heuristic algorithm for photoplethysmography heart rate ... · maximal exercise test56 (imet) on...
Post on 28-Aug-2020
0 Views
Preview:
TRANSCRIPT
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
1
Heuristic Algorithm for Photoplethysmography Heart 1
Rate Tracking during Athletes Maximal Exercise Test 2
3
Sonnia María López-Silva1 Romano Giannetti
2* María Luisa Dotor
1 Juan Pedro Silveira
1 4
Dolores Golmayo3 Francisco Miguel-Tobal
4 Amaya Bilbao
4 Mercedes Galindo
4 Pilar Martín-Escudero
4 5
6
1 Instituto de Microelectrónica de Madrid, CNM-CSIC, C/ Isaac Newton 8, Tres Cantos, Madrid 28760, 7
Spain. 8
2 Departamento de Electrónica y Automática, Universidad Pontificia Comillas, C/ Alberto Aguilera 23, 9
Madrid 28015, Spain. 10
3 Instituto de Ciencias de Materiales de Madrid, CSIC, C/ Sor Juana Inés de la Cruz 3, Madrid 28049, Spain. 11
4 Escuela Profesional de Medicina de la Educación Física y el Deporte, Facultad de Medicina, Universidad 12
Complutense de Madrid, Madrid 28040, Spain. 13
* Corresponding author: Prof. Romano Giannetti 14
Vice-Rector for International Relations 15
Universidad Pontificia Comillas 16
C/ Alberto Aguilera 23, 28015 Madrid (Spain)
17
Tel +34-915422800; Fax +34-915411542 18
E-mail: romano@dea.icai.upcomillas.es
19
20
RUNNING TITLE: PPG Heart Rate Tracking in Athletes21
22
23
* Corresponding author: Romano Giannetti E-mail: romano@dea.icai.upcomillas.es
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
2
Abstract 24
Photoplethysmography (PPG) is a non-invasive optical technique that enables to quantify the arterial 25
blood pulse rate. Signal corruption by motion artifacts is a primary limitation in the practical accuracy and 26
applicability of instruments for monitoring pulse rate during intense physical exercise. Here we develop and 27
validate an algorithm, based on linear filtering, frequency domain and heuristic analysis, capable to extract 28
and display the heart rate from PPG signal in the presence of severe motion artifacts. The basis of the 29
frequency selection is the observed high harmonic content of movement artifact signals with respect to the 30
PPG-derived heartbeat. The algorithm has been implemented in an experimental PPG measurement device 31
and has been developed by analyzing a set of PPG data recorded from a first group of athletes exercise in 32
treadmill. After that, an extensive set of tests has been carried out on real time during maximal exercise tests 33
in treadmill in order to validate the instrument by comparison with a reference electrocardiography 34
measurement system. We have used the Bland and Altman method to compare and evaluate the PPG signal. 35
The accuracy of the heartbeat measurement is better than ± 6.5 bpm (≤ 4.2 %) even under maximal exercise 36
conditions. 37
Keywords: Biomedical signal processing, Photoplethysmography (PPG), Motion artifact, Detection 38
algorithm, Heart rate 39
40
41
42
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
3
43
1. Introduction 44 45
Photoplethysmography (PPG) is a technique offering a simple, useful and compact way to measure several 46
clinical parameters such as oxygen saturation, blood pressure and cardiac output [1,2]. PPG is especially 47
suitable for wearable sensing, that could play an important role in many areas ranging from personal health 48
monitoring to sports medicine. Despite the attractive attributes of PPG and the ease of its integration into 49
wearable devices, the PPG signal is known to be fragile and easily modifiable by motion. In most cases, the 50
noise falls within the same frequency band as the physiological signal of interest, rendering linear filtering 51
ineffective. Research carried on motion artifacts in measured PPG signals has reduced the sensitivity to the 52
artifacts common in the clinical environments, as well as in situations of controlled or moderate motion [1-53
8]. A non-contact, laser emitter remote PPG system, has been proposed [9] and positively evaluated for 54
clinical applications. PPG has been also applied to the evaluation of more severe exercise during incremental 55
maximal exercise test (IMET) on cycle-ergometer, where several parameters are monitored, among them 56
oxygen saturation [10,11] and heart rate [12-15]. Active research efforts are beginning to demonstrate a 57
utility beyond oxygen saturation and heart rate (HR) determination; for instance, the conditions for a correct 58
utilization of PPG HR variability (PPGV) have been recently studied [14] and the authors concluded that 59
PPGV cannot be used during slow walk and cycling exercise. Future trends are being heavily influenced by 60
modern digital signal processing [16]. New commercial developments from the PPG-based prototype [17] 61
support a new interest in the possibilities that PPG technique might offer for HR tracking. 62
Under severe exercise conditions, motion artifacts present one of the most challenging problems for PPG 63
analysis. Design of sensor and packaging can help to reduce the impact of motion disturbances, but it is 64
rarely sufficient for noise removal. Advanced signal processing techniques are often required to discriminate 65
motion artifacts under vigorous activity. Several techniques have been investigated to deal with those 66
artifacts, such as processing context information by additional on-body sensors and light sources [18,19] or 67
adaptive noise cancellation using accelerometers as a noise reference [20]. Very recently, heart rate 68
measurements using a PPG sensor integrated with an adaptive noise cancellation device has been 69
demonstrated at common physical activity, from walking to running up to 8 Km/h [21]. 70
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
4
Spectral analysis, as traditional fast Fourier transform (FFT), is a simple and inexpensive useful tool for 71
separating motion artifact and cardiac physiologic spectra. However, techniques based on spectral analysis 72
will not be applicable to spectra that contain frequency bands close to each other. More sophisticated 73
algorithms have shown significant improvement over FFT for measurements obtained during finger bend 74
manoeuvres [2]. Advanced signal processing techniques can help to overcome this problem. The aim of this 75
work is to present and validate an innovative algorithm able for tracking heart rate from PPG signal during 76
athletes IMET on treadmill. The real-time motion discriminator algorithm, based on linear filtering, 77
frequency domain and heuristic analysis, allows us to extract the heart rate value from PPG signal in the 78
presence of severe motion artifacts. The basis of the frequency selection is the observed high harmonic 79
content of movement artifact signals with respect to the PPG-derived heartbeat. A preliminary version has 80
been already presented, with off-line analysis both of heart rate and oxygen saturation [22]; this paper provides 81
an optimization of the algorithm parameters for the heart rate tracking, as well as real time probes and 82
comparison of the results to the values obtained by a standard technique. The physiological aspects of training 83
are out of the scope of this work. The custom PPG measurement system is based on a transmittance pulse 84
oximetry system [23]. 85
The organization of this paper is as follows. Section 2 describes the materials and methods. Section 3 is 86
devoted to the analysis of the elements for the algorithm design through the analysis of PPG recorded data 87
obtained from a first group of athletes during treadmill IMET. Section 4 is devoted to de development of the 88
algorithm. Section 5 is devoted to present the results of validation of the algorithm both on recorded data of 89
a group T1 of athletes and on real time heart rate tracking during IMET on treadmill of a second group T2 of 90
athletes; we have assessed the agreement between heart rate measured by the new and the reference methods, 91
electrocardiography (ECG), by means of Bland and Altman analysis [24,25] obtaining PPG data that lie 92
within ± 6.5 bpm (beats per minute) with respect to the ECG data. In section 6, we present the discussion of 93
the work. 94
95
2. Materials and Methods 96
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
5
Athletes of both sex genders with a training schedule exceeding an average of 7 - 8 hours per week 97
participated in IMET performed on treadmill during separate sessions. Data presented and analyzed in this 98
work correspond only to the heart rate measurements; others like oxygen consumption, carbon dioxide 99
production and blood pressure are out of the scope of the present work. The studies were conducted at the 100
laboratories of effort physiology of the Professional School of Physical Education of Complutense 101
University, Madrid (Spain), after approval by the local Research Ethical Committee. All the athletes gave 102
their written informed consent. 103
104
2.1. Subjects 105
A first group T1 was composed by 10 white males, endurance runners tested on the treadmill ergometer. A 106
second group T2 includes 20 athletes tested on the treadmill: 9 white females (one runner, 8 football players), 107
9 white male runners and 2 black female basketball players. The development of the algorithm and 108
optimization of its parameters was carried out after data recorded on the first group T1. The algorithm 109
performance has been evaluated on real time during IMET of the independent group of athletes T2. 110
111
2.2. Experimental protocol 112
Heart rate was monitored by electrocardiography (ECG, Burdick, Inc, Model Quest Exercise Stress 113
System) and used as reference. The PPG sensor was placed on one finger with special care to avoid 114
excessive compression of the tissues. The arms movement was not restricted. 115
Next, a complete maximal exercise test was performed, followed by active recovery. The protocol [26] for 116
the treadmill ergometer (HP Cosmos QUASAR 4.0) test established that after one minute at rest (in order to 117
record basal values) and after warming up, the athlete began to walk at 6 Km/h and 1 % slope during 2 min. 118
The athlete then started the effort phase running at 8 Km/h and 1 % slope. During the effort phase, the speed 119
was increased every 2 min by 2 Km/h. The maximum speed achieved varied among individuals. When the 120
speed of 14 Km/h (for female athletes) or 16 Km/h (for male athletes) was reached, the slope was increased 121
to 3 %. Afterwards, the slope was maintained constant, while speed was increased every 2 min by 2 Km/h, 122
until they were unable to continue, then the athlete held the protective bars and jumped off the treadmill. 123
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
6
Active recovery was performed during 2 min at 8 Km/h with a slope of 0 %. At the beginning of the exercise, 124
when walking at 6 Km/h, most of the athletes were adapting to walk on the treadmill and in many cases they 125
took hold of protection bars, which could alter the position of the finger sensor. This will be discussed later. 126
The ECG system computed the heart rate every 10 s, averaging the last eight heart beats. At different stages 127
of the test, once the athlete was running at fixed speed and slope, the full-step rate (SR, in Steps/min) was 128
obtained by counting manually the number of steps in 10 s interval, and the one-foot step was derived from 129
that. When running, steps and waving of the arms are synchronous; hereafter, we will note SR both for the 130
full step and the full arm-waving, which are inseparable issues. 131
132
2.3. PPG measurement system 133
The custom PPG measurement system comprises the transmittance sensor, sensor electronics, data acquisition 134
board (DAQ) and a laptop personal computer (PC). The emitter of the sensor is one laser diode with a peak 135
wavelength in the near infrared, at 850 nm, and matches the usual PPG wavelengths. Three BPW34 p-i-n 136
silicon photodiodes connected in parallel and aligned to increase the detection area are used as photo-detector. 137
The emitter driver, further amplification and filtering stages, timing and sample-and-hold (S&H) circuits are 138
connected to the fingertip probe by cables but could be easily replaced by wireless connections. The 139
photodetectors signal is filtered by an anti-aliasing analog low pass filter at 300 Hz, then sampled with a high 140
speed S&H and fed into the analogue inputs of a 12 bit DAQ (DAQ1200, National Instruments) to be digitized 141
at 1000 Samples/second (Sa/s). Finally, a ten-sample moving average followed by a ten-to-one decimation is 142
performed, which results in 100 Sa/s PPG data. 143
This experimental system is connected to a Pentium-class laptop computer that saves files containing the 144
PPG 100 Sa/s data in volts (V). The next stages of the signal processing are carried out digitally. The post-145
processing of stored signals from the group T1 of athletes is made off-line; once the algorithm has been 146
developed, it has been implemented as a virtual instrument using Lab Windows CVITM
, and the validation is 147
made on-line during IMET of the second group T2 of athletes. 148
149
2.4. Statistical analysis 150
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
7
Comparisons of HR between the two measuring methods (PPG and ECG) were performed using the Student 151
„t‟ test for paired data. Linear Pearson‟s correlation analysis was used to correlate quantitative variables, as 152
an indication that the two methods evolve in parallel. To compare the values of HR obtained by our PPG 153
technique with the standard ECG, we have used the Bland-Altman (B-A) method. Hereafter, we will use the 154
notation PR and ER for the PPG and the ECG heart rate values, respectively. The differences between the 155
methods (PR-ER, in bpm) are plotted against the average value given by the two methods (PR+ER)/2, in 156
bpm. The results are presented as the differences themselves (PR-ER) and mean difference or bias (M) 157
between the two methods. The standard deviation (Sd) value is used to calculate the limits of agreement (LA) 158
computed as M ± 1.96·Sd, providing an interval within which 95 % of differences between measurements by 159
the two methods are expected to lie. The standard errors and confidence intervals (CIs) were determined for 160
the mean bias and for the upper and lower limits of agreement. We defined agreement as a difference of 161
(PR-ER) within ± 10 bpm [3,12]. 162
163
3. Elements for the Algorithm Design 164
Fig. 1 presents the FFT spectra corresponding to PPG signals recorded over 10 s time intervals at rest 165
(time interval finishing at 25.96 s, grey) and when running at 12 Km/h (time interval finishing at 435.56 s 166
dark), during the test performed in treadmill by a male athlete A of group T1. Sampling of ECG- and PPG-167
based measurements are not synchronized, for this reason the match of ECG and PPG readings has been 168
achieved by pairing the nearest values in time. 169
FFT spectra are plotted as power density (V2/Hz) versus frequency in the range 40-240 beats per minute 170
(bpm), corresponding to 0.66 to 4 Hz. For each time interval, the spectra show three main peaks, which are 171
labeled according to their power as P1 (the strongest peak), P2 and P3 (the weakest peak). At rest, the 172
spectrum shows the strongest peak P1 at a frequency coinciding with the recorded ER value for the same 173
time interval, and two other tiny peaks corresponding to harmonics of P1 at frequency 2 x P1 and 3 x P1. 174
When running at 12 Km/h, the spectrum exhibits a quite different peak distribution; for this time interval the 175
recorded ER value coincides with P2 frequency, a peak with an intermediate power; indeed, the one foot-176
step rate coincides with P3 frequency, whereas P1 is the harmonic 2 x P3 corresponding to the full-step rate 177
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
8
SR (steps per minute). 178
Fig. 2 shows the time evolution of peaks P1, P2 and P3 (frequency vs. time) during IMET of athlete A, from 179
rest to maximal effort; time evolution of ER and full-step rate (SR) are also shown, as well as the time 180
interval of walking or running and the stop time. There are three well defined lines with a frequency that 181
increases with time, two of these lines evolves concomitantly, among them a third line evolves steeply with 182
time; the strongest peak is not always associated with the same line but jumps from a line to another The 183
steeper line evolves close to ER values, for this reason we have identified these frequencies as the PPG-heart 184
rate measurements PR. The heartbeat harmonics appears sometimes at a noise level at rest and at the 185
beginning of the IMET and goes out of the plotted range of frequency as the heartbeat increases. Regarding 186
the lower and upper lines, it can be observed that the upper one remains always at twice the frequency of the 187
lower one; furthermore, this upper line evolves close to the full-step rate SR, for this reason we have 188
identified these frequencies as the full-step rate SR, the harmonic of the one-step rate, identified as the lower 189
line. In brief, the time evolution of the spectra peaks defines a pattern composed of three branches, one 190
ascending central branch corresponding to the PR and two branches related to the periodic movement of feet 191
and arms (fundamental and its harmonic, related to the one-foot and the full step rates, respectively) that 192
evolve during exercise at a rate different to the heartbeat. 193
Results obtained from different athletes show that evolution of the peaks follows always the ER and SR 194
evolutions; in some cases, the heartbeat evolves much more steeply than SR and the corresponding line 195
intercepts one of the motion branches at a certain moment of the exercise and there are time intervals where 196
the frequency domains of PR signal and artifacts can overlap. 197
The arterial pulse wave might have higher frequencies but their power content is low or they can be 198
filtered. At low physical activity, the PR value can be obtained by simply looking for the highest peak in the 199
spectrum. The respiratory rate has been detected several times during exercise, at frequencies around 40-50 200
min-1
but it has low power content relative to the heartbeat and movement artifact signals. 201
The algorithm design is based on these observations. Firstly, the algorithm needs a harmonic discriminator, 202
taking into account that movement artifact contribution has higher harmonic content than the heart rate 203
signal. The detection algorithm we have developed is mainly based on analysis of the relative strength of 204
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
9
harmonic content of the PPG signal. Secondly, it is important to have a heuristic algorithm to track the 205
heartbeat, taking into account the past history of the heartbeat itself when PR signal and artifact signal 206
overlap. 207
208
4. Development of the Algorithm 209
We have developed the algorithm by analyzing the PPG data recorded from the first group T1 of athletes and 210
comparing them to the ECG ones. 211
The first step is a linear filtering of the 100 Sa/s PPG signal E(t) through band-pass Bessel filtering (cut-off 212
frequencies f1 and f2, order n = 6), in order to get the pulsating component (Eac(t)) and to suppress high 213
frequency noise and ripple. Cut-off frequencies are chosen in the ranges 0.1~0.3 Hz for f1 and 5~30 Hz for f2. 214
After filtering, a frequency domain analysis is performed by applying a fast Fourier transform (FFT) to a 10 215
second, rectangular-shaped, sliding window of the aforementioned data signal, with an overlap of about 75 216
%, delivering a spectrogram data F(ts), expressed in terms of power density (pi) versus frequency (fi). In this 217
step, the electrical PPG signal is transformed into a sequence of FFT spectra, in which each spectrum is 218
computed every 2.56 s and contains data from a time interval of 10.24 s, which gives a frequency resolution 219
of 0.58 bpm. The instant in time ts to which the spectrum is assigned corresponds to the end of the interval. 220
The spectra stream goes onto a peak search algorithm that selects the peaks above the noise floor. In this 221
step, the Gibbs phenomenon produces lateral lobes that, in theory, could create “phantom” peaks due to 222
side-lobe leakage; this is taken into account in the peak search algorithm, which never selects peaks from 223
adjacent FFT bins. We have explored more complex windowing functions and even more complex 224
approaches [27], but the relatively wide window and the experimental tests shows that leakage, albeit 225
present, did not invalidate the effectiveness of the algorithm. 226
Fig. 3 shows schematically the heuristic algorithm. The peak search algorithm eliminates all the peaks that 227
are lower than the noise floor of F(ts ), (Fig. 3 (a)), and selects those peaks that could be the most relevant. 228
pMax is the highest power value of F(ts ), and KC is the “cut-off” parameter of the algorithm. Values for KC 229
have been tuned in the range 50 - 400 by analyzing and comparing the PPG and ECG data recorded from the 230
group T1 of athletes; KC = 100 was adopted as the default value; it means that the cut-off of minor peaks occurs 231
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
10
when amplitudes are approximately from 7 to 20 times lower than the highest peak. When implemented on the 232
PPG system and in order to simplify the computational cost of the algorithm, the peak search selects only 233
the three main peaks. 234
Next, the algorithm chooses a “best candidate” P(ts, f, p) as the heartbeat PR, with a reliability factor C(ts) 235
which express, in the range (0-1), the degree of confidence that P(ts) was the heart rate value, not a 236
movement artifact. To achieve this, the list of three peaks is fed into three classification engines: energy, 237
harmonic and historic classification. The task of the classification engine is to associate to each peak the 238
reliability factor C(ts) defined as above, using different criteria. In all the following discussion and formula, 239
reliability factors saturates to 0 or 1, respectively, when the computed value exceeds these limits. Fig 4b and 240
4c shows the flow diagram of the heuristic algorithm. 241
The power engine (Fig. 3 (b)) compares each pi value to the highest value pMax of F(ts), and assigns to each 242
(fi, pi) pair a first reliability factor Cpi, which expresses how much the peak has a high power content. Cpi is 243
computed in the range (0-1) as the amplitude of the peak relative to the highest peak. 244
The harmonic engine (Fig. 3 (c)) computes the probability that a peak is part of a signal with high 245
harmonic content, assigning to each of the peaks a second reliability factor Chi. The engine compares pairs 246
of frequencies. For each peak, a sliding mask is superposed on the spectrum. The mask is formed by 247
parabolic segments centered on the frequencies multiple of the peak, with a relative width controlled by a 248
parameter KH, that modulates the tolerance accepted in the frequency comparison. Chi is computed in the 249
range (0-1) based on how much the spectrum energy enters the mask, and its value is low for high harmonic 250
content and high for low harmonic content. When implemented on the PPG system and in order to reduce 251
the computational burden, the algorithm computes how near a frequency is to double another frequency of 252
the list, limiting the analysis to the second harmonic. The parameter KH is one of the main parameter of the 253
algorithm. KH has been tuned in the range 5 - 50, and was optimized for a value of 12. For KH < 5, the mask 254
is too wide and all the peaks appear with a harmonic content. For KH > 50, the mask is too narrow and the 255
algorithm never finds any harmonics. When a peak is found to have a strong second-harmonic counterpart in 256
the spectrum, it is automatically invalidated for the selection process, even if it has a very high confidence 257
value under the power engine point of view. On the other hand, at rest, the main peak corresponds to the 258
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
11
heart rate PR, but very small power harmonics can be present; it is the value of KC of the peak search 259
algorithm which must contribute to eliminate this harmonics of the peak list entering the heuristic algorithm. 260
Finally, the memory or historic engine (Fig. 3 (d)) compares frequency fi(ts) of peaks Pi (ts) of F(ts), with the 261
frequency fi’ (ts-1) of the peak P(ts-1) selected as PR in the previous spectrogram F(ts-1). A third reliability 262
factor Cmi is computed by taking into account the proximity of the frequency fi (ts) to f i‟(ts-1) as well as the 263
former value of the reliability factor Cmi. This factor Cmi is zero for all the peaks but for that with the 264
frequency closest to f i‟(ts-1) selected as PR at previous spectrogram; Cmi is computed in the range 0-1, so that 265
a high value means high proximity to the previous PR peak. 266
Fig 4 shows, in more detail, the flowchart of the algorithm. Fig 4a shows the general data processing for 267
selecting the three main peaks of a spectrogram F(ts). Fig. 4b depicts the flowchart of the “power” and 268
“harmonic” engines; Fig. 4c shows the flowchart of the “memory” engine and the final selection step. 269
In the end, for each spectrogram F(ts), the heuristic algorithm compares all the peaks and selects one of 270
them, according to the assigned reliability factors. The selection is made on the basis of the maximum 271
reliability factor value, that is high Cpi (amplitude of the peak), high Chi (low harmonic content) and high Cmi 272
(high proximity to the previous PR peak). The chosen peak is assigned to the PR(ts) value, and the overall 273
confidence index C (ts) is set to the highest reliability index. 274
As an exception, when no a peak achieves a reliability factor greater than 0.5 (the highest peak could have 275
been marked as non valid by the harmonic content analysis, and the other peaks are not sufficiently good by 276
any of the classification techniques) the algorithm simply “jumps” over, giving the PR value selected for the 277
last spectrogram, and it restarts with the next one. This step is permitted just once in a row to avoid the 278
possibility of sticking to a wrong value; i.e., at the next spectrogram, if the reliability factor is again lower 279
than 0.5, the selection is made as described above, even with bad reliability factors. This step allows the 280
algorithm to ignore especially bad points, without corrupting the memory part of the algorithm. 281
Notice that the memory engine assigns to the peak selected in each spectrogram a reliability factor smaller 282
than the one assigned in the previous spectrogram. In this way, if the peak chosen as PR continues to be 283
chosen by means of the memory algorithm, the reliability factor is always decreasing, and this eventually 284
forces the heuristic algorithm to select the PR peak by a different criterion, so that the memory one could not 285
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
12
overtake the others for a long time. This is very important in the case PR peak is very close to a motion peak, 286
because when the algorithm selects a high-power peak, both the power engine and the memory engine will 287
tend to follow this high peak. If the reliability factor Cmi decreases continuously, the memory engine does 288
not find a solution and the harmonic engine will search for a new peak. 289
The parameters KC and KH have been optimized by analyzing the recorded data of group T1 of athletes. 290
The highest computational cost of the algorithm is found in the generation of the FFT spectra, for which we 291
used the standard FFT subroutines offered by standard software. As mentioned before, in order to reduce the 292
computational cost, some simplifications have been used: a) at the search peak stage, the number of selected 293
peaks is culled to three, those with highest power density; b) at the harmonic analysis only the second 294
harmonic is taken into account, given that higher order harmonics are normally outside the measurement 295
range, and c) the harmonic analysis compares only frequencies, independently of their power density. 296
Once the IMET starts, the harmonic engine is the principal component of the algorithm and allows it to 297
discriminate the movement artifact. When the PR peak is very close to a motion peak, the memory engine 298
helps to follow correct peak, but forcing the harmonic engine to act quickly. When one peak is selected, its 299
frequency is then assumed to be the pulse rate PR, and its value is displayed on the screen. 300
301
4.1. Signal to Noise Ratio 302
An estimation of the signal-to-noise ratio (SNR) of the selected PR data has been done. The selected PR 303
peak has been used as the signal power density and the sum of discarded peaks as the noise power density. 304
The procedure has been repeated for all the PR peaks selected along the exercise and for the whole group T1 305
of athletes. The highest 10 % and the lowest 10% SNR values were discarded and the remaining data were 306
averaged. The SNR estimated in such a way does not take into account the ground floor noise but just 307
discarded peaks; moreover, the noisiest points have been avoided, those where the algorithm has jumped over, 308
so that the resulting figures are probably an overestimation of the real value. Nevertheless, it constitutes a 309
valuable estimation of the SNR of the selected PR data. We have calculated values of SNR, both averaging 310
over the whole IMET and over the last 20 % of IMET time. From the analysis of group T1 recorded data, the 311
averaged values for “full test” and “last 20 %” are - 1.6 dB and - 2.5 dB, respectively. From that, we have 312
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
13
concluded that the algorithm is able to identify the correct cardiac heartbeat PR even with a SNR as low as -313
2.5 dB. 314
315
5. Results 316
5.1. Validation of the algorithm on recorded data of group T1 of athletes. 317
To validate the algorithm, each PR value was compared with the timely nearest ER, within intervals in 318
which the athletes were running, i.e., from 8 Km/h until the maximal effort. The initial periods were not 319
included due to dissimilarities in the athlete actions. The reason is that at 6 Km/h most of the athletes were 320
adapting to walk on the treadmill and in many cases they took hold of protection bars, which could alter the 321
position of the finger sensor and therefore distort the measurement. 322
We have calculated the regression line of ER and PR measurements, for a number of athletes A = 10 and a 323
number of (PR, ER) pairs N = 477, with a correlation R = 0.994 and p < 0.0001. This correlation is an 324
indication that the two parameters evolve in parallel. The histogram of differences between the two methods 325
has confirmed the normal distribution. The Bland and Altman plot of the difference (PR - ER) against the 326
average value given by the two methods ((PR + ER)/2) is shown in Fig. 5. The bias is M = - 1.10 bpm, Sd = 327
2.45 bpm is the standard deviation, and the limits of agreement LA are in the range (- 5.90 bpm, + 3.69 bpm) 328
within the a priori set value of ± 10 bpm. The 95 % confidence intervals are - 1.32 to - 0.88 bpm for the bias, 329
- 6.28 to - 5.52 bpm for the lower limit of agreement (LLA), 3.31 to 4.07 bpm for the upper limit of 330
agreement (ULA). All these intervals are reasonably narrow, the width of limits of agreement is 9.59 bpm 331
and more than 95 % of all (PR-ER) differences fall within LA. The plot reveals that there is no relationship 332
between the differences and the bias. The option of plotting the differences as percentages is useful when 333
there is an increase in variability of the differences as the magnitude of the measurement increases. When 334
plotting as percentages, the values obtained are M = - 0.7 %, LLA = - 4.2 %, ULA = 2.8 %. 335
336
5.2. Results on real time heart rate tracking during IMET on treadmill. 337
The algorithm has been implemented in the experimental PPG system, and has been validated on real time 338
measurements during treadmill ergometer IMET on the second group of athletes T2. The SNR values are - 0.1 339
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
14
dB for “full test” and - 2.1 dB for “last 20 %” respectively. Here again, in spite of the low SNR, the 340
algorithm is able to follow the correct PR value, as shown below. 341
The plot of PR versus ER for the group T2 gives a correlation R = 0.992 and p < 0.0001; the number of 342
athletes is A = 20 and the number of (PR, ER) pairs is N = 1061. We have confirmed the normal distribution 343
of differences between the two methods by their histograms. The (PR, ER) pairs obtained have been 344
displayed in Bland-Altman plots in Fig. 6, as the differences (PR - ER) versus mean value (M) of the 345
differences; the value M = - 0.70 bpm indicates a low negative bias, the standard deviation is Sd = 2.92 bpm. 346
The width of limits of agreement is 11.42 bpm, between - 6.41 bpm (LLA) and + 5.01 bpm (ULA), within 347
the a priori set value of ± 10 bpm. More than 95 % of all the differences obtained for the athletes tested on 348
the treadmill fall within ± 6.5 bpm, even for HR as high as about 200 bpm. The 95 % confidence intervals 349
are - 0.88 to - 0.53 bpm for bias, - 6.72 to - 6.12 bpm for LLA and 4.71 to 5.31 bpm for ULA. For group T2, 350
the values expressed as percentage of the average between the two techniques are M = - 0.4%, LLA = - 4.2 351
% and ULA = 3.4 %. 352
353
6. Discussion 354
The algorithm for the automatic detection of the HR from the PPG signals was found to be efficient, despite 355
the small signal-to-noise ratio. Comparison to ECG measurements, the reference standard, shows very high 356
agreement. The low negative bias obtained on recorded data of group T1 of athletes may be caused by slight 357
differences in beat-to-beat averaging times used in the ECG monitor and the PPG system; if the subject's 358
heart rate is changing rapidly, the different averaging times would influence the results and amplify the 359
calculated errors. Moreover, the small bias might be due to possible errors when collecting data from a 360
variety of sources in which time is recorded by independent and unsynchronized clocks [28] (independent 361
computers for ECG and PPG). The algorithm selects PR values that show high agreement with ER 362
measurements in activities as IMET. The narrow width of limits of agreement, narrower than ± 10 bpm, is 363
an acceptable LA for interchangeability between PR and ER. It should be noted that parameters optimization, 364
carried out on the recorded data, has been performed in a “first step” way. Optimization with an iterative 365
mathematical program should give the best values, but it is not the aim of the present work 366
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
15
Overall, the system shows a general success to track heart rate during the treadmill probe. The system has 367
also shown good results for IMET on cycle-ergometer [29]. When compared to traditional FFT methods the 368
presented algorithm shows significant improvements for separating motion artifact and cardiac physiologic 369
spectra. The limits of agreement are comparable to the mean absolute error obtained using an algorithm 370
including time-frequency techniques based on the smoothed Wigner Ville distribution [2,8], where six 371
subjects participated in the study realizing four kinds of motions at time, frequency and intensity fixed. 372
When comparing our method with results from a PPG sensor integrated with an adaptive noise cancellation 373
device, we obtained LA values between - 6.41 bpm and + 5.01 bpm (or - 4.2 % and 3.4 %), better than 374
values (-21.15% and 20.52 %) obtained for subjects running at 8 Km/h [21]. 375
The algorithm has been developed by analyzing the recorded data of a group of white men, all of them 376
runners; and it has been evaluated on a very dissimilar group, composed of men and women, black and 377
white skins. The algorithm parameters we have used have been the same for female subjects or for black 378
skins and the accuracy of the technique might be higher when adapted for each case. But in the current form 379
the algorithm is suitable to people without discrimination neither gender or skin color. 380
The basic premise of the algorithm is that the motion artifact created in the PPG signal by human running 381
has a high harmonic content with respect to the PPG derived heart rate. The use of graded treadmill running 382
(or cycling) in IMET may bias the motion artifact in the PPG signal by making it quite regular and rhythmic 383
in nature. Although this seems to be a limitation of the PPG-based technique, athletes running are a regular 384
and stable movement. A significant advancement of the algorithm utility would be to show that it can extract 385
representative heart rates during road running or running on uneven ground. Other apparent limitation is that 386
our PPG system is a fingertip probe; the algorithm can be implemented in other PPG systems, as earlobe 387
probes or ring sensor, provided that a rhythmic movement of the body and /or extremities will have high 388
harmonic contents. It is worth to note that other error sources, like movement of the sensor on the finger, do 389
not cause important signals in the FFT analysis once the athlete arms have free movement; otherwise, ECG-390
derived HR values can also give some errors, when the athlete is sweating and the fixation of the ECG 391
electrodes could be poor and produce wrong readings. 392
393
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
16
7. Conclusions 394
This work addresses the problem that motion artifacts limit the use of photoplethysmography (PPG) to 395
quantify key cardiovascular variables such as heart rate during human exercise. During human movement, 396
the PPG signal is corrupted at frequencies that fall within the range of interest. With appropriate filtering and 397
signal processing, the portion of the PPG signal caused by motion artifact could be removed and the heart 398
rate identified. We have described and evaluated a new algorithm to process PPG signals collected during 399
treadmill exercise into heart rate. The algorithm identifies three frequencies of interest from the PPG signal 400
and uses decision making logic to identify the two harmonic frequencies that most likely represent motion 401
artifact leaving one frequency to represent PPG derived heart rate. The algorithm is tested using a 402
heterogeneous athletic population transitioning from rest through a graded maximal intensity treadmill test. 403
Electrocardiography is used to validate the PPG-derived heart rates. This would allow for significant 404
advancement in the area of biosensor development as PPG technology is ideally suited to be incorporated 405
into wearable devices. 406
Acknowledgment 407
This work was supported by the Spanish Consejo Superior de Deportes under CSD Grants 03/UPB10/06, 408
01/UPB10/07, and 01/UPB10/08. S. M. López-Silva has been supported by CSIC I3P Program financed by 409
the Fondo Social Europeo. The authors gratefully acknowledge the assistance of Dr. Isabel Gómez Caridad 410
in revising the present manuscript. 411
412
413
414
415
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
17
References 416
[1] A. A. R. Kamal, J. B. Harness, G. Irving and A. J. Mearns, “Skin photoplethysmography: a review,” 417
Comput. Methods Programs Biomed., 28: 257-269, 1989. 418
[2] J. Allen, “Photoplethysmography and its applications in clinical physiological measurement,” Physiol. 419
Meas., 28: R1-R39, 2007. 420
[3] J. Townshend, B. J. Taylor, B. Galland and S. Williams, “Comparison of new generation motion-421
resistant pulse oximeters,” J. Paediatr. Child Health, 42: 359-365, 2006. 422
[4] B. S. Kim and S. K. Yoo, “Motion artifact reduction in photoplethysmography using independent 423
component analysis,” IEEE Trans. Biomed. Eng., 53: 566-568, 2006. 424
[5] M. J. Hayes and P. R. Smith, “A new method for pulse oximetry possessing inherent insensitivity to 425
artifact,” IEEE Trans. Biomed. Eng., 48: 452-461, 2001. 426
[6] S. Rhee, B. H Yang and H. H. Asada, “Artifact-resistant power-efficient design of finger-ring 427
plethysmographic sensors,” IEEE Trans. Biomed. Eng., 48: 795-804, 2001. 428
[7] Q. Li, R. G Mark and G. D. Clifford, “Artificial arterial blood pressure artifact models and an evaluation 429
of a robust blood pressure and heart rate estimator,” Biomed. Eng. Online, 8: 13, 2009. 430
[8] Y. S. Yan, C. C. Poon and Y. T. Zhang, “Reduction of motion artifact in pulse oximetry by smoothed 431
pseudo Wigner–Ville distribution,” J. Neuroeng. Rehabil., 2: 3, 2005. 432
[9] P. Shi, V. A. Peris, A. Echiadis, J. Zheng, Y. Zhu, P. Y. S. Cheang and S. Hu, “Non-contact reflection 433
photoplethysmography towards effective human physiological monitoring,” J. Med. Biol. Eng., 30: 434
161-167, 2010. 435
[10] H. Benoit, F. Costes, L. Feasson, J. R. Lacour, F. Roche, C. Denis, A. Geyssant and J. C. Barthélémy, 436
“Accuracy of pulse oximetry during intense exercise under severe hypoxic conditions,” Eur. J. Appl. 437
Physiol., 76: 260-263, 1997. 438
[11] Y. Yamaya, H. J. Bogaard, P. D. Wagner, K. Niizeki and S. R. Hopkins, “Validity of pulse oximetry 439
during maximal exercise in normoxia, hypoxia and hiperoxia,” J. Appl. Physiol., 92: 162-168, 2002. 440
[12] K. Nakajima, T. Tamura and H. Miike, “Monitoring of heart and respiratory rates by 441
photoplethysmography using a digital filtering technique,” Med. Eng. Phys., 18: 365-372, 1996. 442
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
18
[13] Y. Iyriboz, S. Powers, J. Morrow, D. Ayers and G. Landry, “Accuracy of pulse oximeters in estimating 443
heart rate at rest and during exercise,” Br. J. Sports Med., 25: 162-164, 1991. 444
[14] K. Charlot, J. Cornolo, J. V. Brugniaux, J. P. Richalet and A. Pichon, “Interchangeability between heart 445
rate and photoplethysmography variabilities during sympathetic stimulations,” Physiol. Meas., 30: 446
1357-1369, 2009. 447
[15] S. H. Liu, K. M. Chang and T. H. Fu, “Heart rate extraction from photoplethysmogram on fuzzy logic 448
discriminator,” Eng. Appl. Artif. Intell., 23: 968-977, 2010. 449
[16] K. H. Shelley, “Photoplethysmography: beyond the calculation of arterial oxygen saturation and heart 450
rate,” Anesth. Analg., 105: S31-S36, 2007. 451
[17] Centre Suisse d‟Electronique et Microtechnique (CSEM), Swiss Center for Electronics and 452
Microtechnology, 2008, available: http//www.csem.ch. 453
[18] H. H. Asada, P. Shaltis, A. Reisner, S. Rhee and R. C. Hutchinson, “Mobile monitoring with wearable 454
photoplethysmographic biosensors,” IEEE Eng. Med. Biol. Mag., 22: 28-40, 2003. 455
[19] G. Cennini, J. Arguel, K. Aksit and A. van Leest, “Heart rate monitoring via remote 456
photoplethysmography with motion artifacts reduction,” Opt. Express, 18: 4867-4875, 2010. 457
[20] H. Han, M. J. Kim and J. Kim, “Development of real-time motion artifact reduction algorithm for a 458
wearable photoplethysmography,” Conf. Proc. IEEE Eng. Med. Biol. Soc., 1538-1541, 2007. 459
[21] M. Z. Poh, N. C. Swenson and R. W. Picard, “Motion-tolerant magnetic earring sensor and wireless 460
earpiece for wearable photoplethysmography,” IEEE Trans. Inf. Technol. Biomed., 14: 786-794, 2010. 461
[22] R. Giannetti, J. P. Silveira, M. L. Dotor, D. Golmayo, P. Martín, A. Bilbao, F. Miguel-Tobal and S. M. 462
López-Silva, “Oxygen saturation measurements in athletes attaining maximal exertion conditions,” 463
Proc. IEEE Instrum. Meas. Technol. Conf. (IMTC), 1: 740-744, 2004. 464
[23] S. M. López-Silva, M. L. Dotor and J. P. Silveira, “NIR transmittance pulse oximetry with laser diodes,” J. 465
Biomed. Opt., 8: 525-533, 2003. 466
[24] J. M. Bland and D. G. Altman, “Measuring agreement in method comparison studies,” Stat. Methods 467
Med. Res., 8: 135-160, 1999. 468
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
19
[25] J. M. Bland and D. G. Altman, “Statistical methods for assessing agreement between two methods of 469
clinical measurement,” Int. J. Nurs. Stud., 47: 931-936, 2010. 470
[26] M. Rabadán and J. C. Segovia, “Pruebas de esfuerzo directas,” in: J. C. Segovia, J. C. Legido and F. J. 471
López-Silvarrey (Eds.), Manual Valoración Funcional: Aspectos Clínicos y Fisiológicos, Madrid: 472
Elsevier, 267-278, 2008. 473
[27] C. Pan, “Gibbs phenomenon removal and digital filtering directly through the fast Fourier transform”, 474
IEEE Trans. Signal Process., 49: 444-448, 2001. 475
[28] G. D. Clifford, W. J. Long, G. B. Moody and P. Szolovits, “Robust parameter extraction for decision 476
support using multimodal intensive care data,” Philos. Trans. R. Soc. A, 367: 411-429, 2009. 477
[29] S. M. López-Silva, M. L. Dotor, J. P. Silveira, R. Giannetti, F. Miguel-Tobal, A. Bilbao, M. Galindo and 478
P. Martín, “Pulse rate measurement from transmittance photoplethysmography in cycle ergometer 479
test,” Proc. IEEE Instrum. Meas. Technol. Conf. (IMTC), 1039-1043, 2010. 480
481
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
20
Figure captions 482 483 484 Fig. 1. FFT spectra corresponding to PPG signals recorded over 10 s time intervals (grey at rest, dark during 485
exercise, running at 12 Km/h); the three main peaks P1, P2 and P3 are labeled following decreasing 486
power density; the values of ER (ECG-derived heart-rate) and SR (full-step rate) for the same time 487
intervals are also shown. 488
Fig. 2. Frequency time evolution of the main peaks (P1, P2, P3) over the whole test; the evolution of ER 489
(ECG-derived heart-rate) and SR (full-step rate) are also plotted. 490
Fig. 3. Description of the heuristic algorithm. After the band-pass linear filtering, in (a), a set of significant 491
peaks is computed, represented by a list of frequency-peak value (power density PD). The selection 492
is made by listing all the peaks that are higher than a noise floor, computed by dividing the highest 493
peak by a parameter Kc. The list is then fed into the three mechanisms that assign a reliability value 494
to each pair. In (b), peaks are classified in function of their relative amplitude; in (c), their harmonic 495
relationship is computed (see a more thorough explanation in the text); finally, in (d), the historic 496
record is taken into account, assigning higher reliability values to the peaks that are near to the 497
previously selected value. 498
Fig. 4. Flowchart of the proposed algorithm. In Fig 4a the general data processing is shown. Fig. 4b depicts 499
the “power” and “harmonic” engines, and Fig. 4c the “memory” engine and the final selection step. 500
Fig. 5. Bland and Altman plot of differences in recorded heart rate measurements versus average values for 501
group T1, with bias value (M) and limits of agreement (M ± 1.96 Sd). Sd is the standard deviation, A 502
is the number of athletes and N is the number of pairs (PR, ER). 503
Fig. 6. Bland and Altman plot of differences in real time heart rate measurements versus average values for 504
group T2, with bias value (M) and limits of agreement (M ± 1.96 Sd). Sd is the standard deviation, A 505
is the number of athletes and N is the number of pairs (PR, ER). 506
507
508
509
510
511
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
21
512
513
514
515
516
517
518
519
520
521
522
523
524
Fig. 1 525 526
527
528
529
530
531
40 60 80 100 120 140 160 180 200 220 2400.000
0.001
0.002
0.003
SR
435.5
6=
192
bp
m
ER
435.5
6=
150
bp
m
ER
25.9
6=
75 b
pm
P2
P3
P3P2
P1
P1
Athlete A
25.96 s
435.56 s
Po
wer
den
sit
y (
V2/H
z)
Frequency (bpm)
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
22
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
Fig. 2 549
550
551
1080 1200 1320 1440 1560 1680 180040
80
120
160
200
240
280
1080 1200 1320 1440 1560 1680 180040
80
120
160
200
240
280
stop
runningwalking
7206004803602401200
athlete A ER
P1
P2
P3
STEP
freq
uen
cy (
bp
m)
time (s)
stop
runningwalking
7206004803602401200
Athlete A P1
P2
P3
SR
ER
Fre
qu
en
cy (
bp
m)
Time (s)
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
23
552
553
554
555
556
557
558
559 560 561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
24
577
578
579
Fig. 3 580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
25
600 601
602
Fig. 4a 603
604
605
606
607
608
609
610
611
612
613
614
615
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
26
616
617
618
619
620
621 622
623
Fig. 4b 624
625
626
627
628
629
630
631
632
633
634
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
27
635
636
637
638
639
640 641
642
643
644
Fig. 4c 645
646
647
648
649
650
651
652
653
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
28
654
655
656
657
658
659
660
661
662
663
664
665
Fig. 5 666
667
668
669
670
671
672
80 100 120 140 160 180 200 220 240-30
-20
-10
0
10
20
30
M - 1.96Sd = - 5.90 bpm
M + 1.96Sd = 3.69 bpm
M = - 1.10 bpm
T1:
A = 10
N = 477
Sd = 2.45 bpm
Dif
f in
heart
rate
, P
R-E
R (
bp
m)
Average by PPG and ECG, (PR+ER)/2 (bpm)
Articles in Press, J. Med. Biol. Eng. (November 14, 2011), doi: 10.5405/jmbe.898
29
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687 688
689
Fig. 6 690
80 100 120 140 160 180 200 220 240-30
-20
-10
0
10
20
30
M = - 0.7 bpm
M - 1.96Sd = - 6.41 bpm
M + 1.96Sd = 5.01 bpm
T2:
A = 20
N = 1061
Sd = 2.92 bpm
Dif
f in
heart
rate
, P
R-E
R (
bp
m)
Average heart rate by PPG and ECG, (PR+ER)/2 (bpm)
top related