A METHOD FOR MEASURING INSPIRATION AND/OR EXPIRATION
FIELD OF INVENTION
The present invention generally relates to a method for measuring inspiration and/or expiration in photoplethysmographic (PPG) respiratory rate monitoring.
Particularly, the present invention relates to utilisation of at least two signal components of a PPG signal for detection of inspiration and/or expiration.
BACKGROUND
Patients, who need to be observed, could among other things have their respiratory rate monitored.
One method of monitoring respiratory rate is to detect inspirations and expirations with photoplethysmographic (PPG). In SE 465551 it is described how to monitor the respiratory rate by using a PPG signal. However, this method is connected with a certain degree of erroneously detected inspirations and expirations, and therewith the respiratory rate is uncertain.
A photoplethysmographic (PPG) signal is obtained by measuring the absorption of light in blood and tissue. The dynamic part of the signal is caused by variations in the blood volume content of the tissue, but also, significantly, by blood perfusion [1]. The non-invasive character and the minimal disturbance of the variables under observation have made the technique widely used. The most widespread application of PPG is in monitoring of arterial oxygen saturation by pulse oximetry [2-3] wherein the cardiac synchronous component of the PPG signal is used. Other applications of the cardiac related component include cutaneous blood flow measurement and heart rate monitoring [1, 4-5].
A respiratory induced intensity variation (RIIV) of the PPG signal baseline is present [5]. The origin of the RIIV signal is not fully understood, but is believed to originate from the respiratory synchronous variations in intrathoracic pressure transmitted to the measurement site by the venous system [6- 8 ] . Respiratory related components originating from the arterial side of the circulatory system are also present in the PPG signal. These include the well-known respiratory sinus arrhythmia (RSA) and the amplitude modulation of the cardiac pulse [9-11].
During respiration, the fluctuation of the intrathoracic pressure is affecting central venous pressure and peripheral venous pressure. This causes a blood volume displacement due to the high compliance of the venous system, which is seen in the PPG signal as a respiratory variation in its baseline
(RIIV) . As the venous return to the right ventricle increases during inspiration, the left ventricle output decreases due to interventricular dependence and blood pooling in the pulmonary circulation. After a delay of a few heartbeats, the blood reaches the left ventricle and causes the stroke volume to rise. This effect is seen as an amplitude modulation of the cardiac component of the PPG signal. These two effects are relatively invariable with time, but somewhat obscured by the RSA. In general, the heart rate is increasing after an inspiration. RSA is considered to reflect vagal control over heart rate, but its underlying mechanisms are not entirely clear.
In a neural network (NN), a structure of mathematical units (neurons) is constructed. Each neuron calculates the sum of its inputs and uses an activation function for its output. The neurons are connected by weights that are adjusted in a training process, aiming at minimizing the network error
function. During the training, the backward propagation algorithm is often used [14]. The main advantage of neural networks is the possibility to generate a complex decision boundary without knowing the statistical properties of the data under observation. This has made the method attractive in medical applications, and neural networks are used in fields such as diagnosing vascular disease from pulse waveforms [15] and ECG classification [16]. In ventilation monitoring, neural networks are used to detect breathing circuit problems [17], in ventilation mode recognition [18] and to classify effective or ineffective respiration in neonates [19].
SUMMARY OF THE INVENTION
An object of the present invention is to improve reliability in breath detection for respiratory rate monitoring.
This object among others is attained by a method as claimed in claim 1 and in claim 11, and a neural network as claimed in claim 9.
The respiratory synchronous signal deduced from at least two different signal components of a PPG signal exhibit an overall improved respiratory detection accuracy compared to a corresponding respiratory synchronous signal deduced from a single signal component, due to the different characteristics of the signal components.
Further characteristics of the invention and advantages thereof will be evident from the following detailed description of embodiments of the invention.
BRIEF DESCRIPTION OF THE DRAWINGS
The present invention will become more fully understood from the detailed description of embodiments of the present
invention given below and the accompanying figures, which are given by way of illustration only, and thus are not limitative of the present invention, wherein:
FIG. 1 shows a table of falsely detected respirations, detected with a zero-crossing algorithm,
FIG. 2 shows a table of falsely detected respirations, detected with a peak algorithm,
FIG. 3 shows a table of falsely detected respirations, detected with a neural network
FIG. 4 shows a measurement set-up,
FIG. 5 shows respiratory components of a PPG-signal,
FIG. 6 shows a neural network,
FIG. 7 shows an example of signal components,
FIG. 8 shows patterns for neural network analysis, and
FIG. 9 shows a neural network output.
DETAILED DESCRIPTION OF EMBODIMENTS
In the following description, for purposes of explanation and not limitation, specific details are set forth, such as particular techniques and applications in order to provide a thorough understanding of the present invention. However, it will be apparent to one skilled in the art that the present invention may be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known methods and apparatuses are omitted so as not to obscure the description of the present invention with unnecessary details.
A photoplethysmographic (PPG) signal includes respiratory
components seen as frequency modulation of the cardiac pulse (respiratory sinus arrhythmia, RSA), amplitude modulation (PULSE), systolic waveform (SYST), diastolic waveform (DIAST), and respiratory induced intensity variations (RIIV) in the PPG baseline.
The accuracy of these components (SYST, DIAST, RSA, PULSE, and RIIV) in respiratory rate measurements has been evaluated with the result according to the tables in Figs. 1-2, where the evaluation was performed with an arrangement as shown in Fig. 4. A PPG reflection mode sensor was positioned laterally on the forehead of a test object by using a non-elastic velcro headband, asserting negligible pressure on the sensor. The sensor included eight light emitting diodes emitting infrared light of 940 nm and two photo detectors. The components were mounted in a silicone bed and light barriers were included to avoid direct transmission of light trough the substrate. The detector-diode distances were 3.5 and 5 mm respectively. The detected signal was processed by offset balancing and amplification [5]. As a respiratory reference, airflow humidity was sensed by a fiber-optic sensor. This sensor detects respiratory variations in the humidity accumulating on an optical fiber end positioned in front of the airways [20]. The reference signal was delivered from the device as a square wave. The test object has the sensor in his nose to certify a correct detection of inspirations and expirations, and yet not impede respiration. Both the PPG and the reference signal were AD-converted and stored in the memory of a personal computer.
The respiratory induced intensity variations (RIIV) in the baseline of the PPG signal were extracted by using a 16th order band-pass Bessel filter (0.13-0.48 Hz), and the cardiac related component by using a 5th order band-pass Butterworth filter (0.50-2.0 Hz). Both filters were digital filters. From
the cardiac signal, each peak and valley were detected by using a simple algorithm based on the zero crossings of its derivative. For each cardiac pulse (denoted i), the following components were formed:
Cji SYST, = Peak value (Systolic waveform)
C2: DIAST, = Previous valley value (Diastolic waveform)
C3: RSA, = Time since last peak _1 (Respiratory Sinus Arrhythmia )
C4: PULSE, = SYST,- DIAST, (pulse amplitude)
Furthermore, a discrete version of the RIIV signal was signed as:
C5: RIIV, = RIIV at time of SYST,
Analyses of the five respiratory components C_ (j=l,...,5) were performed separately. The baseline trend of a 2-minute period was removed by subtracting the smoothed function of the signal, yielding C'-fl (where i is the heart beat number) according to:
C'3.1=C.,1-(C43fl.4+...+C3fl+4)/9
Each period was also normalized to a mean value = 0 and standard deviation (SD) = 1 (yielding C'_(1). Two simple algorithms were then applied for breath detection:
(I) Zero crossing algorithm inspiration (expiration) when:
C '3,,>0 and C '3fl_ι<0
and the opposite for expiration (inspiration),
(II) Peak algorithm inspiration (expiration) when:
C " 3f l> C ' -i and C " .„>C " _„+1
and the opposite for expiration (inspiration).
Whether a true expression represents inspiration or expiration depends on the specific component. From the reference, each inspiration and expiration was noted, and the false positive (FP) and false negative (FN) breath detection rates were calculated for j=l,...,5 and for both algorithms.
The forehead was chosen as measurement position as:
1) The photoplethysmograms from the forehead are accurately reflecting arterial and venous blood volume variations, as nervous effects are minimal [21]
2) In reflectance pulse oximetry, the forehead is an appropriate choice [22]
3) As the forehead is highly vascularised and the frontal bone acts as a reflector, photoplethysmograms of high amplitude and signal-to-noise ratio are obtained [23-24], and
4) The sensitivity to patient motion is reduced at this position.
For these reasons, a traditional transmission measurement approach was not considered. Measurements were performed during spontaneous and metronome controlled breathing. The latter was used to produce different respiratory rates and breath-to-breath intervals, and thus, a more variable and realistic material.
When attempting to combine the respiratory components , pattern recognition in the form of a neural network was chosen [14], which has two advantages :
1) As the relation between the different respiratory
components of the PPG signal is complex and their statistical properties are not fully known, the decision boundary can be formed by the network, and
2 ) The implementation in a monitoring system is straightforward and the speed of processing (once trained) is high.
In order to structure the network, empirical rules and experience is required. The inputs to the present network consist of five well-known components of the PPG signal at five consecutive heartbeats after each respiratory event (inspiration/expiration or neither of the two). Five heartbeats represent a reasonable amount of time for the effects after each event to be fully captured. One hidden layer of neurons was assumed sufficient for this problem. This layer used 5 neurons, chosen from the rule of thumb saying that the number of hidden neurons should be the 2-logarithm of the number of inputs [19]. Standard error function minimization was performed by backward propagation. The training was stopped after a fixed number of epochs , and restarted if the error was above a specific level. This procedure was used to avoid local minima and to not over-fit the network to the training set. By shifting the signals one heartbeat a new pattern was formed. This caused each value to be seen by the network five times. However, as a new pattern was generated after each shift, effects of redundancies are supposed to have a minor influence on the results .
The PULSE parameter is actually a linear combination of two other inputs. This parameter was included anyhow, as it was plausible that this form of pre-processing would make it easier for the network to converge. This was confirmed when this parameter made the least number of errors in the separate analysis .
The patterns following expiration were less clear than those following inspiration. Therefore, only the first output of the network, representing inspiration, was used in the network evaluation. It is possible to further improve the system by combining the two outputs. Another important part was the determination of a suitable threshold to interpret the network output automatically. In the present evaluation, this threshold was set manually.
In the table in Fig. 1 the first column shows the number of errors, for false positives FP, for all test subjects as a percentage of the total number of reference breaths . The second column shows the error rate (SD), for false positives FP, as calculated on individual basis. The third and fourth columns show corresponding percentages for false negatives FN. The calculations in the table in Fig. 1 were analyzed with the Zero crossing algorithm.
The table in Fig. 2 shows corresponding values to those shown in Fig. 1 but in this case the calculations were analyzed with the Peak algorithm.
The best result, i.e. the lowest overall error rate, was found for the PULSE component analyzed by the Zero crossing algorithm (the PULSE row). In a preferred combination of PPG signal components, the PULSE component is one of the components .
Fig. 3 shows a table where the result is evaluated with a neural network. The number of false positive FP and false negative FN breath detections are presented for both natural and metronome measurements together with the total number of reference breaths n and the threshold tr used in the detection process.
In Fig. 5 it is shown an example of a PPG signal and how the
different signal components are defined. Extraction of the respiratory components of the PPG signal with five values from each cardiac pulse (denoted i) : the systolic value SYST,, the diastolic value DIAST,, the respiratory sinus arrhythmia RSA,, the pulse amplitude PULSE,, and the RIIV signal RIIV,.
The general effects of an inspiration seem to be a decrease in SYST and an increase in DIAST. This causes the PULSE parameter to fall. RSA is increasing after an inspiration and RIIV is decreasing. The opposites are seen at expiration.
A fully connected feed-forward neuron network is shown in Fig. 6, in which all five signal-components are used as inputs. Each signal component has five input neurons, the hidden layer has five neurons, and there are two output neurons Yl, Y2. The top output neuron Yl indicates, with a 1, if inspiration occurs, and the lower output neuron Y2 indicates, with a 1, if expiration occurs. At a specific time (i) and event (inspiration, expiration, or neither of the two) five consequent values of each component were normalized and used as inputs.
The components Cj-C.- were presented to the neural network as the values at five consecutive heartbeats at a specific time (i). The five values of each component were normalized to a range of 0-1. An output of [1 0] was assigned when the inputs followed inspiration, [0 1] when the inputs followed expiration and [0 0] when following neither of the two. Shifting the five signals one heartbeat, performing a new normalization and assigning a new output formed the next pattern. In this way, typically 60 patterns per minute measurement time were obtained (or more accurately, the number of heartbeats during the minute minus four) . Consequently, a network with twenty-five inputs and two outputs was obtained. The network was a fully connected feed-forward network, with
sigmoid activation functions and bias values used throughout.
In situations where some components are weak or disturbed (e.g. a less pronounced RSA in an elderly patient) the network will recognize the respiration from the other components. To do this effectively, it is important that the neural network has been trained by relevant data recorded in the specific environment (less pronounced RSA in the exemplary case).
Fig. 7 shows characteristic one-minute example (in metronome controlled breathing) of a cardiac signal, a RIIV "raw signal", and a reference signal (time as x-axis). Inspirations are marked as dots in the diagram. The five calculated respiratory components Cx-C5 are included (heartbeat number as x-axis) along the y-axis in arbitrary units.
The patterns for the neural network analysis are shown in Fig. 8. The boxes show for each component lower quartile, median and upper quartile values. The patterns are divided according to the output (inspiration, expiration, and others). The x- axes are heartbeat number after the event.
A typical two-minute neural network output (solid line) is shown in Fig. 9, i.e. a respiratory synchronous signal, where peaks above the threshold (dashed line) are regarded as inspirations . Reference inspiration points are marked in the figure, and the x-axis shows the heartbeat number.
Advantage of applying the PULSE component is that it originates from the "high-pressure" arterial system and is therefore less affected by environmental and physiological disturbances. In comparison, the RIIV signal, originating from the "low-pressure" venous system, is less accurate. However, it is still included as it is largely different in origin to the other components, which is a positive aspect in the pattern recognition. Further developments, focused on the most
accurate components will optimize a network structure.
To achieve a reliable respiratory rate the different PPG signal components are utilized in dependence on how well each respective component contributes to an accurate detection of inspirations or expirations. Detection of respirations is resolved in time in several parts, where each different component part is weighed against its probability of accurate detection of inspiration or expiration. The PULSE component, as identified above as the most accurate component, will probably provide the most often used component part. As complement, the other components will be used, when better identifying inspirations or expirations. A possible way of achieving this is to utilize a neural network as above described.
An advantage with the estimation of a respiratory rate by means of PPG signal components is the possibility to combine this with the estimation of arterial oxygen saturation (pulse oximetry) and/or cardiac pulse frequency by means of PPG.
It is also to be noted that the amplitude of the respiratory rate signal components may be used in order to estimate tidal volumes .
It has above been described that the respiratory rate has been estimated by means of optical signals. It is also possible to utilize the same signal components by means of other physiological signals, e.g. skin impedance measurements or blood pressure measurements.
The following references are hereby incorporated by reference:
[1] Challoner, A. V. J. Photoelectric plethysmography for estimating cutaneous blood flow. In "Noninvasive Physiological Measurements" (ROLFE, P. Ed), pp. 125-151. Academic Press,
London , 1979 .
[2] Aoyagi, T., Kishi, M. , Yamaguchi, K., and Watanable, S. Improvement of the earpiece oximeter. In Proc . Abstracts 13th Ann . Meet . Japanese Soc . Med. Electron . Biologic Eng. , Japan, 1974.
[3] Tremper, K. K., and Barker, S. J. Pulse oximetry. Anesthesiology 70,98-108 (1989).
[4] Eldrup-Jorgensen S., Schwartz, A. I., and Wallace, J. D. A method for clinical evaluation of peripheral circulation: Photoelectric hemodensitometry. Surgery 59,505-513 (1966).
[5] Lindberg, L-G., Ugnell, H., and Oberg, P. A. Monitoring of respiratory and heart rates using a fibre-optic sensor. Med . Biol . Eng. Comp . 30,533-537 (1992).
[6] Dorlas, J. C, and Nijboer, J. A. Photo-electric plethysmography as a monitoring device in anaesthesia. Br. J. Anaesth . 57,524-530 (1985).
[7] Ugnell, H. The respiratory synchronous photoplethysmographic signal. Its dependence on light wavelength and sample volume. Med . Bioi . Eng. Comput . 34, Suppl. 1, Part 1, 275-276 (1996).
[8] Johansson, A., and Oberg, P. A. Estimation of respiratory- volumes from the photoplethysmographic signal Part II: a model study. Med. Bioi . Eng. Comput . 37,48-53 (1999b).
[9] Kamath, M. V ., and Fallen, E. L. Power spectral analysis of heart rate variability: A noninvasive signature of cardiac autonomic function. Crit . Rev. Biomed . Eng. 21,245-311 (1993).
[10] Dornhorst, A. C, Howard, P., and Leathart, G. L. Respiratory variations in blood pressure. Circulation 6,553-
558 ( 1952 ) .
[11] Buda, A. J., Pinsky, M. R., Ingels, N. B. et al. Effect of intrathoracic pressure on left ventricular performance. N. Engl . J. Med. 301,453-459 (1979).
[12] Johansson, A., and Oberg, P. A. Estimation of respiratory volumes from the photoplethysmographic signal Part 1: experimental results. Med . Bioi . Eng. Comput . 37,42-47 (1999a) .
[13] Johansson, A., Νilsson, L., Kalman, S., and Oberg, P. A. Respiratory monitoring using photoplethysmography - Evaluation in the postoperative care unit. Proc . 20th An . Int . Conf . IEEE Eng . Med . Bioi . Soc . , Hong Kong, 3226, 1998.
[14] Bishop, C. M. Neural networks for pattern recognition', Oxford University Press, Oxford, 1995.
[15] Allen, J., and Murray, A. Development of a neural network screening aid for diagnostic lower limb peripheral vascular disease from photoelectric plethysmography pulse waveforms. Physiol . Meas . 14, 13-22 (1993).
[16] Watrous, R. , and Towell, G. A patient-adaptive neural network ECG patient monitoring algorithm. Computers in cardiology (Cat. No. 95CH35874), IEEE, New York, USA, 229-232, 1995.
[17] Sittig, D. F., and Orr, J. A. A parallel implementation of the backward error propagation neural network training algorithm: experiments in event identification. Comput . Biomed . Res . 25,547-561 (1992).
[18] Leon, M. A., and Lorini, F. L. Ventilation mode recognition using artificial neural networks. Comput . Biomed. Res . 30,373-378 (1997).
[19] Wilks, P. A. D., and English, M. J. A system for rapid identification of respiratory abnormalities using a neural network. Med . Eng. Phys . 17,551-555 (1995).
[20] Vegfors, M., Lindberg, L-G., Pettersson, H., and Oberg, P. A. Presentation and evaluation of a new optical sensor for respiratory rate monitoring. Int . J. Clin . Monit . Comput . 11,151-156 (1994).
[21] Ziege, S., Schmid-Schδnbein, H., Grebe, R. , and Martin, E. Long-term registration of cutaneous microcirculation during general anaesthesia. Int . J. Microcirc . 17,385-394 (1997).
[22] Mendelson, Y. Pulse oximetry: Theory and applications for noninvasive monitoring. Clin . Chem . 38,1601-1607 (1992).
[23] Tur, E., Tur, M. , Maibach, H. I., and Guy, R. H. Basal perfusion of the cutaneous microcirculation: Measurements as a function of anatomic position. J. Investigat . Dermatol . 81,442-446 (1983).
[24] Nijboer , J. A., and Dorlas, J. C. The origin of inverted waveforms in the reflection plethysmogram. Br. J. Anaesth . 54,1289-1293 (1982).
[25] Gilbert, R., Auchincloss, J. H., Brodsky, J., and Boden, W. Changes in tidal volume, frequency, and ventilation induced by their measurement. J. Appl . Physiol 33 , 252-254 (1972).