METHOD AND APPARATUS FOR PHYSIOLOGICAL MONITORING
1. A method of detecting the frequency of a periodic physiological signal of a subject from a video image of the subject, comprising the steps of:
- fitting autoregressive models of one or more orders to a signal derived from the video image to detect spectral components in the signal;
for each spectral component generating a synthetic signal of the same frequency and calculating its similarity in the time domain with the signal derived from the video image;
outputting as the physiological signal frequency the frequency of the synthetic signal with the maximum similarity with the signal derived from the video image.
Autoregressive modelling is used to identify periodic physiological signals such as heart rate or breathing rate in an image of a subject. The colour channels of a video signal are windowed and normalised by dividing each signal by its mean. The ratios of the normalised channels to each other are found and principal component analyses conducted on the ratio signals. The most periodic of the principal components is selected and autoregressive models of one or more different orders are fitted to the selected component. Poles of the fitted autoregressive models of different orders are taken and pure sinusoids corresponding to the frequency of each pole are generated and their cross-correlation with the original component is found. Whichever pole corresponds to the sinusoid with the maximum cross-correlation is selected as the best estimate of the frequency of periodic physiological information in the original video signal. The method may be used in a patient monitor or in a webcam-enabled device such as a tablet computer or smart phone.
|SIGNAL PROCESSING METHOD AND APPARATUS|
Patent #US 20170332978A1
Current AssigneeOxford University Innovation Limited
Sponsoring EntityOxford University Innovation Limited
|RESPIRATION MOTION STABILIZATION FOR LUNG MAGNETIC NAVIGATION SYSTEM|
Patent #US 20180055576A1
Current AssigneeCovidien PLC
Sponsoring EntityCovidien PLC
|SLEEP APNEA THERAPY ENHANCEMENT METHOD AND APPARATUS|
Patent #US 20180206762A1
Current AssigneeIntel Corporation
Sponsoring EntityIntel Corporation
- 1. A method of detecting the frequency of a periodic physiological signal of a subject from a video image of the subject, comprising the steps of:
fitting autoregressive models of one or more orders to a signal derived from the video image to detect spectral components in the signal; for each spectral component generating a synthetic signal of the same frequency and calculating its similarity in the time domain with the signal derived from the video image; outputting as the physiological signal frequency the frequency of the synthetic signal with the maximum similarity with the signal derived from the video image.
- View Dependent Claims (2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25)
The present invention relates to the remote monitoring of a human or animal subject to detect periodic physiological signals such as the heart rate or breathing rate.
The detection of periodic physiological signals, such as heart rate or breathing (respiration) rate, of a subject is a vital part of patient monitoring. Traditionally contact sensors have been used so that the heart rate can be detected by means of an ECG monitor or PPG finger probe and the breathing rate can either be derived from the ECG or PPG signals or measured directly by impedance pneumography. A PPG finger probe also gives the blood oxygen saturation (SpO2). However contact-based sensors are not ideal for monitoring patients who move or are mobile, or for long-term monitoring. More recently, therefore, proposals have been made for non-contact monitoring. One such proposal is to derive measures of the heart rate and breathing rate from a photoplethysmographic image (PPGi). As discussed in the applicant'"'"'s own published patent application WO-A2-2013/027027, a video image of the patient'"'"'s skin can be analysed to detect the variation in transmittance or reflectance of light at certain wavelengths as the volume of blood in the skin capillaries varies with the cardiac cycle. Although invisible to normal sight, the skin can effectively be regarded as pulsing more red and less red with each heart beat. This change can be detected in a standard colour video image of the skin taken with a normal video camera such as a webcam. WO-A2-2013/027027 explains how the heart rate and breathing rate can be derived from such a colour video signal.
The aim is then to spectrally analyse the signal to find a periodic variation corresponding to the heart rate (or breathing rate). In this prior art this is done in steps 105 to 107 by fitting autoregressive models to the signal. Thus in step 105 several autoregressive models of different orders are fitted to the signal (for example orders 8 to 20). As explained in WO-A2-2013/027027 autoregressive modelling can be regarded as filtering with an all-pole infinite impulse response filter (IIR) with a white noise input in which the number of poles is determined by the model order. In spectral analysis it is the properties of these poles which are of interest. The poles can be viewed as points lying in the complex plane in an area conscribed by the unit circle, having a phase angle that is related to a pure frequency component of the signal and a magnitude that is related to the strength of that component. The analysis of the signals in step 105 generates, for each model order, several poles representing different frequencies present in the signal. In order to find which pole represents the heart rate (or breathing rate) poles representing frequencies outside the allowed range for heart rate (or breathing rate) can be ignored and then the pole with the largest magnitude and/or lowest frequency, can be selected. The phase angle of the selected pole corresponds to a particular frequency assumed to correspond to the heart rate estimate from that model. The estimates from each of the different model orders are taken and averaged in step 107, and in step 108 the heart rate estimate based on that average is output.
It will be appreciated that because the frequency estimate from each model order is based on the selected pole, the selection of that pole is of critical importance. However it can be difficult, in practice, to select the correct poles from the fitted autoregressive models. Selection of the wrong pole will result in an erroneous frequency estimate. Further, although it is anticipated that remote monitoring will be better than contact-based sensors in the case of moving subjects, movement of the subject in the video image can create movement artefacts which make signal processing difficult. It is very well known from video encoding technologies how to track parts of a video image from one frame to the next, and such tracking can be used to keep the region of interest on the skin of the subject. Nevertheless it would be useful to improve the motion compensation in order to improve the signal before spectral analysis. PPGi analysis can also be sensitive to lighting changes and it would also be useful to improve the robustness of the method to such changes, e.g. flicker in ceiling lights.
One aspect of the present invention provides a method of detecting the frequency of a periodic physiological signal of a subject from a video image of the subject, comprising the steps of: fitting autoregressive models of one or more orders to a signal derived from the video image to detect spectral components in the signal; for each spectral component generating a synthetic signal of the same frequency and calculating its similarity in the time domain with the signal derived from the video image; outputting as the physiological signal frequency the frequency of the synthetic signal with the maximum similarity with the signal derived from the video image.
Thus with the present invention the pole selection from the fitted autoregressive model or models is conducted by a time domain analysis of the frequency that each pole represents with the signal to which the autoregressive model was fitted. Thus the spectral analysis using the autoregressive model is performed in the frequency domain, but the selection of the resulting pole is performed by reference back to the time domain. The spectral component (pole) which most closely represents the desired periodic physiological signal will have the highest cross-correlation in the time domain with the original signal.
The measure of similarity may be the cross-correlation. Any of the standard cross-correlation techniques well-known from signal analysis can be used.
The synthetic signal is preferably sinusoidal.
The method may further comprise the step of defining a similarity threshold and inhibiting the outputting step if the similarity of the synthetic signal with the maximum similarity with the selected ratio signal is below the threshold. This avoids outputting poor estimates of the frequency of the physiological signal.
The signal derived from the video image may be a single colour channel of video source data, and the colour channel may be visible or infra-red (IR). The signal derived from the video image can be a ratio of two colour channels of video source data, the channels being in the visible or IR region.
Alternatively the signal derived from the video image can be a sequence of co-ordinates obtained by tracking movement in the image, e.g. a sequence of co-ordinates of a physical feature or region of high contrast being tracked.
The signal derived from the video image is derived by the following steps: receiving multiple colour channels of video source data representing said video image, each channel comprising a time series of intensity data for that colour channel; for each different pairing of the colour channels calculating the ratio of the intensity at each time point in a first of the pairing to the intensity at the same time point in the second of the pairing to produce multiple ratio signals; performing source separation on the ratio signals and selecting the output component which is most periodic. The autoregressive models are fitted not to the colour channel signals themselves but to the most periodic component in the ratio signals resulting from dividing the signal in one colour channel by the signal in another. Preferably the ratios are the red to green ratio, the red to blue ratio, and the green to blue ratio (or their reciprocals). Each of the colour channels can be normalised by dividing by its mean value before the ratio calculation. Because changes in lighting of the image or movement of the subject tend to affect all channels at once, taking the ratios of the colour channels eliminates to some degree the effect of such lighting changes or movement.
It is possible to generate a synthetic time-domain signal for each spectral component from each model and calculate its cross-correlation with the signal to which the autoregressive model was fitted, or alternatively only poles in a selected frequency range (e.g. the allowed range for the physiological signal of interest) can be selected, or only the dominant pole (i.e. pole with the greatest magnitude) from each model is used. Choosing only the dominant pole of each model speeds-up the signal processing.
In an alternative embodiment the signal derived from the video image can be derived by the following steps: receiving multiple sequences of co-ordinates of one or more physical features being tracked; performing source separation on the sequences and selecting the output component which is most periodic.
The source separation step mentioned above may be by Principal Component Analysis. The selection of the most periodic output component is preferably by selecting the component having the greatest peakiness of frequency spectrum. In the method the three ratio signals resulting from dividing one of the colour channel signals by another are subject to principal component analysis and whichever component is the most periodic is selected as the signal for autoregressive analysis for that window. The selection as “most periodic” can be achieved by selecting the signal with the peakiest spectrum within the physiological frequency range of interest, i.e. which has a peak with the highest power as a proportion of the power at all frequencies. The selection of the most periodic of the signals results in a further improvement in movement compensation. Other methods of choosing the most periodic signal, such as analysis of the autocorrelation of each signal, can be used.
Autoregressive models of order 8-20, more preferably 7 to 11, can be fitted to the signal derived from the video image.
Preferably a synthetic signal is generated corresponding to the frequency of only the dominant spectral component for each order model.
Each colour channel is preferably normalised by dividing by its mean before the step of calculating the ratio.
The signal derived from the video image is preferably temporally windowed, e.g. into overlapping windows, for example from 4-30 seconds long with an overlap from 0.5 to 10 seconds.
The video source data may be a time series of intensity data for each colour channel for a region of interest defined in the video image. The region of interest is preferably on the skin of the subject and the video source data is a photoplethysmographic image. The region of interest may include a periodically moving part of subject.
The periodic physiological signal may be the heart rate or respiration rate of the subject.
The invention can be embodied as a computer program comprising program code means for executing the method on a computer system. Alternatively it can be incorporated into a dedicated patient monitor.
Thus another aspect of the invention provides apparatus for detecting the frequency of a periodic physiological signal of a subject from a video image of the subject comprising: an input for receiving one or more colour channels of video source data representing said video image, each channel comprising a time series of intensity data for that colour channel; a processor for processing video source data; the processor being configured to execute the steps of the method above; the apparatus further comprising an output to output as the physiological signal frequency the frequency of the synthetic signal with the maximum cross-correlation with the selected ratio signal. As above the colour channel may be visible or IR.
The invention will be further described by way of example with reference to the accompanying drawings in which:—
However alternatively the mode of the distribution of intensities for each of the three colour channels within the region of interest can be used, or another representative intensity for each channel in the region of interest.
Once the representative intensity for each colour channel for each frame has been obtained, a time series of these intensities is assembled for a series of frames in a time window of, for example, 15 seconds. The length of the time window can be different, for example from 8 seconds to 1 minute. Each window overlaps its neighbour by a small time period, for example 1 second, though different overlaps, for example from 0.5 seconds to 5 seconds are usable.
In step 26 the signal values for each channel in each window are normalised by dividing each value by the mean of the representative intensities for that colour channel over that time window.
In step 28 the ratio of the normalised channels to each other is obtained. That is to say for each time point in the sequence, the normalised value for the red channel is divided by the normalised value for the green channel, the normalised value for the red channel is divided by the normalised value for the blue channel and the normalised value for the green channel is divided by the normalised value for the blue channel. This generates three different ratio signals consisting of a ratio value for each of the frames (each time point) in the window.
For each sequence of ratios Principal Component Analysis (PCA) is carried out in step 30 and the output components are detrended and filtered (e.g. by a bandpass digital filter whose pass band is the range of physiological signal frequencies) in step 32.
Then in step 34 whichever of the detrended and filtered components is most periodic is judged. This is achieved in this embodiment by selecting the signal with the peakiest spectrum, i.e. that which has the peak with the highest power as a proportion of the total power at all physiologically-possible frequencies. For example, the criterion can be to be maximise the value of:—
Where F is the signal in the frequency domain, max peak (F) is the peak power, and area (F) is the area underneath the power spectral density curve in the frequency range where the physiological signal may exist. This can easily be calculated from a Fast Fourier Transform of the detrended and filtered PCA output.
As a result whichever of the three components is regarded as most periodic is then the subject of autoregressive modelling in step 36 by fitting autoregressive models of multiple orders to the sequence of values for that window. In this embodiment AR models of order 7-11 are fitted to each sequence. However the order and number of model orders can be varied for different applications.
In the prior art method of
Then in step 42 the detrended and filtered component from step 34 is cross-correlated with the synthesised signal to find the coefficient of correlation c, this being repeated for all possible non-identical phase differences. In this repetition the synthesised signal may be moved relative to the ratio signal by one sample step each time, or a larger step can be used. Whichever phase difference gives the highest cross-correlation c is taken as the best fit for that model order. This procedure is repeated for poles from each of the different model orders. In step 44 the pole with the largest cross-correlation coefficient c is retained.
Steps 40, 42 and 44 can be conducted for only the dominant pole for each model order (i.e. the pole with the largest magnitude, possibly within an allowed frequency range for the physiological signal of interest), or it can be repeated for all poles within the allowed frequency range, or all poles from the model. The fewer poles processed, the quicker the processing.
In step 46 the frequency of the synthetic signal with the highest cross-correlation with the original ratio signal is selected as the estimated frequency of the physiological signal and is outputted. The output is preferably presented as, for example, a heart rate in beats per minute or a respiration rate in breaths per minute (by calculating the frequency in Hz times 60).
Steps 42 and/or 44 may be supplemented by a check that the cross-correlation c is above a predetermined threshold. If the cross-correlation is not above the predetermined threshold then the frequency may not be selected. Alternatively step 46 can be supplemented by the step of checking the cross-correlation coefficient c of the selected frequency against the predetermined threshold and if it is less than the threshold then the estimate is not output. This avoids outputting an estimate based on a poor level of periodic information in the input signal.