WO2024033067A1

WO2024033067A1 - Method for detection of characteristic regions in a time-continuous signal, corresponding device and wearable

Info

Publication number: WO2024033067A1
Application number: PCT/EP2023/070424
Authority: WO
Inventors: Andreas Philipp HASSLER; Klaus Pieper; Oelweiner EVA-MARIA
Original assignee: ams Sensors Germany GmbH
Priority date: 2022-08-11
Filing date: 2023-07-24
Publication date: 2024-02-15

Abstract

Computer-implemented method for detection of characteristic regions in a time-continuous signal, comprising the steps of • obtaining sampled values of the signal by measuring a signal with a sampling rate (100, 102); • conducting a coarse detection of at least one characteristic regions (106); • for each characteristic region, building a current template (108); • building a matching template (110); • interpolating the signal at least in a characteristic region and the matching template with an interpolating rate, yielding an interpolated signal segment and an interpolated matching template (104, 112); • matching the interpolated template with the interpolated signal segment to identify the position of characteristic regions in the signal (114).

Description

METHOD FOR DETECTION OF CHARACTERISTIC REGIONS IN A TIME- CONTINUOUS SIGNAL , CORRESPONDING DEVICE AND WEARABLE

DESCRIPTION

Technical Field of the Disclosure

The invention relates to problem areas , where there is a time continuous signal which has to be sampled . It more speci fically relates to applications in which characteristic regions in the signal have to be identi fied with high accuracy .

Background

Time-continuous signals can stem from multiple sources , for example electrical and optical signals . The sampling rate is usually adj usted to satis fy certain requirements related to time and/or signal value resolution .

Often there is a benefit in reducing the sampling rate regarding energy consumption . This especially is true for optical sensing, as the ON time of the light emitting unit ( e . g . , laser, LED) can hereby be reduced . By reducing the sampling rate , also the time resolution and accuracy is directly reduced . For example , at 100Hz a sample each 10ms is recorded . Hence , it is not possible to determine an appearing event with an accuracy of 1ms , without making use of additional signal processing and certain techniques .

Many physiological signal sources produce repetitive cyclic signals , like the phonocardiogram ( sounds and murmurs made by the heart ) , the electrocardiogram ( electrical signal from the heart ) and the photoplethysmogram ( optically obtained plethysmogram that can be used to detect blood volume changes in the microvascular bed of tissue ) . Often one wants to detect the appearing of these cyclic signal segments with a high precision in time . This normally requires high sampling rates . However, especially for battery powered devices such as wearables this is a problem, as a high sampling rate requires more energy and hence, the battery drain is increased .

Timely accurate detection of repetitive features requires usually high sampling rates. However, some applications (like wearables) , which have quite limited energy resources (battery driven or energy harvesting systems) can't allow such high sampling rates for long. Hence, there is a tradeoff between time-accuracy and sampling rate.

Some signals such as bio signals change their characteristics over time and have a lot of artifacts and noise present, which makes robust and accurate detection of repetitive features (e.g., PPG valley) very hard.

There is a huge market for sports, health & wellness wearables. A trending parameter is the heart rate variability (HRV) which can be derived also using PPG signals (pulse rate variability) . HRV parameters are related to activities of the autonomic nervous system and give information about the sympathetic ( fight/ flight ) and parasympathetic (relax) nervous tonus. Hence, this parameter can give insights about stress state, diseases, exercise recovery, well-being.

It is very common that features of biological signals are detected using peak detection techniques. This includes local maxima and minima detection in continuous time signals. Examples: R-peak detection in ECG signals, peak detection in PPG signals.

The main problem currently with known methods of feature detection such as the PRV approach is the low accuracy of the derived parameters. This is related to the fact that for example PPG signals don't have repetitive "sharp" features like ECG has the R-Peak Summary

The obj ect of the invention is to provide an accurate computer-implemented feature detection method with low energy consumption . Moreover, a corresponding device and wearable are provided .

With respect to the method, this obj ect is solved by method comprising the steps of

• obtaining sampled values of a signal by measuring a signal with a sampling rate ;

• conducting a coarse detection of at least one characteristic region;

• for each characteristic region, building a current template ;

• building a matching template ;

• interpolating the signal at least in a characteristic region and the matching template with an interpolating rate , yielding an interpolated signal segment and an interpolated matching template ;

• matching the interpolated template with the interpolated signal segment to identi fy the , especially exact , position of characteristic regions in the signal .

Preferred embodiments of the method are described in the dependent claims and in the figure description .

The invention is based on the consideration that to reduce energy consumption and possibly other resources , the sampling frequency of a signal of interest should be as low as possible or at least rather low . This low sampling frequency in known methods sets a limit on the accuracy of detection of signal features in time . The lower the sampling rate the greater the inaccuracy of determining the point in time corresponding to this feature . I f a higher accuracy would be needed, the sampling rate would have to be increased which is not always possible or desirable . Applicant has found that the accuracy of feature detection can be enhanced without increasing the sampling frequency by creating a template from a region of interest and by interpolating the template and the corresponding signal region with a frequency / interpolating frequency or rate which is quite larger than the sampling rate . The template is matched with the signal , and then in the high-resolution template , relevant features such as maxima or minima can be determined by high accuracy . The demand for higher accuracy is thereby shi fted into the computational domain . As the sample rate can remain rather low, the energy consumption can be reduced without a drop in accuracy . The signal segment is a signal segment of interest which comprises characteristic regions / features .

Applicant has thereby found as a key feature the combination of a very low sampling measurement system with a sophisticated technique , which allows to provide highly timing-accurate determinations of regions of interest and at the same time allows reducing energy consumption on di f ferent applications by a huge factor .

The method according to the invention is a general method applicable to a broad area of signals , namely to all time continuous signals where a region of interest or characteristic region has to be detected with high time accuracy while using a very low sampling rate .

The term "computer-implemented" encompasses reali zation in hardware or software or combined in hardware and software .

Advantageously, the matching template is the current template . In this case the matching procedure is always performed based on a template built from the current characteristic region .

In a preferred embodiment of the method, the matching template is a weighted average of the template of the current template and at least one previous template and/or whereby based on a number of ( several ) characteristic regions an initial template of a characteristic region is built .

By composing the matching template in such a way that it comprises a part of one or more previous templates and the current template , a change of the signal in time is considered . Advantageously, the first or initial template is built when a number or plurality of characteristic regions have been found already . In such a way, the template with which the matching starts incorporates already knowledge of several characteristic regions . The proposed method thereby includes robust adaptation to the signal ; this feature is introduced to robustly adapt continuously to the signal using the signal history ( " ageing function" ) .

Additionally, it comprises individual adaptation to the signal , which is very useful in case the time continuous signal is changing its characteristics over time ( e . g . region of interest is changing shape , amplitude , ... ) . This for example allows sampling at 20Hz ( time accuracy 1 /20 of a second, equals 50ms ) while having an accuracy for the detection of regions of interest within a few milliseconds .

In a preferred variant , the matching template is a weighted average of a certain percentage of the current template and a certain percentage of the previous template or certain percentages of previous templates , such that the sum of the percentages is 100 . In an advantageous embodiment , the matching template is a weighted average of 25% of the current template and 75% of the previous template . Depending on how quickly the signal might be changing and how strong the noise and artefacts are , the proportion of the current template and the previous template or templates has/have to be adapted . Fast adaptation but less artefact and noise stability is achieved by a weighted average of 50% of the current template and 50% of the previous template . In contrast , a lot of robustness is achieved by a weighted average of 10% of the current template and 90% of the previous template . Advantageously, the matching is conducted by employing a similarity function, especially cross correlation . This approach ensures matching at the point of highest correlation, which in most cases is also the match showing highest similarity and therefore provides highly time- accurate matching .

The initial low sampling rate ("coarse" ) detection is preferably conducted by a search for at least one extremum in the characteristic region . The coarse detection serves to generally find / identi fy the characteristic region in the signal . It is performed on the signal sampled at the sampling rate and therefore does not comprise a very high accuracy compared to the accuracy with which the method finally can detect features .

Preferably, after matching of interpolated matching template and interpolated signal , an extremum is computed in the interpolated signal . The extremum can be a ( local ) minimum or maximum . Often minima or maxima are points in time that need to be located to generate useful output signals , for example the di f ferences of minima or maxima can yield information on the regularity of signals in time .

In a preferred embodiment , the interpolating rate is larger than the sampling rate by a factor between 1 and 1000 , especially 10 .

The sampling rate preferably lies between 5 and 50 Hz , especially 20 Hz .

In a preferred embodiment , after matching of interpolated template and signal , a feature or features ( at least one feature ) is/are extracted from the signal . This feature is for instance an extremum, i . e . , a minimum or maximum .

Using the obtained time-accurate points by this method, frequency analysis can be conducted and as a resulting feature , frequency characteristics can be obtained . In a preferred embodiment of the method, based on the matched templates , an output signal is generated . The output signal can, for instance , be further processed in a wearable and display the information to the user .

In a preferred variant , a PPG signal is sampled by an optical sensor, whereby the characteristic region comprises a a local extremum or a repetitive local characteristic morphology of the signal .

The method described above can be implemented advantageously by means of a computer program running in a computing environment/ hardware . A computer program ( also known as a program, software , software application, script , or code ) can be written in any form of programming language , including compiled or interpreted languages , and it can be deployed in any form, including as a stand-alone program or as a module , component , subroutine , or other unit suitable for use in a computing environment . A computer program does not necessarily correspond to a file in a file system . A program can be stored in a portion of a file that holds other programs or data ( e . g . , one or more scripts stored in a markup language document ) , in a single file dedicated to the program in question, or in multiple coordinated files ( e . g . , files that store one or more modules , sub programs , or portions of code ) .

A computer program can be deployed to be executed on one computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected by a communication network .

With respect to the device the obj ect is solved by a device comprising a sensor for sensing/measuring a signal and a computing unit which is configured to conduct a method described above . The device is computational device with a computing unit such as a processing unit , an embedded computer, a wearable , a smartphone etc . The processes and logic flows described in this specification can be performed by one or more programmable processors / processing units / CPUs executing one or more computer programs to perform functions by operating on input data and generating output.

Preferably, the sensor of the device is an optical sensor. Depending on the signal to be measured, other sensors can be employed. An example for electrical signals is the electrocardiogram (ECG) . It comprises of many characteristic and repetitive features like the P wave, the QRS complex and the T wave. One of its most prominent features is the R peak. Time-accurate detection of it is often desired for different purposes (heart rate, heart rate variability) .

Also, the seismocardiogram (SCG) which shows body vibrations induced by the heartbeat are a good field of application. Especially when someone is interested in the timely-accurate determination of the blood ejection from the heart in the blood stream.

With respect to the wearable, the object is solved by a wearable which is built as or which comprises a device described above.

The overall principle of a preferred embodiment of the method is to conduct first a rough or course detection on the low sampled signal (e.g., via minima/maxima detection) . Then some (>=1) of the previous regions of interest are aligned and a template is created (e.g., by averaging) . Subsequently, the current region of interest and the template is interpolated. Using cross correlation, the best position of the template on the region of interest is found and thereby the precise position in time.

Hence, template matching combined with interpolation allows the timely precise detection of events (e.g., PPG valley, R- peak of ECG) in low sampled signals. Adaptive template matching, where the template is changing according to the changing characteristics of the signal, introduces detection stability and makes it robust against noise and artifacts.

The advantages of the invention are especially as follows. This invention enables robust detection of regions of interest (e.g., R-peak, PPG valley) using a very low sampling rate and hence, a very low energy consumption (energy harvesting and battery driven systems, e.g., wearables) , while still maintaining high time resolution and accuracy. Hence, there is no relevant trade-off between accuracy and sampling rate. Combined with individual adaptation to the signal over time, which is very useful in case the time continuous signal is changing its characteristics over time, this individual adaptation over time makes the approach also very robust against noise and artifacts.

The main technical benefits are that the proposed method allows solutions with low sampling rate, whereby the low sampling allows solutions with low energy consumption. Processing first the low sampled signal is computationally cheap .

The method according to the invention can be applied to all time continuous signals (e.g., PPG, EEG, telecommunication signals, ...) and is not limited to ECG signals only. The template ageing function renders the method robust against noise and artifacts.

The invention is addressing all problem areas with a time continuous signal which has to be sampled. These time continuous signals can stem from multiple sources, for example electrical and optical signals. The sampling rate is usually adjusted in order to satisfy some requirement related to time and/or signal value resolution. Brief Description of the Preferred Embodiments

A preferred embodiment of the invention is described in connection with a drawing . In the drawing :

FIG . 1 an exemplary photoplethysmogram signal ;

FIG . 2 an exemplary ECG signal ;

FIG . 3 exemplary original and up-sampled photoplethysmogram signals ;

FIG . 4 a signal with characteristic regions and templates ;

FIG . 5 a diagram of a preferred method, and

FIG . 6 a detection device in a preferred embodiment .

Identical parts are labelled with the same reference signs .

Detailed Description of the Preferred Embodiments

In FIG . 1 , an exemplary photoplethysmogram, measured from the wrist of a person, is shown . On an x-axis 2 , time is plotted, while on a y-axis 6 a signal 10 is plotted . The signal 10 comprises several maxima 14 and minima 18 .

In such kinds of signals with repetitive features , it is often needed to detect prominent points such as maxima 14 and minima 18 with high temporal accuracy . In the example shown, an accurate detection of the minima 18 is needed . This is often done using high sampling rates , some signal filtering and minima/maxima detection algorithms . In FIG . 2 , an ECG signal 20 is shown . The ECG signal 20 comprises very sharp peaks 24 which are easy to detect by standard maximum detection algorithms .

A very trending parameter is the heart rate variability (HRV) which can be derived also using PPG signals (pulse rate variability) . HRV parameters are related to activities of the autonomic nervous system and give information about the sympathetic ( fight/ flight ) and parasympathetic ( relax ) nervous tonus . Hence , this parameter can give insights about the stress state , diseases , exercise recovery, and wellbeing .

The Heart Rate Variability (HRV) therefore is one of the most promising and widely used marker describing the activity of autonomic nervous system and represents the variation of time intervals between heartbeats . HRV is a trending parameter as it is used a lot for tracking training recovery, well-being, stress levels and even as indicator for diseases and disease onset . Traditionally, HRV is derived from the series of instantaneous cycle intervals obtained from the electrocardiogram (ECG) . However, it is also possible to estimate the variation in heart rate from a photoplethysmography ( PPG) signal , this then is called pulse rate variability ( PRV) . Advantage is here , that there is not the need to have a closed electrical loop through the body, as in the ECG . The PPG can be measured on di f ferent body locations ( finger, wrist , chest , head, ear, ... ) by single spot measurement .

The main problem currently with known PRV approaches is the low accuracy of the derived parameters . This is related to the fact that PPG signals do not comprise repetitive sharp features as the ECG signal 20 comprises sharp peaks 24 (R-

Peaks ) .

In FIG . 3 , an exemplary diagram with several measured PPG signals 30 , 34 , 36 is shown . On the x-axis 2 , the time is plotted, while the signals are plotted on the y-axis 6 . The PPG signal 30 , 34 , 36 which are sampled with a finite resolution comprise characteristic regions with respective minima 40 . Of special interest is the time di f ference or interval 42 between two adj acent minima 40 . Other regions of interest can also be of possible interest .

The minima 40 in the characteristic regions 38 are not as distinct as the extrema in an ECG signal 20 as shown above in FIG . 2 . In order to obtain a high precision detection with standard methods , a very high sample rate would be needed which results in large energy consumption . In mobile devices such smartwatches , such a high sample rate is not possible due to the limited battery capacity . The sample rate is therefore reduced, leading to a loss of accuracy .

The present invention tackles this problem and provides a solution, which provides a detection of characteristic regions 38 with high accuracy and at the same time only requires a rather low sample rate . To achieve this goal , as will be explained below, a combination of a filter approach, interpolation and individual ageing templates for prominent regions of the PPG and to find the best fit (highest cross correlation) is employed . This leads to performance boosts to very high accuracy .

In FIG . 4 , a sample signal 50 is shown which is plotted on the y-axis 6 , while on the x-axis 2 again time is plotted . The signal 50 comprises repetitive characteristic regions 38 which, respectively, comprise a decreasing or falling flank 54 , a minimum 56 , and an increasing or rising flank 58 .

A key aspect of the present invention is that a ( changing) template 60 is generated which represents the characteristic regions 38 . In the present example , the template 60 comprises the flanks 54 , 58 and the minimum 56 encompassed by the flanks 54 , 58 .

In FIG . 5 , a diagram is shown which represents a method according to the invention in a preferred embodiment . A first box 100 represents an optical sensor which is employed to optically measure a signal, especially a biological signal, in the present case a PPG signal as shown in FIG. 3.

A second box 102 represents the sampling of the signal, which in the present embodiment is conducted at a sampling frequency of 20 Hz. If the PPG signal sampled at this frequency would be directly fed into a standard minima detection algorithm, the accuracy of the detection of minima and therefore the computation of time differences of minima would be rather inaccurate. As mentioned above, a way to increase the accuracy would be to increase the sampling rate, leading to a much higher energy consumption.

The invention allows a detection with higher accuracy and still keeping the low sampling rate.

To this end, in a box 104, at least a part / segment of the signal, especially the characteristic region, is interpolated at an interpolation rate which is larger than the sampling rate. In the present example, the interpolation rate is 200 Hz, i.e., it is larger than the sampling rate by a factor of 10.

In a box 106, several characteristic regions 38 are detected in the signal. In the present example shown, 6 characteristic regions are identified by a minimum detection algorithm which is applied to the signal sampled by the sampling rate of 20 Hz .

The rough detection of these characteristic regions can be determined by a primitive local extremum detection (lowest / highest value in defined region) . To improve performance, the signal can be slightly smoothed (e.g., small windows size rolling mean or median filter) .

In a box 108, from the signal regions of the minima / characteristic regions identified, a template is composed. This is a first or initial template which the method uses for the detection of characteristic regions as will be explained below .

For each characteristic region, a current template is created . Once the initial template has been created in box 108 , as the method progresses in a box 110 , a weighted template is created .

The weighted template is an averaged sum of one or more previous templates and the recent or current template . In the present preferred embodiment shown, the weighted template comprises 25% of the current template and 75% of the previous template . The weighting or weighted averaging is especially done like this : for each point in time , the template value is the sum of 0 . 25 times the signal value of the current template and 0 . 75 times the signal of the previous template . In general , with Tx ... T_N being N templates and ai ... a_N being N coef ficients of which the sum is 1 , the averaged template T_A is built as T_A = sum a_N * T_N, whereby as explained above , the summation is conducted for each point of the template

In a box 112 , the weighted template which, as the original signal , comprises a time resolution of the sampled signal , here 20 Hz , is interpolated with an interpolation rate which in the present case is 200 Hz , i . e . , 10 times larger than the sampling rate . The weighted template is a matching template which is used for the matching of template and signal .

In a box 114 , a matching is performed between the weighted template and the characteristic region of interest for which the current template has been created .

The matching of the weighted interpolated template and the interpolated characteristic region of the signal is in the present embodiment and preferably conducted using cross correlation, i . e . , the template is moved across the characteristic region whereby for each relative positioning a correlation function is computed, and whereby the best match is identi fied at an extremum of this correlation function . Once the template has been correlated / matched with the characteristic region, in the signal the minimum is determined . As the interpolated template has an interpolation rate which is 10 times larger compared to the initial sampling rate of the signal , the minimum detection is very accurate .

The use of the weighted templates takes into account the fact that the signal ages , i . e . , it slightly changes its characteristics over time . This is especially the case for biological signals due to the physiology of the (human) body . For example , the characteristic regions can change their shape over time due to artery relaxation .

The method is summari zed again below which helps to boost the accuracy for low sampled signals . First an optical sensor samples at a low sampling rate a PPG signal (boxes 100 , 102 ) . Then 6 minima peaks are roughly detected and used as basis for the creation of a template of these minima regions (boxes 106 , 108 ) . This is done by taking the aligned mean of these 6 minima regions . There is also the possibility to add an ageing function and to use the weighted mean of 25% of the currently created template and 75% of the previously created template (box 110 ) . For the start of the method, the initial detection, a global template can be used instead of the previous template .

When the template has been created, both the PPG minima region of interest and the template are interpolated (here for example from 20Hz to 200Hz , see boxes 104 and 112 ) . Then, using, cross-correlation the best match of the template and the signal segment is found (box 114 ) . The resulting position represents the precise minimum position .

In FIG . 6 , a device 140 , especially a detection device is shown schematically . The device 140 comprises a sensor 144 and a computing unit 146 , whereby the sensor 144 is connected to the computing unit 146 on its input side . The computing unit 146 is configured to conduct a method described above . The device 140 can especially be built as a wearable or be part / be integrated in a wearable such as a smart watch or bracelet .

LIST OF REFERENCE SIGNS x-axis y-axis signal maximum minimum ECG signal peak PPG signal PPG signal PPG signal characteristic region minimum interval signal falling flank minimum rising flank template box box box box box box box box device sensor computing unit

Claims

1. Computer-implemented method for detection of characteristic regions in a time-continuous signal, comprising the steps of

• obtaining sampled values of the signal by measuring a signal with a sampling rate (100, 102) ;

• conducting a coarse detection of at least one characteristic region (106) ;

• for each characteristic region, building a current template ( 108 ) ;

• building a matching template (110) ;

• interpolating the signal at least in a characteristic region and the matching template with an interpolating rate, yielding an interpolated signal segment and an interpolated matching template (104, 112) ;

• matching the interpolated template with the interpolated signal segment to identify the position of characteristic regions in the signal (114) .

2. Method according to claim 1, whereby the matching template is the current template.

3. Method according to claim 1, whereby the matching template is a weighted average of the current template and at least one previous template and/or whereby based on a number of characteristic regions, an initial template of a characteristic region is built.

4. Method according to claim 2, whereby the matching template is a weighted average of a certain percentage of the current template and a certain percentage of the previous template or certain percentages of previous templates, such that the sum of the percentages is 100.

5. Method according to one of the claims 1 to 4, whereby the matching is conducted by employing a similarity function, especially cross correlation.

6. Method according to one of the claims 1 to 5, whereby the coarse detection is conducted by a search for at least one extremum in the characteristic region.

7. Method according to one of the claims 1 to 6, whereby after matching of the interpolated matching template and the interpolated signal, an extremum is computed in the interpolated signal (114) .

8. Method according to one of the claims 1 to 7, whereby the interpolating rate is larger than the sampling rate by a factor between 1 and 1000, especially 10.

9. Method according to one of the previous claims, whereby the sampling rate lies between 1 and 100 Hz, especially 20 Hz .

10. Method according to one of the previous claims, whereby after matching of interpolated template and signal, at least one feature is extracted from the interpolated signal.

11. Method according to claim 10, whereby based on the matched templates, an output signal is generated.

12. Method according to one of the previous claims, whereby a PPG signal is sampled by an optical sensor, and whereby the characteristic region comprises a local extremum or a repetitive local characteristic morphology of the signal.

13. Device (140) , comprising

• a sensor (144) for sensing a signal;

• a computing unit (146) which is configured to conduct a method according to one of the previous claims.

14. Device (140) according to claim 13, whereby the sensor (144) is an optical sensor.

15. Wearable, built as or comprising a device (140) according to claim 13 or 14.