METHOD AND APPARATUS FOR MONITORING A HUMAN OR ANIMAL SUBJECT

Abstract

Methods and apparatus for monitoring a human or animal subject are disclosed. In one arrangement, measurement data representing a time series of measurements on a subject is received. The measurement data is represented as a mathematical expansion comprising a plurality of expansion components and expansion coefficients. First and second partial reconstructions are performed using first and second subsets of the expansion components. First and second spectral analyses are performed on the first and second partial reconstructions to determine first and second dominant frequencies. A frequency of a periodic physiological process is derived based on either or both of the first and second dominant frequencies.

Claims

1. A computer-implemented method of monitoring a human or animal subject, comprising: (a) receiving measurement data representing a time series of measurements on a subject performed by a sensor system; (b) processing the measurement data to represent the measurement data as a mathematical expansion comprising a plurality of expansion components and a corresponding plurality of expansion coefficients representing the respective strength of each expansion component; (c1) forming a first partial reconstruction of the measurement data using a first subset of the expansion components and corresponding expansion coefficients; (c2) forming a second partial reconstruction of the measurement data using a second subset of the expansion components and corresponding expansion coefficients, the second subset being different from the first subset; (d1) performing a first spectral analysis on the first partial reconstruction to determine a first dominant frequency, the first dominant frequency representing a frequency component of largest amplitude in the first partial reconstruction; (d2) performing a second spectral analysis on the second partial reconstruction to determine a second dominant frequency, the second dominant frequency representing a frequency component of largest amplitude in the second partial reconstruction; and (e) providing an output representing a frequency of a periodic physiological process based on either or both of the first dominant frequency and the second dominant frequency.

2. The method of claim 1, wherein the sensor system comprises one or more accelerometers.

3. The method of claim 1, wherein the frequency of the periodic physiological process comprises a respiration rate.

4. The method of claim 1, further comprising: determining a degree of similarity between the first dominant frequency and the second dominant frequency, wherein: the output representing the frequency of the periodic physiological process is selectively provided depending on the determined degree of similarity.

5. The method of claim 4, wherein steps (a)-(d2) are repeated for plural units of the measurement data, and step (e) comprises outputting the frequency of the periodic physiological process only for units of the measurement data for which the determined degree of similarity between the first dominant frequency and the second dominant frequency is above a predetermined threshold.

6. The method of claim 1, wherein each expansion component comprises an eigenvector and each expansion coefficient comprises an eigenvalue.

7. The method of claim 6, wherein the eigenvectors and eigenvalues are determined using singular spectrum analysis.

8. The method of claim 1, wherein the first subset of expansion components comprises the N components respectively having the N largest expansion coefficients.

9. The method of claim 8, wherein 2N5.

10. The method of claim 8, wherein the second subset of expansion components does not comprise the expansion component having the largest expansion coefficient.

11. The method of claim 10, wherein the second subset of expansion components comprises the expansion components corresponding to the N largest expansion coefficients except the largest expansion coefficient.

12. The method of 1, wherein the measurement data is processed to at least partially removing noise from the measurement data in a filtering step prior to the processing to represent the measurement data as a mathematical expansion.

13. The method of claim 12, wherein the filtering step comprises applying a low-pass filter.

14. The method of claim 12, wherein the filtering step comprises applying an adaptive line enhancer.

15. The method of claim 12, wherein: the measurement data comprises multiple channels, each channel representing measurements made using a different measurement mode; and the method comprises determining a signal-to-noise ratio of the measurement data in each channel, after the filtering step, and using the determined signal-to-noise ratio to select one of the channels to use in subsequent steps to provide the output representing the frequency of the periodic physiological process.

16. The method of claim 15, wherein the different measurement modes each comprise measurements of acceleration relative to a different respective axis.

17. A computer program comprising computer-readable instructions that cause a computer to perform the method of claim 1.

18. A computer program product storing the computer program of claim 17.

19. An apparatus for monitoring a human or animal subject, comprising: a data receiving unit configured to receive measurement data representing a time series of measurements on a subject performed by a sensor system; a data processing unit configured to: process the measurement data to represent the measurement data as a mathematical expansion comprising a plurality of expansion components and a corresponding plurality of expansion coefficients representing the respective strength of each expansion component; form a first partial reconstruction of the measurement data using a first subset of the expansion components and corresponding expansion coefficients; form a second partial reconstruction of the measurement data using a second subset of the expansion components and corresponding expansion coefficients, the second subset being different from the first subset; perform a first spectral analysis on the first partial reconstruction to determine a first dominant frequency, the first dominant frequency representing a frequency component of largest amplitude in the first partial reconstruction; perform a second spectral analysis on the second partial reconstruction to determine a second dominant frequency, the second dominant frequency representing a frequency component of largest amplitude in the second partial reconstruction; and provide an output representing a frequency of a periodic physiological process based on either or both of the first dominant frequency and the second dominant frequency.

20. The device of claim 19, further comprising a sensor system configured to perform the measurements on the subject.

Description

[0014] Embodiments of the invention will now be described, by way of example only, with reference to the accompanying drawings in which corresponding reference symbols indicate corresponding parts, and in which:

[0015] FIG. 1 schematically depicts a method of monitoring a subject according to an embodiment;

[0016] FIG. 2 schematically depicts an apparatus for monitoring a subject;

[0017] FIG. 3 depicts data flow through functional modules for implementing methods according to embodiments;

[0018] FIG. 4 is a graph depicting a spectrum obtained by applying FFT to measurement data from subject #27 to which low-pass filtering has been applied;

[0019] FIG. 5 is a graph depicting a spectrum obtained by applying FFT to measurement data from subject #27 to which ALE has been applied;

[0020] FIG. 6 depicts graphs depicting spectra obtained by applying FFT to first and second partial reconstructed versions of measurement data from subject #27 to which ALE has been applied;

[0021] FIG. 7 is a graph depicting a spectrum obtained by applying FFT to first example measurement data from subject #27 to which low-pass filtering has been applied;

[0022] FIG. 8 is a graph depicting a spectrum obtained by applying FFT to the first example measurement data from subject #27 to which ALE has been applied;

[0023] FIG. 9 depicts graphs depicting spectra obtained by applying FFT to first and second partial reconstructed versions of the first example measurement data from subject #27 to which ALE has been applied;

[0024] FIG. 10 is a graph depicting a spectrum obtained by applying FFT to second example measurement data from subject #27 to which low-pass filtering has been applied;

[0025] FIG. 11 is a graph depicting a spectrum obtained by applying FFT to the second example measurement data from subject #27 to which ALE has been applied;

[0026] FIG. 12 depicts graphs depicting spectra obtained by applying FFT to first and second partial reconstructed versions of the second example measurement data from subject #27 to which ALE has been applied;

[0027] FIG. 13 is a graph depicting a spectrum obtained by applying FFT to third example measurement data from subject #27 to which low-pass filtering has been applied;

[0028] FIG. 14 is a graph depicting a spectrum obtained by applying FFT to the third example measurement data from subject #27 to which ALE has been applied;

[0029] FIG. 15 depicts graphs depicting spectra obtained by applying FFT to first and second partial reconstructed versions of the third example measurement data from subject #27 to which ALE has been applied;

[0030] FIG. 16 is a graph depicting a spectrum obtained by applying FFT to fourth example measurement data from subject #27 to which low-pass filtering has been applied;

[0031] FIG. 17 is a graph depicting a spectrum obtained by applying FFT to the fourth example measurement data from subject #27 to which ALE has been applied;

[0032] FIG. 18 depicts graphs depicting spectra obtained by applying FFT to first and second partial reconstructed versions of the fourth example measurement data from subject #27 to which ALE has been applied;

[0033] FIGS. 19(a)-(c) depict the results of measuring respiration rate using a series of segments of measurement data from accelerometers applied to subject #27; and

[0034] FIG. 20 is a graph comparing measurements of RR using a method of an embodiment with measurements of RR obtained from PPG signals and measurements of RR recorded manually by nurses in an average of every 6 hours, for subject #18.

[0035] Methods of the present disclosure are computer-implemented. Each step of the disclosed methods may therefore be performed by a computer. The computer may comprise various combinations of computer hardware, including for example CPUs, RAM, SSDs, motherboards, network connections, firmware, software, and/or other elements known in the art that allow the computer hardware to perform the required computing operations. The required computing operations may be defined by one or more computer programs. The one or more computer programs may be provided in the form of media, optionally non-transitory media, storing computer readable instructions. When the computer readable instructions are read by the computer, the computer performs the required method steps. The computer may consist of a self-contained unit, such as a general-purpose desktop computer, laptop, tablet, mobile telephone, smart device (e.g. smart TV), etc. Alternatively, the computer may consist of a distributed computing system having plural different computers connected to each other via a network such as the internet or an intranet.

[0036] FIG. 1 schematically depicts a framework for methods of monitoring a human or animal subject according to embodiments of the disclosure. The methods may be performed by an apparatus 1 for monitoring a human or animal subject as depicted in FIG. 2. The terms human or animal subject or subject may be used interchangeably with the term patient in the following description.

[0037] The method comprises a step S1 of performing physiological measurements on a subject to obtain measurement data. The physiological measurements may be performed using a sensor system 5 as depicted schematically in FIG. 2. In an embodiment, the sensor system 5 comprises a plurality of sensors, each sensor capable of obtaining measurement data using a different measurement mode. As depicted schematically in FIG. 3, the measurement data may thus comprise plural channels 7A-C, each channel 7A-7C representing measurements made using a different measurement mode. In embodiments described in detail below, the different measurement modes respectively comprise measurements of acceleration relative to a different respective axis. The sensor system 5 may thus comprise plural accelerometers, for example three accelerometers 5A-C. Each accelerometer 5A-C may be configured to detect acceleration relative to a respective one of three different orthogonal axes. In examples described below, these axes are referred to respectively as lateral, vertical and longitudinal.

[0038] In step S2, measurement data obtained by performing the physiological measurements in step S1 is received by a data receiving unit 3. The data receiving unit 3 may form part of a computing system 2 (e.g. laptop computer, desktop computer, smart device, wearable smart device, etc.). The computing system 2 may further comprise a data processing unit 4 configured to carry out steps of the method. FIG. 3 depicts functional modules that may be implemented by the computing system 2 when performing the method.

[0039] In step S3, the measurement data is filtered to at least partially remove noise from the measurement data. The filtering may comprise applying a band-pass filter or low-pass filter. Alternatively, as described in further detail below, an adaptive line enhancer (ALE) may be applied to the measurement data. As depicted in FIG. 3, the filtering may be applied separately to each channel 7A-C of the measurement data via respective filter modules 10A-C.

[0040] In step S4, a signal-to-noise ratio of the filtered measurement data in each channel 7A-C is obtained. In FIG. 3, the signal-to-noise ratios are determined by respective signal-to-noise ratio modules 12A-C. The determined signal-to-noise ratios are used (in selection module 13 of FIG. 3) to select one of the channels 7A-C to use in subsequent steps to provide the output representing the frequency of the periodic physiological process. In an embodiment, the channel 7A-C with the highest signal-to-noise ratio is selected for subsequent processing. The method may be implemented in such a way as to repeatedly output a measurement of the frequency of the periodic physiological process. This may be achieved for example by processing windows of time series that are centred at different time stamps (each window providing a unit of the measurement data received in step S2 and processed in subsequent steps). The windows may overlap with each other or may not overlap with each other. The windows may be referred to as segments. The selection of the best channel 7A-C may be different for different windows. Thus, the selection of the best channel 7A-C may be dynamically updated in time, switching continuously between different channels for different windows (units of measurement data) according to whichever has the highest signal-to-noise ratio at the time.

[0041] In step S5, spectral analysis is performed to extract a frequency of a periodic physiological process of interest. In some embodiments of the present disclosure, the frequency of the periodic physiological process comprises a respiration rate (RR). Further details about the spectral analysis are given below.

[0042] In step S6, a determination is made based on the spectral analysis of step S5 about whether the received measurement data is such as to allow a reliable estimate of the frequency of the periodic physiological process to be obtained. If yes, the method continues to step S7, an output is provided representing the extracted frequency of the periodic physiological process for the received measurement data, and the method loops back to step S2 to receive a next unit of measurement data to be processed. If no, the method loops straight back to step S2 to receive a next unit of measurement data to be processed, without providing any output. This methodology thus advantageously reduces the extent to which unreliable readings of the frequency of the periodic physiological process are output.

FilteringAdaptive Line Enhancer (ALE)

[0043] As mentioned above, in some embodiments an adaptive line enhancer (ALE) is used to at least partially remove noise from measurement data. This process is now described in further detail.

[0044] The ALE, introduced in [8], has been used in many applications [9] for separation of a low level sinusoid or a narrow-band oscillation from broad-band noise. The core concept of the ALE is based on linear prediction where the nearly-periodic signal is believed to be perfectly predicted based on the past samples while a non-periodic signal cannot be predicted. A delayed version of the input signal is used as a reference signal into a least-mean-square (LMS) adaptive filter and the desired signal is considered as the original signal. An error signal of the LMS filter provides an estimate of the noise in the input signal and the output signal corresponds to the desired separated signal (i.e. separated from the noise). In specific examples of the present disclosure where it is desired to estimate RR using acceleration signals, the ALE can be applied to each accelerometer axis separately. The ratio of the power of signal to the power of noise provides a signal-to-noise ratio (SNR). The SNR can be used, as described above, to select an accelerometer axis to use in subsequent processing steps (e.g. an axis with the highest SNR is selected).

Spectral AnalysisSSA

[0045] After selection of the desired accelerometer axis, Fast Fourier Transform (FFT) could be applied to estimate the highest peak amplitude of the spectrum of ALE signal output in the range of 0.05 Hz to 1 Hz. Then, a frequency with maximum peak amplitude could be obtained, which could correspond to the frequency of the periodic physiological process being monitored (e.g. respiration rate). However, it has been found that the frequency obtained in this way does not accurately reflect the underlying frequency of the periodic process in all situations. For some segments (units) of measurement data, multiple local maximum peaks in the spectral domain are observed, which makes discrimination of the physiological frequency difficult. Embodiments of the present disclosure use the following spectral analysis in step S5 of the method to improve reliability.

[0046] In some embodiments, the spectral analysis of step S5 comprises processing the measurement data to represent the measurement data as a mathematical expansion. The measurement data may thus be transformed by the processing of the measurement data, the transformation resulting in the measurement data being represented as a mathematical expansion. The mathematical expansion comprises a plurality of expansion components and a corresponding plurality of expansion coefficients. Each expansion coefficient represents the respective strength of each expansion component in the mathematical expansion. Thus, expansion components which contribute to the measurement data to a large extent will have a large expansion coefficient and expansion components which contribute to the measurement data to a smaller extent will have smaller expansion coefficients. In some embodiments, each expansion component comprises an eigenvector and each expansion coefficient comprises an eigenvalue. In some embodiments, the eigenvectors and eigenvalues are determined using singular spectrum analysis (SSA) [10]. It has been shown in [11] that after applying SSA into acceleration signals and using the resulting eigenvectors and their corresponding eigenvalues, it is possible to extract the trend of the data or the lowest possible dominant frequency (considering the largest eigenvalues) in the signals.

[0047] The inventors have found that by performing partial reconstructions of the measurement data using different combinations of the eigenvectors mentioned above, it is possible to extract the frequency of the physiological process being monitored more accurately and/or identify segments of measurement data from which it is not possible to obtain a reliable estimate of the frequency of the physiological process.

[0048] In an embodiment, a first partial reconstruction of the measurement data is formed using a first subset of the expansion components (e.g. eigenvectors) and corresponding expansion coefficients (e.g. eigenvalues). A second partial reconstruction of the measurement data is formed using a second subset of the expansion components (e.g. eigenvectors) and corresponding expansion coefficients (e.g. eigenvalues). The second subset is different from the first subset.

[0049] A first spectral analysis (e.g. using a fast Fourier transform, FFT) is then performed on the first partial reconstruction to determine a first dominant frequency. The first dominant frequency represents a frequency component of largest amplitude in the first partial reconstruction.

[0050] A second spectral analysis (e.g. using a fast Fourier transform, FFT) is performed on the second partial reconstruction to determine a second dominant frequency. The second dominant frequency represents a frequency component of largest amplitude in the second partial reconstruction.

[0051] The first and second partial reconstructions may be referred to respectively as sub-band1 and sub-band2. The provision of an output representing a frequency of a periodic physiological process of step S7 can be performed based on either or both of the first dominant frequency and the second dominant frequency. In a case where the measurement data is able to yield a reliable value for the physiological frequency of interest, the first dominant frequency and the second dominant frequency should both correspond closely to the physiological frequency (and not some other artefactual frequency present in the measurement data). The output may be obtained for example by taking an average of the first dominant frequency and the second dominant frequency, or simply by taking one or the other of the first dominant frequency and the second dominant frequency (as they are both similar).

[0052] The inventors have found that in cases where the measurement data is unable to yield a reliable value for the physiological frequency of interest, for example because there are several spectral peaks in the frequency range of interest, the situation can be detected automatically and reliably by detecting when the first dominant frequency and the second dominant frequency differ from each other to a significant extent (e.g. by more than a predetermined threshold amount). Based on this recognition, in some embodiments a degree of similarity between the first dominant frequency and the second dominant frequency is determined and the output representing the frequency of the periodic physiological process is selectively provided (i.e. provided or not provided) in step S7 depending on the determined degree of similarity. For example, the output is provided if the degree of similarity is determined to be higher than a predetermined threshold.

[0053] In some embodiments, the first subset of expansion components (corresponding to sub-band1) comprises the N components respectively having the N largest expansion coefficients. N may satisfy the following inequality: 2N5. In an embodiment, N=3. In an embodiment, the second subset of expansion components (corresponding to sub-band2) does not comprise the expansion component having the largest expansion coefficient. In an embodiment, the second subset of expansion components comprises the expansion components corresponding to the N largest expansion coefficients except the largest expansion coefficient. In an embodiment, the second subset of expansion components comprises the 2.sup.nd and 3.sup.rd largest.

DETAILED EXAMPLES

[0054] In the detailed examples described below, the spectral analysis of step S5 was performed on measurement data to which ALE had been applied in step S3 and to an accelerometer axis selected using the ALE output in step S4 to decompose the measurement data into corresponding eigenvectors and eigenvalues. The eigenvectors were sorted in descending order of associated eigenvalues and grouped using a set of indices as {1,2,3} (denoting eigenvectors related to first, second, and third largest eigenvalues) and {2,3} (denoting eigenvectors related to the second and third largest eigenvalues). After grouping the eigenvectors into these two sub-bands (sub-band1 for {1,2,3} indices and sub-band2 for {2,3} indices), a final elementary matrix was formed by summation of all elementary matrices in each group. Diagonal averaging was applied to the final elementary matrix and a resulting signal formed for each group. For each sub-band, one narrowband signal was reconstructed (an example of the partial reconstruction referred to above). An FFT algorithm was then applied to each narrowband signal to estimate and compare the frequency spectrum for the two selected sub-bands.

[0055] The FFT was applied to each reconstructed signal corresponding to the sub-bands and the frequency with maximum peak amplitude was considered as a potential respiratory component. Then, for each sub-band, the RR was calculated as breaths per minute (based on the maximum peak). If the absolute difference in estimated RR from the two sub-bands was less than 4 bpm (an example of the predetermined threshold mentioned above), an average RR was used as the final estimated RR, otherwise, it was determined that a reliable RR could not be obtained from the measurement data (the no loop in FIG. 1). The threshold of 4 bpm was found appropriate for the particular dataset being considered. It is expected that a lower threshold could be used by increasing the frequency resolution.

Dataset for Detailed Examples

[0056] Waveform data were collected from voluntary participants in the Post Intensive Care Risk Alerting and Monitoring (PICRAM)II trial. This waveform data was collected using continuous monitoring equipment provided by Hidalgo Ltd. with an attached transmittance photoplethysmograph sensor provided by Nonin Medical Inc. The data was collected at both the John Radcliffe and the Reading Berkshire hospital. In this research only the data collected at Oxford John Radcliffe was used for analysis.

[0057] Time-stamped observations for patients admitted to the ward at the John Radcliffe hospital were recorded within the experiment. The data corresponds to admissions between 2013 and 2014. Sampling frequencies were 256 Hz for an electrocardiogram (ECG), 75 Hz for a PPG, and 25 Hz for an accelerometer.

[0058] Ten patients (five male, five female) were selected from a larger dataset. These selected patients' age range was from 45 to 77 years old at the time of discharge from the ICU. The select patients' mean age was calculated as 63.7 years old. The length of their stay at the ICU ranged from 1 day to 15 days. The hospital length of stay for these patients ranged from 7 days to 47 days. All patients were white except one with black/British black ethnicity.

Results for Detailed Examples

[0059] For estimation of RR from accelerometer signals, the ALE was applied to each axis of the accelerometer signal to separate the oscillatory and noise component of the accelerometer axes. A delay () of 10 samples was used to generate the reference signal for the adaptive filter. The normalised least mean square (NLMS) adaptive filter with an order of (M=20), and an offset of (N=50) was applied. Units of measurement data were obtained by forming 64-second windowed data segments and discarding the first 16 seconds of each data segment. Each resulting 48-seconds long data segment corresponded to one unit of measurement data to be processed. A reduced window size of 48-seconds promotes convergence of the NLMS. The ALE separates the noise and signal component of each accelerometer axis.

[0060] FIGS. 4-6 compare the processed signal for each accelerometer axis in the spectral domain using FFT for an example unit of measurement data. The SNR as the power of signal to the power of noise was calculated as 0.19, 0.11 and 0.32 for axis 1 (lateral), 2 (vertical), and 3 (longitudinal) respectively. The highest SNR corresponded to axis 3 (longitudinal axis). This axis was selected for further analysis to estimate RR.

[0061] FIG. 4 shows the result of applying an FFT to the measurement data for each accelerometer axis after a low-pass filter is applied in step S3 of the method. The low-pass filtered spectrum shows a similar maximum peak (considering frequencies more than 0.1 Hz) for each of the three accelerometer axes.

[0062] FIG. 5 shows the result of applying the FFT to the measurement data for each accelerometer axis after an ALE is applied in step S3 of the method. The resulting spectra are visibly enhanced relative to FIG. 4. The selected accelerometer axis with the largest SNR (third axis for this segment) was further subjected to the spectral sub-band decomposition procedure using SSA, with the results being shown in FIG. 6. This procedure improves analysis of the acceleration signals, particularly where, for example, a sharp change of amplitudes in the acceleration signal creates a low frequency in the spectral domain that is not related to the respiratory component. As explained above, two new signals are reconstructed in sub-band1 and sub-band2. Then, FFT is applied to output an estimated RR where the difference of estimated RR in the two specified sub-bands is less than 4 bpm. For patient #27, at Jun.06.2013 13:50, a manual observation of RR as 19 bpm was recorded. In the midnight of Jun.07.2013, after about 10 hours and 10 minutes, a manual observation of RR as 22 bpm was recorded. In this time interval, the maximum, minimum and mean recorded RR was obtained as 22 bpm, 18 bpm and 19.92 bpm. The selection of this patient is based on low variability of the manual observations to demonstrate cases where applying FFT alone to the accelerometer axes is not effective. In addition, applying SSA sub-band decomposition to select an appropriate frequency band to estimate RR is necessary. A wrongly selected frequency sub-band could lead to major errors of 10 bpm or more in estimation of RR. FIG. 6 shows the result of applying the FFT to each sub-band using SSA, where the first sub-band related to grouping {1,2,3} eigenvectors has greater peak amplitude comparing to the second sub-band {2,3}. In both sub-bands the maximum spectral peaks are in close agreement. For this segment (unit) of measurement data, the estimated RR was obtained as 17.62 bpm at June.06.2013 14:03:23. The closest manual observation was recorded as 19 bpm at June.06.2013 13:50:00. In FIGS. 7-18, four examples from subject #27 are provided to further demonstrate that applying both ALE and SSA sub-band decomposition assists with providing a reliable estimation of RR. For each example, three graphs corresponding respectively to the graphs of FIGS. 4-6 are provided.

[0063] Example 1: FIGS. 7-9 show that all three accelerometer axes provide measurement data containing a dominant low frequency peak around a frequency of 0.35 Hz (21 bpm), both when processing is based on low pass filtering in step S3 (FIG. 7) and when processing is based on ALE in step S3 (FIG. 8). The segmented window corresponding to the measurement data for Example 1 starts at Jun.06.2013 19:30:28 where a manual RR of 22 was recorded at Jun.06.2013 19:30:00. The estimated RR using the proposed method was calculated as 21.39 bpm. The maximum spectral peak is in a good agreement to the recorded manual observation.

[0064] Example 2: FIGS. 10-12 show results for an example in which two of the accelerometer axes demonstrate a dominant spectral peak around a frequency of 0.38 Hz (22.8 bpm), while the other axis shows a maximum peak around 0.3 Hz (15 bpm). This example shows that selection of the best accelerometer axis is advantageous. The ALE can be used to find the best axis using SNR information, as described above.

[0065] Example 3: FIGS. 13-15 show results for an example in which there is a strong low frequency component which has affected the results after low-pass filtering. This example shows that FFT applied to raw acceleration produces a maximum peak in frequencies of less than 0.2 Hz (FIG. 13relating to a respiratory rate of about 12 bpm). A band-pass filter with lower and upper cut-off frequencies as 0.05 Hz and 1 Hz, can also been applied to raw acceleration where the maximum peak happens before 0.2 Hz. The output of the ALE producing the spectrum of FIG. 14 yields SNRs of 0.31, 0.17, 1.46 for axis 1 (lateral), 2 (vertical), and 3 (longitudinal) respectively. The FFT applied to the third axis (which has the highest SNR) produced a spectral peak around 0.4 Hz after applying FFT to the signal ALE output (see FIG. 14). The sub-band decomposition of the ALE signal output of the third axis is shown in FIG. 15, where the largest peaks in both sub-bands are in close agreement, thereby indicating that an RR estimate based on the largest peaks is reliable. The average of these peaks was used to estimate the final RR. For this segment of measurement data, the estimated RR was 23.91 bpm at Jun.06.2013 19:33:36, where the closest bpm observed manually was 22 bpm at Jun.06.2013 19:30:00. For this example, only a band-pass filter with cut-off frequencies of [0.2 Hz-0.5 Hz] could derive the frequency spectrum with a peak around 0.4 Hz, while our method is able to find a good match in filtered signals in two sub-bands without using such a narrow band-pass filter, which would not be appropriate when seeking to detect transitions between normal and abnormal states of the subject.

[0066] Example 4: FIGS. 16-18 show results for an example in which a low frequency edge is present in the measurement data, which can produce uncertainty in the RR estimation. For example, an edge frequency of about 0.2 Hz (12 bpm) will be picked after applying low-pass filtering and also the SSA algorithmsub-band1as the maximal spectral peak (see arrows), while there is also a dominant frequency at 0.38 Hz (22.8 bpm), which appeared in sub-band2 as well. Based on prior knowledge, investigating manual observations and matching to PPG-based estimates, the frequency of about 0.38 (22.8 bpm) is the actual RR related frequency. For this example, our method thus failed to select the correct peak corresponding to the RR. However, the lack of agreement in the frequency domain between the two reconstructed sub-band signals identifies the problem. The uncertainty in the estimation of RR is identified and no RR is returned by the proposed algorithm. The peak corresponding to 12 bpm is most probably related to a frequency edge and very likely a result of an induced transient motion. Therefore, it is invalid.

[0067] Uniform acceleration signals are expected to provide clear spectra. By detailed examination of the acceleration signals, it is found that many types of motions, especially transient motion interferences not related to respiration, provide low frequency edges or artefacts in the accelerometer spectra. Examples of spectra containing such low frequency artefacts are shown in Examples 3 and 4 and it is described how our method provides sufficient sensitivity to extract the correct RR values and/or reliably recognise when a reading is probably invalid.

[0068] FIGS. 19(a)-(c) depict the results of measuring respiration rate using a series of segments (units) of measurement data from accelerometers applied to subject #27. FIG. 19(a) depicts the result of applying an FFT to segments of measurement data to which a band pass filter in the range of [0.1 Hz-1 Hz] has been applied (compared to manual observations and measurements using PPG). FIG. 19(b) depicts the result of applying an FFT to segments of measurement data to which a band pass filter in the range of [0.2 Hz-1 Hz] has been applied (compared to manual observations and measurements using PPG). For both FIGS. 19(a) and 19(b), a maximum spectral peak in the low-frequency range of [0.08 Hz-0.5 Hz] among all the axes after applying FFT was used to estimate RR.

[0069] FIG. 19(c) depicts the result of using a method according to an embodiment to determine RR. The method included use of an ALE in step S3 and SSA in step S5. It is evident from FIGS. 19(a)-(c) that the estimated RR does vary continuously from 10 to 20. By detailed investigation of segments of the measurement data, it was found that transient motion induced quantities affect the signal spectrum in the low frequency domain, creating (invalid) RR readings of 10 bpm related to a picked frequency edge. By testing several segments of measurement data producing an RR of 10 bpm and slightly reducing or shifting windowed segments of measurement data to include only uniform accelerations, RRs of around 20 bpm were re-calculated. This confirmed continuous change of breathing rate from 20 bpm to 10 bpm is mostly invalid. As FIG. 19(c) demonstrates, our method minimizes the effects of low frequency artefacts in respiration derived spectra by using the SSA method which attempts to find a shared low frequency component in the reconstructed signals of two sub-bands. Good agreement is seen between manual observation, accelerometer and PPG-based estimates for subject #27, highlighting the advantages of our method.

[0070] For the patients selected in this study, measurements of RR using a method of an embodiment were compared with measurements of RR obtained from PPG signals and measurements of RR recorded manually by nurses in an average of every 6 hours. The results of such a comparison are shown in FIG. 20 for subject #18. This subject shows continuous recorded signals for more than 2 days. The accelerometry-based estimations of RR overlap strongly with both the PPG-based estimations of RR and the manual observations in the plot of FIG. 20, demonstrating strong agreement between RR estimates for most parts of the recorded signals.

[0071] In Table I, for 10 selected patients, detailed statistics are provided. Based on this table, the number of PPG segments to estimate RR is shown in the second column for each subject. These segments are converted to their time duration in terms of the number of hours of recorded PPG data. Then, the number of segments with a PPG Signal Quality Index (SQI), indicated a quality of the signal for the segment in question, of greater than or equal to 0.85 is provided in the third column. The percentage of segments produced RR estimates based on smart fusion of PPG data (agreement of auto-regression, AR, spectrum in two modulations) are shown for each subject. For accelerometer-based estimation of RR, the initial number of segments of measurement data and the corresponding time duration in terms of the hours of data are provided. Also the percentage of obtained segments to estimate RR based on the spectral analysis approach involving agreement of sub-bands (as discussed above) is provided.

[0072] The time-stamps have been used to compare the errors of pairwise estimated RRs where both PPG and accelerometer-based RR estimates are available. The mean absolute error (MAE) is shown in the last column of Table I. As it can be seen from Table I, all the resulted MAEs are less than 2.60 bpm. It should be noted that the sample size to calculate the MAE is not the same for all the subjects. Meanwhile, very good MAEs have been achieved for most subjects (less than 2 bpm). The accelerometers have been able to produce RR estimates for about 70 hours of the continuous data while this has been recorded for about 27 hours of PPG data. Although after data fusion, for error measurements, the number of segments are reduced to consider highly reliable RR outputs of the algorithms.

TABLE-US-00001 TABLE 1 COMPARISON OF PPG AND ACCELEROMETER RR ESTIMATES PPG Accelerometer Difference patient segments No. segment No. ratio (%) of segments No. ratio (%) of MAE ID (hours) SQI 0.85 smart fusion (hours) spectral fusion (bpm) 18 1543 (13.72) 1281 20.69 3989 (70.92) 28.48 1.34 16 2403 (21.36) 2018 12.00 1397 (24.84) 12.46 1.81 34 2699 (23.10) 812 23.40 1450 (25.78) 18.00 1.24 43 3121 (27.75) 2930 17.62 3993 (70.99) 12.28 1.56 45 3007 (26.73) 1923 15.61 1540 (27.38) 19.81 1.52 36 502 (4.47) 351 12.54 1488 (26.46) 18.08 1.84 13 1514 (13.46) 946 11.89 3976 (70.69) 6.19 2.56 15 1519 (13.51) 200 22.00 3925 (69.78) 9.69 0.90 23 533 (4.74) 404 10.15 1375 (24.45) 25.10 1.73 27 637 (5.67) 507 5.13 1655 (29.43) 16.02 1.86

REFERENCES

[0073] [1] P. Jiang and R. Zhu, Dual tri-axis accelerometers for monitoring physiological parameters of human body in sleep, IEEE Sensors, 2016. [0074] [2] M. Haescher, D. J. C. Matthies, J. Trimpop, and B. Urban, A study on measuring heart- and respiration-rate via wrist-worn accelerometer-based seismocardiography (SCG) in comparison to commonly applied technologies, Proceedings of the 2nd international Workshop on Sensor-based Activity Recognition and Interaction, 2015. [0075] [3] D. S. Morillo, L. F. C. F. J. L. Rojas Ojeda, and A. Jimenez, An accelerometer-based device for sleep apnea screening, IEEE Trans. Inf. Technol. Biomed., vol. 14, pp. 491-499, 2010. [0076] [4] Z. Zhang and G. Z. Yang, Monitoring cardio-respiratory and posture movements during sleep: What can be achieved by a single motion sensor, 2015 IEEE 12th International Conference on Wearable and Implantable Body Sensor Networks (BSN), 2015. [0077] [5] S. P. Jeelani, P. Maniyar, J. Joseph and M. Sivaprakasam, Accelerometer based system for continuous respiratory rate monitoring, IEEE International Symposium on Medical Measurements and Applications (MeMeA), pp. 171-176, 2017. [0078] [6] A. Bates, M. J. Ling, J. Mann and D. K. Arvind, Respiratory Rate and Flow Waveform Estimation from Tri-axial Accelerometer Data, International Conference on Body Sensor Networks, Singapore, pp. 144-150, 2010. [0079] [7] A. M. Chan, N. Ferdosi and R. Narasimhan, Ambulatory respiratory rate detection using ECG and a triaxial accelerometer, Conf Proc IEEE Eng Med Biol Soc., pp. 4058-4061, 2013. [0080] [8] B. Widrow, J. Glover, J. McCool, J. Kaunitz, C. Williams, R. Hern, J. Zeidler, E. Dong, and R. Goodlin, Adaptive noise canceling principles and applications, Proceedings of IEEE, vol. 63, pp. 1692 1716, 1975. [0081] [9] R. M. Ramli, A. O. A. Noor, and S. A. Samad, A review of adaptive line enhancers for noise cancellation, Australian Journal of Basic and Applied Sciences, vol. 6, no. 6, pp. 337-352, 2012. [0082] [10] N. G. V. Nekrutkin and A. Zhigljaysky, Analysis of Timeseries Structure: SSA and Related Techniques. Chapman and Hall/CRC, 2001. [0083] [11] D. Jarchi, C. Wong, R. M. Kwasnicki, B. Heller, G. A. Tew, and G. Z. Yang, Gait parameter estimation from a miniaturized ear-worn sensor using singular spectrum analysis and longest common subsequence, IEEE Trans. Biomed. Eng., pp. 1261-1273, 2014.