Journal of Clinical Monitoring and Computing (2004) 18: 309–312 C© Springer 2005

AN ALGORITHM FOR THE DETECTIONOF INDIVIDUAL BREATHS FROM THE PULSEOXIMETER WAVEFORMPaul Leonard, MBChB, MRCP,1 Neil R. Grubb,MBChB, MD, MRCP,2 Paul S. Addison, MEng, PhD,3

David Clifton, BEng, PhD,3 and James N. Watson, BSc,PhD3

From the 1Department of Accident and Emergency Medicine, TheRoyal Hospital for Sick Children, Sciennes Rd, Edinburgh, EH91LF, Scotland, UK, 2Department of Cardiology, The Royal Infir-mary of Edinburgh, Edinburgh, Scotland, UK, 3CardioDigital Ltd,Elvingston Science Centre, Glasdmuir, East Lothian, Scotland, UK.

Received 31 March 2004. Accepted for publication 30 August 2004.

Address correspondence to Paul Leonard, MBChB, MRCP, Depart-ment of Accident and Emergency Medicine, The Royal Hospital forSick Children, Sciennes Rd, Edinburgh, EH9 1LF, Scotland, UK.E-mail: paul leonard@blueyonder.co.uk

Leonard P, Grubb NR, Addison PS, Clifton D, Watson JN. An algorithmfor the detection of individual breaths from the pulse oximeter waveform.

J Clin Monit 2004; 18: 309–312

ABSTRACT. Objectives. To determine if wavelet analysis tech-niques can be used to reliably identify individual breaths from thephotoplethysmogram (PPG). Methods. Photoplethysmogramswere obtained from 22 healthy adult volunteers timing their res-piration rate in synchronisation with a metronome. A secondarytiming signal was obtained by asking the volunteers to actuate asmall push button switch, held in their right hand, in synchro-nisation with their respiration. Each PPG was analyzed usingprimary wavelet decomposition and two new, related, secondarydecompositions to determine the accuracy of individual breathdetection. Results. The optimal breath capture was obtainedby manually polling the three techniques, allowing detection of466 out of the 472 breaths studied; a detection rate of 98.7%with no false positive breaths detected. Conclusion. Our tech-nique allows the accurate capture of individual breaths from thephotoplethysmogram, and leads the way for developing a simplenon-invasive combined respiration and saturation monitor.

KEY WORDS. Photoplethysmography, breath detection, respirationrate, noninvasive measurement.

INTRODUCTION

Measurement of respiratory rate is an important part ofthe initial and continuing assessment of any unwell patient.Continuous monitoring of respiratory rate can be doneclinically, but is extremely labor intensive. Current prac-tice for the automatic measurement of patient respirationrequires the monitoring of CO2 levels using a capnograph,which is obtrusive (particularly in the non-ventilated pa-tient) and cumbersome to set up quickly; or the monitoringof transthoracic impedance, which is prone to movementartefact and requires the attachment of ECG leads. There isa clear need for a robust automatic continuous non-invasivemeasure of respiratory rate, that is suitable for patients ofall ages who are breathing spontaneously or with supporteither and breathing air or supplemental oxygen.

Pulse oximetry is a commonly used non-invasive tech-nique for monitoring oxygen saturation in blood based onthe differential absorption of red and infra-red light by oxy-genated haemoglobin and deoxygenated haemoglobin. Ithas been shown to provide an accurate measure of bothoxygen saturation and heart rate [1]. Most clinicians arefamiliar with the appearance of the photoplethysmogram(PPG), the plot of light absorption against time, displayedby most pulse oximeters although most only use it to de-termine the quality of the signal the machine is reading[2]. More recently, it has been shown that it is possibleto use signal analysis techniques to determine the respira-tory rate from the PPG [3–5], although these methods do

310 Journal of Clinical Monitoring and Computing Vol 18 Nos 5–6 2004

not have sufficient robustness for implementation within aclinically useful device as evidenced by their absence fromthe marketplace. In our previous work, we have deter-mined the respiration rate using wavelet transform-basedmethods [6] and more recently using novel secondary trans-form methods [7, 8] which we have developed to augmentour methods.

In this paper, we detail the results of an experimentalstudy to test our new methods in the detection of respira-tion features from within the PPG signal. Specifically, weassess their ability to locate individual breaths.

METHODS AND MATERIALS

Time frequency analysis in wavelet space

Wavelet transforms are a family of relatively new signalprocessing strategies that enable time-frequency unfoldingof time-domain signals. These techniques have particularadvantages over for example, Fourier-based techniques, forfocusing upon and extraction of pertinent features withinambiguous data sets [9]. The wavelet transform of a signalx(t) is defined as

T(a , b ) = 1√a

∫ +∞

−∞x(t )ψ∗

(t − b

a

)d t (1)

where ψ∗(t ) is the complex conjugate of the wavelet func-tion ψ(t ), a is the dilation parameter of the wavelet andb is the location parameter of the wavelet. The time-scalerepresentation given by Equation (1) can be converted to atime-frequency representation where the characteristic fre-quency associated with the wavelet is inversely proportionalto the scale a [9]. By interrogating the wavelet transformof the PPG signal, we can extract breathing information ina robust manner [6].

Experimental protocol

For the study 22 healthy volunteers (14 male, 8 female),aged 27 to 82, breathing air, were monitored in an uprightsitting position. Approval was obtained from the local ethicscommittee with informed consent given by all volunteers.An Oxisensor©R II probe was attached firmly to the in-dex finger of the left-hand with the optical source directedthrough the fingernail. In each trial, the volunteers wereasked to time their respiration rate in synchronisation withan on-screen bar-graph “metronome” set at a constant rateof 15 breaths per minute (0.25 hz). In order to record asecondary timing signal as a confidence measure for themetronome signal the volunteers were asked to actuate a

small push button switch, held in their right hand, in syn-chronisation with their inhalation (switch depressed) andexhalation (switch released) movements.

A Nellcor N100 pulse oximeter was used to monitorthe probe output in all trials. The red/infrared multiplexedphoto diode signal was tapped at the input stage (i.e., priorto any processing electronics) within the Nellcor machine.This signal was then de-multiplexed into separate red andinfrared signals prior to data logging. We used the infraredsignal for the work reported here. Such an arrangementwas deemed necessary to ensure that the recorded photo-plethysmogram was as “pure” and unprocessed as possible,in the order to avoid the possibility that breathing modu-lations would otherwise be corrupted by signal processingstages within the Nellcor machine.

The first 30 s of each patient trace was discarded to allowfor the initial transient (‘settling down’) period of the vol-unteer. After this initial period, 90 s of trace were analyzedusing a primary decomposition method (as described in [6])as well as two secondary transform methods, namely ridgeamplitude (RAP) and ridge frequency perturbation (RFP)analysis (as described in [7, 8]). The optimal breath capturewas determined by manually polling the three techniques,and selecting the most accurate at any particular point intime.

Individual breath detection

Breaths were detected from the phase signal output fromour ridge following algorithm [6]. We define a detectedbreath as a drop in phase of at least 90% of 2π in thealgorithm. This counts as a positive detection if it corre-sponds with the metronome phase within half a cycle ofthe metronome phase. If the phase curve leading to the−2π drop is erroneous, i.e., contains a significant miss-ing portion or has an obviously erroneous slope then thesebreaths were discounted (D). We disregard all such breaths,whether apparently detected or not as it is difficult to becertain that the detection is true (i.e., not simply due tothe co-incidental timing of a false positive breath with themetronome signal). Thus, for our technique,

Ba = Bc + FN + D (2)

where Ba is the number of actual breaths, Bc is the numberof correctly identified breaths, FN is the number of falsenegatives and D is number of the discounted breaths in theerroneous region. The number of false positives is separateto this relationship.

Previous studies have simply identified whether the to-tal number of breaths counted in a particular section of thePPG signal matches the number of breaths taken during the

Leonard et al: Individual Breath Detection 311

same period. This technique does not confirm individualbreath detection as false positives (FP) and false negatives(FN) may cancel each other out: a problem outlined byJohansson et al. [4] in their study concerning the moni-toring of neonatal respiration using the PPG. In our ap-proach we achieve an accurate temporal location of eachbreath through the metronome signal, verified by the pa-tient switch signal, and can therefore check the validity ofeach of our detected breaths in turn. Although our methodof counting FP, FN and D breaths gives a more conservativepicture, we believe it gives a more accurate representationof the robustness of the technology.

RESULTS

Table 1 contains a summary of the results from the breathdetection study for all 22 subjects using the methods out-lined above. Each method was used in turn to detect in-dividual breaths from the PPG signals from each patient.The optimal method(s) for each signal is presented in thetable for each patient together with the hit rate obtainedfor that method. As can be seen from the table the stan-dard breathing band algorithm (Method 1) is the optimumdetector for 15 out of the 22 traces. For eight out of these15, the breathing band algorithm can be matched by oneor more of the SWFD based algorithms (Methods 2 and3). However, in seven of the 22 cases the breathing bandalgorithm is surpassed in breath detection by one or both ofthe SWFD-based methods. By polling the optimal methodin each case (right hand column of Table 1), we detected466 out of the 472 breaths; a detection rate of 98.7%. Table1 also contains details of the false positive (FP), false neg-ative (FN) and discounted breaths (D), as defined above.Through manual polling, the number of false positives de-tected was reduced to zero and the percentages of falsenegatives and discounted breaths were both 0.6%.

Note that for one of the signals (volunteer 12) the signalsaturated during the experiment. This was not realized untilthe off line analysis phase. Hence, the results for this signalpertain to fewer breaths.

DISCUSSION

This study clearly shows that our techniques allow the ac-curate detection on individual breaths from the PPG signal.The lack of false positive detections is particularly encour-aging, as the clinical application of this technology wouldrequire the robust detection of low respiratory rates and ap-noea. Our results compare favorably with a previous study[10] who obtained a mean of 11.1% FP and 3.7% FN,

Table 1. Results of pulse oximetry breathing algorithm validationtrials

Optimal breathPatient No. Optimal method∗ capture results†

1 1 22/220 – 0 – 0

2 1 & 3 22/220 – 0 – 0

3 1 22/220 – 0 – 0

4 1 22/220 – 0 – 0

5 1 & 3 22/220 – 0 – 0

6 1 22/220 – 0 – 0

7 1 22/220 – 0 – 0

8 2 22/220 – 0 – 0

9 1 & 2 22/220 – 0 – 0

10 3 18/220 – 2 – 2

11 1, 2 & 3 22/220 – 0 – 0

12 1 & 2 10/100 – 0 – 0

13 1 22/220 – 0 – 0

13 1 22/220 – 0 – 0

14 2 22/220 – 0 – 0

15 2 & 3 22/220 – 0 – 0

16 1 & 3 22/220 – 0 – 0

17 1 20/220 – 1 – 1

18 2 22/220 – 0 – 0

19 3 22/220 – 0 – 0

20 1 & 2 & 3 22/220 – 0 – 0

21 2 & 3 22/220 – 0 – 0

22 1 & 2 & 3 22/220 – 0 – 0

Total captured breaths 466/472 = 98.7%Total FP 0/472 = 0.0%Total FN 3/472 = 0.6%Total D 3/472 = 0.6%

∗Method 1 is the primary decomposition method, method 2 is the ridge amplitude perturbation

(RAP) analysis and method 3 the ridge frequency perturbation (RFP) analysis†The top figures in each box are the correctly detected breaths/actual breaths, the lower figures

are FP – FN – D (False positives – False negatives – Discounted) Breaths.

312 Journal of Clinical Monitoring and Computing Vol 18 Nos 5–6 2004

respectively for their automated breath detection algorithmapplied to the PPG signal from a group of 16 patients re-covering from general anesthesia. They increased their de-tection by visually identifying each breath obtaining 4.6%FP and 5.8% FN, respectively.

The advantage of the CWT is that, due to its variablewindow width it can track rapid changes in frequencies intime unlike the short time-Fourier transform (STFT). Inaddition, it does not suffer from the cross terms inherentin other time-frequency methods (e.g. Wigner-Ville)which makes accurate ridge following impossible. Themethod relies on high resolution in wavelet space andhence the continuous wavelet transform is the preferredmethod as the time-frequency discretization employedby the related discrete wavelet transform (DWT) andthe stationary wavelet transform (SWT) is, in general,too coarse for the useful application of the method [9].The continuous wavelet transform is implemented in themethod through a fine discretization in both time andfrequency which the authors have found particularly goodat revealing hidden signal information from a variety ofother biosignals [11–13]. We are currently working on theautomation of the polling algorithm to ensure the optimaldetection of respiratory rate in real time. In addition, con-temporary signal analysis hardware is capable of supportingthe real-time wavelet analysis required by the method. Thehardware implementation of the algorithm is currentlybeing assessed with a view to produce a real-time breathingmonitor.

Current practice for the automatic measurement ofpatient respiration requires additional equipment andis obtrusive in nature. We believe that measurement ofrespiration directly from the photoplethysmogram wouldgreatly simplify the monitoring of patient respiration and,it is envisaged, lead to wider use of respiration monitoringin general.

The Authors would like to acknowledge the support of the WelcomeTrust through grant number 069078/Z/02/Z.

REFERENCES

1. Trafford J, Lafferty K. What does the photoplethysmograph mea-sure? Med Biol Eng Comput 1984; 22: 479–480.

2. Moyle JT. Pulse oximetry. London: BMJ Publishing Group,1994.

3. Ugnell H. Photoplethysmographic heart and respiratory mon-itoring: Instrument designa and evaluation. Linkoping Studiesin Science and Technology Dissertations No. 386, Departmentof Biomedical Engineering, Linkoping University, Linkoping,Sweden, 1995.

4. Johansson A, Oberg PA, Sedin G. Monitoring of heart and res-piratory rates in newborn infants using a new photoplethysmo-graphic technique. J Clin Monit Comput 1999; 15: 461–467.

5. Lindberg L-G, Ugnell H, Oberg PA. Monitoring of respiratoryand heart rates using a fibre-optic sensor. Med Biol Eng Comput1992; 30: 533–537.

6. Leonard P, Beattie TF, Addison PS, et al. Standard pulse oxime-ters can be used to monitor respiratory rate. Emerg Med J 2003;20: 524–525.

7. Addison PS, Watson JN. Secondary wavelet feature decoupling(SWFD) and its use in detecting patient respiration from the pho-toplethysmogram. 25th Annual International Conference of theIEEE Engineering in medicine and Biology Society, Cancun,Mexico, 17–21 September 2003.

8. Addison PS, Watson JN. Secondary transform decouplingof shifted nonstationary signal modulation components: Ap-plication to photoplethysmography. International Journal ofWavelets, Multiresolution and Information Processing, 2004. (Inprint.)

9. Addison PS. The Illustrated wavelet transform handbook, Insti-tute of Physics Publishing, Bristol; 2002.

10. Nilsson L, Johansson A, Kalman S. Monitoring of respiratoryrate in postoperative care using a new photoplethysmographictechnique. J Clin Monit Comput 2000; 16: 309–315.

11. Addison PS, Watson JN, Clegg GR, et al. Finding Coordinatedatrial activity during ventricular fibrillation using wavelet de-composition. IEEE Eng Med Biol Mag 2002; 21: 58–65.

12. Watson JN, Addison PS, Clegg GR, et al. Evaluation of ar-rhythmic ECG signals using a novel wavelet transform method.Resuscitation 2000; 43(2): 121–127.

13. Watson JN, Addison PS, Grubb N, Clesgg G, Robertson C, FoxKAA. Wavelet-based filtering for the clinical evaluation of atrialfibrillation. 29th Annual International Conference of the IEEEEngineering in medicine and Biology Society, Istanbul, Turkey;25–28 October, 2001.