Secondary Logo

Journal Logo

Technology, Computing, and Simulation: Research Report

Performance of Masimo Rainbow Acoustic Monitoring for Tracking Changing Respiratory Rates Under Laryngeal Mask Airway General Anesthesia for Surgical Procedures in the Operating Room

A Prospective Observational Study

Atkins, Joshua H. MD, PhD; Mandel, Jeff E. MD, MS

Author Information
doi: 10.1213/ANE.0000000000000362
  • Free

Continuous monitoring of changes in respiratory rate under a variety of clinical conditions may be useful in identifying patients at risk for adverse events. Changes in respiratory rate have been associated with intensive care unit admission in hospitalized patients,1 and continuous monitoring of heart rate and respiratory rate has been demonstrated to reduce total length of stay, intensive care unit admission days, and code-blue rates.2,3 The Masimo Pulse CO-Oximeter (Masimo Corporation, Irvine, CA) with acoustic monitoring technology measures respiratory rate based on analysis of acoustic signals generated across the upper airway during turbulent flow with breathing and has been compared with capnography for accuracy.4–7 Monitors designated for unattended use in awake patients typically use signal processing features to reduce false alarms, and such features may delay or even preclude detection of clinically significant events. We sought to determine the accuracy of this monitor under conditions of rapid transitions in respiratory rate.

Monitors that derive respiratory rates from counting breaths use information from a window that includes many breaths. Thus, the rate reported at any instant in time is correlated with the rates existing over the entire window. Signals with changing respiratory rates are termed nonstationary, and signal processing methods can recover the frequency at a single instant in time from nonstationary signals. The Hilbert–Huang Transform8 is such a method, and this method has been applied to a wide variety of signals, including respiratory signals, to assess coupling between respiratory flow and heart rate variability.9–11 The Hilbert–Huang Transform separates signals into independent basis functions of differing time scales. The separation is useful in signals that are composed of multiple physiologic processes, such as respiratory flow, which include respiration, cardiogenic oscillations, and various motion artifacts. When assessing the performance of a device measuring respiratory rate, agreement at stable respiratory rates may not be predictive of the ability to track changing respiratory rates. To investigate this issue, we examined the performance of the Masimo respiratory acoustic monitoring (RRa™) under conditions of general anesthesia with spontaneous ventilation through a laryngeal mask airway (LMA) because this permitted us to safely produce changes in respiratory rate by altering the variables of pressure support ventilation. We examined 3 issues: the delay of RRa in reporting a change of 1 breath per minute (bpm), the precision and bias of RRa compared to instantaneous rate estimates (IREs), and the rapidity with which disparities between RRa and IREs are resolved.


With approval of the IRB of the Perelman School of Medicine of the University of Pennsylvania and written patient consent, 53 patients scheduled for elective urological surgical procedures were prospectively enrolled in this prospective observational study from November 16, 2011, to March 14, 2012, in the Surgicentre of the Perelman Center for Advanced Medicine. In all patients, spontaneous ventilation with sevoflurane via LMA was the planned anesthetic. No muscle relaxants were used in any of the anesthetics. In all anesthetics, transitions between pressure support and manual ventilation modes, using the Drager Fabius (Drager USA, Telford, PA) anesthesia machine, were used to modulate ventilatory patterns. We had no reliable pilot data with which to perform a power analysis, but our experience with this mode of ventilation suggested that in a study of 50 subjects, we would produce a sufficient number of rate changes to permit analysis. All elements of the STROBE checklist (version 4) were addressed.

Before induction of anesthesia, an adhesive bioacoustic sensor (RAS-125, rev C) and oximetry sensor (M-LNCS Adtx) were placed on the patient’s neck and finger, respectively, following the manufacturer’s directions for use. These sensors were connected to a pulse CO-oximeter with rainbow acoustic monitoring technology (Rad-87, version 7805, Masimo). The default alarm setting for detection of respiratory pause was left unchanged at 30 seconds. A Hans Rudolph 4700B pneumotachometer (Hans Rudolph, Inc., Shawnee, KS) was placed proximal to the heat- and moisture-exchange filter in the anesthesia circuit. Data were acquired by a 12-bit A/D converter at 120 Hz using custom-built software developed in LabVIEW (National Instruments, Austin, TX). Data from the Rad-87 were acquired with proprietary software (ADC, Masimo) that reported the integer respiratory rate at 1-second intervals.

Induction of general anesthesia was performed with 8% sevoflurane supplemented with up to 40 mg IV propofol as needed. An AirQ Masked Laryngeal Airway (Cookgas, St. Louis, MO) was placed once a stable plane of deep anesthesia was achieved. Anesthesia was maintained with sevoflurane in oxygen titrated by the attending anesthesiologist to clinical variables. Ventilation was maintained with alternating spontaneous ventilation and pressure support delivered by a Drager Fabius ventilator. After completion of the procedure, sevoflurane delivery was halted and the LMA removed when deemed clinically appropriate by the attending anesthesiologist.

Pneumotachograph signals were visually inspected. Periods in which the investigators could not reliably count the number of breaths in a 60-second epoch by examining the raw waveforms were excluded from analysis. These included motion artifacts, ventilator valve chatter, and loss of sensor signals. In addition, periods in which no RRa data were obtained due to sensor detachment were excluded. Pneumotachograph signals were analyzed to yield IREs as detailed in the Appendix. Briefly, ensemble empirical mode decomposition using 450 realizations and noise scaling of 30% of the signal standard deviation was used on epochs of 90-second duration. The dominant intermediate mode function of the decomposition was identified, and the Hilbert–Huang Transform was applied to yield IREs. These IREs were filtered using a zero-phase low-pass Butterworth filter with cutoff frequency of 0.025 Hz before downsampling to 1 Hz to match the frequency response of the RRa.

The paired RRa and IREs samples were compared by 3 methods. The first analysis considered segments in which an increase or decrease of >4 bpm was observed. For each of these segments, IREs were quantized to the nearest integer, as depicted in Figure 1. The time difference between when IREs made a rate transition and when RRa made the same rate transition was determined for each rate transition. This permitted determination of the median delay (and interquartile range) between the IREs and the RRa. The RRa signal was shifted in time by this delay so the respiratory rates had the minimum phase difference.

Figure 1
Figure 1:
Respiratory rate determined by respiratory acoustic monitoring (RRa™) (blue), instantaneous rate estimates (IREs) (solid red), and IRE quantized to nearest integer (dashed red) from a representative epoch. Thirteen transitions in rate are depicted.

The second analysis used the segments identified in the previous analysis. Bland–Altman analysis for repeated measurements when the true value is changing was performed. Ninety-five percent confidence limits were calculated using 1-way analysis of variance scaled by degrees of freedom, as detailed on page 576 of the cited reference.12 Autocorrelation was assessed for the differences between measures.

The third analysis was the probability of a persistent disparity between the measures as a function of time. For each observation, the difference between the 2 methods was compared with 4 bpm. When this threshold was exceeded, the length of a run until the difference returned to within the threshold was noted. The probability of persistent error is the percentage of runs that exceed a given duration; the duration below which 90% of errors had resolved was determined. This is a conditional (Bayesian) probability because only segments in which the error exceeds the threshold are considered. For this measure, bootstrap analysis was performed with 1000 sets of randomly selected patients from the existing data sets to determine the 95% confidence intervals.

All statistical analysis was performed with MATLAB version 2013b (Mathworks, Natick, MA) with the Statistical and Econometric toolboxes.


Complete data sets were obtained from 50 patients; in 2 data sets no ADC records were obtained, and in 1 data set the RRa sensor was disconnected for the majority of the case. A total of 1416 minutes of respiratory data were analyzed; 147 minutes (median 3.8, interquartile range 4.8, range 0–33.9 minutes per patient) were excluded due to inability to determine respiratory rate from the raw waveforms, as described in Methods. Analysis of rate transitions in the first analysis identified 136 segments in which a change of respiratory rate >4 bpm was observed. For these 136 segments, the average rate change was 8.7 ± 4.6 bpm, and the average duration was 236 ± 196 seconds. A total of 1302 changes of 1 bpm were noted in these segments. For these 1302 changes, the median delay from a change in quantized IREs to the change in RRa was 45 seconds (interquartile range 20 seconds). A histogram of delay values is shown in Figure 2. Further comparisons of respiratory rates were made with this temporal adjustment.

Figure 2
Figure 2:
Time delay from rate transitions in quantized instantaneous rate estimates to rate transitions in respiratory acoustic monitoring (RRa™) for the 1302 rate transitions identified in phase 1 analysis.

Autocorrelation of the differences in respiratory rate was assessed for the 136 segments identified in the previous analysis using partial autocorrelation, which indicated the data fit an autoregressive model with 2 lags, consistent with the low-pass filtering applied to the instantaneous frequency data. Bland–Altman analysis is presented in Figure 3. A total of 32,091 paired observations of respiratory rate are represented in this bubble plot. Linear regression of error versus average of respiratory rate indicated no relationship between error and average of respiratory rates (R2 = 0.003, P < 0.001). The mean of the respiratory rates ranged from 5.5 to 48.5 bpm. The 95% limits of agreement were −2.1 to 2.2 bpm, with a mean error of 0.05 bpm. It should be noted that the limits of agreement are dependent on the filtering applied to the instantaneous respiratory rates. This is unavoidable, as the instantaneous rates are available at the sampling rate of the A/D converter (120 Hz) and must be downsampled for comparison to the RRa rates, which are provided once per second. The bias and linear regression are only trivially affected because the respiratory rates are within the pass band of the Butterworth filter.

Figure 3
Figure 3:
Bubble plot of the Bland–Altman analysis of respiratory data. The radius of the bubbles is proportional to bin count; bubbles with radius below 10% of maximum are thinned for visual clarity. Gray bars represent 95% limits of agreement (−2.1 to 2.2 bpm); the red line is the linear regression of error versus respiratory rate, error = −0.004 + 0.13 × RR (R 2 = 0.003, P < 0.001).

In general, RRa and IREs were in close agreement, and when RRa overestimated the respiratory rate, it was typically due to continuing to report the last reliable respiratory rate when the instantaneous rate was zero. This is illustrated in Figure 4. In the first 20 seconds, 3 breaths are noted, then a single breath 15 seconds later, and then no discernible breath for <30 seconds. Breathing resumes at twice the original rate. The IREs report the rate during the pause as zero; the RRa does not show a significant decrease in respiratory rate during this period because the threshold for detecting respiratory pause (set at 30 seconds) had not been reached. This is the expected performance of the RRa and can be modified by altering the respiratory pause threshold.

Figure 4
Figure 4:
Upper panel, Pneumotachograph signal. Lower panel, instantaneous rate estimates (IREs) rate (red) and respiratory acoustic monitoring (RRa™) rate (blue) during a respiratory pause induced by discontinuing pressure support ventilation. Respiratory rates are filtered with zero-phase low-pass filter with cutoff frequency of 0.025 Hz. Masimo RRa will display the most recently acquired respiratory rate until the pause detection algorithm is triggered after approximately 30 seconds.

Figure 5 depicts the probability of persistent error >4 bpm versus duration of error. Two hundred forty-three periods of such error were identified. The median time for resolution of 90% of these errors was 33 seconds; bootstrap analysis indicated a 95% confidence interval of 23 to 48 seconds. The longest error duration seen in the 50 patients was 160 seconds.

Figure 5
Figure 5:
Probability of persistent error (exceeding 4 bpm) versus error duration.


The significant finding of this study is that the RRa-derived respiratory rate tracks the rate obtained from the pneumotachograph over a wide range of rates, including rates consistent with those seen in sepsis and opioid-induced respiratory depression. When respiratory rate changed, RRa responded in approximately 45 seconds, which is likely to be sufficient for detection of changes in respiration in the majority of clinical settings. Although the rates frequently differed by a small number of bpm, and occasionally by a large number of bpm, these errors resolved quickly. The differences in rate were often the result of 2 different approaches to measuring respiratory rate. The RRa is determined in real time, with no access to future data and is based on a count of the number of breaths seen over a preceding time window. The IREs have full access to past and future data and extracts the respiratory rate from the first derivative of the signal phase at a single instant in time. This method was selected to permit an accurate estimate of the ability of RRa to rapidly detect changes in respiratory rate, and results obtained from this method can be expected to differ from other breath-counting methods to which RRa has been compared. Because IREs will have a low degree of correlation with past respiratory rates (attributable to filtering and downsampling of the signal), it will preserve chaos in respiratory rate variation, while breath-counting methods will suppress it. Although there may be useful information in the variation of respiratory period, the intent of RRa is to measure respiratory rates in patients using time windows, which are less sensitive to breath-to-breath variability. A value that is continually changing is less likely to be clinically useful, and most current clinical decisions are based on respiratory rates determined over a period of time ranging from 20 seconds to 1 minute. The other clinical decision variable that RRa provides, detection of respiratory pauses, is determined from continuous monitoring of breath sounds and is an independent algorithm from the respiratory rate calculation. This variable was not assessed in this manuscript but has been characterized elsewhere.4,7

This study has the limitation of being a single-center observational study. An additional limitation is the use of intraoperative data with LMAs. Although we had a small number of episodes of obstruction with the LMA, we neither attempted to assess the utility of the RRa in the presence of obstruction typically seen with natural airways nor the effects of behaviors such as chewing, talking, or coughing. A further limitation is the use of filtering of the IREs data, which may artificially narrow the 95% confidence intervals.


Under conditions of spontaneous ventilation during general anesthesia, RRa provides accurate estimates of respiratory rate changes over a wide range of respiratory rates. Although erroneous values can be obtained under conditions of transient apnea, these errors rarely persist for longer than 30 seconds. RRa is able to track changes in respiratory rate with minimal delay. To the extent that immediate knowledge of changes in respiratory rate is beneficial in early identification of patients at risk for adverse outcomes, RRa may be a useful clinical monitoring indicator.


Respiration is a chaotic process, and the breath period is not stable from one breath to the next. Such signals are neither stationary nor linear and thus are not adequately represented by Fourier decomposition. A new method for decomposing such signals is empirical mode decomposition (EMD).8 This method decomposes a signal into modes of differing time scales locally symmetric about zero. This decomposition is performed by obtaining the signal envelope by connecting the local maxima and minima by spline curves and then subtracting the average of these 2 curves from the original signal. The process is repeated until the energy of the subtracted signal is below a set threshold. The subtracted signal is termed an intermediate mode function. The process is repeated on the residual signal until the resulting residual no longer oscillates. A well-described problem with EMD is termed mode mixing, in which signals of different time scales are represented in a single mode. To avoid mode mixing, the system is implemented using ensemble EMD,13 which adds white noise to multiple copies of the signal (termed realizations) and then averages these to attenuate the noise. An example of decomposition of 16 seconds of spirometer signal is shown in Figure 6.

Figure 6
Figure 6:
Decomposition of spirometer signal into 11 intermediate mode function (IMF) signals. The dominant mode is IMF 8.

Modes 1 to 5 are largely composed of noise, modes 6 and 7 may represent cardiogenic oscillations, and modes 10 and 11 are slower processes than respiration. Intermediate mode function 8 contains the largest signal at what appears to be the respiratory rate. Although this signal may appear to be a sine wave, it is not, but it shares enough properties with a sine wave to allow us to perform the Hilbert transform, which recovers the corresponding cosine wave. We create a 3D plot with the original signal along real axis, the Hilbert transformed signal along the imaginary axis, and time along the z-axis, resulting in a spiral-shaped curve, as shown in Figure 7. At any point in time, the angle to the curve is the instantaneous phase of the signal. The first derivative of the phase is the instantaneous frequency, and the length of the chord is the instantaneous magnitude.

Figure 7
Figure 7:
The Hilbert representation of intermediate mode function 8. The original signal is plotted on the real axis, the Hilbert transform along the imaginary axis, and time along the z-axis.

Instantaneous frequency varies over the respiratory cycle but has the advantage of being uniformly sampled, while measures such as peak-to-peak interval are irregularly sampled and must be resampled at regular intervals for analysis. Before downsampling from the acquisition rate of 120 samples per second to once per second, the signal must be low-pass filtered. For this, we use a zero-phase filter to avoid delaying the signal. The filtered instantaneous frequency estimate for the sample signal is depicted in Figure 8.

Figure 8
Figure 8:
Instantaneous frequency estimate for intermediate mode function 8.


Name: Joshua H. Atkins, MD, PhD.

Contribution: This author helped design the study, conduct the study, and write the manuscript.

Attestation: Joshua H. Atkins has seen the original study data, reviewed the analysis of the data, and approved the final manuscript.

Conflicts of Interest: The author has no conflicts of interest to declare.

Name: Jeff E. Mandel, MD, MS.

Contribution: This author helped design the study, conduct the study, analyze the data, and write the manuscript.

Attestation: Jeff E. Mandel has seen the original study data, reviewed the analysis of the data, and approved the final manuscript and is the author responsible for archiving the study files.

Conflicts of Interest: Jeff E. Mandel served as a consultant for Masimo Corporation from October 2011 to October 2012 and served on a clinical advisory panel during that period.

This manuscript was handled by: Maxime Cannesson, MD, PhD.


Statistical review was provided by Rebecca Speck, PhD. Masimo Corporation provided funding and equipment used in this study. Masimo employees, Fatima Traore and Serop Gharibian, provided assistance in extracting respiratory rates from the ADC files, and Mark Holody and Elie Sarraf reviewed the manuscript prior to submission.


1. Cuthbertson BH, Boroujerdi M, McKie L, Aucott L, Prescott G. Can physiological variables and early warning scoring systems allow early recognition of the deteriorating surgical patient? Crit Care Med. 2007;35:402–9
2. Brown H, Terrence J, Vasquez P, Bates DW, Zimlichman E. Continuous monitoring in an inpatient medical-surgical unit: a controlled clinical trial. Am J Med. 2014;127:226–32
3. Churpek MM, Yuen TC, Park SY, Meltzer DO, Hall JB, Edelson DP. Derivation of a cardiac arrest prediction model using ward vital signs*. Crit Care Med. 2012;40:2102–8
4. Goudra BG, Penugonda LC, Speck RM, Sinha AC. Comparison of acoustic respiration rate, impedance pneumography and capnometry monitors for respiration rate accuracy and apnea detection during GI endoscopy anesthesia. Open J Anesthesiol. 2013;3:74–9
5. Mimoz O, Benard T, Gaucher A, Frasca D, Debaene B. Accuracy of respiratory rate monitoring using a non-invasive acoustic method after general anaesthesia. Br J Anaesth. 2012;108:872–5
6. Patino M, Redford DT, Quigley TW, Mahmoud M, Kurth CD, Szmuk P. Accuracy of acoustic respiration rate monitoring in pediatric patients. Paediatr Anaesth. 2013;23:1166–73
7. Ramsay MA, Usman M, Lagow E, Mendoza M, Untalan E, De Vol E. The accuracy, precision and reliability of measuring ventilatory rate and detecting ventilatory pause by rainbow acoustic monitoring and capnometry. Anesth Analg. 2013;117:69–75
8. Huang NE, Shen Z, Long SR, Wu MC, Shih HH, Zheng Q, Yen N-C, Tung CC, Liu HH. The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis. P Roy Soc Lond A Mat. 1998;454:903–95
9. Balocchi R, Menicucci D, Santarcangelo E, Sebastiani L, Gemignani A, Ghelarducci B, Varanini M. Deriving the respiratory sinus arrhythmia from the heartbeat time series using empirical mode decomposition. Chaos Soliton Fract. 2004;20:171–7
10. Bu N, Ueno N, Fukuda O. Monitoring of respiration and heartbeat during sleep using a flexible piezoelectric film sensor and empirical mode decomposition. Conf Proc IEEE Eng Med Biol Soc. 2007;2007:1362–6
11. Kuo TB, Yang CC, Huang NE. Quantification of respiratory sinus arrhythmia using Hilbert-Huang Transform. Adv Adapt Data Anal. 2009;1:295–307
12. Bland JM, Altman DG. Agreement between methods of measurement with multiple observations per individual. J Biopharm Stat. 2007;17:571–82
13. Torres ME, Colominas MA, Schlotthauer G, Flandrin P. A complete ensemble empirical mode decomposition with adaptive noise. IEEE Int Conf Acoustics Speech Sig Proc. 2011:4144–7
© 2014 International Anesthesia Research Society