Patent Application
20210334566
The present invention relates to a method, apparatus and computer program for detecting a biometric event in a noisy signal using Principle Component Analysis (PCA). More particularly, but not exclusively, the present invention relates to detecting one or more heartbeats in an input signal.
In many signal processing applications, it is necessary to determine whether or not a certain event is present in a noisy input signal. The presence of noise can cause errors in conventional threshold-based methods. For example, in the field of healthcare, a pulse rate is conventionally measured by counting heartbeats in an input signal such as a photoplethysmography (PPG) signal or electrocardiograph (ECG) signal, using a threshold-based method in which a heartbeat is counted when the signal crosses a certain threshold. Noise in the signal can result in errors, for example by causing the peak of an actual heartbeat to fall below the threshold, or by causing a spurious peak which exceeds the threshold and triggers the heartbeat detection algorithm when a heartbeat has not actually occurred. There is therefore a need in the art for an improved method of detecting biometric events such as heartbeats in noisy input signals.
The invention is made in this context.
According to a first aspect of the present invention, there is provided a method of detecting a biometric event in an input signal, the method comprising: performing principal component analysis PCA on samples of a plurality of model signals to generate a transformation matrix having more informative components and less informative components, each of the model signals comprising a known signal which includes the biometric event to be detected; reducing a dimensionality of the transformation matrix by discarding one or more of the more informative components; transforming a plurality of samples of the input signal using the reduced dimensionality transformation matrix; determining a probability that the biometric event is present in the plurality of samples of the input signal, by calculating a predefined probability function for the transformed samples; and determining that the input signal includes the biometric event if the probability is higher than a threshold.
In some embodiments according to the first aspect, the plurality of samples of the input signal are selected by applying a time window to the input signal, the time window having the same duration as the plurality of model signals, and the method further comprises moving the window in time through the input signal and recalculating the probability function for each one of a plurality of positions of the window, to determine whether the biometric event is present at different times in the input signal.
In some embodiments according to the first aspect, the plurality of model signals are each arranged to have a peak amplitude at the same position within the signal, and in response to a determination that the input signal includes the biometric event the method further comprises identifying a time index of one of the plurality of samples of the input signal at an equivalent position to the position of the peak amplitude within the model signals, and recording the time index of the identified sample for the detected biometric event.
In some embodiments according to the first aspect, the biometric event comprises one of a heartbeat, a variation in a heartbeat and a user's activity. Where the biometric event comprises a heartbeat the plurality of model signals comprising a plurality of known heartbeat signals.
In some embodiments according to the first aspect in which the biometric event to be detected is a heartbeat, determining that the input signal includes a heartbeat comprises: identifying a probable heartbeat, in response to the probability being higher than the threshold; determining a time period between the probable heartbeat and an immediately preceding heartbeat in the input signal; and determining whether the probable heartbeat is an actual heartbeat based on a comparison between the determined time period and a known pulse rate.
In some embodiments according to the first aspect in which the biometric event to be detected is a heartbeat, determining whether the probable heartbeat is an actual heartbeat comprises: determining an expected interval between heartbeats based on the known pulse rate, and determining that the probable heartbeat is not an actual heartbeat if the determined time period differs by more than a threshold amount from the expected interval. For example, in some embodiments the threshold amount is ±30% of the expected interval.
In some embodiments the method of the first aspect comprises a further step prior to determining the probability, of setting the probability of the biometric event occurring to zero for a predefined time following each detection of a biometric event.
One way of reducing computational complexity and saving processing power is to set the probability to zero prior to performing the step of determining the probability during a time period where one know that any biometric event detected will not or is extremely unlikely to be the biometric event. In this regard a heartbeat for example has a maximum rate and therefore following detection of a heart beat there will be a short period of time where another heart beat will not occur. Setting the probability to zero for this period reduces the time period needing to be analysed and therefore saves processing power while also increasing accuracy, as false positives that may occur due to noise in this period are eliminated.
Alternatively, in some embodiments according to the first aspect in which the biometric event to be detected is a heartbeat, determining whether the probable heartbeat is an actual heartbeat comprises determining that the probable heartbeat is not an actual heartbeat if the determined time period is less than a predefined minimum time period. For example, in some embodiments the predefined minimum time period is 200 milliseconds. This is an alternative way of reducing the number of false positives by ruling out signals that occur in a period where one assesses a heartbeat cannot occur.
In some embodiments according to the first aspect in which the biometric event to be detected is a heartbeat, the method further comprises identifying a subject from which the input signal was obtained by comparing the transformation matrix to a plurality of stored transformation matrices, each associated with a particular subject.
In some embodiments according to the first aspect, the method further comprises validating the input signal by determining the standard deviation of the standard deviation of the input signal, wherein the input signal is rejected if the standard deviation of the standard deviation is higher than a preset threshold.
According to a second aspect of the present invention, there is provided a computer-readable storage medium arranged to store computer program instructions which, when executed, perform a method according to the first aspect.
According to a third aspect of the present invention, there is provided apparatus for detecting an biometric event in an input signal, the apparatus comprising a principal component analysis PCA unit configured to perform PCA on samples of a plurality of model signals to generate a transformation matrix having more informative components and less informative components, each of the model signals comprising a known signal which includes the biometric event to be detected, and to reduce a dimensionality of the transformation matrix by discarding one or more of the more informative components, a sample transformation unit configured to transform a plurality of samples of the input signal using the reduced dimensionality transformation matrix, a probability determining unit configured to determine a probability that the biometric event is present in the plurality of samples of the input signal, by calculating a predefined probability function for the transformed samples, and an biometric event detecting unit configured to determine that the input signal includes the biometric event if the probability is higher than a threshold.
According to a fourth aspect of the present invention, there is provided apparatus for detecting an biometric event in an input signal, the apparatus comprising a processing unit comprising one or more processors, and memory arranged to store computer program instructions which, when executed by the processing unit, cause the apparatus to: perform principal component analysis PCA on samples of a plurality of model signals to generate a transformation matrix having more informative components and less informative components; reduce a dimensionality of the transformation matrix by discarding one or more of the more informative components; transform a plurality of samples of the input signal using the reduced dimensionality transformation matrix; determine a probability that the biometric event is present in the plurality of samples of the input signal, by calculating a predefined probability function for the transformed samples; and determine that the input signal includes the biometric event if the probability is higher than a threshold.
In some embodiments according to the third or fourth aspect, the biometric event to be detected is a heartbeat and the plurality of model signals comprise a plurality of known heartbeat signals, and the apparatus further comprises a sensor configured to obtain the input signal by recording values of a physiological parameter over time. In some embodiments the sensor may be a photoplethysmography sensor.
Embodiments of the present invention will now be described, by way of example only, with reference to the accompanying drawings, in which:
In the following detailed description, only certain exemplary embodiments of the present invention have been shown and described, simply by way of illustration. As those skilled in the art would realize, the described embodiments may be modified in various different ways, all without departing from the scope of the present invention. Accordingly, the drawings and description are to be regarded as illustrative in nature and not restrictive. Like reference numerals designate like elements throughout the specification.
Referring now to
First, in step S101, principal component analysis (PCA) is performed on a plurality of heartbeat samples. Each one of the plurality of heartbeat samples can be referred to as a model heartbeat, and comprises a signal which is known to each contain a single heartbeat. For example, the model heartbeats can be extracted from a suitable biometric signal, such as a photoplethysmography (PPG) signal or electrocardiograph (ECG) signal, which contains heartbeats at known points in time within the signal. The model heartbeats can be recorded in advance and stored in suitable non-volatile computer-readable memory for analysis at a later time.
An example of a plurality of heartbeat samples is illustrated in
The model heartbeats can be stored in an array with a number of rows i equal to the number of model heartbeats and a number of columns j equal to the number of samples in each model heartbeat. The element aij of the array therefore contains the jth sample of the ith model heartbeat. An example of an array containing ten samples for each of the eleven model heartbeats plotted in
The indexes in an array of model heartbeat samples taken from PPG measures are not independent, and so the covariance matrix will not be diagonal. However, they can be transformed into a space in which dimensions are orthogonal using PCA. In the present embodiment, in step S101 PCA is performed on the plurality of model heartbeats to generate a transformation matrix. The elements of the PCA transformation matrix are ordered according to variance, with the first element having the most variance and the last element having the least variance. The elements with higher variance can be referred to as more informative components, and the elements with lower variance can be referred to as less informative components. That is, the less informative components have lower variances than the more informative components.
It is known to perform dimensionality reduction on a PCA matrix by retaining the components with most variance and discarding the components with less variance, on the basis that the components with most variance contain more information about the differences between the signals that were used to calculate the PCA matrix. In contrast however, in embodiments of the present invention the dimensionality is reduced by discarding the more informative components, that is, the components of the PCA matrix which have higher variances. As explained above, the inventors of the present invention have noted that the components with less variance (i.e., the less informative components) give a better indication of whether a particular signal belongs to the distribution than components with more variance, since points lying far from the mean values on dimensions which have less variance will indicate that the sample does not belong to the distribution.
Therefore in the present embodiment, in step S102 the dimensionality of the PCA transformation matrix is reduced from n to k by discarding the (n−k) most informative components, where n is the size of the original PCA transformation matrix, and k is the number of retained components. This is equivalent to transforming the samples of the model heartbeats from the original PPG space into a space where the dimensions are orthogonal, ensuring that only those points with small variance are retained. For example, by performing dimensionality reduction the number of dimensions in the transformation matrix may be reduced from 100 to 10. In some embodiments, dimensionality reduction can be performed by discarding a fixed number (n-k) of the more informative components. In other embodiments, dimensionality reduction can be performed by discarding any components with a variance higher than a certain threshold.
Next, in step S103 the reduced-dimensionality transformation matrix is applied to samples of the input signal. This has the effect of transforming the input signal into a space in which dimensions are orthogonal, and where the dimensions are ordered by the amount of variance. In embodiments in which the model heartbeats were normalised before performing PCA, the same transformation in terms of centering and/or scaling the amplitude values may also be applied to the samples of the input signal, before applying a rotation using the reduced-dimensionality transformation matrix.
Then, in step S104 a predefined probability function is calculated for the transformed samples. The probability function calculates the probability that the samples of the input signal belong to the distribution that was used to create the transformation matrix, specifically, the distribution of sample values for a plurality of model heartbeats. The output of the probability function is therefore related to the probability that the input signal includes a heartbeat.
Next, in step S105 the probability that was calculated in step S104 is compared against a threshold. If the probability is higher than the threshold, it is determined that the input signal contains a heartbeat. On the other hand, if the probability is lower than the threshold, it is determined that the input signal does not contain a heartbeat.
By reducing the dimensionality of the PCA transformation matrix, as described above in relation to step S102 of
For example, if a smartphone sensor is used to record a PPG signal with a total duration of 1 second at a sampling rate of 240 Hz, the full covariance matrix would have a size of 240×240. Without dimensionality reduction, this would result in 240×240=57,600 multiplications having to be performed 240 times each second. By comparison, if dimensionality reduction is performed by retaining the 10 least informative components and discarding the remaining more informative components, then it is only necessary to perform 240×10=2,400 multiplications during each sampling period, followed by calculating the disjoint probability of 10 points, an operation which is O(n) and therefore involves a number of operations in the same order of magnitude as the number of reduced dimensions. Therefore embodiments of the present invention may be particularly advantageous in applications where the available processing resources are limited, for example in wearable devices or other types of mobile device such as smartphones.
Referring now to
In some embodiments, a sliding probability function can be calculated by moving the window in time through the input signal and recalculating the probability function for each one of a plurality of positions of the window, to determine whether a heartbeat is present at different times in the input signal. An example of a sliding probability function calculated for the signal of
In the example shown in
In the present embodiment the model heartbeats are arranged so as to have the amplitude peak at sample index j=5, which occurs four sampling intervals after the first sample. The probability function calculated at step S104 of
In
Embodiments of the present invention can also reliably detect heartbeats in noisy input signals.
Referring now to
First, in step S201 the time at which the probable heartbeat occurs in the input signal is noted. For example, when a sliding probability function is used as described above, the time of the probable heartbeat can be determined based on the current starting point of the time window and the known position of the heartbeat in the model heartbeats.
Next, in step S202 the time period between the probable heartbeat and the immediately preceding heartbeat in the input signal is determined. If the probable heartbeat is an actual heartbeat, then this time period represents the interval between consecutive heartbeats. In step S203, it is checked whether the determined period is greater than a predefined minimum time period, which can be referred to as a minimum pulse interval. The minimum pulse interval may be set to be lower than the shortest interval that would be expected for a realistic maximum heart rate. If the determined time period is found to be less than the minimum pulse interval, then in step S204 it is determined that the probable heartbeat cannot be an actual heartbeat.
In the present embodiment the minimum pulse interval is set to 200 ms (milliseconds), which is equivalent to a heart rate of (1000/200)*60=300 bpm (beats per minute). Since the maximum heart rate of a human is generally expected to be around 200-220 beats per minute, it can be assumed that if the time period calculated in step S202 is less than 200 ms, the probable heartbeat cannot be an actual heartbeat since the pulse rate could not be that high. It will be understood that in other embodiments a different minimum pulse interval may be set. For example, in some embodiments a value of less than or equal to 270 ms may be used, equivalent to a heart rate of approximately 220 bpm.
If the determined time period is found to be greater than the minimum pulse interval, then the probable heartbeat may be an actual heartbeat. Accordingly, in step S205 the time period that was determined in step S203 is compared to an interval between consecutive heartbeats that would be expected based on a current pulse rate. For example, the current pulse rate can be determined based on the total number of heartbeats that have been detected within a preceding predefined time period or can be determined based on the average interval between a predefined number of heartbeats.
In step S205, the time period is determined to be consistent with the expected interval if it differs from the expected interval by less than a threshold amount. If the time period is not found to be consistent with the expected interval, then in step S206 it is determined that the probable heartbeat cannot be an actual heartbeat. On the other hand, if the time period is consistent with the expected interval, then in step S207 it is determined that the probable heartbeat is an actual heartbeat.
In step S205, the threshold for determining whether or not the time period is consistent with the expected interval can be defined in relative or absolute terms, for example as a percentage of the expected interval or as a fixed time difference. In the present embodiment the time period determined in step S202 is deemed to be consistent with the expected interval if it is within ±30% of the expected interval. However, in other embodiments a different threshold may be used.
The checks provided in steps S203 and S205 may be applied in order to verify whether or not a probable heartbeat detected using a method such as the one shown in
In some embodiments a similar logic may be applied before using a process such as the one shown in
Referring now to
Depending on the embodiment, the processing unit 310, memory 320 and sensor 330 may be embodied in the same physical device or may be physically separate from one another. For example, the processing unit and memory may be included in one device, such as a smartphone, and the sensor 330 may be included in a physically separate device that can communicate with the processing unit 310 via a suitable wired or wireless connection, for example in a wearable device such as a smartwatch which includes an integrated PPG sensor, or a chest strap with integrated heart rate sensor.
As shown in
The PCA unit 311 is configured to perform PCA on samples of a plurality of known heartbeat signals to generate a transformation matrix, and to reduce a dimensionality of the transformation matrix by discarding one or more of the more informative components, as described above in relation to steps S101 and S102 of
Embodiments of the present invention have been described which can be used to determine whether an input signal contains a heartbeat. In some embodiments, the input signal can be validated before proceeding to check whether a heartbeat is present, to avoid unnecessarily expending processing resources when the input signal is unsuitable for detecting a heartbeat. For example, in one embodiment the input signal can be validated by determining the standard deviation of the standard deviation of the input signal, which may also be referred to as the second momentum of the input signal.
Finally, embodiments of the present invention have been described in which a PCA transformation matrix is derived from a plurality of model heartbeats. In some embodiments, the system can adapt to a particular individual's characteristics by updating the model heartbeats using heartbeats extracted from the input signal. This can improve the accuracy for that particular individual, by training the system to recognize the characteristic waveform of that user's heartbeat. Furthermore, in some embodiments a plurality of PCA transformation matrices may be stored for different users, enabling the system to identify a subject from which the input signal was obtained by comparing the transformation matrix to the plurality of stored transformation matrices.
Embodiments of the present invention have been described in relation to detecting heartbeats in physiological signals such as PPG or ECG signals. However, in other embodiments of the invention the same principles disclosed above may be applied to process different types of biometric signals. In general, the PCA-based techniques disclosed herein can be used to detect any type of biometric event in a noisy signal. For example, in some embodiments the PCA-based event detection method may be applied to detect a user performing a certain activity such as a step from a noisy signal indicative of a user's movement.
Whilst certain embodiments of the invention have been described herein with reference to the drawings, it will be understood that many variations and modifications will be possible without departing from the scope of the invention as defined in the accompanying claims.
A person of skill in the art would readily recognize that steps of various above-described methods can be performed by programmed computers. Herein, some embodiments are also intended to cover program storage devices, e.g., digital data storage media, which are machine or computer readable and encode machine-executable or computer-executable programs of instructions, wherein said instructions perform some or all of the steps of said above-described methods. The program storage devices may be, e.g., digital memories, magnetic storage media such as a magnetic disks and magnetic tapes, hard drives, or optically readable digital data storage media. The embodiments are also intended to cover computers programmed to perform said steps of the above-described methods.
Features described in the preceding description may be used in combinations other than the combinations explicitly described.
| Number | Date | Country | Kind |
|---|---|---|---|
| 1816386.5 | Oct 2018 | GB | national |
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/GB2019/052841 | 10/8/2019 | WO | 00 |
