methods which involve only PPG signal, obtaining a respiratory rate error of -0.17 ± 6.67 % (-2.16 ± 12.69 mHz). The presented methods, PWV and combination of PPG derived respiration methods, avoid the need of ECG to derive respiration without degradation of the obtained estimates, so it is possible to have reliable respiration rate estimates from just the PPG signal. An erratum for this article can be found under doi: 10.1007/s11517-012-0997-2
Conference Paper
Recently, it has become easier and more common to measure physiological signals through wearable devices such as smart watches. Extracting emotional states of individuals with problems expressing it, such as autistic individuals, can help their parents, friends, and therapists to obtain a better understanding of what they feel throughout their day. Although emotion recognition methods based on physiological signals have been studied for many years, there is a smaller body of literature about systems working with data obtained from wearable devices. In this paper, we present an emotion recognition system with a small footprint suitable for limited resources of wearable devices. Other than identifying the emotions (with a success rate of 65%), The proposed system also tags each recognition with a confidence value (on average 57%).
Conference Paper
With the advancement of technology, non-intrusive monitoring of some physiological signals through smart watches and other wearable devices are made possible. This provides us with new opportunities of exploring newer fields of information technology applied in our everyday lives. One application which can help individuals with difficulty in expressing their emotions, e.g. autistic individuals, is emotion recognition through bio-signal processing. To develop such systems, however, a significant amount of measurement data is necessary to establish proper paradigms, which enable such analyses. Given the sparsity of the available data in the literature, specifically the ones using portable devices, we conducted a set of experiments to help in enriching the literature. In our experiments, we measured physiological signals of various subjects during four different emotional experiences; happiness, sadness, pain, and anger. Measured bio-signals are Electrodermal activity (EDA), Skin Temperature, and Heart rate. In this paper, we share our measurement results and our findings regarding their relation with happiness, sadness, anger, and pain.
Conference Paper
Early Warning Score (EWS) systems are utilized in hospitals by health-care professionals to interpret vital signals of patients. These scores are used to measure and predict amelioration or deterioration of patients’ health status to intervene in an appropriate manner when needed. Based on an earlier work presenting an automated Internet-of-Things based EWS system, we propose an architecture to analyze and enhance data reliability and consistency. In particular, we present a hierarchical agent-based data confidence evaluation system to detect erroneous or irrelevant vital signal measurements. In our extensive experiments, we demonstrate how our system offers a more robust EWS monitoring system.