Estimation of Respiratory Rate from Smartphones Acceleration DataThanakij PechprasarnSuporn Pongnumkulthanakij.pechprasarn@nectec.or.thsuporn.pongnumkul@nectec.or.th*

OutlineMotivationRelated workOur methodAlgorithmData setExperimentResultSummary*

Motivation (1)Vital signsTemperatureRespiratory rate (RP)Heart rate (pulse)Blood pressure

*

Motivation (2)*

RP EstimationDirect MethodIPRIPaccelerometerlaser-basedultrasoundaudio processingvideo processingIndirect MethodECGPPGABPPAT*

Related Work(accelerometer approach)2008 Hung et al. band-pass filter2009 Jin et al. PCA2010 Bates et al. rotation and movement2011 Dehkordi et al. evaluated RP signal2011 Liu et al. adaptive filter with BSN2012 Dehkordi et al. models with ensemble learning

*

Our Method*

Algorithm (1)*pre-processingspectral analysisderivationRPXYZselect one

Algorithm (2)Pre-processingSmoothingDetrendingSpectral analysisFast Fourier transform (FFT)RP derivation

*RP = f * 60

Algorithm (3)Time series:

Tt = trend componentSt = seasonal componentIt = irregular component

*

Algorithm (3)Time series:

Tt = trend componentSt = seasonal componentIt = irregular component

*estimated with moving average (MA)

Algorithm (3)Time series:

Tt = trend componentSt = seasonal componentIt = irregular component

For smoothing, *estimated with moving average (MA)

Algorithm (3)Time series:

Tt = trend componentSt = seasonal componentIt = irregular component

For smoothing, For detreding, *estimated with moving average (MA)

Data Set*RP unit = breaths per minute

Data SetBreathingRP12slowRP18normalRP30fast

ExperimentEach data set is labeled according to the counting done by practitioner referred as HCRPIntroduce another RP estimate called WCRP by counting the waves in the actual acceleration data*RP = count/time * 60

HCRP, WCRP*

Human Count (HCRP)Wave Count (WCRP)121218183040

Result*

Data SetRespiratory Rate (BPM)WCRPALRPRP121211.72RP181818.75RP304037.50bedN/A0.00tableN/A0.00

SummaryPropose a method to estimate respiratory rate using only smartphone

*

Future WorkExperiment with more dataCompare with existing methodsImprove on the algorithm such as automatic selection of 1D from 3D data*

*

WCRPall, WCRPrange*

Data SetWCRPfWCRPTWCRP3*TWCRPRP12120.205.0015RP18180.303.3310RP30400.671.504.5

Real-time Plot*

*

WorkHWSampling FrequencyPlacementOrientationTarget RPGold StandardHungADXL204 (2D acc) with belt, power supply1 kHz (14-bit DAQ-6024E)chestsit, stand, lie downnormal, apnea, deepRIPJin3D acc10 Hzdiaphragm musclesit, stand, lean, lie down *simulated5 breathing cyclesIPBatesMMA7260QT (3D acc), Orient-3 (wireless), ADXRS300 (gyro)64 Hz (12-bit dsPIC30F3014)left lower costal margin, below the ribsnormal, irregularnasal cannula pressure transducersDehkordiADXL335z (3D acc)2 kHz (8-bit NI9205)suprasternal notchlie down (supine, prone, left side)deep, normal, shallowpressure spirometry and SGRLiuSCA3000 (3D acc), BSN25 Hzwaist-wornairflow CO2 analysis with BIOPAC CO2100CDehkordiADXL335z (3D acc)500 Hz (8-bit NI9205)suprasternal notch, the thorax and the abdomenlie down (supine, prone, left side)apneanasal/oral airflow

Acceleration Waveform*

PPG vs ABP vs ECG*

ReferenceP. D. Hung, et al. "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, et al. "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, et al. "Respiratory rate and flow waveform estimation from tri-axial accelerometer data." In Body Sensor Networks (BSN), 2010 International Conference on, pp. 144-150. IEEE, 2010.P. Kh. Dehkordi, et al. "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.G. Z. Liu, et al. "Estimation of Respiration Rate from Three-Dimensional Acceleration Data Based on Body Sensor Network." Telemedicine and e-Health 17, no. 9 (2011): 705-711.P. Dehkordi, et al. "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.*

Round-off ErrorIn case of HCRP and WCRP, counting actually introduces round-off error as it can only accept integer.On the other hand, ALRP produces real value, which would eliminate round-off error.*

Device (1)Impedance pneumography (IP)*

Device (2)Respiratory inductive plethysmography (RIP)*

Device (3)Electrocardiogram(ECG, EKG)*

Device (4)Photoplethysmogram (PPG)*

Device (5)Arterial blood pressure (ABP)*

Device (6)Accelerometer*

Data CollectionWe use SensorLog and iOS app to help collect the acceleration dataWith a sampling rate (fsampling) of 10 HzThe actual smartphone is iPhone 4*https://itunes.apple.com/us/app/sensorlog/id388014573?mt=8

Moving AverageMA is a common technique in drawing a trend line for time series data.*MA = average of nearby terms

SignalCope with noisy, shifts in frequency, a trend found in some data sets*

3D-axis Accelerometer*

Reasons for Detrending SubtaskThe signal occasionally contains a trend which corrupt FFTAverage value of zero, non-zero average (constant factor) will contribute to the final frequency of FFT*

Good morning. My name is Thanakij Pechprasarn and I am from NECTEC.I am delighted to be here today to talk about my research work,which is to estimate respiratory rate from smartphones acceleration data.*Here is an outline of my slide.*To provide you some background, first of all vital signs are very important in clinical medicine.The word vital in vital sign mean critical as it can help doctor quickly justify the level of an illness.The more degraded in the vital sign, the more sicker the patient is. (http://meded.ucsd.edu/clinicalmed/vital.htm)There are many vital signs. For example, temperature, heart rate, blood pressure, and also respiratory rate or breathing rate.I became interested in respiratory rate because most of the time it is the vital sign that is being neglected.One reason is that because it is a kind of tedious tasks as it requires a personnel like a nurse to frequently keep checking.However, respiratory rate is particularly important in a case of sleep-disordered breathing (or sleep apnea.)*Patients with sleep apnea have a lower level of breathing or even actually stop breathing at all while sleeping.Therefore, these patients have a significant higher risk of death.Yet, sleep apnea is a common physical disorder that affects millions of adults and children. (http://www.hss.energy.gov/sesa/corporatesafety/feosh/docs/Sleep_Apnea.pdf)This suggests that we need an automatic approach to help monitor the breathing rate of the patients.

*There are many methods being proposed to automatically estimate the respiratory rate.The methods can be classified into 2 categories: direct method and indirect method.The direct method makes use of the actual effect of breathing to estimate the rate.For example, IP (impedance pneumography - changes in the electrical impedance) , RIP (or respiratory inductive plethysmography - changes in size of part of the body) and accelerometer (changes in acceleration of part of the body).On the other hand, indirect method makes use of non-respiratory signals to estimate the respiratory rate. For example, heart rate (ECG) or blood pressure (ABP)Nevertheless, today our approach is using an accelerometer-based approach.

*Next, I will provide related work in this particular idea.In 2008, Hung established a fundamental work with band-pass filter.In 2009, Jin made use of PCA to select 1D out of 3D data.In 2010, Bates dealt with rotation and movement.In 2011, Dehkordi evaluated the respiratory signal.Also, in 2011, Liu employed an adaptive band-pass filter with body sensor network.In 2012, Dehkordi proposed models with an ensemble learning approach.*Nowadays people are using smartphones in their daily livesand these smartphones have already equipped with an accelerometer.It would be useful and very convenient if users can use their smartphones to measure their respiratory rates.Therefore, we propose a method that can estimate respiratory rate using only smartphones.*Next, this diagram is showing an overview of our algorithm.Firstly, out of 3D acceleration data, XYZ, only 1 dimension will be selected.Note that we manually select the strongest signal in this part.Next, we try to reduce the noise in the input signal with our pre-processing unit.We then employ spectral analysis technique to transform the data in time domain into frequency domain.After obtaining the fundamental frequency, we calculate the final RP*There are 3 main tasks in our algorithm.The first one is the pre-processing step which includes 2 subtasks: smoothing and detrending.The smoothing is for removing the noise in the signalwhereas the detrending is for removing the trend.Next, for the spectral analysis, we employ fast Fourier transform to extract fundamental frequency,which would be corresponding to the desired breathing rate.And the last task is the derivation of the RP.*Time series data can be decomposed into a trend component,a seasonal component and an irregular component.*We estimate a trend component with moving average.*For smoothing, we think of the meaning of this subtask which is to remove the noise.So we set the signal output to be the estimated trend.*Next, for detrending, now we want to remove a trend.So this time we subtract the trend instead.*Next, for the data set, we to have RP18 for representing the normal breathing rate of an healthy adult.We also have RP12 and RP30 for slow and fast breathing rates.

For the detail of our data collecting process,firstly, we ask a patient to lie down on a bed attached with a smartphone on his/her chest.Next, we ask a patient to simulate one breathing rate either slow, normal, or fast breathing rate.and at the same time we have a practitioner to count the breaths for about 30 seconds.*Then, each data set is labeled according to the counting done by practitioner, which will be referred as HCRP.

However, this can lead to many errors such as human error (counting, timing) and round-off error.So in our experiment, we introduce another RP estimate called WCRP by counting the number of waves in the actual acceleration data.

(HCRP is estimated from counting the chest movementswhereas WCRP is estimated from counting the number of waves.)*Using the same period, we calculate the value of WRP andfind that the value of WCRP is as similar as HCRP.which shows that in general our acceleration data can be used to estimate RP.

Yes, there is a big difference in RP30, which became very hard to sayAs we have little clue for further investigation.

(Note that WCRP can still introduce round-off error.)*RP estimates from our algorithm, ALRP, are shown in the rightmost column.The result shows that our method reasonably estimates the RP in all cases.*Our future work will include a greater number of sets of data which would let us know how well our algorithm perform on different data sets.In addition, we could give audience a clearer view of our proposed methods performance by comparing it with existing methods.The last would be to improve on our algorithm itself such as automatic selection of 1D out of 3D axis data.*In our experiment, we find that we need 3 waves for our algorithm to work well with all data sets.So we find WCRP of 3-wave that have the value as similar as WCRP of whole period.Those selected range will be used by our algorithm to estimate RP called ALRP.***