ecti2013-respiration_rate_20130402_submission.pdf
Paper submitted at ECTI-CON 2013, "Estimation of Respiratory Rate from Smartphone’s Acceleration Data"
Content
Estimation of Respiratory Rate from Smartphone’s Acceleration Data
Thanakij Pechprasarn1, Suporn Pongnumkul2
National Electronics and Computer Technology Center 112 Thailand Science Park, Phahonyothin Road, Klong 1, Klong Luang, Pathumthani 12120, Thailand
1
[email protected] 2 [email protected]
Abstract— Abnormal respiratory rates have been shown to be an important predictor of serious clinical illness, but respiratory rate is a vital sign that is often not recorded because methods for measuring respiratory rates are cumbersome. We propose an approach to record and monitor respiratory rate of a patient that is lying down by placing an accelerometer-equipped smartphone on the patient’s chest. We develop an algorithm based on fast Fourier transform (FFT) to estimate the respiratory rate from the noisy acceleration data. The main contribution of this paper is that our proposed algorithm can estimate respiratory rates using only tri-axial acceleration data from sensor in commodity smartphones without any other special equipment. Preliminary results show that our method can reasonably estimate the respiratory rate. Keywords— respiratory rate, accelerometer, time series, moving average, Fourier transform
The algorithm we developed is based on the fast Fourier transforms. The data is first pre-processed to reduce noise using smoothing and detrending. Then we perform FFT to find the highest frequency. After that, we derived the respiratory rate from the three axes of signals by choosing the strongest frequency. We conducted an experiment to verify our methods, where we asked a healthy adult to lie down and had a general practitioner to measure the respiratory rate by counting the chest movements for about 30 seconds. The subject was asked to simulate fast breathing, normal breathing and slow breathing. We compare the observed respiratory rate from the general practitioners, the wave counts from the visual signals and the answer from our algorithm and found that our method can reasonably estimate the respiratory rate in all three cases. Our main contribution is the algorithm for estimating respiratory rates from just the acceleratory signals from commodity smartphones. The rest of this paper is organized as follows. Section II reviews the related work. Section III describes our algorithm. Section IV describes the experiment including the setup and the result. Then we conclude in Section V. II. RELATED WORK Respiratory rate is one of a vital sign that can be used to monitor the wellness of an individual [3-6]. It helps detect any malfunction of breathing activities, especially during sleeping. Respiratory rate is normally estimated from a recorded waveform. There are 2 main categories of waveform recording methods. In first category, a waveform of respiratory efforts is directly recorded. For example, impedance pneumography (IP) measures a respiratory activity through differential changes in capacitance. Respiratory inductive plethysmography (RIP) uses stretch sensors on the chest wall whereas flow thermography utilizes the changes in temperature of air flow when breathing. These methods can be viewed as a mainstream for the direct category. IP is the most widespread method used in a hospital whereas RIP is the most common method for overnight monitoring [4]. Another class, also classified to be in a direct category but much less common, includes the use of an accelerometer, a laser-based device, ultrasound, and audio and/or video processing. On the other hand, it is also possible to record a respiratory waveform
I. INTRODUCTION Respiratory rate is the number of breaths that a person takes in one minute while at rest. In practice, medical practitioners measure respiratory rate by counting how many times the chest moves up and down within one minute, or within 30 seconds and multiply the count by two. The method for measuring respiratory rate is tedious and time-consuming; therefore, it is a vital sign that is often neglected [1]. However, as respiratory rates may increase with fever, illness, or other medical conditions, it is an important predictor of serious clinical illnesses. The technologies for measuring respiratory rates are still an active area of research [2]. Most smartphones nowadays offer various built-in sensors and often include the tri-axial accelerometer, which measures the acceleration in three orthogonal directions. An accelerometer can be used to sense vibration, e.g. the vibration of a machine, orientation, e.g. in human activity monitoring. The tri-axial accelerometer is used as an inclinometer to reflect the abdomen or chest movement caused by respiration. We aim at creating a tool that can be used by medical practitioners and non-practitioners alike; therefore, we keep the device and the measurement methods simple. The device we use is an iPhone 4, a commodity smartphone that has triaxial accelerometers, and the measurement method is simply placing the smartphone on the patient’s chest for 30 seconds while the patient is in a lying down position.
978-1-4799-0545-4/13/$31.00 ©2013 IEEE
in an indirect fashion. However, this category would involve an extra step to rebuild and extract a respiratory waveform from other related waveforms. For instance, a respiratory waveform can be derived from electrocardiogram (ECG), photoplethysmogram (PPG), arterial blood pressure (ABP) and the peripheral arterial tonometry (PAT). Derivation of respiratory rate from ECG is an area that has been studied extensively compared to the others in the category. Moreover, recently researchers tend to include various signal sources simultaneously in order to improve a final estimate of respiratory rate. For example, a novelty proposed by Nemati et al. [7] is also based on data fusion which includes ECG, PAT, IP and PPG. Due to an emergence of microelectromechanical systems (MEMS)-based accelerometer [8], there is more recent published work that utilizes an accelerometer to estimate respiratory rate. Primitive reasons for the usage of an accelerometer, compared to traditional methods, are that it is a cheaper, non-invasive approach, and also viable to be used outside hospitals without a supervision of a professional. An accelerometer can be attached to a different part of the body, for example, a chest/thorax [3,8], a thorax-abdomen wall (including diaphragm muscle and lower costal margin) [4,5,9], an abdomen/waist [8,10] and even suprasternal notch [6,8]. Yet, currently there is no consensus on the best placement of the sensor. In our work, we decide to place a sensor on a patient’s chest. Besides placement, the posture of a person is reported to greatly affect the conducted breathing activity [5]. Thus, many require a patient to sit or lie down steadily to be able to successfully extract respiratory rate [3-6,8-9]. On the other hand, A novelty from Liu et al. [10] has studied the effects of posture changes like walking and running and still be able to compute respiratory rate out of that particular activities. Nevertheless, activity detection is not our main focus of this paper so we collect data only when a patient lies down steadily. It is proposed by previous publication [3] that breathing frequency is ranging between 0-1 Hz. Based on this knowledge, it becomes very common to employ a band-pass or low-pass filter like Butterworth allowing only a certain range of frequency to persist [3-6,8-11]. In addition, some groups improve on Butterworth filter by involving an adaptive computation of the cut off frequencies [3,10]. After applying the filter, the signal becomes cleaner. Our proposed method includes an original technique that can also be used for the purpose of cleansing the noisy data. Next, generally, a high dimensional accelerometer like a tri-axial accelerometer would be used. Therefore, many [4,5] have proposed a way to derive 1D respiratory signal out of a 2D/3D accelerometer. The process can be a manual selection of one dimension to be a representative of the signal. Bates et al. [5] suggest that we can compute an angular rate of breathing motions to deal with the problem of axis fusion. Some researchers prefer that a method based on principal component analysis (PCA) should be used to reduce the dimensionality [4,10]. In this paper, because we want to focus on our proposed algorithm, we employ a manual selection for simplification; however, our view is that a method like either angular rate derivation or
PCA can cooperate with our algorithm and would help improve a signal quality before operating. After cleaning, according to Bates [5], there are 3 respiratory derivation methods including peak findings (counting) [6], threshold crossing [5,11] and spectral analysis (Fourier analysis or autoregression) [3-4,10]. Our work involves the use of Fourier transformation to calculate the respiratory rate. It would be useful to state that our approach makes use of only a smartphone without any other special hardware so our approach fully takes an advantage of mobility provided by the smartphone. In addition, in general the capability of a built-in accelerometer equipped with a smartphone cannot be compared with a dedicated high quality (i.e. sensitivity) accelerometer being used in other publications. To the best of our knowledge, we are the first who reports the work that utilizes only a smartphone’s accelerometer. Nevertheless, we found related work by Ono et al. [9] which also makes use of an accelerometer of an iPod touch, but with some extra hardware. In addition, their approach seems to be far obviated from the common knowledge of the community and their result is shown to have limited success. III. OUR METHOD There are three main phases in our approach. The first step can be viewed as a pre-processing step as it is mainly for cleaning up the noisy input sensory data, a time series from an accelerometer. For the second step, the smoothed signal in a time domain will be transformed into a frequency domain using a Fourier transform algorithm. After that, the algorithm will continue to estimate a value of respiratory rate (RP) based on the power spectrum given by the output of the Fourier transformation. A. Pre-Processing Step In our first pre-processing part, there are two subtasks: 1) smoothing and 2) detrending. Firstly, for the smoothing task, our primary intention is to decrease any obscure or less important features found in the input and retain dominant characteristics or important features of the signal. We employ a technique called, moving average (MA), for this task. Theoretically, a time series can be decomposed additively into
yt Tt S t I t .
(1)
Where Tt = a trend component, St = a seasonal component and It = an irregular component. Then, a trend component is extracted using MA and the signal becomes
yt Tt .
(2)
We employ 13-term moving average for all experiments ranging from slow to fast breathing rate. We find that we do not have to alter a number of terms used in our experiments as it still can successfully estimate a trend from each data set.
After the noisy signal has been smoothed, we then proceed to the second subtask which is to detrend the signal. In order to have a clearer view of this particular step, we define the input for this step to be yt*, which would be the result of (2), i.e. yt* = yt. We repeat the MA technique again as we find that it has served us well for this detrending purpose. A trend component has been extracted, but this time is from yt*. Finally, to remove a trend, we proceed as follows
To compute the respiratory rate, firstly, as we knew that the respiratory frequency would be in a range of 0-1 Hz, so only the frequencies within this range are preserved. After that, there are still many peaks in the spectrum with different heights (power); however, only the highest peak is considered to be our fundamental frequency. Next, after we extract the dominant peak, we compute an estimate of the respiratory rate based on the breathing frequency of the selected peak. IV. EXPERIMENT AND RESULTS A. Experimental Setup We conducted an experiment where we recruited a healthy adult as a subject for measuring respiratory rate. The subject was asked to lie down with an iPhone 4 placed on her chest. We collected three sets of data where she was asked vary her breathing rate—breathe normally, breathe fast, and breathe slowly. Each data has 30 seconds of data. The accelerometer is set to log sensory data with a sampling frequency of 10 Hz. To get a baseline, we also asked a general practitioner to observe the chest movements and report the respiratory rate of each of these data sets. Lastly, we also include placements of a smartphone on a bed and a table for the purpose of being a control case. B. Data Collected The baseline collected from the countings by a general practitioner reported that the subject’s breathings are 12, 18 and 30 BPMs for slow, normal and fast breathing respectively. The result is consistent with reports for normal breathings in the literatures. The range of normal breathing rate of an adult slightly differs from source to source. For instance, [12] pointed out that 12-16 BPM is normal. On the other hand, 1220 BPM becomes normal according to [13]. These counting reports from the practitioner are for having a general sense of the actual rate of the data set. Although the operation is conducted by a professional, we agree that the counting could introduce some kind of human errors (which would include rounding errors.) Fig. 1 displays a plot of a time series of the sensory data. According to the figure, the chart is noisy; nevertheless, it clearly exhibits a sine-wave pattern. This is a positive sign for us as this suggests that our acceleration data may correspond to the breathing and would potentially infer the respiratory rate of a subject. We proceed by counting the number of waves in the signal for each data set and derive an estimate of respiratory rate based on this counting. We call the estimate from this method as WCRP. We find that the number falls within about the same range of an estimate from the chestmovement counting (will be referred to as HCRP) done by a professional. Further detail is reported in Table I. This finding would support our argument in using an accelerometer to estimate the respiratory rate.
yt* yt* Tt* .
(3)
At first it may be not obvious why this second subtask is required. The reason is because occasionally we find a trend either upwards or downwards in our certain data sets and this trend can collapse the outcome of Fourier transform in the next step. Therefore, the detrending process has been involved for eliminating the situation. B. Spectral Analysis We make use of a fast Fourier transform (FFT) algorithm to transform the data given in a time domain and convert it into a frequency domain. The output of FFT or power spectrum is composed of a series of frequency components associated with their powers. After that, the fundamental frequency, or a respiratory frequency, can then be revealed in the output spectrum. Now it is a good time to discuss about another reason for the detrending subtask in the previous step. It is for ensuring that the input signal has an average value of zero before transforming it with Fourier’s algorithm. We find that this step becomes essential in our work because an input signal without an average value of zero will yield perplexing output because the result, frequency components, are not only from a desired frequency but also from a constant factor silently resided in the signal. Moreover, in our case, we find that these two tend to be mixing together as we are calculating respiratory rate; the targeted frequency could fall in a very small range (i.e. between 0-1 Hz) and typical values would be as close as zero like 0.2 Hz which has RP of 12 breaths per minute (BPM). With these low frequency components, it can be easily obfuscated with zero and almost zero frequencies from a constant term hidden in the signal due to its non-zero average. This would yield higher complexity than necessary for further processing. Hence, detrending is applied in order to obtain a meaningful output and be able to do further analysis. C. Derivation of Respiratory Rate In our last step, we try to derive the final respiratory rate from the power spectrum generated from the previous step. Before we begin, all the frequencies with very low powers will be filtered out using a certain value of a threshold. The value of the threshold is selected in a heuristic fashion after conducting several experiments. By setting a threshold, we find that the algorithm performs better in differentiating a control case like bed and table from typical respiratory cases.
TABLE I COMPARISON OF HUMAN COUNT AND WAVE COUNT
Human Count (HCRP) 12 18 30
Wave Count (WCRP) 12 18 40
Fig. 3 RP18 after smoothing
Fig. 1 A trend found in RP30
In addition to the noisiness of the data, there are other incurring problems. For example, the signal can also contain either an upward or downward trend as appeared in Fig. 1. Moreover, Fig. 2 indicates that the breathing may not be done at a constant rate for a whole period of time. All of these would contribute into the more difficulty of the problem.
Furthermore, as we knew that the subject may not breathe with the same frequency all the time. The difference of breathing rates in the same signal would actually reconcile provided that the practitioner still find that the condition for HCRP is met. Then, to show the effectiveness of our algorithm, we carefully choose a specific range within a signal that can be a good representative of that particular data set. The criterion is that the WCRP from this shorter selected range should be similar to the WCRP shown in Table I. After this arrangement, we then can show that our proposed algorithm can estimate the value of respiratory rate as close as WCRP using the same period. For instance, in a case of fast breathing set RP30, we have WCRP of 40, which results in a frequency of 0.67 Hz (a period of 1.5 second.) In our trials, we find that we need at least 3 waves in order to obtain a reasonable result out of this data set using FFT. Therefore, we look for a range of 4.5 seconds that can fully cover 3 waves. For RP12 and RP18, we calculate the required number of seconds in the same fashion and find that they are 15 and 10 seconds, respectively. Illustrations of this selection are given in Fig. 1 for RP30 and in Fig. 3 for RP18 with a pair of red lines indicating a selected range. C. Results Next, we executed our algorithm on the selected portion of the signal. The final result, an estimate of respiratory rate from our method, ALRP, is shown in Table II.
TABLE II RESULTS OF OUR PROPOSED ALGORITHM
Fig. 2 A signal with different breathing frequencies
To cope with the noisy data, we employ a smoothing technique. Next, a detrending process has been used for removing a trend from the signal. We use 13-term moving average to handle the situation in both cases. Details of the smoothing and detrending techniques are given in part III. An example of results after smoothing is shown in Fig. 3.
Data Set RP12 RP18 RP30 bed table
Respiratory Rate (BPM) WCRP ALRP 12 11.72 18 18.75 40 37.50 N/A 0.00 N/A 0.00
V. CONCLUSIONS This paper proposes a method and an algorithm for measuring and monitoring human’s respiratory rates using the accelerometer data from smartphones. The device setup is designed to be simple—using only a commodity smartphone with no other devices. An algorithm that we developed, based on fast Fourier transform shows promising results in the experiment we conducted. This approach can be easily deployed and assist in measuring and monitoring respiratory rate by medical practitioners and normal users alike. REFERENCES
[1] [2] Fig. 4 A power spectrum of RP18 J. Hogan. "Why don’t nurses monitor the respiratory rates of patients?." British Journal of nursing 15, no. 9 (2006): 489-492. C. Takano, and Y. Ohta. "Heart rate measurement based on a timelapse image." Medical engineering & physics 29, no. 8 (2007): 853857. P. D. Hung, S. Bonnet, R. Guillemaud, E. Castelli, and P. T. N. Yen. "Estimation of respiratory waveform using an accelerometer." In Biomedical Imaging: From Nano to Macro, 2008. ISBI 2008. 5th IEEE International Symposium on, pp. 1493-1496. IEEE, 2008. A. Jin, B. Yin, G. Morren, H. Duric, and R. M. Aarts. "Performance evaluation of a tri-axial accelerometry-based respiration monitoring for ambient assisted living." In Engineering in Medicine and Biology Society, 2009. EMBC 2009. Annual International Conference of the IEEE, pp. 5677-5680. IEEE, 2009. A. Bates, M. J. Ling, J. Mann, and D. K. Arvind. "Respiratory rate and flow waveform estimation from tri-axial accelerometer data." In Body Sensor Networks (BSN), 2010 International Conference on, pp. 144150. IEEE, 2010. P. Kh. Dehkordi, M. Marzencki, K. Tavakolian, M. Kaminska, and B. Kaminska. "Validation of respiratory signal derived from suprasternal notch acceleration for sleep apnea detection." In Engineering in Medicine and Biology Society, EMBC, 2011 Annual International Conference of the IEEE, pp. 3824-3827. IEEE, 2011. S. Nemati, A. Malhotra, and G. D. Clifford. "Data fusion for improved respiration rate estimation." EURASIP journal on advances in signal processing 2010 (2010). P. Dehkordi, M. Marzencki, K. Tavakolian, M. Kaminska, and B. Kaminska. "Monitoring torso acceleration for estimating the respiratory flow and efforts for sleep apnea detection." In Engineering in Medicine and Biology Society (EMBC), 2012 Annual International Conference of the IEEE, pp. 6345-6348. IEEE, 2012. T. Ono, H. Takegawa, T. Ageishi, M. Takashina, H. Numasaki, M. Matsumoto, and T. Teshima. "Respiratory monitoring with an acceleration sensor." Physics in medicine and biology 56, no. 19 (2011): 6279. G. Z. Liu, Y. W. Guo, Q. S. Zhu, B. Y. Huang, and L. Wang. "Estimation of Respiration Rate from Three-Dimensional Acceleration Data Based on Body Sensor Network." Telemedicine and e-Health 17, no. 9 (2011): 705-711. G. B. Drummond, A. Bates, J. Mann, and D. K. Arvind. "Validation of a new non-invasive automatic monitor of respiratory rate for postoperative subjects." British journal of anaesthesia 107, no. 3 (2011): 462-469. S. Moses, Family practice notebook. 2013. Can be accessed online via http://www.fpnotebook.com/lung/exam/RsprtryRt.htm. W. Lindh, M. Pooler, C. Tamparo, and B. M. Dahl. Delmar's comprehensive medical assisting: administrative and clinical competencies. Cengage Learning, 2009.
We also display an example of a power spectrum of RP18 after the FFT step in Fig. 4. This intermediate output is used for determining a breathing frequency. Lastly, we plot our estimates in a real-time manner to reveal any inconsistency in breathing of the subject. In addition, this plot would make a clearer view that our algorithm is not restricted to only specifically selected range. We select the range just to establish a solid ground for our experiment that our algorithm can give a similar result to WCRP. An example of a real-time plot is displayed in Fig. 5. The plot is updated at every 2 second using last 15, 10 and 4.5 seconds for RP12, RP18 and RP30, respectively.
Real-time Respiratory Rate
[3]
[4]
[5]
[6]
Respiratory Rate
[7]
40 20 0
[8]
5 10 15 20 25
Respiratory Rate
40 20
[9]
[10]
0 5 10 15 20 25
Respiratory Rate
40 20 0
[11]
[12]
5 10 15 Time (sec) 20 25
[13]
Fig. 5 A real-time plot of RP12, RP18 and RP30 respectively
Thanakij Pechprasarn1, Suporn Pongnumkul2
National Electronics and Computer Technology Center 112 Thailand Science Park, Phahonyothin Road, Klong 1, Klong Luang, Pathumthani 12120, Thailand
1
[email protected] 2 [email protected]
Abstract— Abnormal respiratory rates have been shown to be an important predictor of serious clinical illness, but respiratory rate is a vital sign that is often not recorded because methods for measuring respiratory rates are cumbersome. We propose an approach to record and monitor respiratory rate of a patient that is lying down by placing an accelerometer-equipped smartphone on the patient’s chest. We develop an algorithm based on fast Fourier transform (FFT) to estimate the respiratory rate from the noisy acceleration data. The main contribution of this paper is that our proposed algorithm can estimate respiratory rates using only tri-axial acceleration data from sensor in commodity smartphones without any other special equipment. Preliminary results show that our method can reasonably estimate the respiratory rate. Keywords— respiratory rate, accelerometer, time series, moving average, Fourier transform
The algorithm we developed is based on the fast Fourier transforms. The data is first pre-processed to reduce noise using smoothing and detrending. Then we perform FFT to find the highest frequency. After that, we derived the respiratory rate from the three axes of signals by choosing the strongest frequency. We conducted an experiment to verify our methods, where we asked a healthy adult to lie down and had a general practitioner to measure the respiratory rate by counting the chest movements for about 30 seconds. The subject was asked to simulate fast breathing, normal breathing and slow breathing. We compare the observed respiratory rate from the general practitioners, the wave counts from the visual signals and the answer from our algorithm and found that our method can reasonably estimate the respiratory rate in all three cases. Our main contribution is the algorithm for estimating respiratory rates from just the acceleratory signals from commodity smartphones. The rest of this paper is organized as follows. Section II reviews the related work. Section III describes our algorithm. Section IV describes the experiment including the setup and the result. Then we conclude in Section V. II. RELATED WORK Respiratory rate is one of a vital sign that can be used to monitor the wellness of an individual [3-6]. It helps detect any malfunction of breathing activities, especially during sleeping. Respiratory rate is normally estimated from a recorded waveform. There are 2 main categories of waveform recording methods. In first category, a waveform of respiratory efforts is directly recorded. For example, impedance pneumography (IP) measures a respiratory activity through differential changes in capacitance. Respiratory inductive plethysmography (RIP) uses stretch sensors on the chest wall whereas flow thermography utilizes the changes in temperature of air flow when breathing. These methods can be viewed as a mainstream for the direct category. IP is the most widespread method used in a hospital whereas RIP is the most common method for overnight monitoring [4]. Another class, also classified to be in a direct category but much less common, includes the use of an accelerometer, a laser-based device, ultrasound, and audio and/or video processing. On the other hand, it is also possible to record a respiratory waveform
I. INTRODUCTION Respiratory rate is the number of breaths that a person takes in one minute while at rest. In practice, medical practitioners measure respiratory rate by counting how many times the chest moves up and down within one minute, or within 30 seconds and multiply the count by two. The method for measuring respiratory rate is tedious and time-consuming; therefore, it is a vital sign that is often neglected [1]. However, as respiratory rates may increase with fever, illness, or other medical conditions, it is an important predictor of serious clinical illnesses. The technologies for measuring respiratory rates are still an active area of research [2]. Most smartphones nowadays offer various built-in sensors and often include the tri-axial accelerometer, which measures the acceleration in three orthogonal directions. An accelerometer can be used to sense vibration, e.g. the vibration of a machine, orientation, e.g. in human activity monitoring. The tri-axial accelerometer is used as an inclinometer to reflect the abdomen or chest movement caused by respiration. We aim at creating a tool that can be used by medical practitioners and non-practitioners alike; therefore, we keep the device and the measurement methods simple. The device we use is an iPhone 4, a commodity smartphone that has triaxial accelerometers, and the measurement method is simply placing the smartphone on the patient’s chest for 30 seconds while the patient is in a lying down position.
978-1-4799-0545-4/13/$31.00 ©2013 IEEE
in an indirect fashion. However, this category would involve an extra step to rebuild and extract a respiratory waveform from other related waveforms. For instance, a respiratory waveform can be derived from electrocardiogram (ECG), photoplethysmogram (PPG), arterial blood pressure (ABP) and the peripheral arterial tonometry (PAT). Derivation of respiratory rate from ECG is an area that has been studied extensively compared to the others in the category. Moreover, recently researchers tend to include various signal sources simultaneously in order to improve a final estimate of respiratory rate. For example, a novelty proposed by Nemati et al. [7] is also based on data fusion which includes ECG, PAT, IP and PPG. Due to an emergence of microelectromechanical systems (MEMS)-based accelerometer [8], there is more recent published work that utilizes an accelerometer to estimate respiratory rate. Primitive reasons for the usage of an accelerometer, compared to traditional methods, are that it is a cheaper, non-invasive approach, and also viable to be used outside hospitals without a supervision of a professional. An accelerometer can be attached to a different part of the body, for example, a chest/thorax [3,8], a thorax-abdomen wall (including diaphragm muscle and lower costal margin) [4,5,9], an abdomen/waist [8,10] and even suprasternal notch [6,8]. Yet, currently there is no consensus on the best placement of the sensor. In our work, we decide to place a sensor on a patient’s chest. Besides placement, the posture of a person is reported to greatly affect the conducted breathing activity [5]. Thus, many require a patient to sit or lie down steadily to be able to successfully extract respiratory rate [3-6,8-9]. On the other hand, A novelty from Liu et al. [10] has studied the effects of posture changes like walking and running and still be able to compute respiratory rate out of that particular activities. Nevertheless, activity detection is not our main focus of this paper so we collect data only when a patient lies down steadily. It is proposed by previous publication [3] that breathing frequency is ranging between 0-1 Hz. Based on this knowledge, it becomes very common to employ a band-pass or low-pass filter like Butterworth allowing only a certain range of frequency to persist [3-6,8-11]. In addition, some groups improve on Butterworth filter by involving an adaptive computation of the cut off frequencies [3,10]. After applying the filter, the signal becomes cleaner. Our proposed method includes an original technique that can also be used for the purpose of cleansing the noisy data. Next, generally, a high dimensional accelerometer like a tri-axial accelerometer would be used. Therefore, many [4,5] have proposed a way to derive 1D respiratory signal out of a 2D/3D accelerometer. The process can be a manual selection of one dimension to be a representative of the signal. Bates et al. [5] suggest that we can compute an angular rate of breathing motions to deal with the problem of axis fusion. Some researchers prefer that a method based on principal component analysis (PCA) should be used to reduce the dimensionality [4,10]. In this paper, because we want to focus on our proposed algorithm, we employ a manual selection for simplification; however, our view is that a method like either angular rate derivation or
PCA can cooperate with our algorithm and would help improve a signal quality before operating. After cleaning, according to Bates [5], there are 3 respiratory derivation methods including peak findings (counting) [6], threshold crossing [5,11] and spectral analysis (Fourier analysis or autoregression) [3-4,10]. Our work involves the use of Fourier transformation to calculate the respiratory rate. It would be useful to state that our approach makes use of only a smartphone without any other special hardware so our approach fully takes an advantage of mobility provided by the smartphone. In addition, in general the capability of a built-in accelerometer equipped with a smartphone cannot be compared with a dedicated high quality (i.e. sensitivity) accelerometer being used in other publications. To the best of our knowledge, we are the first who reports the work that utilizes only a smartphone’s accelerometer. Nevertheless, we found related work by Ono et al. [9] which also makes use of an accelerometer of an iPod touch, but with some extra hardware. In addition, their approach seems to be far obviated from the common knowledge of the community and their result is shown to have limited success. III. OUR METHOD There are three main phases in our approach. The first step can be viewed as a pre-processing step as it is mainly for cleaning up the noisy input sensory data, a time series from an accelerometer. For the second step, the smoothed signal in a time domain will be transformed into a frequency domain using a Fourier transform algorithm. After that, the algorithm will continue to estimate a value of respiratory rate (RP) based on the power spectrum given by the output of the Fourier transformation. A. Pre-Processing Step In our first pre-processing part, there are two subtasks: 1) smoothing and 2) detrending. Firstly, for the smoothing task, our primary intention is to decrease any obscure or less important features found in the input and retain dominant characteristics or important features of the signal. We employ a technique called, moving average (MA), for this task. Theoretically, a time series can be decomposed additively into
yt Tt S t I t .
(1)
Where Tt = a trend component, St = a seasonal component and It = an irregular component. Then, a trend component is extracted using MA and the signal becomes
yt Tt .
(2)
We employ 13-term moving average for all experiments ranging from slow to fast breathing rate. We find that we do not have to alter a number of terms used in our experiments as it still can successfully estimate a trend from each data set.
After the noisy signal has been smoothed, we then proceed to the second subtask which is to detrend the signal. In order to have a clearer view of this particular step, we define the input for this step to be yt*, which would be the result of (2), i.e. yt* = yt. We repeat the MA technique again as we find that it has served us well for this detrending purpose. A trend component has been extracted, but this time is from yt*. Finally, to remove a trend, we proceed as follows
To compute the respiratory rate, firstly, as we knew that the respiratory frequency would be in a range of 0-1 Hz, so only the frequencies within this range are preserved. After that, there are still many peaks in the spectrum with different heights (power); however, only the highest peak is considered to be our fundamental frequency. Next, after we extract the dominant peak, we compute an estimate of the respiratory rate based on the breathing frequency of the selected peak. IV. EXPERIMENT AND RESULTS A. Experimental Setup We conducted an experiment where we recruited a healthy adult as a subject for measuring respiratory rate. The subject was asked to lie down with an iPhone 4 placed on her chest. We collected three sets of data where she was asked vary her breathing rate—breathe normally, breathe fast, and breathe slowly. Each data has 30 seconds of data. The accelerometer is set to log sensory data with a sampling frequency of 10 Hz. To get a baseline, we also asked a general practitioner to observe the chest movements and report the respiratory rate of each of these data sets. Lastly, we also include placements of a smartphone on a bed and a table for the purpose of being a control case. B. Data Collected The baseline collected from the countings by a general practitioner reported that the subject’s breathings are 12, 18 and 30 BPMs for slow, normal and fast breathing respectively. The result is consistent with reports for normal breathings in the literatures. The range of normal breathing rate of an adult slightly differs from source to source. For instance, [12] pointed out that 12-16 BPM is normal. On the other hand, 1220 BPM becomes normal according to [13]. These counting reports from the practitioner are for having a general sense of the actual rate of the data set. Although the operation is conducted by a professional, we agree that the counting could introduce some kind of human errors (which would include rounding errors.) Fig. 1 displays a plot of a time series of the sensory data. According to the figure, the chart is noisy; nevertheless, it clearly exhibits a sine-wave pattern. This is a positive sign for us as this suggests that our acceleration data may correspond to the breathing and would potentially infer the respiratory rate of a subject. We proceed by counting the number of waves in the signal for each data set and derive an estimate of respiratory rate based on this counting. We call the estimate from this method as WCRP. We find that the number falls within about the same range of an estimate from the chestmovement counting (will be referred to as HCRP) done by a professional. Further detail is reported in Table I. This finding would support our argument in using an accelerometer to estimate the respiratory rate.
yt* yt* Tt* .
(3)
At first it may be not obvious why this second subtask is required. The reason is because occasionally we find a trend either upwards or downwards in our certain data sets and this trend can collapse the outcome of Fourier transform in the next step. Therefore, the detrending process has been involved for eliminating the situation. B. Spectral Analysis We make use of a fast Fourier transform (FFT) algorithm to transform the data given in a time domain and convert it into a frequency domain. The output of FFT or power spectrum is composed of a series of frequency components associated with their powers. After that, the fundamental frequency, or a respiratory frequency, can then be revealed in the output spectrum. Now it is a good time to discuss about another reason for the detrending subtask in the previous step. It is for ensuring that the input signal has an average value of zero before transforming it with Fourier’s algorithm. We find that this step becomes essential in our work because an input signal without an average value of zero will yield perplexing output because the result, frequency components, are not only from a desired frequency but also from a constant factor silently resided in the signal. Moreover, in our case, we find that these two tend to be mixing together as we are calculating respiratory rate; the targeted frequency could fall in a very small range (i.e. between 0-1 Hz) and typical values would be as close as zero like 0.2 Hz which has RP of 12 breaths per minute (BPM). With these low frequency components, it can be easily obfuscated with zero and almost zero frequencies from a constant term hidden in the signal due to its non-zero average. This would yield higher complexity than necessary for further processing. Hence, detrending is applied in order to obtain a meaningful output and be able to do further analysis. C. Derivation of Respiratory Rate In our last step, we try to derive the final respiratory rate from the power spectrum generated from the previous step. Before we begin, all the frequencies with very low powers will be filtered out using a certain value of a threshold. The value of the threshold is selected in a heuristic fashion after conducting several experiments. By setting a threshold, we find that the algorithm performs better in differentiating a control case like bed and table from typical respiratory cases.
TABLE I COMPARISON OF HUMAN COUNT AND WAVE COUNT
Human Count (HCRP) 12 18 30
Wave Count (WCRP) 12 18 40
Fig. 3 RP18 after smoothing
Fig. 1 A trend found in RP30
In addition to the noisiness of the data, there are other incurring problems. For example, the signal can also contain either an upward or downward trend as appeared in Fig. 1. Moreover, Fig. 2 indicates that the breathing may not be done at a constant rate for a whole period of time. All of these would contribute into the more difficulty of the problem.
Furthermore, as we knew that the subject may not breathe with the same frequency all the time. The difference of breathing rates in the same signal would actually reconcile provided that the practitioner still find that the condition for HCRP is met. Then, to show the effectiveness of our algorithm, we carefully choose a specific range within a signal that can be a good representative of that particular data set. The criterion is that the WCRP from this shorter selected range should be similar to the WCRP shown in Table I. After this arrangement, we then can show that our proposed algorithm can estimate the value of respiratory rate as close as WCRP using the same period. For instance, in a case of fast breathing set RP30, we have WCRP of 40, which results in a frequency of 0.67 Hz (a period of 1.5 second.) In our trials, we find that we need at least 3 waves in order to obtain a reasonable result out of this data set using FFT. Therefore, we look for a range of 4.5 seconds that can fully cover 3 waves. For RP12 and RP18, we calculate the required number of seconds in the same fashion and find that they are 15 and 10 seconds, respectively. Illustrations of this selection are given in Fig. 1 for RP30 and in Fig. 3 for RP18 with a pair of red lines indicating a selected range. C. Results Next, we executed our algorithm on the selected portion of the signal. The final result, an estimate of respiratory rate from our method, ALRP, is shown in Table II.
TABLE II RESULTS OF OUR PROPOSED ALGORITHM
Fig. 2 A signal with different breathing frequencies
To cope with the noisy data, we employ a smoothing technique. Next, a detrending process has been used for removing a trend from the signal. We use 13-term moving average to handle the situation in both cases. Details of the smoothing and detrending techniques are given in part III. An example of results after smoothing is shown in Fig. 3.
Data Set RP12 RP18 RP30 bed table
Respiratory Rate (BPM) WCRP ALRP 12 11.72 18 18.75 40 37.50 N/A 0.00 N/A 0.00
V. CONCLUSIONS This paper proposes a method and an algorithm for measuring and monitoring human’s respiratory rates using the accelerometer data from smartphones. The device setup is designed to be simple—using only a commodity smartphone with no other devices. An algorithm that we developed, based on fast Fourier transform shows promising results in the experiment we conducted. This approach can be easily deployed and assist in measuring and monitoring respiratory rate by medical practitioners and normal users alike. REFERENCES
[1] [2] Fig. 4 A power spectrum of RP18 J. Hogan. "Why don’t nurses monitor the respiratory rates of patients?." British Journal of nursing 15, no. 9 (2006): 489-492. C. Takano, and Y. Ohta. "Heart rate measurement based on a timelapse image." Medical engineering & physics 29, no. 8 (2007): 853857. P. D. Hung, S. Bonnet, R. Guillemaud, E. Castelli, and P. T. N. Yen. "Estimation of respiratory waveform using an accelerometer." In Biomedical Imaging: From Nano to Macro, 2008. ISBI 2008. 5th IEEE International Symposium on, pp. 1493-1496. IEEE, 2008. A. Jin, B. Yin, G. Morren, H. Duric, and R. M. Aarts. "Performance evaluation of a tri-axial accelerometry-based respiration monitoring for ambient assisted living." In Engineering in Medicine and Biology Society, 2009. EMBC 2009. Annual International Conference of the IEEE, pp. 5677-5680. IEEE, 2009. A. Bates, M. J. Ling, J. Mann, and D. K. Arvind. "Respiratory rate and flow waveform estimation from tri-axial accelerometer data." In Body Sensor Networks (BSN), 2010 International Conference on, pp. 144150. IEEE, 2010. P. Kh. Dehkordi, M. Marzencki, K. Tavakolian, M. Kaminska, and B. Kaminska. "Validation of respiratory signal derived from suprasternal notch acceleration for sleep apnea detection." In Engineering in Medicine and Biology Society, EMBC, 2011 Annual International Conference of the IEEE, pp. 3824-3827. IEEE, 2011. S. Nemati, A. Malhotra, and G. D. Clifford. "Data fusion for improved respiration rate estimation." EURASIP journal on advances in signal processing 2010 (2010). P. Dehkordi, M. Marzencki, K. Tavakolian, M. Kaminska, and B. Kaminska. "Monitoring torso acceleration for estimating the respiratory flow and efforts for sleep apnea detection." In Engineering in Medicine and Biology Society (EMBC), 2012 Annual International Conference of the IEEE, pp. 6345-6348. IEEE, 2012. T. Ono, H. Takegawa, T. Ageishi, M. Takashina, H. Numasaki, M. Matsumoto, and T. Teshima. "Respiratory monitoring with an acceleration sensor." Physics in medicine and biology 56, no. 19 (2011): 6279. G. Z. Liu, Y. W. Guo, Q. S. Zhu, B. Y. Huang, and L. Wang. "Estimation of Respiration Rate from Three-Dimensional Acceleration Data Based on Body Sensor Network." Telemedicine and e-Health 17, no. 9 (2011): 705-711. G. B. Drummond, A. Bates, J. Mann, and D. K. Arvind. "Validation of a new non-invasive automatic monitor of respiratory rate for postoperative subjects." British journal of anaesthesia 107, no. 3 (2011): 462-469. S. Moses, Family practice notebook. 2013. Can be accessed online via http://www.fpnotebook.com/lung/exam/RsprtryRt.htm. W. Lindh, M. Pooler, C. Tamparo, and B. M. Dahl. Delmar's comprehensive medical assisting: administrative and clinical competencies. Cengage Learning, 2009.
We also display an example of a power spectrum of RP18 after the FFT step in Fig. 4. This intermediate output is used for determining a breathing frequency. Lastly, we plot our estimates in a real-time manner to reveal any inconsistency in breathing of the subject. In addition, this plot would make a clearer view that our algorithm is not restricted to only specifically selected range. We select the range just to establish a solid ground for our experiment that our algorithm can give a similar result to WCRP. An example of a real-time plot is displayed in Fig. 5. The plot is updated at every 2 second using last 15, 10 and 4.5 seconds for RP12, RP18 and RP30, respectively.
Real-time Respiratory Rate
[3]
[4]
[5]
[6]
Respiratory Rate
[7]
40 20 0
[8]
5 10 15 20 25
Respiratory Rate
40 20
[9]
[10]
0 5 10 15 20 25
Respiratory Rate
40 20 0
[11]
[12]
5 10 15 Time (sec) 20 25
[13]
Fig. 5 A real-time plot of RP12, RP18 and RP30 respectively
Sponsor Documents
Recommended
No recommend documents