Clinical Quality Guaranteed Physiological DataCompression in Mobile Health Monitoring

Sungwon Yang, Jihyoung Kim, and Mario Gerla

Department of Computer Science

University of California, Los Angeles

{swyang, jhkim, gerla}@cs.ucla.edu

ABSTRACTData compression is essential for continuously collected phys-iological signals in mobile health monitoring applicationsin order to prolong battery lifetime and reduce transmis-sion costs. Transformation-based compression techniqueshave been widely used due to their high compression ra-tio; however, distortion caused during compression processdegrades clinical quality of decompressed signals. In thispaper, we propose a simple method called “Critical Mark-ers” method that is based on detection of peaks and valleysin the original signal. When used in conjunction with ex-isting transformation-based compression methods, the crit-ical markers corrects the distortions without compromisingthe fidelity of the compressed output. The critical markerscan also be used standalone to replace existing compres-sion methods in certain types of diagnosis, thus reducingline and processor overhead. We have implemented the pro-posed method on a smartphone and have tested it with realECG and PPG data sets. The experimental results confirmthat our method maintains high compression performancewhile also guaranteeing high clinical quality.

Categories and Subject DescriptorsJ.3 [Life and Medical Sciences]: Medical information sys-tems

KeywordsPhysiological Signal Compression, ECG Compression, Clin-ical Compression Quality, Mobile Health Monitoring

1. INTRODUCTIONRecent advancements in miniaturization of portable med-

ical sensors have enabled continuous health monitoring indaily life. Over the past few years, many mobile health-care platforms have been proposed and most of them haveadopted the general architecture illustrated in Figure 1 [4,6,

Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.MobileHealth’12, June 11, 2012, Hilton Head Island, SC, USA.Copyright 2012 ACM 978-1-4503-1292-9/12/06 ...$10.00.

Gateway'

Raw Sensor Data

Body'Sensors'

Pre-processed Sensor Data/

Event

Internet Servers

Tier'1'BAN/PAN'

Tier'2'WLAN/GPRS'

Tier'3'Network'

Infrastructure'

Emergency Message

Analyzed Medical

Information Analyzed Medical

Information

Figure 1: Mobile Health Monitoring Systems

10,11,13,15–17,19]. Body-wearable sensors measure biomed-ical signals and transmit them wirelessly to a portable gate-way device such as a smartphone. The gathered data ispre-processed locally on the gateway and transferred to aremote monitoring center for further in-depth analysis.

Since physiological signals such as electrocardiograms(ECG), electroencephalograms (EEG), or photoplethysmo-graphs (PPG) are collected continuously at a high samplingrate, they tend to produce a large amount of data. This largevolume of data can be an obstacle to mobile health commer-cial implementations, since large data implies high batteryconsumption and high cellular fees for wireless transmissionto a remote monitoring center. Therefore, data compressionis very attractive in such systems.

Compression of physiological data has been extensivelystudied in the literature. Among many compression meth-ods, transformation based schemes are widely used due totheir high Compression Ratio (CR). In particular, the wavelettransform has attracted great attention because of its sim-plicity and versatility. Existing popular transformation-basedcompression schemes may achieve a high compression ratio,but they may not guarantee the quality requirement for ac-curate diagnosis because signals can be distorted during thecompression steps (e.g. transformation, thresholding, andquantization). Since a small change in physiological signalscan result in misdiagnoses, critical diagnostic informationmust be preserved during the compression procedure. Inthis paper, we propose a simple yet e↵ective method that

satisfies both high compression ratio and high clinical qual-ity requirements.

The proposed technique, which is based on the detec-tion of peak and valley points in signals, can be used aloneto directly compress the data or can also be used in com-bination with any existing transformation based compres-sion methods to correct the distorted signals. We conductperformance evaluation using a smartphone as a gatewaydevice in mobile monitoring systems. The experimental re-sults show that our method dramatically improves the clini-cal quality at the expense of a small decrease in compressionratio when used together with a transformation-based com-pression scheme. The proposed method also provides highquality of restored signals as well as adequate compressionperformance even when it is used alone.

2. RELATED WORKThis section briefly reviews previous works that focused

on clinical quality in biomedical data compression. Severalstudies proposed quality controlled compression schemes forECG data. Jie Chen et al. proposed an orthonormal wavelettransform based ECG compression scheme that employs anadaptive quantization strategy, by which a predeterminedPercent Root-Mean-Square Di↵erence (PRD) can be guar-anteed [3]. In the same context, Shaou-Gang Miaou etal. proposed an algorithm that searches for an appropri-ate bit rate in an automatic, smooth, and fast manner forthe wavelet-based compression to meet the given PRD re-quirement [14]. However, it is shown that determining theclinically acceptable numerical PRD value is not trivial [18].

Alvaro Alesanco et al. investigated the advantages anddrawbacks of PRD and Root Mean Square (RMS) in orderto guarantee quality in threshold wavelet-based compressionschemes that segment the signal into blocks [2]. The authorsfound that RMS could be a better indicator than PRD if thesignal presents many low energy blocks after segmentation;however, PRD threshold could lead to better clinical resultsthan RMS when the signal is placed inside a low energyblock. In spite of those findings, the paper concluded thatthere is no optimum choice between PRD and RMS whenone target the quality of biomedical signals.

There also were e↵orts to devise new distortion mea-suring methods for compressed physiological data from thepoint of view of diagnosis. Yaniv Zigel et al. introduceda distortion measure for ECG compression called WeightedDiagnostic Distortion (WDD) [20]. The WDD is based onPQRST complex diagnostic features (i.e. P wave duration,QT interval, T shape, and ST elevation) of the original ECGsignal and the reconstructed one. Unlike PRD or RMS,WDD contains direct diagnostic information and thus it isclinically a more meaningful indicator.

3. BIO-SIGNAL COMPRESSIONThis section explains why previous physiological signal

compression methods do not guarantee clinical quality andpresents how transformation-based compression schemes dis-tort the original signals, possibly resulting in misdiagnosesor making precise analysis impossible.

3.1 Transform-based compressionCompression of physiological data, such as ECG and

EEG, has been extensively studied in the literature. In gen-

eral, the compression schemes can be classified into two ma-jor categories: 1) direct data compression techniques whichtreat the sample signals directly in the time domain, and 2)transformation based approaches in which the samples aretransformed into another domain resulting in concentratinga large amount of energy into a small number of coe�cients.Review and comparison of some techniques in the two cate-gories are well presented in [9].

Transformation-based methods are widely used due totheir high CR. Among many transforms (e.g. Karhunen-loôeve Transform, Discrete Cosine Transform, Fast FourierTransform), the wavelet transform has attracted great at-tention because of its simplicity and versatility [5, 7, 8, 12].The coe�cients obtained from transformation process usu-ally consist of a small number of non-zero values and a largenumber of near-zero values, following power-law distribu-tion. In order to reduce the number of non-zero values andyield a higher compression ratio, a certain number of near-zero coe�cients are suppressed. This thresholding processmainly causes distortion in the original signal. Since the dis-tortion can impact any part of the data, clinically importantdata is not guaranteed to be preserved.

3.2 Compression accuracy assessmentRMS and PRD are commonly used for measurement of

signal distortion, which is calculated as defined in (1) and(2), respectively.

RMS =

sPN

n=1

(x[n]� x[n])2

N

(1)

PRD =

sPN

n=1

(x[n]� x[n])2P

N

n=1

x

2[n]⇥ 100(%) (2)

where x [n] is the original signal, x[n] is the reconstructedsignal, and N is the number of samples.

Both RMS and PRD are usually viewed as good indica-tors of the accuracy of the reconstructed signal. However,neither guarantees the accuracy requires in clinical studies.Namely, these indicators equally weigh the reconstructionerror in all portions of the signal; thus, they cannot expressthe quality of the parts of the signal that are diagnosticallysalient.

The examples in Figure 2(a) and (b) show that PRD doesnot always accurately reflect clinical quality. We took sam-ple ECG segments from the MIT-BIH database [1] (recordnumber: 112) and compressed the samples using Fast WaveletTransform (FWT) with uniform quantization on an HTCIncredible smartphone. We then used the Lempel-Ziv andHaruyasu method (LZH) to actually compress the coe�-cients. The resulting CR was around 11. Figure 2(a) showsone portion of the original and the reconstructed signals.The PRD of this portion is 3.1934. Figure 2(b) shows an-other portion of original and reconstructed signals of whichthe PRD is 3.1308. Based on PRD, the reconstructed signalin 2(b) is more accurate than one in 2(a). However, fromcardiologists’ perspectives, the signal in 2(b) is more dis-torted than 2(a). This is because the P and T waves (markedwith A and B) are distorted, leading to possible misdiagnosisof a 2:1 type heart block or of a Premature Atrial Contrac-tion (PAC). The S-T segment is also distorted (marked withC), and can be a obstacle to accurate diagnosis.

0 100 200 300 400 500 600 700-1.5

-1

-0.5

0A

mpl

itude

(mV

)

Original ECG Signal

0 100 200 300 400 500 600 700-1.5

-1

-0.5

0

Am

plitu

de (m

V)

Sample Number

Reconstructed ECG Signal (PRD=3.1934)

(a) PRD=3.1934 (b) PRD=3.1308

Figure 2: Original and Reconstructed ECG

4. CRITICAL MARKERSWhen it comes to clinical diagnosis of physiological sig-

nals, in general, three factors are taken into account: 1)interval between particular points, 2) magnitude of signals,and 3) contour-shape of signals. Transformation-based com-pression techniques may well preserve intervals of waves;however, they do not guarantee preservation of magnitudesof peaks and detailed contour shapes. This section proposesa simple yet e↵ective method that guarantees the clinicalquality of compressed physiological data. The proposedmethod e↵ectively preserve the three factors above. Ourscheme can be used alone to achieve high CR or it can becombined with any existing transformation-based compres-sion methods to provide perfect clinical accuracy.

4.1 Clinical Quality Improvement by CriticalMarkers

Peaks and valleys are the most important points in phys-iological signals, which determine magnitudes and contourshapes of signals. Unfortunately, the most distortion is likelyto occur at peak/valley points during transformation-basedcompression. Correctly recorded peak and valley values cancorrect records of magnitudes and enable accurate calcula-tion of intervals or drawing of contour shapes. Therefore,the quality of signals can be greatly improved if peak/valleypoints are well preserved in the compressed data.

One well-known peak detection method is using first orsecond derivatives; however, it is known to yield many falsepositives due to noise. Other methods, based on transfor-mations or low-pass filters, can also make the signal di↵erentfrom the original shape. To guarantee the accuracy of themedical diagnosis, we propose a simple algorithm that ex-tracts peaks and valleys from the original raw signal in realtime. The main idea is based on the simple fact that a peakfollows steadily increasing values and precedes steadily de-creasing values. For a valley, the opposite applies.

Figure 3 depicts our peak/valley detection algorithm.First, the algorithm sets two window sizes, W

out

and W

in

.W

out

involves in determining if a certain point is a peak/valleyandW

in

involves in determining if a certain point is on an in-creasing/decreasing line. Then, a slope vector SV is createdat each point as signal samples arrive. The size of vectorSV is W

out

� 1. Inside W

out

, VLT

[n] and V

RT

[n] (0nN,N = bW

out

2

c) are calculated as follows:

V

LT (RT )

[n] =

n+

Win2X

n�Win2

x

i

(3)

out

Win

VLT[N]VLT[N-1]

VLT[1]VLT[0]

...

...

...

...

Win

Slope Vector

(SV)

VRT[N]VRT[N-1]

VRT[1]VRT[0]

SVLT SVRT

Figure 3: Peak Detection Algorithm

Once V

LT

[n] and V

RT

[n] are obtained, SV is filled. Eachvector value is a 1-bit flag, which is determined as follows:

SV

LT (RT )

[n] =

(1 if V

LT (RT )

[n] V

LT (RT )

[n+ 1]

0 otherwise

(4)At each point x, if the elements of SV

LT

are all 1s and theelements of SV

RT

are all 0s, x is marked with a peak. Simi-larly, if the elements of SV

LT

are all 0s and the elements ofSV

RT

are all 1s, then it is marked with a valley. Otherwise,it is a point on an increasing or decreasing line. Since thisdouble-window mechanism works as a linear smoothing fil-ter, it naturally drops false-positives caused by noise. Thefiltering level is adjusted based on the two window sizes.

4.2 Peak/Valley detection accuracyTo evaluate the accuracy of our peak/valley detection

algorithm, we tested it with ECG and PPG signals. Werandomly picked 20 sample ECG segments from the MIT-BIH database and also took 20 sample PPG data sets fromthe authors. First, we recorded clinically important peaksand valleys in the samples by visual inspection, and then, wecompared those points with ones detected by the algorithm.Since the proposed algorithm has two parameters, W

out

andW

in

, we measured the performance as functions of them.Figure 4 shows how accurately the algorithm detects

peaks and valleys for the both signal samples. In case ofECG signals, for two combinations (i.e. (W

out

: 3, Win

: 1)and (W

out

: 7, Win

: 3)), the algorithm perfectly detects thedesignated peaks and valleys. It provides over 90% accu-racy in cases that W

out

is under 9. For all Wout

values, theaccuracy improves as W

in

also increases, reaching 97-98%when W

out

is 9 and 11. Although larger W

in

improves theaccuracy, the accuracy dramatically degrades once W

out

ex-ceeds 13; it never reaches 90%. For PPG signals, (W

out

: 3,W

in

: 1) and (Wout

: 5, Win

: 1) pairs achieve 100% detectionaccuracy. For other combinations, a similar tendency to theECG cases is seen, accuracy degradation over larger W

out

.Small W

out

and W

in

(e.g W

out

: 3 and W

in

: 1) per-fectly catch peaks and valleys; however, as expected theyalso cause false-positives. There should be many “ripples”in the raw signal due to noise, especially in ECG signals.They are technically peaks and valleys, but they do not

Win

1 3 5 7

Wout

3 100 (100) - - -5 96.90 (100) - - -7 93.80 (86.3) 100 (86.37) - -9 91.20 (81.81) 97.93 (84.09) - -11 69.43 (77.27) 92.73 (75) 96.37 (79.54) -13 47.18 (70.45) 77.72 (75.02) 80.31 (77.67) -15 32.14 (70.05) 53.36 (74.56) 60.13 (77.45) 69.43 (78.58)

Table 3: Accuracy of Peak/Valley Detection for ECG (PPG) (%)

Win

1 3 5 7

Wout

3 21.15 (14.16) - - -5 10.38 (8.96) - - -7 8.30 (7.12) 10.97 (7.82) - -9 6.39 (7.02) 8.43 (7.41) - -11 2.86 (6.43) 5.32 (6.43) 6.05 (7.03) -13 1.64 (6.05) 3.28 (6.45) 3.15 (6.64) -15 1.02 (5.84) 1.82 (6.44) 1.96 (6.61) 2.21 (6.82)

Table 4: Data Size of Peaks/Valleys for ECG (PPG) (%)

This is further discussed in Section 7.7.

7.5. Time complexity

The proposed algorithm uses only addition/subtraction operations and itdoes not require any iterations. Therefore, it is fast and its complexity is com-pletely linear. Figure 10 presents the processing time, captured on an HTCIncredible smart phone, for 5 (Wout, Win) pairs that provide more than 95%accuracy as a function of data size in case of ECG signals. The processing timeincreases linearly as the data size grows. Thus, the proposed algorithm workswell in real time on any modern mobile phones and even on low-end embeddedgateway devices.

7.6. Wavelet-based compression with critical markers

Once a certain number of critical markers is extracted from the raw signalbefore transformation-based compression, they are transmitted together withthe compressed data to a remote center. Then they can be displayed over thedecompressed signal to visually correct errors due to compression procedures.Alternatively, the distorted values can simply be replaced with the critical mark-ers to plot calibrated signals. The proposed approach inevitably decreases CRas it uses additional information. However, the increments of data size maytake second place to the improved signal quality. Our method extracts critical

19

Figure 4: Detection Accuracy for ECG (PPG) (%)

Win

1 3 5 7

Wout

3 100 (100) - - -5 96.90 (100) - - -7 93.80 (86.3) 100 (86.37) - -9 91.20 (81.81) 97.93 (84.09) - -11 69.43 (77.27) 92.73 (75) 96.37 (79.54) -13 47.18 (70.45) 77.72 (75.02) 80.31 (77.67) -15 32.14 (70.05) 53.36 (74.56) 60.13 (77.45) 69.43 (78.58)

Table 3: Accuracy of Peak/Valley Detection for ECG (PPG) (%)

Win

1 3 5 7

Wout

3 21.15 (14.16) - - -5 10.38 (8.96) - - -7 8.30 (7.12) 10.97 (7.82) - -9 6.39 (7.02) 8.43 (7.41) - -11 2.86 (6.43) 5.32 (6.43) 6.05 (7.03) -13 1.64 (6.05) 3.28 (6.45) 3.15 (6.64) -15 1.02 (5.84) 1.82 (6.44) 1.96 (6.61) 2.21 (6.82)

Table 4: Data Size of Peaks/Valleys for ECG (PPG) (%)

This is further discussed in Section 7.7.

7.5. Time complexity

The proposed algorithm uses only addition/subtraction operations and itdoes not require any iterations. Therefore, it is fast and its complexity is com-pletely linear. Figure 10 presents the processing time, captured on an HTCIncredible smart phone, for 5 (Wout, Win) pairs that provide more than 95%accuracy as a function of data size in case of ECG signals. The processing timeincreases linearly as the data size grows. Thus, the proposed algorithm workswell in real time on any modern mobile phones and even on low-end embeddedgateway devices.

7.6. Wavelet-based compression with critical markers

Once a certain number of critical markers is extracted from the raw signalbefore transformation-based compression, they are transmitted together withthe compressed data to a remote center. Then they can be displayed over thedecompressed signal to visually correct errors due to compression procedures.Alternatively, the distorted values can simply be replaced with the critical mark-ers to plot calibrated signals. The proposed approach inevitably decreases CRas it uses additional information. However, the increments of data size maytake second place to the improved signal quality. Our method extracts critical

19

Figure 5: Data Size for ECG (PPG) (%)

have to be considered as peaks and valleys as they are usu-ally generated by noise. Figure 5 shows how many pointsare detected as peaks/valleys. The numbers in the tablerepresent the number of detected peaks/valleys against thesize of sample data in percentage terms. Although (W

out

: 3and W

in

: 1) combination never misses major peaks/valleys,it also detects many minor peaks/valleys. Over 20% and14% of points are classified into peaks/valleys with theseparameter settings for the ECG and the PPG cases, re-spectively. It is because small W

out

and W

in

cannot fil-ter out noise e↵ectively. As the windows increase, thoseminor peaks/valleys are filtered out as the ripples are aver-aged over longer intervals; thus, false positives decrease. ForECG samples, when W

out

is 7 and W

in

is 3, the accuracyof detecting peaks/valleys and the number of false positivesbecomes balanced. The balanced window sizes are W

out

:5and W

in

:1 for PPG samples. We call these balanced pointsCritical Markers. The appropriate values for the two win-dows need to be determined adaptively depending on thesampling rates of signals. We found that the optimum win-dow pair value for signals with 100-400Hz sampling rate isobtained as follows: (W

in

,W

out

) = bSamples per second

(100,50)

c (if

the value is even number, add one). The sampling rates forthe ECG and PPG samples are 360Hz and 100Hz, respec-tively. The exact data size of peaks/valleys checked by visualinspection was approximately 4% for the both cases. Ourmethod does not filter all minor peaks/valleys. However,those minor peaks/valleys are also parts of the original sig-nal. Therefore, those unfiltered points never degrade clinicalquality. Moreover, those unfiltered points contribute greatlyto the compression quality when the algorithm is used aloneto compress signals. This is further discussed in Section 5.2.

4.3 Time complexityThe proposed algorithm uses only addition/subtraction

operations and it does not require any iterations. Therefore,it is fast and its complexity is completely linear. Figure 6presents the processing time, captured on an HTC Incrediblesmart phone, for 5 (W

out

, Win

) pairs that provide more than95% accuracy as a function of data size. The processing time

2 4 6 8 100

100

200

300

400

500

600

700

Data Size (KB)

Ela

psed

Tim

e (m

sec)

Time Complexity of Proposed Algorithm

Wout:3 Win:1Wout:5 Win:1Wout:7 Win:3Wout:9 Win:3Wout:11 Win:5

Figure 6: Time Complexity of the Algorithm

increases linearly as the data size grows. Thus, the proposedalgorithm works well in real time on any modern mobilephones and even on low-end embedded gateway devices.

5. EVALUATIONThis section presents how well the proposed scheme cor-

rects distorted ECG and PPG signals. We also show thatour scheme works well alone to compress physiological datain terms of the quality for clinical diagnosis.

5.1 Compression with critical markersOnce a certain number of critical markers is extracted,

the markers are transmitted together with the compresseddata to a remote center. Then they can be displayed overthe decompressed signal to visually correct errors. Alterna-tively, the distorted values can simply be replaced with thecritical markers to plot calibrated signals. The proposedapproach inevitably decreases CR as it uses additional in-formation. However, the increments of data size may takesecond place to the improved signal quality. As shown inFigure 5, for example, W

out

: 7 and W

in

: 3 settings for ECGrequire 10% of the data size against the original sample data,and it decreases CR from 11.263 to 7.714 when combinedwith Wavelet transformation-based compression method.

Figure 7 presents two examples of how our method im-proves the compressed ECG and PPG signals from the clin-ical perspective. Figure 7(a) and 7(b) are parts of the rawECG and PPG signals with critical markers detected by thealgorithm. At every sharp peak/valley, exactly one point isaccurately picked. On the other hand, at relatively bluntpoints, more than one point is selected. These multiplemarkers at minor peaks/valleys slightly increase the com-pressed data size. However, they never obstruct clinical di-agnosis, moreover they improve the clinical quality. Fig-ure 7(c) and 7(d) show how the proposed method can pre-vent misdiagnosis. In ECG case, at points A and B, a peakthat does not exist in the original signal is generated duringthe Wavelet transformation-based compression. However,with the aid of the critical markers, (i.e. there are no mark-ers on the peak), physician/cardiologist can know this peakis an error. For another example, at points C and D, thereconstructed signal flattened valleys so it made T and Pwaves unclear. However, one can clearly infer not only thata valley does exist at that area but also its exact value due tothe critical markers. The final example is found at point E.The reconstructed valley at point E is located significantlyhigher than the original point; this error is corrected by thecritical marker. Similarly, at F, in PPG case, valley points

(a) Original ECG Signal with Critical Markers100 200 300 400 500 600 700 800 900 1000

0

0.5

1

1.5

2

2.5

Am

plitu

de

Original PPG Signal

100 200 300 400 500 600 700 800 900 10000

0.5

1

1.5

2

2.5

Sample Number

Am

plitu

de

Reconstructed PPG Signal(b) Original PPG Signal with Critical Markers

(c) Reconstructed Signal (CR: 11.263) (d) Reconstructed Signal (CR: 12.031)

Figure 7: Visualization of ECG and PPG Signals with Critical Markers (Wout

: 7 and W

in

: 3)

can be corrected by the markers. PPG signals in the circlemarked with G would be wrongly classified due to notchescaused by distortion during the Wavelet based compression.The critical markers prevents this misdiagnosis.

5.2 Compression using critical markers onlyIn the previous sub-section, critical markers are used

to help physicians to analyze the compressed signals cor-rectly. The proposed algorithm does not filter out all minorpeaks/valleys. Thus, it can be a drawback as it increasesthe compression data size. However, we can take advantageof the non-filtered minor markers under circumstance thatthe marker information is solely used to compress physio-logical signals. Both the major and minor markers can beused to compress the raw signal directly, which improvesclinical quality. Since peaks and valleys are the key factorsfor clinical diagnosis, accurately extracted peaks/valleys al-ready involve most significant information. In addition, theminor peaks/valleys provide information for more accuratesignal reconstruction. Figure 8 presents examples showinghow well the restored signals, using only the critical mark-ers, represent the original signal. The solid lines representthe original signals and the dotted lines represent the recon-structed signals.

Figure 8(a) and 8(b) presents the reconstructed ECGand PPG signals with (W

out

: 3, Win

: 1) pair, which pro-vides 100% of major peak/valley detection and some minormarkers. The restored signal is so close to the original sig-nal; thus, one may not visually find remarkable di↵erencesbetween them. The PRDs for both examples are 5.2563and 9.4321, which are worse than the compressed signalby Wavelet transformation (i.e. PRD: 3.8433 and 4.1724).However, the proposed method provides more accurate clin-ical information because it never distorts peaks/valleys, un-like the transformation-based compression. As we changethe parameters to (W

out

: 7, Win

: 3) pair for ECG and (Wout

:5, W

in

: 1) for PPG, 100% of major markers and less minormarkers are detected, resulting in less accuracy in the sig-

nal reconstruction. The restored signals are presented inFigure 8(c) and 8(d). The PRDs are degraded into 7.0488and 14.1231, and few visually-discoverable errors are found.However, they still provide su�cient and accurate clinicalinformation. The CR of our method depends on both the(W

out

, Win

) parameter sets and the types of signals. In thecase of ECG data from the MIT-BIH database, our schemeprovides approximately 14.32 CR (W

out

: 7, W

in

: 3) and8.73 CR (W

out

: 3, W

in

: 1) after the critical markers arecompressed by LZH technique. For our PPG sample data,the CR is 15.61 with (W

out

: 5, Win

: 1) pair and 8.94 with(W

out

: 3, Win

: 1) setting.

6. CONCLUSIONData compression of physiological signals is essential in

mobile health monitoring systems in order to reduce batteryconsumption and wireless transmission overhead. Althoughmany compression schemes have been studied over decades,few of them take clinically important reconstruction qual-ity of the compressed data into account. This is becausemost previous works used PRD or RMS as indicators ofthe compression accuracy. This paper shows that PRD andRMS are inappropriate for physiological signals. It then pro-poses a simple yet very e↵ective compression technique thatis based on detection of clinically important points in con-tinuously collected physiological signals. Experiments con-ducted on a smartphone platform using real ECG and PPGdata sets confirm that the proposed compression methodprovides clinically satisfactory signal reconstruction as wellas high compression ratio. As a future work, we are workingon the development of an interpolation model that providesmore realistic physiological signal reconstruction.

7. REFERENCES[1] http:

//www.physionet.org/physiobank/database/mitdb/.[2] A. Alesanco, J. Garcia, P. Serrano, L. Ramos, and

R. Istepanian. On the guarantee of reconstructionquality in ecg wavelet codecs. In EMBS ’06.

200 400 600 800 1000 1200 1400

-1

-0.5

0A

mpl

itude

(mV

)

Wout:3, Win:1 (PRD=5.2563)

Reconstructed SignalOriginal Signal

200 400 600 800 1000 1200 1400

-1

-0.5

0

Am

plitu

de (m

V)

Sample Number

Wout:7, Win:3 (PRD=7.0488)

Reconstructed SignalOriginal Signal

(a) ECG Signal, Wout

: 3 and W

in

: 1100 200 300 400 500 600 700 800 900 1000

0

0.5

1

1.5

2

2.5

Am

plitu

de

Wout:3, Win:1 (PRD=9.4321)

Reconstructed SignalOriginal Signal

100 200 300 400 500 600 700 800 900 10000

0.5

1

1.5

2

2.5

Sample Number

Am

plitu

de

Wout:5, Win:1 (PRD=14.1231)

Reconstructed SignalOriginal Signal

(b) PPG Signal, Wout

: 3 and W

in

: 1200 400 600 800 1000 1200 1400

-1

-0.5

0

Am

plitu

de (m

V)

Wout:3, Win:1 (PRD=5.2563)

Reconstructed SignalOriginal Signal

200 400 600 800 1000 1200 1400

-1

-0.5

0

Am

plitu

de (m

V)

Sample Number

Wout:7, Win:3 (PRD=7.0488)

Reconstructed SignalOriginal Signal

(c) ECG Signal, Wout

: 7 and W

in

: 3

100 200 300 400 500 600 700 800 900 10000

0.5

1

1.5

2

2.5

Am

plitu

de

Wout:3, Win:1 (PRD=9.4321)

Reconstructed SignalOriginal Signal

100 200 300 400 500 600 700 800 900 10000

0.5

1

1.5

2

2.5

Sample Number

Am

plitu

de

Wout:5, Win:1 (PRD=14.1231)

Reconstructed SignalOriginal Signal

(d) PPG Signal, Wout

: 5 and W

in

: 1

Figure 8: Signal Reconstruction with Critical Markers in cooperation with Hermite Interpolation

[3] J. Chen and S. Itoh. A wavelet transform-based ecgcompression method guaranteeing desired signalquality. Biomedical Engineering, IEEE Transactionson, 45(12):1414 –1419, dec. 1998.

[4] M. Chen, S. Gonzalez, A. Vasilakos, H. Cao, andV. C. Leung. Body area networks: A survey. Mob.Netw. Appl., 16(2):171–193, Apr. 2011.

[5] C. K. Chui. An introduction to wavelets. AcademicPress Professional, Inc., San Diego, CA, USA, 1992.

[6] V. Gay and P. Leijdekkers. A health monitoringsystem using smart phones and wearable sensors.International Journal of ARM, 8(2), June 2007.

[7] M. Hilton. Wavelet and wavelet packet compression ofelectrocardiograms. Biomedical Engineering, IEEETransactions on, 44(5):394 –402, may. 1997.

[8] C.-B. J, L.-G. J, and R.-V. E. A wavelet-packets basedalgorithm for eeg signal compression. Medicalinformatics and the Internet in medicine, 29(1):15–27,March 2004.

[9] S. Jalaleddine, C. Hutchens, R. Strattan, andW. Coberly. Ecg data compression techniques-aunified approach. Biomedical Engineering, IEEETransactions on, 37(4):329 –343, apr. 1990.

[10] S. Lam, K. L. Wong, K. O. Wong, W. Wong, andW. H. Mow. A smartphone-centric platform forpersonal health monitoring using wireless wearablebiosensors.

[11] P. Leijdekkers and V. Gay. Personal heart monitoringsystem using smart phones to detect life threateningarrhythmias. Computer-Based Medical Systems, IEEESymposium on, 0:157–164, 2006.

[12] Z. Lu, D. Y. Kim, and W. Pearlman. Waveletcompression of ecg signals by the set partitioning inhierarchical trees algorithm. Biomedical Engineering,IEEE Transactions on, 47(7):849 –856, jul. 2000.

[13] R. Magjarevic, O. Medvedev, A. Kobelev, S. Schookin,M. Jatskovsky, G. Markarian, and I. Sergeev.Smartphone-based approach for monitoring vitalphysiological parameters in humans. In R. Magjarevicand J. H. Nagel, editors, World Congress on MedicalPhysics and Biomedical Engineering 2006, IFMBEProceedings. Springer Berlin Heidelberg.

[14] S.-G. Miaou and C.-L. Lin. A quality-on-demandalgorithm for wavelet-based compression of

electrocardiogram signals. Biomedical Engineering,IEEE Transactions on, 49(3):233 –239, mar. 2002.

[15] C. Otto, A. Milenkovi, C. Sanders, and E. Jovanov.System architecture of a wireless body sensor networkfor ubiquitous health monitoring. In Journal of MobileMultimedia, 2006.

[16] S. Yang and M. Gerla. Energy-e�cient accelerometerdata transfer for human body movement studies. InSensor Networks, Ubiquitous, and TrustworthyComputing (SUTC), 2010 IEEE InternationalConference on, pages 304 –311, june 2010.

[17] S. Yang and M. Gerla. Personal gateway in mobilehealth monitoring. In Pervasive Computing andCommunications Workshops (PERCOM Workshops),2011 IEEE International Conference on, pages 636–641, march 2011.

[18] S. K. Yoo, K. Lee, and M. H. Lee. Empiricaldetermination of an ecg compression ratio for mobiletelecardiology applications. Telemed J E Health,14(2):156–63, 2008.

[19] L. Zhong, M. Sinclair, and R. Bittner. Aphone-centered body sensor network platform: Cost,energy e�ciency & user interface. In BSN ’06,Washington, DC, USA. IEEE Computer Society.

[20] Y. Zigel, A. Cohen, and A. Katz. The weighteddiagnostic distortion (wdd) measure for ecg signalcompression. Biomedical Engineering, IEEETransactions on, 47(11):1422 –1430, nov. 2000.