Scholarly article on topic 'Gyroscope vs. accelerometer measurements of motion from wrist PPG during physical exercise'

Gyroscope vs. accelerometer measurements of motion from wrist PPG during physical exercise Academic research paper on "Chemical sciences"

CC BY-NC-ND
0
0
Share paper
Academic journal
ICT Express
OECD Field of science
Keywords
{"Photoplethysmography (PPG)" / Gyroscope / Accelerometer / "Heart rate monitoring" / "Physical exercise"}

Abstract of research paper on Chemical sciences, author of scientific article — Alexander J. Casson, Arturo Vazquez Galvez, Delaram Jarchi

Abstract Many wearable devices include PPG (photoplethysmography) sensors for non-invasive heart rate monitoring. However, PPG signals are heavily corrupted by motion interference, and rely on simultaneous motion measurements to remove the interference. Accelerometers are used commonly, but cannot differentiate between acceleration due to movement and acceleration due to gravity. This paper compares measurements of motion using accelerometers and gyroscopes to give a more complete estimate of wrist motion. Results show the two sensor signals are very different, with low correlations present. When used in a wrist PPG heart rate algorithm gyroscope motion estimates obtain better performance in half of the cases.

Academic research paper on topic "Gyroscope vs. accelerometer measurements of motion from wrist PPG during physical exercise"

Accepted Manuscript

Gyroscope vs accelerometer measurements of motion from wrist PPG during physical exercise

Alexander J. Casson, Arturo Vazquez Galvez, Delaram Jarchi

PII: S2405-9595(16)30162-X

DOI: http://dx.doi.org/10.10167j.icte.2016.11.003

Reference: ICTE 65

To appear in: ICT Express

Received date: 26 September 2016 Revised date: 25 November 2016 Accepted date: 28 November 2016

Please cite this article as: A.J. Casson, A.V. Galvez, D. Jarchi, Gyroscope vs accelerometer measurements of motion from wrist PPG during physical exercise, ICT Express (2016), http://dx.doi.org/10.1016/jicte.2016.11.003

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

=xpress

In format ion & Communications Technology

Gyroscope vs Accelerometer Measurements of Motion from Wrist PPG During Physical Exercise

Alexander J. Casson 1*, Arturo Vazquez Galvez 1, Delaram Jarchi 1

1 School of Electrical and Electronic Engineering, The University of Manchester, Manchester, UK (e-mail: alex.casson@manchester.ac.uk) * Corresponding author

Abstract

Many wearable devices include PPG (photoplethysmography) sensors for non-invasive heart rate monitoring. However, PPG signals are heavily corrupted by motion interference, and rely on simultaneous motion measurements to remove the interference. Accelerometers are used commonly, but cannot differentiate between acceleration due to movement and acceleration due to gravity. This paper compares measurements of motion using accelerometers and gyroscopes to give a more complete estimate of wrist motion. Results show the two sensor signals are very different, with low correlations present. When used in a wrist PPG heart rate algorithm gyroscope motion estimates obtain better performance in half of the cases.

Index Terms: Photoplethysmography (PPG); Gyroscope; Accelerometer; Heart rate monitoring; Physical exercise.

I. Introduction

Very long term portable heart rate monitoring is quickly emerging as one of the leading uses of wearable technology. Most smart watches, including the Apple Watch, Samsung Gear S2, and Fitbit Surge include a heart rate monitor which can be used for optimizing workouts [1] and potentially for automatically detecting serious events such as atrial fibrillation [2]. As such they are enabling the study of heart function in the general population in a way not possible with traditional Holter monitors and they are seen as an important part of future personalized and preventative healthcare. It is estimated that 80% of heart diseases and 70% of strokes could be avoided with suitable preventative techniques [3].

These monitors are based upon Photoplethysmography (PPG), which shines a light into the wrist and measures the amount of light reflected back, which changes with blood flow. Unlike the Electrocardiogram (ECG) the PPG does not require a sticky conductive gel based electrode and so is highly suited to wearable applications. However, PPG signals are significantly corrupted by artefacts due to physical activity which have historically limited PPG to relatively motion free clinical settings [4]. Today, several different signal processing methods have been proposed for removing the interference, allowing heart rate to be extracted in the presence of motion. Nevertheless, in practical use the performance of these methods is still under investigation [5].

Recently many different motion artifact removal algorithms have been proposed based upon the 2015 'IEEE Signal Processing cup' [6, 7]. This provided a database of: wrist PPG; co-located 3-axis accelerometer measurements to give a reference recording of wrist motion; and a chest ECG to give a gold standard comparison of the subject's heart rate; for 23 subjects. As a result, many algorithms are based upon using the

accelerometer signal as a reference of the motion present, and subtracting it in some way from the PPG trace, for example by using an adaptive filter [6]. While 3-axis accelerometers have been widely applied in fitness trackers, as they measure acceleration they do not give a complete picture of the movement of the wrist and PPG sensor node. In particular, accelerometers alone cannot differentiate between acceleration components due to movement of the subject/sensor and acceleration components due to gravity. Using accelerometers in isolation thus gives only a partial estimate of the true motion present.

We propose that this can be overcome by using co-located gyroscopes in addition to the co-located accelerometers. Results from fitness trackers have shown how this combination of sensors can extract both the orientation and angular velocity of the sensor node [8, 9]. However, to our knowledge, gyroscope information has not yet been used in algorithms for estimating PPG based heart rate during physical activity. We have recorded a new database of wrist based PPG measurements which include both a 3-axis accelerometer and a 3-axis gyroscope while participants walk, run and cycle. In this article we characterize the gyroscope information and compare it to that collected by the accelerometers, demonstrating the potential use of gyroscopic data for improving the performance of future algorithms.

II. Methods

A. Data Collection

PPG recordings were collected from 8 subjects for approximately 5 minutes each as they undertook a range of physical exercise activities on a treadmill and exercise bike, different for each person as detailed in Table 1.

ICT Express

Table 1. Summary of Exercises Performed by Each Subject. Green = Subject Did That Activity, Red = Subject Did Not.

Subject Walk Run

Low resistance bike (lrb)

High resistance bike (hrb)

PPG and motion were recorded using a Shimmer 3 GSR+ [10], with 3-axis gyroscope, 3-axis low noise (±2 g) accelerometer, 3-axis wide range (±16 g) accelerometer, and PPG input. A green LED PPG sensor was glued to the main Shimmer unit to give a rigid connection and allow the accelerometers and gyroscopes inside the Shimmer unit to accurately record the movement of the PPG sensor. The combined unit was connected to the wrist using a continuously adjustable strap (similar to the Scosche Rhythm+ [11]), see Fig. 1. All of the signals (including a simultaneous ECG for gold standard heart rate) have been made public: https://physionet.org/works/WristPPGduringexercise/. Note in Table 1 (and Table 3) there is no subject 7 as this data was not shared publicly.

B. Signal Comparison

For comparing the accelerometer and gyroscope recordings of motion we first present a range of example signals, showing visually the different captures of motion provided. These are considered in both the time and frequency domain, using a standard FFT applied to the data sections. The results are quantified by considering the correlations present between the different motion signals (3-axes of accelerometer data and 3-axes of gyroscope data), demonstrating that across all of the records the accelerometer and gyroscope consistently collect different information.

Finally we use the gyroscope data in a PPG heart rate extraction algorithm, comparing the performance of the algorithm when the recorded motion input comes from the accelerometers and from the gyroscopes. Our algorithm is described in detail in [12], and overviewed

in Fig. 2 where the Acc input is either the acceleromter or the gyroscope depending on the test case. The method is similar to that in [13] using a normalized least mean squared adaptive filter to remove motion interference from the PPG before finding the heart beat frequency using the Short Time Fourier Transform (STFT).

Figure 2. PPG heart rate estimation algorithm. Accx is (accelerometer/gyroscope) input on axis x.

III. Results and Discussion

A. Qualitative Comparisons

Fig. 3 and Fig. 4 show 70 s example sections of the motion signals recorded during walking and high resistance biking respectively. Readily apparent is the 1 g component due to gravity in the accelerometer output

i=xpress

/пformation Л Communications Technology

which manifests mainly in one axis (axis 2 in Fig. 3, axis 3 in Fig. 4), with this axis changing as the subject moves from standing during walking to a seated biking position. Such an offset is not seen in the gyroscope data.

The walking data (Fig. 3) is generally free of large transients in the accelerometer output, with this instead being a periodic signal at the footfall frequency. The FFT of the gyroscope and accelerometer data for axis 1 is shown in Fig. 5. This shows that both sensors extract the dominant footfall frequency (1.27 Hz), and a sub-harmonic at the full gait cycle (0.64 Hz). The dominant frequency in the PPG trace is the footfall frequency of 1.27 Hz. Beyond these, the precise estimation of motion is very different from the two types of sensor. In particular the accelerometer reports much more activity at both high and low frequencies, with a peak in the 79 Hz range common across many recordings.

In contrast in the biking condition many large amplitude transients are seen, for example at 50 s in Fig. 4, occurring when the subject repositions their hands on the handle bar, and an artifact is seen in the time domain PPG at this point. While both the accelerometer and gyroscope capture that a transient occurs at this time, the precise shapes, Fig. 6, are very different.

B. Correlations of Motion Information

These differences in the motion recordings are quantified in Table 2 which shows the correlations between the gyroscope and low noise acclerometer data. (For compactness the correlations with the wide range accelerometer are not shown. These are very highly

correlated, >0.95, with the low noise accelerometers.)

In Table 2 all of the correlation coefficients are low, with few greater than 0.5. Although the accelerometers and gyroscopes are sampled simulatenously and are rigidly connected together inside the same package, the resulting signals have very different morphologies, giving different representations of the motion present.

Frequency (Hz)

Figure 5. Normalized FFT power during walking.

C. Algorithm Performance

The heart rate algorithm accuracy is given in Table 3. This shows the mean absolute deviation of the heart rate measured in each record, using the PPG signal and acceleratometer/gyroscope input, compared to a gold standard heart rate found from a simulatenous chest ECG measurement. ECG heart rate was derived via the R peak locations, identified using the Pan Tompkins algorithm with incorrect/misplaced R peaks removed by eye.

On our challenging PPG database which contains a range of example motions, and many transient events which are known to cause poor performance [6], the current adaptive filter does not always converge (lines highlighted in grey in Table 3), and further algorithm development is necessary to give usable performances in all cases. Using the accelerometer signals the absolute mean error across all records is 6.9 beats per minute. Using the gyroscope data this increases to 9.9 beats per minute, although this is highly variable. For half of the subjects the error rate is improved by using the gryoscope information in place of the accelerometer, for half of the subjects the error rate is increased. This makes it apparent that useful motion information is being

ICT Express

provided by the gyroscopes, but making optimal use of this requires a data fusion stage.

Such fusion techniques are common in activity tracking to find the orientation and rotational velocity of the sensor [8, 9] and such a fused estimate of motion could be used to provide a single, better, motion input to the heart rate algorithm. Alternatively it may be possible to dynamically switch between using accelerometer data or gyroscopic data at different points in time. One of the key challenges in using gyroscope information in a wearable is that gyroscopes use much more power than accelerometers and thus a periodic sampling of gyroscope data to augment the motion estimation provided by the accelerometers might be highly desirable.

Table 2. Correlation Coefficients Between Accelerometer (AX) and Gyroscope (GX) Data on Axis X.

Record G1 vs A1 G1 vs A2 G1 vs A3 G2 vs A2 G2 vs A3 G3 vs A3

si walk 0.21 -0.20 0.55 -0.03 0.14 -0.46

s1_lrb 0.08 0.01 0.02 0.03 -0.03 0.00

si hrb 0.12 0.03 0.08 0.02 -0.04 0.08

s2 walk 0.11 0.00 0.43 0.16 0.25 -0.39

s2_lrb 0.05 0.02 0.04 0.07 -0.04 0.04

s2_hrb 0.03 0.02 0.04 0.01 -0.11 0.07

s3 walk 0.12 -0.20 0.53 0.28 0.10 -0.33

s3 run 0.28 0.02 0.63 0.23 0.15 -0.48

s3_lrb -0.02 0.04 0.02 -0.01 -0.01 -0.02

s3 hrb 0.01 0.13 0.02 -0.07 -0.02 -0.03

s4 run -0.26 0.04 -0.12 -0.29 0.05 -0.05

s5 run 0.43 -0.25 -0.37 0.40 -0.41 -0.49

s5_lrb 0.65 -0.51 -0.45 0.47 0.55 -0.49

s6 walk 0.01 0.01 -0.04 0.03 0.00 -0.02

s6 run 0.06 -0.04 0.06 0.09 0.05 -0.12

s6_lrb 0.06 -0.02 0.02 -0.07 0.06 0.03

s8 walk -0.08 0.15 -0.56 0.03 -0.39 0.53

s8 run -0.21 -0.06 -0.42 -0.01 -0.18 0.32

s9 walk -0.03 0.13 -0.55 -0.23 -0.25 0.43

Table 3. Heart Rate Extraction Performance Comparison Using Different Measures of Motion as the Adaptive Filter Input.

Error rate (bpm)

Subject

Accelerometer Gyroscope

1 12.1 11.4

2 19.9 19.1

5 2.0 10.7

6 5.3 0.1

8 1.2 8.9

9 0.6 9.3

|Mean| 6.9 9.9

IV. Conclusions

This article has, for the first time, investigated simulatenous gyroscope data for remoing motion artifacts from wrist PPG to give improved heart rate estimates during motion compared to using only accelerometer measures as used conventionally. When used in a heart rate extraction algorithm gyroscope motion estimates obtained better performance in half of the cases. The results have characterised the motion recordings obtained from both types of sensor for the first time and demonstrated that both may have a role in further decreasing the error rate in wrist PPG.

Acknowledgments

This work was in part supported by the UK EPSRC Research Council grant number EP/M009262/1.

References

[1] R. Benson and D. Connolly, Heart rate training. Champaign: Human Kinetics, 2011.

[2] A. Carpenter and A. Frontera, "Smart-watches: a potential challenger to the implantable loop recorder?" Europace, no. 4, pp. 1-3, 2016.

[3] A. Honka, K. Kaipainen, H. Hietala, et al., "Rethinking health: ICT-enabled services to empower people to manage their health," IEEE Rev. Biomed. Eng., vol. 4, no.

1, pp. 119-139, 2011.

[4] R. Krishnan, B. Natarajan, and S. Warren, "Two-stage approach for detection and reduction of motion artifacts in photoplethysmographic data," IEEE Trans. Biomed. Eng., vol. 57, no. 8, pp. 1867-1876, 2010.

[5] S. E. Stahl, H.-S. An, D. M Dinkel, et al., "How accurate are the wrist-based heart rate monitors during walking and running activities? Are they accurate enough?," BMJ Open SportExerc. Med., vol. 2, no. 1, pp. 119-139, 2016.

[6] E. Khan, F. A. Hossain, S. Z. Uddin, et al., "A robust heart rate monitoring scheme using photoplethysmographic signals corrupted by intense motion artifacts," IEEE Trans. Biomed. Eng., vol. 63, no. 3, pp. 550-562, 2016.

[7] Z. Zhang, Z. Pi, and B. Liu, "TROIKA: A general framework for heart rate monitoring using wrist-type photoplethysmographic signals during intensive physical exercise," IEEE Trans. Biomed. Eng., vol. 62, no. 2, pp. 522-531, 2015.

[8] K. Tong and M. H. Granat, "A practical gait analysis system using gyroscopes," Med. Eng. Phys., vol. 21, no.

2, pp. 87-94, 1999.

[9] R. E. Mayagoitia, A. V. Nene, and P. H. Veltink, "Accelerometer and rate gyroscope measurement of kinematics: an inexpensive alternative to optical motion analysis systems," J. Biomech., vol. 35, no. 4, pp. 537542, 2002.

[10] Shimmer. [Internet]. Available: ww.shimmersensing.com.

[11] Scosche. [Internet]. Available: ww.scosche.com/rhythm+.

[12] D. Jarchi and A. J. Casson, "Estimation of heart rate from foot worn photoplethysmography sensors during fast bike exercise," IEEE EMBC, Orlando, 2016.

[13] T. Schack, C. Sledz, M. Muma, et al., "A new method for heart rate monitoring during physical exercise using photoplethysmographic signals," EUSIPCO, Nice, 2015.