System and method for measuring blood pressure

Abstract

A system and method of measuring blood pressure behaviour of a subject. One method comprises extracting at least one waveform from a sound recording of operation of the subject's heart, each waveform representing sounds generated by operation of the subject's heart; averaging each of the extracted waveforms over a predetermined interval; and extracting features of the each of the averaged waveforms based on shape of the averaged waveforms. Other methods extract at least one element as the feature, the element(s) being a ratio of of form Xαβ_YδΦ, where each of X and Y is a peak or notch in any of the deriving waveform shapes, each of α and δ identifies one of the plurality of determined waveform shapes and each of β and Φ, relates to the allocated number of the corresponding peak or notch. The method further comprises providing the extracted features to a model trained for the subject to determine a measurement of blood pressure behaviour of the subject.

Description

SYSTEM AND METHOD FOR MEASURING BLOOD PRESSUREReference to Related Application(s)

[0001] This application claims Convention priority from Australian Patent Application No. 2023902904, filed 08 September 2023, the contents of which are incorporated herein by reference.Technical Field

[0002] The present invention relates generally to a method of measuring blood pressure and / or blood pressure variations and, in particular, to a method of measuring blood pressure and / or blood pressure variations based on sound signals such as a phonocardiogram. The present invention also relates to a system and apparatus for measuring blood pressure and / or blood pressure variations, and to a computer program product including a computer readable medium having recorded thereon a computer program for measuring blood pressure.Background

[0003] High blood pressure is a major risk factor for many conditions, including cardiovascular disease and chronic kidney disease. The ability to accurately monitor blood pressure is highly important in terms of monitoring the health of a patient and identifying potential health risks.

[0004] Traditional blood pressure measurement systems may not be tolerable to all patients. One known implementation for measuring blood pressure involves applying pressure to a cuff worn on an arm of a patient and measuring blood pressure as the cuff pressure changes. Implementations involving cuffs are often uncomfortable, invasive and impractical. Some patients can become stressed with the application of a cuff, increasing blood pressure and causing a reading that is not reflective of their normal state. Additionally, cuff based measurements cannot be acquired continuously nor over long periods of time, since such measurements interrupt the blood flow.

[0005] Other methods of measuring blood pressure include a pulse oximeter worn commonly worn on the finger or ear of a patient together with a phonocardiogram. Methods using a pulse oximeter often use pulse transit time in determining blood pressure. Pulse transit time measurements require a photoplethysmography measurement (using a pulse oximeter) located at a remote site on the body (finger tip, ear lobe, etc.) so that the time between the heart soundand the peak oxygen level at a more distant physiological site can be measured. Methods based on pulse transit time can also be unsuitable for acquiring blood pressure measurements continuously or over long periods of time, as the pulse transit time depends also on variables other than blood pressure such as mechanical properties of the arteries.

[0006] Wearable devices have become increasingly popular for monitoring health indicators such as pulse and blood pressure. Wearable devices often suffer difficulties in obtaining accurate measurements due to external changes such as sweat on skin, changes in skin colour and the like, which can take place due to changes of physical activity.

[0007] Another limitation of state-of-the-art wearable devices is related to the lack of calibration required to obtain accurate measurements from a particular subject. Universal calibration techniques are often used, which may be inappropriate for some specific users.

[0008] A non-invasive approach that can remove the discomfort an inconvenience of traditional blood pressure monitoring methods is required. Further, a solution that is less vulnerable to external changes such as sweat on skin, changes in skin colour and the like would be useful. A solution that is suitable for continuous, real-time or near real-time monitoring would also be useful. A solution for continuous monitoring would preferably be easy to use, enabling patients to monitor blood pressure without continuous medical assistance or presence in a medical facility.Summary

[0009] It is an object of the present invention to substantially overcome, or at least ameliorate, one or more disadvantages of existing arrangements, or provide a useful alternative thereto.

[0010] The arrangements described generally relate to using sound waves produced by the body, relating to operation of the heart, to measure blood pressure. The sounds are typically acquired by means of an electronic stethoscope in form of a phonocardiogram. The sounds can be sensed by other devices such as microphone, or any other sound sensor capable of detecting sound signals indicative of heart operation from a human body sufficiently to extract heart sounds such as S1 and S2 sounds.

[0011] In some arrangements, the sensed sounds are used to extract heart sounds such as S1 , S2, S3 and S4 sounds and at least one element is determined using the extracted heart soundsfor each patient. The element relates to a ratio present in any of the peaks and notches of each of the heart sounds that is determined to show behaviour correlating to changes in blood pressure.

[0012] According to a first aspect of the present disclosure, there is provided a computer- implemented method of measuring blood pressure behaviour of a subject, the method comprising: extracting at least one waveform from a sound recording of operation of the subject’s heart, each extracted waveform representing one of S1 and S2 sounds generated by operation of the subject’s heart; averaging each of the extracted waveforms over a predetermined interval; extracting features of each of the averaged waveforms based on a shape of each of the averaged waveforms; and providing the extracted features to a model trained for the subject to determine a measurement of blood pressure behaviour of the subject.

[0013] Another aspect of the present disclosure provides a non-transitory computer-readable medium storing instructions executable on a processor, the instructions configured to implement a method of measuring blood pressure of a subject, the method comprising: extracting at least one of an S1 and an S2 waveform from a sound recording of operation of the subject’s heart, each waveform representing one of S1 and S2 sounds generated by operation of the subject’s heart; averaging each of the extracted waveforms over a predetermined interval; extracting features of the each of the averaged waveforms based on amplitude a shape of each of the averaged waveforms; and providing the extracted features to a model trained for the subject to determine a measurement of blood pressure behaviour of the subject.

[0014] Another aspect of the present disclosure provides a system, comprising: a sound sensor configured to measure a sound recording of operation of a subject’s heart; a processor; and a memory, the memory storing instructions executable on the processor to: extract at least one waveform from the sound recording, each waveform representing one of S1 and S2 sounds generated by operation of the subject’s heart; average each of the extracted waveforms over a predetermined interval; extract features of the each of the averaged waveforms based on a shape of each of the averaged waveforms; and provide the extracted features to a model trained for the subject to determine a measurement of blood pressure behaviour of the subject.

[0015] Another aspect of the present disclosure provides a method of measuring blood pressure of a subject, comprising: capturing sounds made by the subject’s heart; identifying a plurality of waveforms from the captured sounds, each waveform corresponding to operation of the heart and having a plurality of peaks and notches, each of the peaks and notches allocated a label; determining an element for the subject, the element being derived at a training stage as showing correlation with variations in blood pressure, wherein the element is a ratio of absolutemagnitudes of form Xap Ys*, where each of X and Y is a peak or notch in any of the identified waveforms, each of a and 6 identifies one of the plurality of waveforms and each of 0 and (|) relates to the allocated label of the corresponding peak or notch in the corresponding waveform; and obtaining a blood pressure measurement by providing values of the determined element in the identified waveforms to a trained model.

[0016] Another aspect of the present disclosure provides a system for measuring blood pressure of a subject, comprising: a sound sensor configured to capture sounds made by the subject’s heart; a processor; and a memory, the memory storing instructions executable on the processor to: identify a plurality of waveforms from the captured sounds, each waveform corresponding to operation of the heart and having a plurality of peaks and notches, each of the peaks and notches allocated a label; determine an element for the subject, the element being derived at a training stage as showing correlation with variations in blood pressure, wherein the element is a ratio of absolute magnitudes of form Xap Ys*, where each of X and Y is a peak or notch in any of the identified waveforms, each of a and 6 identifies one of the plurality of waveforms and each of 0 and (|) relates to the allocated label of the corresponding peak or notch in the corresponding waveform; and obtain a blood pressure measurement by providing values of the determined element in the identified waveforms to a trained model.

[0017] Another aspect of the present disclosure provides a non-transitory computer-readable medium storing instructions executable on a processor, the instructions configured to implement a method of measuring blood pressure of a subject, the method comprising: receiving captured sounds made by the subject’s heart; identifying a plurality of waveforms from the captured sounds, each waveform corresponding to operation of the heart and having a plurality of peaks and notches, each of the peaks and notches allocated a label; determining an element for the subject, the element being derived at a training stage as showing correlation with variations in blood pressure, wherein the element is a ratio of absolute magnitudes of form Xap Ys*, where each of X and Y is a peak or notch in any of the identified waveforms, each of a and 6 identifies one of the plurality of waveforms and each of 0 and (|) relates to the allocated label of the corresponding peak or notch in the corresponding waveform; and obtaining a blood pressure measurement by providing values of the determined element in the identified waveforms to a trained model.

[0018] Other aspects are also disclosed.Brief Description of the Drawings

[0019] At least one embodiment of the present invention will now be described with reference to the drawings, in which:

[0020] Fig. 1 shows an example system for measuring blood pressure;

[0021] Fig. 2 shows an experimental arrangement for measuring blood pressure based on the system of Fig. 1 ;

[0022] Figs. 3A and 3B collectively form a schematic block diagram representation of an electronic device upon which described arrangements can be practised;

[0023] Figs. 4A and 4B show data from phonocardiogram readings of a subject during exercise and recovery periods respectively;

[0024] Figs. 5A and 5B show examples of variation in S1 and S2 sounds during exercise and recovery periods;

[0025] Fig. 6A to 6D show example average S1 and S2 waveforms over a predetermined interval during exercise and recovery phases;

[0026] Fig. 7 shows features of an example S1 waveform;

[0027] Figs. 8A and 8B show example ratios between selected peaks and notches of an S1 waveform during exercise and recovery periods;

[0028] Figs. 9A to 9F show example S1 waveforms and trends in S1 waveform max peak to next peak ratios in different subjects;

[0029] Figs. 10A to 10D show example frequency responses of S1 and S2 waveforms in exercise and recovery periods

[0030] Figs. 11A to 11 D show example Karhunen-Loeve Transform components of S1 waveforms captured in experimentation;

[0031] Fig. 12 shows a top-level flow chart of a method for measuring relative changes in blood pressure for a subject;

[0032] Fig. 13 shows a method of training a model for measuring relative changes in blood pressure for a subject as used in the method of Fig. 12;

[0033] Fig. 14 shows a method of determining relative changes in blood pressure as used in the method of Fig. 12;

[0034] Fig. 15 shows a top-level flow chart of a method for measuring blood pressure of a subject;

[0035] Fig. 16 shows a method of tuning or calibrating a model for measuring blood pressure of a subject as used in Fig. 15;

[0036] Fig. 17 shows a method of determining a blood pressure measurement as used in Fig. 15;

[0037] Fig. 18A shows features of an example S1 waveform;

[0038] Fig. 18B shows features of an example S2 waveform;

[0039] Figs. 19A to 19D show sets of ratio analysis for different elements individual, each ratio analysis tracking changes in correlation coefficients and blood pressure;

[0040] Fig. 20 shows a method of training a model for measuring relative changes in blood pressure for a subject;

[0041] Fig. 21 shows a method of determining features corresponding to changes in blood pressure measurements using heart sounds measured in the method of Fig. 20;

[0042] Fig. 22 shows a method of measuring relative changes in blood pressure for a subject using a model trained using the method of Fig. 20;

[0043] Fig. 23 shows a method of training a model for measuring blood pressure of a subject;

[0044] Fig. 24 shows a method of determining features corresponding to changes in blood pressure measurements using heart sounds measured in the method of Fig. 23; and

[0045] Fig. 25 shows a method of measuring blood pressure of a subject using a model trained using the method of Fig. 23.Detailed Description including Best Mode

[0046] Where reference is made in any one or more of the accompanying drawings to steps and / or features, which have the same reference numerals, those steps and / or features have for the purposes of this description the same function(s) or operation(s), unless the contrary intention appears.

[0047] It is to be noted that the discussions contained in the "Background" section and that above relating to prior art arrangements relate to discussions of documents or devices which form public knowledge through their respective publication and / or use. Such should not be interpreted as a representation by the present inventor(s) or the patent applicant that such documents or devices in any way form part of the common general knowledge in the art.

[0048] The arrangements described use body generated sound waves related to heart operation to measure blood pressure or variations in blood pressure. The sound waves are typically recorded using a phonocardiograph but may be measured by any sensor sufficiently sensitive to capture patterns of heart sounds in a manner that allows sounds such as S1 and / or S2 sounds to be detected, for example an electronic stethoscope, a wearable microphone, a wearable accelerometer, a parabolic microphone and the like. In some arrangements, the sound waves are typically recorded using a phonocardiograph but may be measured by any sensor sufficiently sensitive to capture patterns of heart sounds in a manner that allows heart sounds (such as S1, S2, S3 and S4 sounds or a subset thereof) to be detected, for example an electronic stethoscope, a wearable microphone, a wearable accelerometer, a parabolic microphone and the like. The sound waves are preferably captured using an external wearable device but may in some implementations be captured using an implantable device. Use of the body generated sounds captured using a sensor such as a phonocardiogram rather than a cuff or oximeter allows blood pressure to be measured in non-invasive manner and typically without discomfort to the patient. Fig. 1 shows an example system 100 for measuring blood pressure. The system 100 comprises a sound sensor (also referred to as a sound sensing device) 110 configured for communication with a processor device 130. In the example described, the sound sensor 110 is an electronic stethoscope. The sound sensor 110 can alternatively be any device configured for detecting and recording body generated sounds relating to heart operation through a patient’s skin with sufficient accuracy to allow features of the patient’s heartbeat to be identified. Example electronic stethoscopes include a Thinklabs One digital stethoscope, a 3M™ Littmann® CORE Digital Stethoscope and the like. Other acoustic sensors or microphones capable of detecting and recording body generated sounds relating to heart operation through a patient’s skin with sufficient accuracy to allow features of the patient’sheartbeat to be identified. While an external sound sensor is preferable, in some cases the body sounds can be recorded using sound sensors inside the body as well as using contactless sensors.

[0049] In the example of Fig. 1, the electronic stethoscope 110 communicates with the processor device 130 via a network 120. The network 120 may be wired or wireless, as described in further detail below. In the example described the processor device 130 is a smartphone. The processor device may be other types of devices such as a laptop computer, a desktop computer, a server or the like. In other implementations the processor device 130 and the sound sensing device 110 may form a single device.

[0050] The sound sensor 110 is used to record heart sounds of a human subject, typically by placement of the stethoscope acoustic sensor (not shown) on the skin in an area suitably near the heart of the subject, for example over skin 140. A phonocardiogram of the heart sounds starts recording when the electronic stethoscope 110 is switched on and stops when the electronic stethoscope 110 is stopped. The recorded phonocardiogram can be stored in a memory of the electronic stethoscope 110 and / or transmitted via the network 120 for storage in a memory of the processor device 130.

[0051] A normal human heartbeat typically comprises a number of known sound patterns. In particular, sounds known as the S1 , S2 are typically present in all individuals, whereas S3 and S4 sounds may be present or recognised in some individuals only. The S1 sound relates to closure of mitral and tricuspid valves of the heart. The S2 sound relates to closure of aortic and pulmonic valves of the heart. Some of the arrangements described relate to detection (or extraction) and use at least one of S1 and S2 sounds from sounds captured using the electronic stethoscope 110. The detected S1 and / or S2 sounds are used in some implementations to determine blood pressure and / or blood pressure changes at the processor device 130. The S3 sound is a relatively low-frequency, shorter vibration that occurs at the end of the rapid diastolic filling period of the right or left ventricle. The S4 sound results from the contraction of the atria pushing blood into a stiff or hypertrophic ventricle. The arrangements described relate to identification of an element represented as a ratio from peaks and / or notches of measured heart sounds. Some arrangements described use at least one of S1 and S2 sounds for a patient. In some implementations, S1, S2 and S3 sounds may be used. In other implementations, S1, S2, S3 and S4 sounds may be used. In other implementations, S1, S2 and S4 sounds may be used. The peaks and notches of the ratio element may be from the same heart sound or from different heart sounds of the subject. The detected sounds and corresponding element are used to determine blood pressure and / or blood pressure changes at the processor device 130.

[0052] In the course of researching the arrangements described a number of experiments were conducted. Fig. 2 shows an example experimental arrangement 200 used in the experiments. The system 200 comprises an electronic stethoscope 210 (corresponding to the sensor 110), a processor device 230, a Biopac MP150 Data Acquisition and Analysis system 235, a two-lead electrocardiogram (ECG) system 250, a force sensing resistor 260 and a blood pressure monitor 270.

[0053] The electronic stethoscope 210 corresponds to the sounds sensor 110 and the device 230 corresponds to the processor device 130. The blood pressure monitor 270 used was a clinically validated Omron IntelliSense HEM-907 Blood Pressure Monitor device having a cuff 270c used to take reference blood pressure readings 275 from a subject. The two-lead ECG system 250 was applied to skin of the subject to collect echocardiogram signals 255 of the subject. The force sensing resistor 260 was placed on an arm of the subject with the cuff 270c of the blood pressure monitor 270 wrapped around both arm and resistor to clarify start and stop times of blood pressure recordings for synchronisation of phonocardiogram (PCG) and blood pressure readings. A blood pressure recording was started when the cuff 270c began to tighten and stopped when the cuff was released.

[0054] The electronic stethoscope 210 used in the experiments conducted was a Thinklabs One digital stethoscope set to a bandwidth of 20-2000 Hz. Other frequency ranges could be used, based on the operating range of the stethoscope.

[0055] The Biopac system 235 was configured to receive and synchronise PCG signals 215 from the electronic stethoscope 210, electrocardiogram signals 255 from the ECG system 250 and signals 265 from the force sensing resistor 260.

[0056] In the experiments conducted, the blood pressure of each participant subject was initially measured over a certain period of time at rest. Then changes of blood pressure of each participant subject were induced by having each participant engage in an isotonic handsqueezing exercise using a dynamometer. Each participant squeezed the dynamometer at approximately 30% of their full strength and held the dynamometer for a fixed period of time. This exercise activity is known to promote an increase of blood pressure. The period when dynamometer was squeezed is referred to hereafter as the exercise period or activity period. After the exercise was completed, there was a recovery period during which time the blood pressure decreases, referred to hereafter as the recovery period. Measurements of both blood pressure 275 and the phonocardiogram 215 were made at certain intervals throughout the activity and recovery periods. Various other signals such as the ECG 255 were also measured simultaneously as control measurements, but the other signals were not used to determineblood pressure. The subject’s blood pressure, ECG and PCG recordings were made for 2- minute exercise and recovery periods in the results discussed below.

[0057] The PCG data 215 recorded by the stethoscope 210 was communicated via the network 120 and analysed at the processor device 230. In particular, some of the experiments conducted provided the phonocardiograms to models implemented on the processor device 130 to identify and analyse S1 and S2 waveforms of each subject’s phonocardiogram. S1 and S2 waveforms were identified using a Hidden Markov model in the experiments conducted.

[0058] Figs. 4A and 4B show graphs 400 and 450 of phonocardiogram amplitudes over time for exercise and recover phases respectively. The graph 400 shows a PCG trace 410, S1 maximum amplitudes 420 and S2 maximum amplitudes 430 for the subject’s exercise period.As shown in the graph 400, amplitudes of each of 410, 420 and 430 generally increase over the exercise period as the subject’s heart must work harder to pump more blood. The graph 450 shows a PCG trace 460, S1 maximum amplitudes 470 and S2 maximum amplitudes 480 for the subject’s recovery period. As shown in the graph 450, amplitudes of each of 460, 470 and 480 generally start to decrease at the start of the recovery period and begin to form a more regular pattern at the end of the recovery period.

[0059] Analysis conducted on experimental data also related to time intervals between consecutive S1 occurrences (referred to as S11 time intervals), consecutive S2 and S1 occurrences (referred to as S21 time intervals) and consecutive S1 and S2 occurrences (referred to as S12 time intervals).

[0060] Figs. 5A and 5B show graphs 500 and 550 of phonocardiogram amplitudes over time for exercise and recovery periods respectively. The graph 500 shows time intervals for each of S11 (510), S21 (520) and S12 (530) for the subject’s exercise period. As shown in the graph 500, a general sawtooth pattern was shown for each of 510, 520 and 530, corresponding to the subject’s breathing pattern during exercise. The graph 550 shows time intervals for each of S11 (560), S21 (570) and S12 (580) for the subject’s exercise period. As shown in the graph 550, the sawtooth pattern for each of 560, 570 and 580, becomes less pronounced as the subject’s breathing pattern evens during the recovery period.

[0061] The S1 and S2 waveforms were also analysed in the experiments conducted. In one set of analysis, each of the S1 and S2 waveforms was peak-to-peak normalised using Equation (1): xX’ —Max(X)-Min{X)(1)

[0062] Each of the normalised S1 and S2 waveforms was averaged over an interval. Figs. 6A and 6B show graphs of normalised S1 and S2 signals, which had been signal averaged over a12 second interval to reduce presence of noise. Fig. 6A shows a graph 600 of amplitude (y-axis) of an S1 signal compared to time (x-axis) averaged across 12 seconds in the exercise period.

[0063] An enlargement 605 of peak amplitude from the graph 600 shows systematic increases in S1 sound amplitude peaks as the exercise period progresses.

[0064] Fig. 6B shows a graph 620 of amplitude (y-axis) of an S2 signal compared to time (x- axis) averaged across 12 seconds in the exercise period. The S2 signal corresponds to the same phonocardiogram as used for the S1 sound in Fig. 6A. An enlargement 625 of peak amplitude from the graph 620 shows a general or systematic increase in S2 sound amplitude peaks as the exercise period progresses.

[0065] Figs. 6C and 6D show graphs of normalised S1 and S2 signals averaged over a 12 second interval for the subject of Figs. 6A and 6B during the recovery period. Fig. 6C shows a graph, 650, of amplitude (y-axis) of an S1 signal compared to time (x-axis). Each waveform is represented with a different colour and has been averaged over a time window of 12 seconds in the recovery period. An enlargement 655 of the highest peak from the graph 650 shows general decreases in the amplitude (or height) of such peak in S1 as the recovery period progresses. Fig. 6D shows a graph 670 of amplitude (y-axis) of 12 seconds average S2 signals in the recovery period. Similarly to 6C, the y-axis represents the amplitude of the signal and the x-axis is time. The S2 signal corresponds to the same phonocardiogram as used for the S1 sound in Fig. 6C. A general trend was also seen in the first notch of the S2 signal, whereby amplitude (or depth) of the first notch decreases as the recovery period continued.

[0066] The S1 and S2 waveforms were also analysed for trends in ratios between features of the S1 and S2 waveforms such as peaks and notches. The shape of an S1 or S2 waveform is typically unique to an individual. Fig. 7 shows a graph 700 of amplitude (y-axis) over time (x- axis) for an S1 waveform for an example subject. A trace 710 in the graph 700 shows an example waveform pattern for the S1 sound, having positive peaks P1-P5 and negative notches N1-N4. While the height and spacing of the peaks and notches can vary per individual, each individual’s S1 waveform will typically have a lowest notch and a highest peak. In the example of Fig. 7, the second peak, P2, is the peak with highest amplitude, whereas the first notch, N1, has the lowest amplitude. As notch and peak patterns can vary for an individual, different definitions of peak and notch numbering can be used. For example, in Fig. 7, notch 1 (N 1 ) relates to a first notch, also being the minimum notch and the peak used relates to the peak with a maximum value, P2.

[0067] The experiments conducted analysed ratios of P2 (second peak in Fig. 7) to N1 (first notch in Fig. 7) for S1 waveforms for a subject having an S1 waveform corresponding to the waveform 700. Figs. 8A and 8B show trends in a ratio of P2 to N1 for S1 sounds recorded inexercise and recovery phases. Fig. 8A shows a graph 800 of a P2 / N1 ratio over the exercise period. As shown in Fig. 8A the P2 / N1 ratio generally increases, with the exception of a limited number of outliers, as the subject continues to exercise. Fig. 8B shows a graph 850 of a P2 / N1 ratio over the recovery period. As shown in Fig. 8B the P2 / N1 ratio generally decreases, again with an exception of 1-2 outliers as the subject continues in the recovery phase, in contrast to Fig. 8A.

[0068] The examples used in Figs. 7 and 8 are for a specific subject. Other experiments for different subjects showed unique waveform patterns for maximum and adjacent peaks of S1 sounds. Fig. 9A shows an example S1 waveform 900 for a subject, having a maximum peak P1 and a next maximum peak P2. Fig. 9B shows a graph 905 of trends in a ratio of P1 to P2 for the subject of Fig. 9A. Fig. 9B is during isotonic exercise (to increase blood pressure). A general decrease of the particular ratio correlates with the increase of blood pressure in the example of Fig. 9B. Other individuals could show opposite or different behavioural trends. While individuals showed different behavioural trends, each individual trend was typically observed to be consistent and correlate well with the blood pressure change. Such trends were generally either positively or negatively correlated with the blood pressure changes depending on the subject.

[0069] Fig. 9C shows an example S1 waveform 910 for a subject, having a maximum peak P1 and a next maximum Peak P2. Fig. 9D shows a graph 915 of trends in ratio of P1 to P2 for the subject of Fig. 9C. Fig. 9D is also during isotonic exercise (to increase blood pressure).However, the example of Fig. 9D is for a different subject to Figure 9B. In the example of Fig. 9D, a general decrease trend of the ratio feature (P1-P2 ratio) is observed, with two outlier points at the end.

[0070] Fig. 9E shows an example S1 waveform 930 for a subject, having a maximum peak P2 and a next maximum Peak P3. Fig. 9F shows a graph 935 of trends in ratio of P2 to P3 for the subject of Fig. 9E in the exercise phase. As shown by each of Figs. 9B, 9D and 9F, each subject showed ratios between the S1 maximum peak to the next peak to follow similar increasing trends once occasional outliers were disregarded.

[0071] The experimental analysis conducted also included frequency analysis of S1 and S2 waveforms. A forward Fourier transform was applied to the 12-second averages of the S1 and S2 waveforms for exercise and recovery phases illustrated in Figs. 6A and 6B. Figs. 10A and 10B show spectrum responses (graphs 1000 and 1010) for S1 and S2 waveforms in the exercise period respectively. Figs. 10C and 10D show the spectrum responses for S1 and S2 waveforms (graphs 1020 and 1030) in the recovery period for the subject of Fig. 10A and 10B. The graphs 1000 and 1010 are shifted to higher frequencies than the graphs 1020 and 1030 in the example of Figs. 10A to 10D. Frequency shifting was seen to occur at particular frequenciesrather than across a full spectrum of frequencies and to vary on a subject by subject basis in the tests conducted.

[0072] Additionally, a Karhunen-Loeve Transform (KLT) analysis was performed on the averaged S1 and S2 waveforms. Similarly to using principal component analysis (PCA), a KLT analysis allows the S1 and S2 waveforms to be represented as sets of orthogonal functions. A KLT analysis extracts features of the waveform, the waveform shape being defined by amplitude. A KLT analysis was found to be particularly suitable in terms of minimising total mean squared error and yield in the experiments conducted. However, other transforms capable of representing the S1 (and / or S2) waveform as an orthogonal function could also be used, for example independent component analysis (ICA), Canonical Correlation Analysis (CCA), Empirical Mode Decomposition (EMD), Sparse Decomposition based on Defined Dictionaries and the like.

[0073] Figs. 11 A and 11 B show first KLT components of each of the averaged S1 and S2 sounds in the exercise phase. The experiments conducted showed the first KLT component weight varying systematically with measured blood pressure, as measured using the blood pressure monitor 270. For example, Fig. 11A shows a set of data 1100 having an average increase in coefficient value in the exercise phase for S1. Fig. 11 B shows a set of data 1105 having an average decrease in coefficient value in the exercise phase for S2.

[0074] The KLT analysis of averaged S1 and S2 sounds from the recovery phase was also performed. Figs. 11 C and 11 D show the first component weight for measurements taken in the recovery phase for the subject of Figs. 11 A and 11 B. The first KLT component weight was again observed to vary systematically with blood pressure, as measured using the blood pressure monitor 270. Fig. 11C shows a set of data 1110 having an average decrease in coefficient value in the recovery phase for S1. Fig. 11D shows a set of data 1115 having an average increase in coefficient value in the recovery phase for S2. As described above, different transforms can be used to represent S1 and S2 waveforms as orthogonal functions. Use of different transforms will involve determining trends in exercise and recovery for different components of the transfer result on a per-patient basis. In each instance, a resultant appropriate coefficient or component will need to be selected. The appropriate coefficient can be determined through experimentation to analyse trends for each coefficient in relation to measured blood pressure changes.

[0075] The experiments conducted, including the analysis described above, demonstrated evidence of relationships between features of each of S1 and S2 waveform sounds and blood pressure measurements. Similarly, consistent trends in behaviour were observed in S2 waveforms as for S1 waveforms, allowing correlation between features of the S2 waveforms and changes in blood pressure to be observed. Based on the experiments conducted, including but not limited to the analysis conducted above, a method of measuring blood pressure usingthe system 100 was developed. The arrangements described use sound recording of a subject’s heart such as a phonocardiogram of a subject’s heartbeat to determine a blood pressure measurement by extracting waveforms from the sound recording. Each waveform represents one of S1 and S2 sounds generated by operation of the subject’s heart and can be referred to as S1 waveforms and S2 waveforms respectively.

[0076] As described above, the following measurements showed corresponding responses or trends in the exercise phase (corresponding to increasing blood pressure) and the recovery phase (corresponding to decreasing blood pressure):• Maximum amplitude of each of S1 and S2 with or without normalisation and features of each of S1 and S2 related to relative amplitudes.• Time intervals between S1 and S2 (or consecutive S1 waveforms).• Frequency response of S1 and / or S2. For example, a FFT response or first components of a KLT performed on each of S1 and S2. Components of other decomposition transforms such as ICA, CCA, EMD and sparse decomposition can similarly be used to show trends in each of exercise and recovery.• Shape and features of each S1 and S2 related to amplitude of the waveform (peaks and notches and ratios between those).

[0077] As above, the shape of each of the S1 and S2 waveforms showed trends corresponding to changes in blood pressure. The shape of each waveform be reflected in the amplitude (and variation in amplitude) and / or shape of the peaks and notches of the waveform over the course of a heartbeat. The trends in shape of the waveform can be determined in the time domain (for example peak amplitude, pattern of notches and peaks or times between peaks for S1-1. S2-1, S2-2, or an analysis of the waveform shape such as KLT) and the frequency domain (for example FFT response).

[0078] In the responses or trends above, signals for one of S1 or S2 may be sufficient for use in determining blood pressure (for example maximum amplitude, frequency response, intervals such as S11 and features such as peaks and notches). While both S1 and S2 can be used for such trends, an S1 signal or an S2 signal alone may be sufficient. Other trends such as time intervals for S12 or S21, require both S1 and S2 sounds.

[0079] Particular features of the S1 and S2 sounds relating to amplitude and pattern of the waveform are the individual peak and notch patterns. The peak and notch patterns were shown to have systematic responses to increasing and decreasing blood pressure, dependent on the subject’s S1 and S2 waveform patterns. The features firstly relate to relationships between peaks and notches. As the peak-notch pattern of each individual can vary, identification of anappropriate profile can be used to determine which features should be used for each subject. For example, a database of waveform shapes (also referred to as specifications or agreements) can be developed. Each waveform shape has associated features for use in determining blood pressure behaviour. The appropriate agreement can be determined using techniques such as best fit matching and / or initial calibration to the individual.

[0080] In some implementations, placement of the sound measuring device (such as the stethoscope 110) can have some effect on the shape of the waveform. In some embodiments, placement of the sound measuring device 110 can be accounted for, for example, by prior calibration based on placement position such as varying vertical positioning on one of the left chest, the right chest or other areas of the chest. Additionally, frequency responses can be analysed on an individual basis and used to verify blood pressure changes.

[0081] The method of measuring blood pressure is described using the electronic stethoscope 110 and the processor device 130. In the example described herein, the processor device is a smartphone. Figs. 3A and 3B collectively form a schematic block diagram of a general-purpose electronic device 301 including embedded components, corresponding to the processor device 130, upon which the methods to be described are desirably practiced. The electronic device 301 can be, for example, a mobile phone, a tablet or another device in which processing resources are limited. Nevertheless, the methods to be described may also be performed on higher-level devices such as desktop computers, server computers, and other such devices with significantly larger processing resources.

[0082] As seen in Fig. 3A, the electronic device 301 comprises an embedded controller 302. Accordingly, the electronic device 301 may be referred to as an “embedded device.” In the present example, the controller 302 has a processing unit (or processor) 305 which is bidirectionally coupled to an internal storage module 309. The storage module 309 may be formed from non-volatile semiconductor read only memory (ROM) 360 and semiconductor random access memory (RAM) 370, as seen in Fig. 3B. The RAM 370 may be volatile, nonvolatile or a combination of volatile and non-volatile memory.

[0083] The electronic device 301 includes a display controller 307, which is connected to a video display 314, such as a liquid crystal display (LCD) panel or the like. The display controller 307 is configured for displaying graphical images on the video display 314 in accordance with instructions received from the embedded controller 302, to which the display controller 307 is connected.

[0084] The electronic device 301 also includes user input devices 313 which are typically formed by keys, a keypad or like controls. In some implementations, the user input devices 313may include a touch sensitive panel physically associated with the display 314 to collectively form a touch-screen. Such a touch-screen may thus operate as one form of graphical user interface (GUI) as opposed to a prompt or menu driven GUI typically used with keypad-display combinations. Other forms of user input devices may also be used, such as a microphone (not illustrated) for voice commands or a joystick / thumb wheel (not illustrated) for ease of navigation about menus.

[0085] As seen in Fig. 3A, the electronic device 301 also comprises a portable memory interface 306, which is coupled to the processor 305 via a connection 319. The portable memory interface 306 allows a complementary portable memory device 325 to be coupled to the electronic device 301 to act as a source or destination of data or to supplement the internal storage module 309. Examples of such interfaces permit coupling with portable memory devices such as Universal Serial Bus (USB) memory devices, Secure Digital (SD) cards, Personal Computer Memory Card International Association (PCM I A) cards, optical disks and magnetic disks.

[0086] The electronic device 301 also has a communications interface 308 to permit coupling of the device 301 to a computer or the communications network 120 via a connection 321. The connection 321 may be wired or wireless. For example, the connection 321 may be radio frequency or optical. An example of a wired connection includes Ethernet. Further, an example of wireless connection includes Bluetooth™ type local interconnection, Wi-Fi (including protocols based on the standards of the IEEE 802.11 family), Infrared Data Association (IrDa) and the like. The device 301 can communicate with the sound sensor 110 via the network 120, for example.

[0087] Typically, the electronic device 301 is configured to perform some special function. The embedded controller 302, possibly in conjunction with further special function components 310, is provided to perform that special function. For example, where the device 301 is a mobile telephone handset. In this instance, the components 310 may represent those components required for communications in a cellular telephone environment. Where the device 301 is a portable device, the special function components 310 may represent a number of encoders and decoders of a type including Joint Photographic Experts Group (JPEG), (Moving Picture Experts Group) MPEG, MPEG-1 Audio Layer 3 (MP3), and the like.

[0088] The methods described hereinafter may be implemented using the embedded controller 302, where the processes of Figs. 11 to 16 and 20 to 25 may be implemented as one or more software application programs 333 executable within the embedded controller 302.The electronic device 301 of Fig. 3A implements the described methods. In particular, with reference to Fig. 3B, the steps of the described methods are effected by instructions in the software 333 that are carried out within the controller 302. The software instructions may be formed as one or more code modules, each for performing one or more particular tasks. The software may also be divided into two separate parts, in which a first part and the corresponding code modules performs the described methods and a second part and the corresponding code modules manage a user interface between the first part and the user.

[0089] The software 333 of the embedded controller 302 is typically stored in the non-volatile ROM 360 of the internal storage module 309. The software 333 stored in the ROM 360 can be updated when required from a computer readable medium. The software 333 can be loaded into and executed by the processor 305. In some instances, the processor 305 may execute software instructions that are located in RAM 370. Software instructions may be loaded into the RAM 370 by the processor 305 initiating a copy of one or more code modules from ROM 360 into RAM 370. Alternatively, the software instructions of one or more code modules may be preinstalled in a non-volatile region of RAM 370 by a manufacturer. After one or more code modules have been located in RAM 370, the processor 305 may execute software instructions of the one or more code modules.

[0090] The application program 333 is typically pre-installed and stored in the ROM 360 by a manufacturer, prior to distribution of the electronic device 301. However, in some instances, the application programs 333 may be supplied to the user encoded on one or more portable storage devices (not shown) and read via the portable memory interface 306 of Fig. 3A prior to storage in the internal storage module 309 or in the portable memory 325. In another alternative, the software application program 333 may be read by the processor 305 from the network 320, or loaded into the controller 302 or the portable storage medium 325 from other computer readable media. Computer readable storage media refers to any non-transitory tangible storage medium that participates in providing instructions and / or data to the controller 302 for execution and / or processing. Examples of such storage media include floppy disks, magnetic tape, CD-ROM, a hard disk drive, a ROM or integrated circuit, USB memory, a magneto-optical disk, flash memory, or a computer readable card such as a PCMCIA card and the like, whether or not such devices are internal or external of the device 301. Examples of transitory or non-tangible computer readable transmission media that may also participate in the provision of software, application programs, instructions and / or data to the device 301 include radio or infra-red transmission channels as well as a network connection to another computer or networked device, and the Internet or Intranets including e-mail transmissions and information recorded onWebsites and the like. A computer readable medium having such software or computer program recorded on it is a computer program product.

[0091] The second part of the application programs 333 and the corresponding code modules mentioned above may be executed to implement one or more graphical user interfaces (GUIs) to be rendered or otherwise represented upon the display 314 of Fig. 3A. Through manipulation of the user input device 313 (e.g., the keypad), a user of the device 301 and the application programs 333 may manipulate the interface in a functionally adaptable manner to provide controlling commands and / or input to the applications associated with the GUI(s). Other forms of functionally adaptable user interfaces may also be implemented, such as an audio interface utilizing speech prompts output via loudspeakers (not illustrated) and user voice commands input via the microphone (not illustrated).

[0092] Fig. 3B illustrates in detail the embedded controller 302 having the processor 305 for executing the application programs 333 and the internal storage 309. The internal storage 309 comprises read only memory (ROM) 360 and random access memory (RAM) 370. The processor 305 is able to execute the application programs 333 stored in one or both of the connected memories 360 and 370. When the electronic device 301 is initially powered up, a system program resident in the ROM 360 is executed. The application program 333 permanently stored in the ROM 360 is sometimes referred to as “firmware”. Execution of the firmware by the processor 305 may fulfil various functions, including processor management, memory management, device management, storage management and user interface.

[0093] The processor 305 typically includes a number of functional modules including a control unit (CU) 351, an arithmetic logic unit (ALU) 352, a digital signal processor (DSP) 353 and a local or internal memory comprising a set of registers 354 which typically contain atomic data elements 356, 357, along with internal buffer or cache memory 355. One or more internal buses 359 interconnect these functional modules. The processor 305 typically also has one or more interfaces 358 for communicating with external devices via system bus 381 , using a connection 361.

[0094] The application program 333 includes a sequence of instructions 362 though 363 that may include conditional branch and loop instructions. The program 333 may also include data, which is used in execution of the program 333. This data may be stored as part of the instruction or in a separate location 364 within the ROM 360 or RAM 370.

[0095] In general, the processor 305 is given a set of instructions, which are executed therein. This set of instructions may be organised into blocks, which perform specific tasks or handle specific events that occur in the electronic device 301. Typically, the application program 333 waits for events and subsequently executes the block of code associated with that event. Events may be triggered in response to input from a user, via the user input devices 313 of Fig. 3A, as detected by the processor 305. Events may also be triggered in response to other sensors and interfaces in the electronic device 301.

[0096] The execution of a set of the instructions may require numeric variables to be read and modified. Such numeric variables are stored in the RAM 370. The disclosed method uses input variables 371 that are stored in known locations 372, 373 in the memory 370. The input variables 371 are processed to produce output variables 377 that are stored in known locations 378, 379 in the memory 370. Intermediate variables 374 may be stored in additional memory locations in locations 375, 376 of the memory 370. Alternatively, some intermediate variables may only exist in the registers 354 of the processor 305.

[0097] The execution of a sequence of instructions is achieved in the processor 305 by repeated application of a fetch-execute cycle. The control unit 351 of the processor 305 maintains a register called the program counter, which contains the address in ROM 360 or RAM 370 of the next instruction to be executed. At the start of the fetch execute cycle, the contents of the memory address indexed by the program counter is loaded into the control unit 351. The instruction thus loaded controls the subsequent operation of the processor 305, causing for example, data to be loaded from ROM memory 360 into processor registers 354, the contents of a register to be arithmetically combined with the contents of another register, the contents of a register to be written to the location stored in another register and so on. At the end of the fetch execute cycle the program counter is updated to point to the next instruction in the system program code. Depending on the instruction just executed this may involve incrementing the address contained in the program counter or loading the program counter with a new address in order to achieve a branch operation.

[0098] Each step or sub-process performed by the device 130 in the processes of the methods of Figs. 12 to 17 and 20 to 25 described below can be associated with one or more segments of the application program 333, and is performed by repeated execution of a fetch-execute cycle in the processor 305 or similar programmatic operation of other independent processor blocks in the electronic device 301.

[0099] Fig. 12 shows a method of 1200 of measuring relative changes in blood pressure of a human subject. The method 1200 can be implemented in part using the sound sensor 110 and in part on the processor device 130. Steps implemented on the processor device 130 can be implemented by executing the application 333, stored on the memory 309, on the processor 305.

[0100] The method 1200 starts at a training step 1210. The training step 1210 operates to take reference measurements and train a model to detect trends in blood pressure using an ECG signal. A method 1300 of training a model, as implemented at step 1210, is described with reference to Fig. 13.

[0101] The method 1200 continues from step 1210 to a capture reference sounds step 1220. At step 1220, reference sounds are captured for the subject’s heart, for example capturing ECG sounds using the sensor 110. The sensor 110 is held against skin of the subject in a location suitable for capturing sounds indicating the subject’s heartbeat. For example, the sensor can be placed on the subject’s chest close to the heart.

[0102] The method 1200 continues from step 1220 to a communicate sounds step 1230. At step 1230 the sounds captured at step 1210 are communicated to the processor device 130.

[0103] The method 1200 continues from step 1230 to reference setting step 1240. The step 1240 operates to set an initial (or reference) time from which relative changes in blood pressure are tracked. For example, the reference time can be a time the measurements start at or starting a timer.

[0104] The method 1200 continues under execution of the processor 305 from step 1240 to a relative measurement step 1250. The step 1250 operates to determine relative rather than quantitative changes in blood pressure, for example whether blood pressure is (i) increasing, (ii) decreasing or (iii) generally holding steady, for example within a designated range. A suitable range may be determined via experimentation or based on attributes of the subject, such as age, medical condition and the like. The relative changes can be measured with reference to the reference time, or periodically after the reference time. The step 1250 uses the model trained at step 1210 to determined relative blood pressure changes. Operation of the step 1250 is described hereafter in relation to Fig. 14.

[0105] The method 1200 continues from step 1250 to a check step 1260. The step 1260 checks if a change has occurred that requires a new reference measurement to be taken. For example,placement of the sensor 110 on the subject’s body may have changed, a certain time interval may have passed, a certain event may have occurred (such as coming out of surgery) or the like. If the change relates to placement, the change may be indicated by stopping recording of sounds related to operation of the heart. Alternatively, the change may be indicated by receiving an input of the user. If a suitable change has occurred (“Y” at step 1260), the method 1200 returns to step 1220 and takes a new reference measurement from which relevant changes in blood pressure are to be measured. In setting a new reference, variations in sounds due to the movement of the sensor can be accounted for, such that new relative measurements are from the same placement as the set reference. The method 1200 continues from step 1220 to the step 1230 as described above.

[0106] If no change is detected (“N” at step 1260), the method 1200 remains at step 1250 measuring changes in relative blood pressure without resetting the reference while readings are available.

[0107] The step 1210 can be implemented once per subject. In some instances, the step 1210 can be implemented some time before the remaining steps of the method 1200. For example, the step 1210 can be implemented at admission, whereas the remaining steps can be implemented later in a clinic.

[0108] Fig. 13 shows a method 1300 of training a model to measure relative changes in blood pressure for a subject, as executed at step 1210 of the method 1200. The method 1300 can be implemented by the software 333, as stored on the memory 309 and controlled by execution of the processor 305.

[0109] The method 1300 starts at a training sound capture step 1310 to capture or generate a sound recording of operation of the subject’s heart according to a training protocol. The training protocol requires examples of patterns of the subject’s heart to be captured during both exercise and recovery periods for a period of time sufficiently long to capture trends in S1 and S2 waveform shape.

[0110] In one implementation, the training protocol relates to assessment of handgrip Maximal Voluntary Contraction (MVC). The participant or subject is in a seated position, holding the dynamometer in their dominant arm maintaining the elbow joint at 90 degrees. The subject squeezes the dynamometer as hard as possible for one second and then relaxes. The exercise is triplicated with 30 seconds of rest between repetitions. The mean of the three readings is the MVC.

[0111] The subject is then asked to remove upper-body clothing for the placement of the ECG leads on their chest and abdomen. The participant is connected to a blood pressure monitor (for example 270), a sound sensor (the sound sensor 110), an ECG (such as 250) and finally a force-sensing resistor (FSR, e.g. 260) on the arm under the cuff (typically the non-dominant arm) which produces a signal marking the start and end of a blood pressure recording. Heart sounds and ECG are continuously recorded during the experiment. The participant is required to sit down upright on a chair with their back supported. The subject’s non-dominant arm will rest on a table, and the subject will rest for 10 minutes before the first measurement. After the 10-minute rest period ends, the subject remains at rest for 15 minutes and blood pressure is measured three times separated by 5 minutes between recordings. The ECG data and heart sounds are also recorded and saved simultaneously.

[0112] After the previous phase, the subject starts the exercise by squeezing the dynamometer at 30% MVC for a period of 120 seconds maintaining the elbow joint at 90°. Blood pressure readings are taken during the exercise with an automated blood pressure monitor (such as 270) at approximately 90 seconds from the onset of the exercise and at approximately 2 min from the onset of the exercise. The blood pressure monitor 270 is actioned approximately 35 seconds before each measurement since that is approximately the time the blood pressure monitor 270 takes to provide a single reading. After 120 seconds of isometric exercise, the subject is asked to stop the exercise entering in a recovery period for 5 min. During the recovery period, 3 additional blood pressure measures are taken at 1, 3, and 5 minutes post exercise. After the recovery period, the second 120-seconds exercise starts (similar protocol) and is followed by another recovery period (similar protocol).

[0113] The above protocol provides an example implementation only. Other exercise forms and timings may be used to implement the exercise and recovery periods. For example, the subject may be placed on a treadmill, similarly to in a stress test, or perform some other form of exercise such as lifting a weight. Moreover, forms of stimulation other than exercise or combined with exercise may be used, such as chemical stimulation using drugs, visual stimulation, acoustic stimulation or any other form than can physically or chemically interact with the subject. The type of exercise or stimulation method may be varied based on considerations such equipment available, location, physical abilities of the subject, and medical conditions of the subject. The MVC example described above can be particularly useful if the subject has limited mobility or energy.

[0114] The period of time over which the training measurements are used can also be varied. The period of time for each or exercise and recovery periods should be sufficiently long to allow heart rate increase and decrease in an observable manner and can vary based on the exercise method or stimulation method used and the condition of the subject. For example, a subject with relatively high blood pressure or a known heart condition, the exercise period may be shortened if the subject’s heart rate could be expected to increase overly quickly. Similarly, the number of blood pressure readings and the timing at which blood pressure readings are taken during each of the exercise and recovery periods can also be varied. For example, for a patient where a shorter recovery period is used may have 1 or 2 blood pressure readings taken rather than 3 or exercise period readings can be taken at times other than 90 second and 2 minutes. Timing of actioning the blood pressure monitor before each measurement can be varied based on factors such as the time taken to make a reading with a particular blood pressure monitor device, exercise interval total time, measurement time being used for the protocol and the like.

[0115] At step 1310 the system 100 is used to capture phonocardiogram readings during the exercise and recovery periods. The sensor 110 is held against skin of the subject in a location suitable for capturing sounds indicating the subject’s heartbeat. For example, the sensor can be placed on the subject’s chest close to the heart. The sound sensor 110 starts records sounds at step 1310. The step 1310 can output a training set comprising ECG readings 1305 and associated blood pressure readings 1307 as the ground truth data.

[0116] The method 1300 continues from step 1310 to a communicating step 1315. At step 1315 the sounds captured at step 1310 (for example sounds 1305) are communicated to the processor device 130. In an example arrangement, the sounds are a phonocardiogram communicated directly from the stethoscope 110 to the processor device 130 via the network 120. In other arrangements, the sounds may be transmitted to a further storage device, for example a cloud server. Each portion of the recorded sounds 1305 is tagged with being for one of exercise and recovery periods. In some arrangements, the simultaneously recorded blood pressure readings (for example for blood pressure monitor 270) are associated with corresponding times of the recorded sounds.

[0117] The method 1300 continues from step 1315 to a waveform extraction step 1320. The step 1320 executes to obtain (extract or identify) S1 and / or S2 waveforms from the reference phonocardiograms 1205 captured during the training protocol at step 1310. The S1 and S2 waveforms can be obtained using a model. Examples of suitable models for obtaining S1 and S2 waveforms are a Deep Neural Network (DNN) or a Hidden Markov Model (HMM) or any other type of model that can be trained to extract complex patterns. In some implementations,one of S1 and 2 waveforms can be extracted. In other implementations, as in the example described, both S1 and S2 waveforms are extracted (identified).

[0118] The method 1300 continues from step 1320 to a synchronisation step 1330. The synchronisation step 1330 operates to synchronise the extracted S1 and S2 waveforms with the blood pressure readings taken at step 1310 (such as 1307) and the associated tagging or ground truth data of whether associated with exercise or recovery or increasing / decreasing blood pressure. The synchronisation allows features of each of the extracted S1 and S2 sounds to be correlated with the blood pressure measurements taken during the training protocol. Synchronisation can be achieved using multiple methods, for example based on timestamps, recorded start time, use of a synchronising system such as the BioPac system 235 or the like. The synchronised waveforms are used to extract individual feature data from the S1 and / or S2 waveforms obtained from execution of the step 1320, as described in the following steps.

[0119] The method 1300 continues from step 1330 to a normalisation step 1340. The step 1340 executes to normalise the synchronised S1 and / or S2 waveforms using techniques such as min- max ratio, peak normalisation, square of min-max ratio and the like. The normalisation step typically operates to remove outliers.

[0120] The method 1300 continues from step 1340 to a noise removal step 1350. A step 1350 the application 333 performs signal averaging of the generated and synchronised S1 and S2 signals over a predetermined interval or window length to reduce background noise. Any suitable signal averaging technique that can increase strength of each S1 or S2 signal relative to noise can be used. In the arrangements described the interval is 12 seconds. However different intervals can be used dependent on factors such as noise present, patient pulse rate, patient breathing cycle and the like, provided at least one cardiac cycle is included in the interval. For example, breathing data may be used to estimate the averaging interval and adapt the interval over time to further individualise use of S1 and / or S2 sounds. For example, a longer breathing cycle may result in a longer averaging interval and vice versa. The steps 1340 and 1350 in combination can be considered to generate averaged waveforms.

[0121] The method 1300 continues from step 1350 to a feature extraction step 1360. The feature extraction step 1360 executes to extract individual feature data of the waveforms based on shape of the S1 and / or S2 waveforms obtained from execution of the step 1350. The feature data can include peak and notch patterns such as P2-N1 ratios, peak to peak ratios and the like. The feature extraction step can select features such as P2 / N1 ratio if the selected profile relates to the profile 700. In some implementations, the step 1350 can operate to select anappropriate profile or agreement by performing best fit with a selected S1 or S2 sound associated with a single heartbeat. Furthermore, different thresholds can be used to identify peaks and notches as well as different agreements can be used to label them. For example, notch one (N1) can be defined as the MIN(X) where X is the signal in S1 or S2, and peak one can be defined as the first peak after N1. Alternatively, peak 1 can be defined as the first peak prior N1. However, some subjects may not show a peak 1 when applying this agreement. Similar agreements to label and identify peaks and notches can be applied to identified the agreement that provides more accurate results. The selected agreement is used to identify the peaks and / or notches to be used for the individual subject.

[0122] In some implementations a Karhunen-Loeve Transform (KLT), PCA analysis or other types of pattern or feature analysis methods can be applied to signals to extract features of the subject’s cardiac S1 and S2 patterns at step 1360. The step 1320 can also operate to convert the S1 and S2 sounds to the frequency domain. The extracted features are synchronised with the reference blood pressure readings 1307 due to operation of the step 1320.

[0123] The method 1300 continues from step 1360 to a training step 1370. The step 1370 operates to select the extracted features and the associated synchronised blood pressure readings and associated training data. In some arrangements, the step 1370 executes to train a deep regression model used to determine blood pressure for the subject. The step 1370 can firstly operate to use feature data for the selected specification or profile most suitable for the individual’s S1 or S2 sounds.

[0124] The step 1370 can also operate to train a deep neural network such as a multi-layer one-dimensional convolutional neural network or a variational autoencoder using the blood pressure measurements (such as 1307) from step 1310 and the synchronised features extracted at step 1360. In some instances, the model may be partially trained. For example, initial layers of the deep neural network may be pre-trained in relation to particular features for S1 or S2 waveforms. For example, the method 1300 may already have been implemented using data from subjects other than the current subject. In one example, the method 1300 may be implemented using a reference database in which the readings 1305 and 1307 are already stored for a number of subjects. The number and characteristics of the subject can vary, for example may reflect a wide range of subject of different demographics and heart or medical conditions. In other arrangements the pre-training may be wherein measurements 1305 and 1307 are collected for subjects with similar (or suspected similar) heart or medical conditions. The features used for training relate to features extracted using the techniques described by operation of steps 1320 to 1360. If a specification or profile is selected, the training can beperformed based on the selection, for example by providing P2 / N1 data for training if the profile is found to correspond to the profile 700. If a frequency analysis or KLT or other analysis has been performed, the results can be used to train the model. Training can be performed using standard machine learning regression and training techniques such as deep learning, random forest, statistical classification, regression analysis and the like.

[0125] In one example, deep network training techniques can be implemented at step 1370 using one or more models with a parameterizable latent variable space, for example using a generative model such as a generative adversarial network (GAN) or a conditional variational autoencoder. Using a parameterizable latent variable space can allow different modes to be determined based on training data or based on particular patient characteristics, instrumentational characteristics, measurement modes or the like. For example, if placement of the sound sensor 110 affects acoustic coupling or impedance and accordingly affects measure S1 or S2 sounds, this may be accounted for during training using a parameterizable latent variable space. Other aspect that may potentially be considered by training using a parameterizable latent variable space may include ration of subcutaneous to visceral fat of the subject, affect of athletic activities on blood pressure calibration readings and the like.

[0126] The step 1370 can be implemented for each placement recorded at step 1310. Where the electronic stethoscope is placed on the subject’s body can affect the phonocardiogram received. For example, placement on different locations at different distances from the heart, such as front left chest versus right-hand back under shoulder-blade can cause some variations in extracted signal features. The step 1370 executes to train the model based on the placement. For example, an input can be provided that measurement was taken at the front left chest of the like. The step 1370 further executes to update the regression model using the features extracted at step 1360.

[0127] The method 1300 outputs trained model 1380 corresponding to the particular subject. Fig. 14 shows a method of 1400 of determining relative changes in blood pressure as implemented at step 1250 of the method 1200. The method 1400 can be implemented on the processor device 130 can be implemented by executing the application 333, stored on the memory 309, on the processor 305.

[0128] The method 1400 starts at a receiving step 1410. The step 1410 receives sounds captured by the sound sensor 110 at 1220, for example by transmission from the sound sensor 110. The length of the sounds can be predetermined, for example relating to several heartbeats,such as 30 seconds or the like. The sounds can be received at discrete intervals or continuously.

[0129] The method 1400 continues under execution of the processor 305 from step 1410 to an extraction step 1420. At step 1420 the software 333 executes to extract at least one waveform signal, each waveform signal representing one of S1 and S2 sounds from the received phonocardiogram signal. The S1 and S2 waveforms can be obtained using a model. Examples of suitable models are a Deep Neural Network (DNN), a Hidden Markov Model (HMM), Convolutional Neural Network, or any other type of model that can be trained to extract complex patterns. In the example discussed below, both S1 and S2 measurements are used to determine relative blood pressure changes. However, in other implementations, one of S1 and S2 measurements can be used.

[0130] The method 1400 continues from step 1420 to a normalising step 1430. The step 1430 operates similarly to the step 1340 and executes to normalise the S1 and S2 waveforms using techniques such as min-max ratio, peak normalisation, square of min-max ratio and the like. The step 1430 operates to remove or reduce outliers in the S1 and S2 signals determined at step 1420.

[0131] The method 1400 continues to noise reduction step 1440. At step 1440 the application 333 performs signal averaging of the generated the S1 and S2 signals over a predetermined interval or window length to reduce background noise, typically in a similar manner to the step 1350. In the arrangements described the interval is 12 seconds. However different intervals can be used dependent on factors such as noise present, patient pulse rate, patient breathing cycle and the like. The averaging window determines the accuracy and resolution of the blood pressure measurements extracted from the heart sounds. The averaging should be long enough to comprise several heart beats but not so long as to comprise significant blood pressure changes. Depending on variable such as noise and patient pulse rates, the interval should typically be under two minutes. The steps 1430 and 1440 in combination can be considered to generate averaged waveforms.

[0132] The method 1400 continues from step 1440 to a feature extraction step 1450. The step 1450 operates to extract features based on shape of the averaged S1 and / or S2 signals for use in determining blood pressure in a similar manner to the step 1360. In terms of measuring relative blood pressure, the step 1450 extracts at least the maximum amplitudes corresponding to each heartbeat of the S1 and S2 signals. The step 1450 can also operate to determine a profile of each of S1 and S2 for a single heartbeat in terms of peaks and notches. The profilerelates to presence of peaks and notches and their order. Once the profile is determined and agreement of specification that correlates most closely in terms of presence of peaks and notched can be selected. For example, if the minimum notch is followed by the maximum peak as in Fig. 7, the N1 / P2 features may be extracted. Alternatively, if the profile is more similar to Fig. 9A, the selected specification may indicate that the peaks P1 and P2 are to be extracted. The step 1450 can also implement a KLT analysis and extract first order coefficients as all or some of the extracted features.

[0133] The method 1400 continues from step 1450 to a relative blood pressure determining step 1460. At step 1460 the features extracted at step 1450 for each of S1 and S2 are provided to the model 1380 trained at step 1220 for determining relative changes in blood pressure. The model used is related to the features extracted at step 1450. For example, if the features extracted at 1450 relate to amplitude, amplitude patterns are analysed to determine if amplitude is increasing or decreasing, indicative of increasing or decreasing blood pressure respectively. A threshold required to indicate a level of amplitude change sufficient to indicate a relative change in blood pressure can be determined via experimentation on a per-subject basis or estimated based on previous data from other subjects. The level of amplitude change can depend on factors relating to the subject, for example age, average pulse rate, relevant medical conditions and the like.

[0134] Additionally, and / or alternatively, at step 1460 the model can use the features extracted based on a specification or agreement selected based on the profile of the user’s S1 and / or S2 sounds. For example, the models can operate to analyse patterns in P2 / N1 ratios or P1-P2 ratios or intervals to determine if blood pressure has increased, decreased or generally remained constant. A threshold required to indicate a level of feature change sufficient to indicate a relative change in blood pressure can be determined via experimentation. The level of change required can depend on factors relating to the subject, for example age, average pulse rate, relevant medical conditions and the like.

[0135] Additionally, and / or alternatively, at step 1460 the model can analyse patterns of KLT coefficients to determine if S1 and S2 sounds indicate that blood pressure is increasing, decreasing or remaining generally constant. A threshold required to indicate a level of coefficient value change sufficient to indicate a relative change in blood pressure can be determined via experimentation. The level of change required can depend on factors relating to the subject, for example age, average pulse rate, relevant medical conditions and the like. The method 1400 can be repeated as long as signals are received at step 1410.

[0136] Fig. 15 shows a method of 1500 of measuring blood pressure of a human subject. The method 1500 can be implemented in part using the electronic stethoscope or sound sensor 110 and in part on the processor device 130. Steps implemented on the processor device 130 can be implemented by executing the application 333, stored on the memory 309, on the processor 305. The example of Fig. 15 uses both S1 and S2 waveforms to determine blood pressure behaviour. In other implementations, one of S1 and S2 measurements can be used.

[0137] The method 1500 receives an obtain pre-trained model step 1505. The step can either execute the method 1300 for training a model for the subject or obtain a pre-existing trained model for the subject. For a pre-existing model, the medical facility may have the model 1380, trained for the subject by execution of the method 1300 from a previous visit, stored in a memory such as the memory 309 or stored on a cloud server.

[0138] The method 1500 starts at a calibration reading step 1510. The step 1540 operates in a similar manner to the step 1310 to collect data for further tuning the model 1380. In some instances, the step 1510 may follow the same protocol as used at step 1310. At 1510 the system 100 is used to take additional phonocardiogram (PCG) readings 1515 and blood pressure readings 1517. The reference phonocardiogram reading should be at least a heartbeat long and preferably several heartbeats long. A sensor of the electronic stethoscope 110 is held against skin of the subject in a location suitable for capturing a phonocardiogram of the subject’s heartbeat. For example, the sensor 110 can be placed on the subject’s chest close to the heart. The electronic stethoscope 110 records the phonocardiogram signal 1515 at step 1510. Blood pressure readings 1517 for the subject corresponding to a time the phonocardiogram is captured are also collected at step 1510. The blood pressure readings 1517 can be collected in traditional manner such as using a sphygmomanometer or an oximeter. The blood pressure readings 1517 and the phonocardiogram 1205 are synchronised, for example using a BioPac system as shown at 235 of Fig. 2. The reference readings taken at step 1510 preferably use the same implementation and protocol as used at step 1310.However, other protocols or means of measuring blood pressure may be used, for example based on equipment available in other areas of the hospital or for intervals based on characteristics of the subject or the subject’s current circumstances (for example receiving medication, current heart rate and the like). For example, in some subjects PCG readings may be less clear or more prone to noise. In subjects where intervals are more prone to noise, longer intervals or a higher number of measurements may be appropriate for calibration.

[0139] In implementations of the step 1510, several instances of the PCG 1515 may be measured. Each instance may be taken at a different area of the subject’s body, for exampleupper or lower from chest or upper or lower back chest. The area of the subject’s body associated with each PCG reading can be stored. Additionally, the different PCG readings may relate to use of different sensors, for example different stethoscopes, a stethoscope and / or a microphone, or an external and / or an internal sensor and the like. Placement of the sound sensor 110 may be optimised on an individual basis, for example based on physiological aspects such as whether the back or front of chest provides a stronger PCG signal. The corresponding blood pressures 1517 are stored for each measured PCG 1515.

[0140] The method 1500 continues from step 1510 to a communication step 1520. At step 1520 the calibration readings (the phonocardiogram 1515 and corresponding blood pressure measurements 1517) captured at step 1510 are communicated to the processor device 130. If multiple placements or sensors were used, corresponding placement information is also transmitted to the device 130. In an example arrangement the phonocardiogram is communicated directly from the stethoscope 110 to the processor device 130 via the network 120. In other arrangements the phonocardiogram 1515 may be transmitted to a further storage device, for example a cloud server. If the sound sensor 110 and the processor device 130 form an integrated device, the step 1520 can relate to saving the phonocardiogram 1515 in a memory of the integrated device, such as the memory 309.

[0141] The method 1500 continues from step 1520 to a calibration step 1530. The arrangements described calibrate or tune the model 1380 (obtained at step 1505) to determine blood pressure. Parameters of the model 1380 can be tuned for each individual subject based on reference measurements made for that subject. Step 1530 operates to use the reference phonocardiogram 1515 to extract features for use in determining blood pressure, providing a first level of tuning or calibration. In some implementations, the step 1330 can operate to tune or calibrate a deep network regression model such as a one-dimensional multi-layer convolutional neural network (CNN) or a variational autoencoder for determining blood pressure, for example based on shape or amplitude features of S1 and / or S2 such as notches or directly based on the waveform itself. At step 1530 the application 333 executes to use the phonocardiogram 1515 and the blood pressure readings 1517 captured at step 1510 to account for individual S1 / S2 waveform shapes of the user using the blood pressure measurements collected at step 1510. The step 1530 outputs a tuned or calibrated model 1580. A method 1600 of calibrating the phonocardiogram for a user, as used at step 1530, is described in further detail in relation to Fig. 16.

[0142] The method 1510 continues under execution of the processor 305 from step 1530 to a measurement step 1540. The step 1540 operates to generate one or more sound recordings ofoperation of the subject’s heart by operation of the sensor 110, for example capturing one or more PCG waveforms 1545 for the subject. The method 1500 continues from step 1540 to a blood pressure determining step 1550.

[0143] The step 1505 obtains or generates the model 1380 using historical data. The step 1505 may be implemented one time, for example at an initial visit or admission of the patient to a medical facility. The steps 1510 to 1530 effectively gather historical data that is used to tune or calibrate the model 1380 features to be used in determining blood pressure measurements for each individual subject. The steps 1510 to 1530 can be implemented occasionally and are not necessarily required every time a new blood pressure measurement is required. For example, the steps 1510 to 1530 can be implemented when a subject is admitted into a clinic, hospital or other healthcare facility, prior to undergoing a surgical procedure, or on an occasional or periodic basis. The calibration step 1530 can be implemented based on general placement area of the electronic stethoscope on the human body, for example the upper left-hand region of the chest or the like.

[0144] In the example of Fig. 15, calibration at 1530 is performed before measurement at step 1540. In other implementations, measurements may be taken first and calibration measurements, those measured at step 1510, may be taken at the end of a measurement interval. In such an implementation, calibration 1530 can be performed after or overlapping with step 1540 and prior to determining any measurements at step 1550.

[0145] The step 1550 executes to receive the one or more measurement phonocardiograms 1545 for a subject. Each phonocardiogram 1545 is usually recorded at a different time to the time the reference phonocardiogram at step 1510 is captured. A corresponding blood pressure signal is not required to be captured at step 1550. Rather, the calibrated model 1580 determined at step 1530 for the subject is used in association with the measurement phonocardiogram signal 1545 to determine the subject’s blood pressure at the time of measurement. The step 1550 can be repeated for each PCG 1545, that is more than once after execution of step 1530, and at any time after the step 1530 is completed.

[0146] Fig. 16 shows a method 1600 of calibrating or tuning a model measure blood pressure for a subject, as executed at step 1530 of the method 1500. The method 1600 can be implemented by the software 333, as stored on the memory 309 and controlled by execution of the processor 305.

[0147] The method 1300 starts at a waveform extraction step 1610. The step 1610 executes to obtain (identify or extract) S1 and / or S2 waveforms from the reference phonocardiograms 1515 captured during the step 1510. The S1 and S2 waveforms can be obtained using a model. Examples of suitable models for obtaining S1 and S2 waveforms are a Deep Neural Network (DNN) or a Hidden Markov Model (HMM) or any other type of model that can be trained to extract complex patterns. In some implementations, one of S1 and 2 waveforms can be extracted. In other implementations, as in the example described, both S1 and S2 waveforms are extracted.

[0148] The method 1500 continues from step 1610 to a synchronisation step 1620. The synchronisation step 1620 operates to synchronise the extracted S1 and S2 waveforms with the blood pressure readings taken at step 1510 (such as 1517) and the associated tagging or ground truth data of whether associated with exercise or recovery or increasing / decreasing blood pressure. The synchronisation allows features of each of the extracted S1 and S2 sounds to be correlated with the blood pressure measurements taken during the training protocol. Synchronisation can be achieved using multiple methods, for example based on timestamps, recorded start time, use of a synchronising system such as the BioPac system 235 or the like. The synchronised waveforms are used to extract individual feature data from the S1 and S2 waveforms obtained from execution of the step 1610, a described in the following steps.

[0149] The method 1600 continues from step 1620 to a normalisation step 1630. The step 1630 executes to normalise the synchronised S1 and / or S2 waveforms using techniques such as min- max ratio, peak normalisation, square of min-max ratio and the like. The normalisation step typically operates to remove outliers.

[0150] The method 1600 continues from step 1630 to a noise removal step 1640. At step 1640 the application 333 performs signal averaging of the generated and synchronised S1 and S2 signals over a predetermined interval or window length to reduce background noise. Any suitable signal averaging technique that can increase strength of each S1 or S2 signal relative to noise can be used. In the arrangements described the interval is 12 seconds. However different intervals can be used dependent on factors such as noise present, patient pulse rate, patient breathing cycle and the like, provided at least one cardiac cycle is included in the interval. For example, breathing data may be used to estimate the averaging interval and adapt the interval over time to further individualise use of S1 and / or S2 sounds. For example, a longer breathing cycle may result in a longer averaging interval and vice versa. The steps 1630 and 1640 in combination can be considered to generate averaged waveforms.

[0151] The method 1600 continues from step 1640 to a feature extraction step 1650. The feature extraction step 1650 executes to extract individual feature of the waveforms based on shape of the noise-adjusted S1 and / or S2 waveforms obtained from execution of the step 1640. The feature data can include peak and notch patterns such as P2-N1 ratios, peak to peak ratios and the like. The feature extraction step can select features such as P2 / N1 ratio if the selected profile relates to the profile 700. In some implementations, the step 1650 can operate to select an appropriate profile or agreement by performing best fit with a selected S1 or S2 sound associated with a single heartbeat. Furthermore, different thresholds can be used to identify peaks and notches as well as different agreements can be used to label them. For example, notch one (N1) can be defined as the MIN(X) where X is the signal in S1 or S2, and peak one can be defined as the first peak after N1. Alternatively, peak 1 can be defined as the first peak prior N1. However, some subjects may not show a peak 1 when applying this agreement.

[0152] A database of agreements to label and identify peaks and notches can be applied to identify the agreement that provides more accurate results. The selected agreement is used to identify the peaks and / or notches to be used for the individual subject.

[0153] In some implementations a Karhunen-Loeve Transform (KLT), PCA analysis or other types of pattern or feature analysis methods can be applied to signals to extract features of the subject’s cardiac S1 and S2 patterns at step 1650. The step 1650 can also operate to convert the S1 and S2 sounds to the frequency domain. The extracted features are synchronised with the reference blood pressure readings 1517 due to operation of the step 1620.

[0154] The method 1600 continues from step 1650 to a tuning step 1660. The step 1660 operates to select the extracted features and the associated synchronised blood pressure readings and associated training data and to tune or calibrate the model 1380 using the selected measurements. In some arrangements, the step 1660 executes to further train 1380 implemented as a deep regression model used to determine blood pressure. The step 1660 can firstly operate to use feature data for the selected specification or profile most suitable for the individual’s S1 or S2 sounds. The step 1660 can operate to further train a deep neural network such as a multi-layer one-dimensional convolutional neural network or a variational autoencoder using the blood pressure measurements (such as 1517) from step 1510 and the synchronised features extracted at step 1650. The model 1380 is partially trained, for example due to operation of the method 1300. Effectively, initial layers of the deep neural network are pretrained in relation to particular features for S1 or S2 waveforms for the subject by operation of the method 1300. The features used for tuning relate to features extracted using the techniques described by operation of steps 1610 to 1650. If a specification or profile is selected, the tuningcan be performed based on the selection, for example by providing P2 / N1 data for training if the profile is found to correspond to the profile 700. If a frequency analysis or KLT or other analysis has been performed, the results can be used to tune the model. Deep learning techniques using a parameterizable latent variable space may also be use, as described in relation to step 1370. The step 1660 can use tuning or calibration techniques for machine learning or model training such as robust regression techniques. Robust regression methods used can include iteratively reweighted least squares, random sample consensus (RANSAC), repeated median regression and the like to tune the model 1380 to the individual.

[0155] The step 1660 can be implemented for each placement recorded at step 1310. Where the electronic stethoscope is placed on the subject’s body can affect the phonocardiogram received. For example, placement on different locations at different distances from the heart, such as front left chest versus right-hand back under shoulder-blade can cause some variations in extracted signal features. The step 1660 executes to tune the model based on the placement. For example, an input can be provided that measurement was taken at the front left chest of the like. The step 1660 further executes to update the model 1380 using the features extracted at step 1650 to generate the tuned model 1580.

[0156] The method 1600 outputs the trained model 1580 corresponding to the particular subject.

[0157] Fig. 17 shows a method 1700 of measuring blood pressure based on a phonocardiogram as executed at step 1550 of the method 1500. The method 1700 can be implemented by the software 333, as stored on the memory 309 and controlled by execution of the processor 305.

[0158] The method 1700 starts at a receiving step 1710. The step 1710 receives an instance of the measurement phonocardiogram signal 1545 as an input, for example by transmission from the electronic stethoscope 110.

[0159] The method 1700 continues under execution of the processor 305 from step 1710 to a signal extraction step 1720. At step 1720 the software 333 executes to extract S1 and S2 waveform signals from the received phonocardiogram signal 1545. The S1 and S2 waveforms can be obtained using a model. Examples of suitable models are a Deep Neural Network (DNN), a Hidden Markov Model (HMM), Convolutional Neural Network, or any other type of model that can be trained to extract complex patterns. In the example discussed below, both S1 and S2 measurements are used to determine blood pressure. However, in otherimplementations, one of S1 and S2 measurements can be used, depending on a model or data available.

[0160] The method 1700 continues from step 1720 to a normalising step 1730. The step 1730 executes to normalise the S1 and S2 waveforms using techniques such as min-max ratio, peak normalisation, square of min-max ratio and the like. The normalisation step operates to typically remove or reduce outliers.

[0161] The method 1700 continues from step 1730 to noise reduction step 1740. At step 1740 the application 333 performs signal averaging of the generated the S1 and S2 signals over a predetermined interval or window length to reduce background noise, typically in a similar manner to steps 1640, 1440 and 1350. In the arrangements described the interval is 12 seconds. However different intervals can be used dependent on factors such as noise present, patient pulse rate, breathing cycle of the patient and the like, provided at least one cardiac cycle is included in the interval. The averaging should be long enough to comprise several heart beats but not so long as to comprise significant blood pressure changes. Depending on variable such as noise and patient pulse rates, the interval should typically be under two minutes. The steps 1730 and 1740 in combination can be considered to generate averaged waveforms.

[0162] The method 1700 continues from step 1740 to a feature extraction step 1750. The feature extraction step 1750 executes to extract individual feature data of the waveforms based on shape of the S1 and S2 waveforms obtained from execution of the step 1720. Similarly to the steps 1650, 1450 and 1360, the step 1750 extracts features of the each of the averaged waveforms based on amplitude. The feature data can include shape features such as peak and notch patterns (e.g., P2-N1 ratios), peak to peak ratios, amplitude and the like. The feature extraction step can extract features based on the features extracted at step 1650. The features can relate to the agreement determined at step 1650, for example to select P2 / N1 ratio, if the selected profile relates to the profile 700. Alternatively, in some implementations, the step 1750 can operate to select an appropriate profile by performing best fit with a selected S1 or S2 sound associated with a single heartbeat. Furthermore, different thresholds can be used to identify peaks and notches as well as different agreements can be used to label them. For example, notch one (N1) can be defined as the MIN(X) where X is the signal in S1 or S2, and peak one can be defined as the first peak after N1. Alternatively, peak 1 can be defined as the first peak prior N1. However, some subjects may not show a peak 1 when applying this agreement. Similar agreements to label and identify peaks and notches can be applied to identify the agreement that provides more accurate results.

[0163] In some implementations a Karhunen-Loeve Transform (KLT), PCA analysis or other types of pattern or feature analysis methods can be applied to signals to extract features (KLT coefficients) of the subject’s cardiac S1 and S2 patterns. The step 1750 can also operate to convert the S1 and S2 sounds to the frequency domain, the frequency response providing some of the extracted features.

[0164] The method 1700 continues from step 1750 to blood pressure determining step 1760. At step 1760 the features extracted at step 1750 for each of S1 and S2 are provided to the tuned model 1580 generated at step 1660 for the same subject. The model 1580 outputs a blood pressure reading based on the extracted features. For example, the step 1760 can operate to receive one or more extracted features such as a P1-N2 interval, amplitude, KLT coefficients as appropriate to the tuned model along with an initial blood pressure reading. The tuned model 1580 uses the extracted features to generate a new blood pressure reading.

[0165] The steps 1720 and 1420 each operate to extract at least one of a first and a second waveform from a sound recording of the subject’s heart. Each extracted waveform represents one of S1 and S2 sounds generated by operation of the patient’s heart. The steps 1750, 1650, 1450 and 1360 operate to extract features of the each of the averaged waveforms based shape of the waveform(s) generate by the previous step. The extracted features may relate to extracting maximum amplitudes of the noise reduced waveform over the predetermined interval or window, performing an analysis function such as a KLT analysis on the waveforms and extracting first coefficient values of the KLT analysis, extracting to the frequency domain or identifying a specification or agreement. As described above, identifying a specification or agreement relates to matching a portion of one of the extracted waveforms corresponding to a single heartbeat to one of a plurality of different waveform shapes and selecting a pattern related to notches and peaks of the waveform as the extracted features. The steps 1750, 1650, 1450 and 1360 can accordingly extract at least one of maximum or minimum amplitude, time intervals between S1 and / or S2 sounds, frequency responses, waveform profile and the like.

[0166] The extracted features are provide to a model trained for the subject at each of steps 1250 and 1550 (for example models 1370 and 1580 respectively). The 1760 and 1460 each operate to determine, based on the extracted features, a measurement of blood pressure behaviour of the subject. The blood pressure behaviour may be measured as one of (i) determining a measurement value (step 1460), or (ii) determining a relative measurement (step 1760).

[0167] The shape of a heart sound waveform (any of S1, S2, S3 and S4) waveform is typically unique to an individual. Fig. 18A shows a graph 1800 of amplitude (y-axis) over time (x-axis) for an S1 waveform for an example subject. Each waveform has a number of peaks and notches, each allocated a label. The label is typically a number but may also be another label such as an alphanumeric character. A trace 1810 in the graph 1800 shows an example waveform pattern for the S1 sound, having positive peaks P11 to P14 and negative notches N11 to N14. While the height and spacing of the peaks and notches can vary per individual, each individual’s S1 waveform will typically have a lowest notch and a highest peak. In the example of Fig. 18A, the second peak, P12, is the peak with highest amplitude, and the second notch, N12, has the lowest amplitude.

[0168] Fig. 18B shows a graph 1820 of amplitude (y-axis) over time (x-axis) for an S2 waveform for an example subject, being the same subject whose measurements resulted in the S1 waveform of Fig. 18A. A trace 1830 in the graph 1820 shows an example waveform pattern for the S2 sound, having positive peaks P21 to P24 and negative notches N21 to N24. While the height and spacing of the peaks and notches can vary per individual, each individual’s S2 waveform will typically have a lowest notch and a highest peak. In the example of Fig. 18B, the second peak, P22, is the peak with highest amplitude, and the second notch, N22, has the lowest amplitude (notch N21 having a second-lowest amplitude).

[0169] The number and amplitude (height) of peaks and notches will vary for each individual. As notch and peak patterns can vary for an individual, different definitions of peak and notch labelling can be used. For example, a threshold minimum amplitude may be required in identifying a peak or a notch in a heart sound. The threshold may be based both on height (the absolute height of a peak or a notch) and time (for example minimum time separation from a past peak / notch or from an identified starting time of the heart sound). The threshold may be determined though experimentation and may vary based on aspects associated with a group of patients. For example, elderly patients or patients with a particular health condition may have a lower threshold amplitude than other patients. In another example, at least one of a minimum time separation or threshold difference may be required between consecutive peaks and between consecutive notches, or a minimum time after a start of a detected heart sound (whether S1 , S2, S3 or S4). For example, P21 is selected as a peak rather than a first, lower peak. The minimum time separation and / or threshold difference may be determined though experimentation and may vary based on aspects associated with a group of patients. In some implementations, peak detection algorithms may also be used such as a findpeaks function in MatLab and find_peaks in SciPi.

[0170] The inventors have determined that the peaks and notches of each individual’s heart sounds can be used to determine at least one element that show a correlation with measured blood pressure and in turn may be used to determine changes in blood pressure.

[0171] Such elements are in the form of Xap Ys*, where:• X and Y can take any of the two values P or N, where P represents a peak and N represents a notch.• The subindexes in X and Y can take any positive integer value. The first subindex in each of these letters identifies a heart sound waveform. For instance, a value of 1 for a or 6 identifies S1, 2 identifies S2, 3 identifies S3 and 4 identifies S4.• The second subindex in X and Y identifies the particular peak or notch using a label allocated to the peak or notch. For example, a value for p or 4> of 1 identifies a first peak or notch, a value of 2 identifies a second peak or notch, a value of 3 identifies a third peak or notch and so forth.

[0172] An example of an element is Pi2_N23, which corresponds to the ratio between the second peak in S1 (e.g., P12 in Fig. 18A) and the third notch in S2 (e.g., N23 in Fig. 18B), based on a normalised signal (to be described). Another example of an element is Ni4_P23 which corresponds to the ratio between the fourth notch in S1 (e.g., P14 in Fig. 18A) and the third peak in S2 (e.g., P23 in Fig. 18B)

[0173] X and Y can be any of P (peak) and N (notch) and may both be peaks or both be notches. The first subindices can be from any of the heart sounds and may be from the same heart sound or different heart sounds (e.g., may both be from one of S1 , S2, S3 or S4) determined for the individual subject. The second subindices can be for any peak or notch in the relevant sound. Two different measurements must be used to form the ratio of the element. For example, X and Y cannot both indicate the same peak or notch of the same sound.

[0174] Figs. 19A to 19D show an example set of results obtained during experiments conducted by the inventors. Fig. 9A shows an example frequency histogram 1900 for a number of individual patients who took part in the experiments conducted. The histogram 1900 plots correlation coefficient occurrence (y-axis) for different element ratios (x-axis), where each correlation coefficient is above a threshold such as 0.4. For the individuals whose readings Fig. 9A relates to, changes in value of element Pi3_Pis (1902) show a highest correlation withsystolic blood pressure (SBP), occurring for almost 70% of the recordings. Elements Pi2_Nn and Ni2_Nu show a next highest correlation with systolic blood pressure (SBP), occurring for about 60% of the recordings.

[0175] Fig. 19B shows an example histogram 1910 using the same measurements for the same individual as Fig. 19A. This histogram 1910 plots different recordings for the participating individuals (y-axis) for different element ratios (x-axis). A grid entry in the histogram 1910 in dark shading (for example 1912) indicates that the correlation coefficient is above 0.4. A grid entry in the histogram 1910 in light shading (for example 1914) indicates that the correlation coefficient is below 0.4. This histogram 1910 indicates for all the recordings the systolic blood pressure (SBP) shows high correlation (0.4 or above) with at least one of the following three features: Pl 3_Pl 5 ; Pl 2_N l 1 ; N l 5_N22.

[0176] Figs. 19A and 19B relate to systolic blood pressure. Figs. 19C and 19D provide results for the same individuals using diastolic blood pressure (DBP). Fig. 19C shows an example frequency histogram 1920 for the same patients of Figs. 19A and 19B. The histogram 1920 plots correlation coefficient occurrence (y-axis) for different element ratios (x-axis), where each correlation coefficient is above 0.4. For the individuals of Fig. 9C, element Pi2_N22 1922 (ratio of peak 2 of S1 to notch 2 of S2) correlates with DBP for almost 70% of the recordings. Elements NH_N22 and Nu_N22 show similar correlations.

[0177] Fig. 19D shows an example histogram 1930 using the same measurements for the same individuals as Fig. 19C. This histogram 1930 plots different recordings (y-axis) for different element ratios (x-axis). A grid entry in the histogram 1930 in dark shading (for example 1932) indicates that the correlation coefficient is above 0.4. A grid entry in the histogram 1930 in light shading (for example 1934) indicates that the correlation coefficient is below 0.4. This histogram 1930 indicates for all the recordings the DBP shows high correlation (0.4 or above) with at least one of the following three elements or features: Pi 2_N22, N n_N22 and N u_N22.

[0178] Other individuals showed that elements could be determined (with each individual having unique elements or ratios) that correlated sufficiently with changes in blood pressure. As shown in Figs. 19A to 19D, each individual can be determined to have at one or more elements having sufficiently high correlation with changes in blood pressure, such that despite changes in the positioning of the stethoscope and volunteer, features with the structure of Xap Ys* that correlate with blood pressure can be determined. Based on Figs. 19A and 19B for systolic blood pressure and Figs. 19C and 19D for diastolic blood pressure, for each individual, at least one element can be found that has a suitably high level of correlation with changes in bloodpressure. Further, at least one common element can typically also be found in larger sets of data measured for several individuals, the at least one common element having a suitably high correlation with changes in blood pressure for most individuals in that data set.

[0179] Further description of methods using measurements of heart sounds to determine blood pressure, or changes in blood pressure, are described in relation to Figs. 20- 25, to be described, using the sound sensor 110 and the processor device 130.

[0180] Fig. 20 shows a method 2000 of training a model to measure relative changes in blood pressure for a subject. The method 2000 can be implemented by the software 333, as stored on the memory 309 and controlled by execution of the processor 305.

[0181] The method 2000 starts at a training sound capture step 2010 to capture or generate a sound recording of operation of the subject’s heart according to a training protocol. The training protocol requires examples of patterns of the subject’s heart to be captured during both exercise and recovery periods for a period of time sufficiently long to capture trends in heart sound (such as S1 , S2, S3 and S4) waveform shape. The training protocol may be similar to inventor experiments described in relation to Fig. 2 in some implementations.

[0182] In one implementation, the training protocol relates to assessment of handgrip Maximal Voluntary Contraction (MVC). The participant or subject is in a seated position, holding the dynamometer in their dominant arm maintaining the elbow joint at 90 degrees. The subject squeezes the dynamometer as hard as possible for one second and then relaxes. The exercise is triplicated with 30 seconds of rest between repetitions. The mean of the three readings is the MVC.

[0183] The subject is then asked to remove upper-body clothing for the placement of the ECG leads on their chest and abdomen. The participant is connected to a blood pressure monitor (for example 270), a sound sensor (the sound sensor 110), an ECG (such as 250) and finally a force-sensing resistor (FSR, e.g. 260) on the arm under the cuff which produces a signal marking the start and end of a blood pressure recording. Heart sounds and ECG are continuously recorded during the experiment. The participant is required to sit down upright on a chair with their back supported. The subject’s non-dominant arm will rest on a table, and the subject will rest for 10 minutes before the first measurement. After the 10-minute rest period ends, blood pressure is measured three times separated by 5 minutes between recordings. The ECG data and heart sounds are also recorded and saved simultaneously.

[0184] The subject starts the exercise by squeezing the dynamometer at 30% MVC for a period of 90 seconds maintaining the elbow joint at 90 degrees. The blood pressure readings are taken during the last 20 seconds by starting the automated BP monitor 270 around 55 seconds into the exercise. This is because the HEM907 takes around 35 seconds to provide a single reading. After 90 seconds, the subject is asked to stop the exercise. During the recovery period, 3 additional BP measures will be taken with 1 , 3, 5 minutes post exercise. At ten minutes post recovery, the second 90-sec exercise starts (similar protocol).

[0185] In experiments conducted, each patient participated in providing sets of readings for training on at least two separate days. Each patient provided at least 2 sets of 120 s and 5 min time of readings in each of the exercise or recovery zones respectively. A total of 16 patients were used for capturing training sounds and readings. In measuring relative blood pressure changes, a minimum training set could be 1 set of exercise and recovery, over a total time of 5 to 7 minutes. However, shorter or longer times may be used for each individual depending on factors such as age, existing medical conditions, physical ability and the like.

[0186] The above protocol provides an example implementation only. Other exercise forms may be used to implement the exercise and recovery periods. For example, the subject may be placed on a treadmill, similarly to in a stress test, or perform some other form of exercise such as lifting a weight. The type of exercise may be varied based on considerations such equipment available, location and physical abilities of the subject. The MVC example described above can be particularly useful if the subject has limited mobility or energy.

[0187] The period of time over which the training measurements are used can also be varied. The period of time for each or exercise and recovery periods should be sufficiently long to allow heart rate in increase and decrease in an observable manner and can vary based on the exercise method used and the condition of the subject. For example, a subject with relatively high blood pressure or a known heart condition, the exercise period may be shortened if the subject’s heart rate could be expected to increase overly quickly.

[0188] At step 2010 the system 100 is used to capture phonocardiogram readings during the exercise and recovery periods. The sensor 110 (for example the stethoscope 210) is held against skin of the subject in a location suitable for capturing sounds indicating the subject’s heartbeat. For example, the sensor can be placed on the subject’s chest close to the heart. The sound sensor 110 starts records sounds at step 2010. The step 2010 can output a training set comprising ECG readings 2005 and associated blood pressure readings 2007 (for example 275 captured using a device such as the monitor 270), the training set typically including groundtruth data indicating blood pressure measurements taken using the cuff device (for example 2007 taken using the device 270) and whether blood pressure is increasing, constant or decreasing based on exercise or recovery periods. In some implementations, both the blood pressure measurement and whether increasing, decreasing or constant may both be used as ground truth data. In other implementations, only one of the blood pressure measurements or whether increasing, constant or decreasing may be used as the ground truth data.

[0189] The method 2000 continues from step 2010 to a communicating step 2015. At step 2015, the sounds captured at step 2010 (for example sounds 2005 corresponding to 215) are communicated to the processor device 130. In an example arrangement, the sounds are a phonocardiogram communicated directly from the stethoscope 110 to the processor device 130 via the network 120. In other arrangements, the sounds may be transmitted to a further storage device, for example a cloud server. Each portion of the recorded sounds 2005 is tagged with being for one of increasing, decreasing or constant and the simultaneously recorded blood pressure readings (for example for blood pressure monitor 270) are associated with corresponding times of the recorded sounds.

[0190] The method 2000 continues from step 2015 to a waveform extraction step 2020. The step 2020 executes to identify or extract a set a plurality heart sound waveform shapes (S1 and S2 waveforms and, in some implementations, S3 and S4 waveforms) from the reference phonocardiograms 2005 captured during the training protocol at step 2010. Each waveform shape corresponds to operation of the heart and has a plurality of peaks and notches. The heart sound waveforms can be obtained using a model. Examples of suitable models for identifying S1 and S2 waveforms are a Deep Neural Network (DNN) or a Hidden Markov Model (HMM) or any other type of model that can be trained to identify complex patterns. In some implementations, only S1 and S2 waveforms can be identified. In other implementations, as in the example described, all of S1, S2, S3 and S4 waveforms can be identified if present. Similar processes are used to identify S3 and S4 waveforms as S1 and S2 waveforms. Each waveform shape corresponding to operation of the heart and having a plurality of peaks and notches, each of the peaks and notches allocated a number or label.

[0191] The method 2000 continues from step 2020 to a synchronisation step 2030. The synchronisation step 2030 operates to synchronise the heart sound waveforms extracted at step 2020 with the blood pressure readings taken at step 2010 (such as 2007) and the associated tagging or ground truth data of whether associated with exercise or recovery or increasing / decreasing blood pressure. The synchronisation allows features of each of the extracted heart waveform sounds to be correlated with the blood pressure measurements takenduring the training protocol. Synchronisation can be achieved using multiple methods, for example based on timestamps, recorded start time, use of a synchronising system such as the BioPac system 235 or the like. The synchronised waveforms are used to extract individual feature data from the heart sound waveforms obtained from execution of the step 2020, as described in the following steps.

[0192] The method 2000 continues from step 2030 to a normalisation step 2040. The step 2040 executes to normalise the synchronised heart sound waveforms using techniques such as min- max ratio, peak normalisation, square of min-max ratio and the like. The normalisation step typically operates to remove outliers.

[0193] The method 2000 continues from step 2040 to a noise removal step 2050. A step 2050 the application 333 performs signal averaging of the generated and synchronised heart sound waveform over a predetermined interval or window length to reduce background noise. Any suitable signal averaging technique that can increase strength of each heart sound waveform’s signal relative to noise can be used. In the arrangements described, the interval is 12 seconds. However different intervals can be used dependent on factors such as noise present, patient pulse rate, patient breathing cycle and the like. For example, breathing data may be used to estimate the averaging interval and adapt the interval over time to further individualise use of heart sound waveforms. For example, a longer breathing cycle may result in a longer averaging interval and vice versa. The steps 2040 and 2050 can be considered to “clean” the heart sound waveforms.

[0194] The method 2000 continues from step 2050 to a feature determining step 2060. The feature determining step 2060 executes to extract or determine a feature or element of the heart waveform determined which shown behaviour that corresponds with fluctuations in blood pressure. The extracted feature (also referred to as an element) can also be considered to provide a ratio in form of Xap Ys* as described above. Fig. 21 shows an example method 2100, described hereafter, as implemented at the step 2060.

[0195] The method 2000 continues from step 2060 to a training step 2070. The step 2070 operates to select the extracted features and the associated synchronised blood pressure readings and associated training data. In some arrangements, the step 2070 executes to train a deep regression model used to determine blood pressure for the subject. The step 2070 can firstly operate to use feature data for the selected specification or profile most suitable for the individual’s heart sound waveforms. The step 2070 can also operate to train a deep neural network such as a multi-layer one-dimensional convolutional neural network or a variationalautoencoder using the blood pressure 2060 (such as 2007) from step 2010 and the synchronised features extracted at step 2060. In some instances, the model may be partially trained. For example, initial layers of the deep neural network may be pre-trained in relation to changes in peaks and notches of heart sound waveforms relative to blood pressure, for example based on training samples obtained using the step 2010 for multiple individuals. The training at step 2070 in such instances relates to training a subset of layers, being final (later) layers of a convolutional neural network for example. Training is based on the selected one or more elements, for example by providing Pi3_Pi5 data for training in the example of Fig. 19A. Training can be performed using standard machine learning regression and training techniques. For example, regression techniques can include linear, logistic, ridge, lasso, polynomial and Bayesian regression techniques. Example supervised learning techniques can include convolutional neural network or a variational autoencoders, or other similar techniques. Typically, some training is performed at step 2070 to tune or calibrate a model for an individual, for example training upper layers of a pre-existing model only. However, for relative measurements, performing additional training for an individual may not be required and previously trained model, generated based on results from a database of individuals, may be obtained at step 2070 from a database. The previously trained model would typically be developed over time using the previous iterations of steps 2010 to 2070 and / or experimentation. The previously trained model and may relate to training data from one or more individuals having one or more similar characteristics, such as age range, cardiac characteristics, medical conditions and the like.

[0196] The method 2000 outputs a trained model 2080 (whether trained specifically for an individual or obtained from a database).

[0197] Fig. 21 shows the method 2100 of determining an element of a subject’s heart sounds, providing an example implementation of the step 2060. The method 2100 can be implemented by the software 333, as stored on the memory 309 and controlled by execution of the processor 305.

[0198] The method 2100 starts at an identify waveform step 2110. The step 2110 operates to receive the noise removed waveforms generated at step 2050, to identify a start and end of a waveform using the waveforms over the interval window. The step 2110 can operate in a similar manner to the step 2020 and use models to identify waveforms such as a Deep Neural Network (DNN) or a Hidden Markov Model (HMM) or any other type of model that can be trained to identify complex patterns may be used to determine the start and end of each waveform. Thestep 2110 can be implemented for each heart sound, for example for each of S1, S2, and, in some implementations, S3 and / or S4.

[0199] The method 2100 continues under execution of the processor 333 from step 2110 to an identifying step 2120. The step 2120 executes to detect peaks and notches in the waveform identified in step 2110. The peaks and notches can be determined using known peak detection algorithms and may be based on factors such as changing slope, absolute magnitude and the like. Peak and notch detection techniques as described in relation to Figs. 18A and 18B may be used. A threshold minimum absolute magnitude may be used to in identifying peaks and notches in some implementations. A peak is a local maximum with a magnitude over a given threshold, and a notch is a local minimum over a given threshold. The threshold can also include a time threshold, as described in relation to Figs. 18A and 18B in relation to minimum time between peaks / notches or a start time of an identified heart sound. The threshold may be determined though experimentation by trial and error and initial human supervision and may vary based on aspects associated with a group of patients. For example, elderly patients or patients with a particular health condition may have a lower threshold amplitude than other patients. In another example, at least one of a minimum time separation or threshold difference may be required between consecutive peaks and between consecutive notches. For example, in Fig. 18B, P21 is selected as a peak rather than a first, lower peak shortly after 0.02 seconds. The minimum time separation and / or threshold difference may be determined though experimentation and may vary based on aspects associated with a group of patients.

[0200] The step 2120 executes to find peaks and notches for each of the heart sound waveforms for the subject. For example, the step 2120 can execute to find peaks and notches for an S1 waveform, peaks and notches for an S2 waveform, peaks and notches for an S3 waveform and peaks and notches for an S4 waveform if all 4 waveforms have been extracted. In the example of Figs. 18A and 18B, where only S1 and S2 waveforms are detected, the step 2120 executes to identify (i) the peaks P11, P12, P13 and P14 and the notches N11, N12, N13 and N14 for the S1 waveform 1810, and (ii) the peaks P21, P22, P23, and P24 and the notches N21, N22, N23 and N24 for the S2 waveform 1830.

[0201] The step 2120 is typically executed for all samples including exercise and recovery. While the arrangement described focus on exercise and recovery other available heart sound data with an associated blood pressure measurement may also be used. The intention of step 2120 is to identify peak and notches in the heart sounds over time, so that continuous blood pressure monitoring may be enabled.

[0202] The method 2100 continues from step 2120 to a ratio determining step 2130. The step 2130 executes to determine potential element ratios present in the heart sound waveforms using the peaks and notches identified in operation of the step 2120 over the recorded time. The step 2130 effectively operates, for each peak or notch, to determine a set of ratios with each of the other peaks and notches identified at step 2120. Each ratio is determined using the absolute magnitude of the selected peak or notch and the absolute magnitude of the other peak or notch. Duplicate ratios, for example inverse ratios, can be discarded. Similarly to the step 2120, the step 2130 is executed for all samples available.

[0203] Using the example of Figs. 18A and 18B and using the peak P11 as a selected peak, the ratios are determined as in Table 1:Table 1

[0204] The step 2130 operates to determine ratio combinations for each park or notch, generating data similar to that of Table 1 for each of P11, P12, P13 and P14 and the notches N11, N12, N13 and N14 for the S1 waveform 1810, and (ii) the peaks P21, P22, P23, and P24 and the notches N21 , N22, N23 and N24 in the example of Figs. 18A and 18B.

[0205] The method 2100 continues from step 2130 to a ratio determining step2140. The step 2140 executes to calculate values for each ratio determined in step 2130 across each of the set of samples collected in training over time, for example across exercise and recovery periods. The ratios are determined as the element of magnitudes for Xap Ys*, where each of X and Y can be a peak or notch in any of the derived sounds, the first subindex (a or 6) relates to the number of S1 to S4 and the second subindex (0 or (|)) is the peak / notch label or number, as described above.

[0206] The method 2100 continues under execution of the processor 333 from step 2140 to a score determining step 2150. The step 2150 compares fluctuation in the set of values determined for each ratio in step 2130 with fluctuation in the blood pressure measurements taken at step 2010 (i.e. the measurements 2007) over time. The step 2150 operates to determine a correlation score for each ratio based on changes in the ratio values compared to changes in synchronised measured blood pressure. The correlation may be determined using correlation or pattern matching algorithms such as linear or non-linear regression algorithms with consideration of noise and outliers based on experimentation.

[0207] The method 2100 continues under execution of the processor 333 from step 2150 to an element determining step 2160. The step 2160 executes to select at least one element (also referred to as a feature) 2170 based on the correlation scores determined at the step 2150. The step 2160 may select the score based on any of a number of criteria. Typically, the one or more elements are determined to be a highest score. The one or more elements may also be based on a highest score across a longest period of time and / or a threshold. In experiments conducted, the threshold used was a correlation equal to or higher than 0.4. In other implementations, a different threshold may be used. The threshold may be a predetermined value (for example based on a required level of accuracy or an expected level of noise), or determined based on experimentation, for example determined as different groups of subjects with different medical or cardiac histories or conditions are used for training.

[0208] The method 2100 ends after execution of the step 2160. In some implementations, the method 2100 may return a single determined element as 2170, for example P _Ni5 of Fig. 19A. In other implementations, the method 2100 may be configured to return more than one element as output 2170. As shown in Figs. 19A to 19D, more than one element may have a suitable correlation to changes in blood pressure for an individual or group of individuals. While use of a single element for the output 2170 can be sufficient, use of more than one element may provide benefits of redundancy or double checking or to account for effects of noise.

[0209] The methods 2000 and 2100 may be implemented for one or both of systolic and diastolic blood pressure.

[0210] Fig. 22 shows a method 2200 of determining relative blood pressure in a subject. The method 2200 can be implemented by the software 333, as stored on the memory 309 and controlled by execution of the processor 305. The method 2200 starts at a sound capturing step 2210. At the step 2210, the sound sensor 110 is held against the subject skin close to the subject’s heart and captures sounds made by operation of the subject’s heart. The step 2210 can include the sound sensor transmitting the sounds to the processor device 130, similarly to operation of step 2015 in implementations where the sound sensor 110 and the processor device 130 are separate devices. The captured sounds can be communicated at discrete intervals or continuously. The method 2200 may be implemented for one or both of systolic and diastolic blood pressure, corresponding to whether the methods 2000 and 2100 were implemented for one or both of systolic and diastolic blood pressure for the individual patient.

[0211] The method 2200 continues under execution of the processor 305 from step 2200 to an extraction step 2220. At step 2220 the software 333 executes to extract or identify a plurality of heart sound waveforms (also referred to as waveform shapes from the sounds captured at step 2210. The step 2220 operates in a similar manner to the step 2020. Each waveform corresponds to operation of the heart and has a plurality of peaks and notches. Each of the peaks and notches is allocated a number or label, as described in step 2120 and described in relation to Figs. 18A and 18B. The extracted (identified) waveform signals can include at least S1 and S2 waveforms. In some implementations, the extracted waveform signals can also include S3 and / or S4 waveforms. The heart sounds waveforms can be identified or extracted using a model at step 2220. Examples of suitable models are a Deep Neural Network (DNN), a Hidden Markov Model (HMM), Convolutional Neural Network, or any other type of model that can be trained to extract complex patterns. In the example discussed below, both S1 and S2 waveforms are used to determine relative blood pressure changes. However, in other implementations, one or more of S1 , S2, S3 and S4 waveforms can be used. The waveforms identified at step 2220 include at least the waveforms identified or extracted at training, that is the waveforms extracted at step 2020.

[0212] The method 2200 continues from step 2220 to a normalising step 2230. The step 2230 operates similarly to the step 2040 and executes to normalise the heart sound waveforms extracted at step 2220 using techniques such as min-max ratio, peak normalisation, square of min-max ratio and the like. The step 2230 operates to remove or reduce outliers in the heart sound waveforms extracted at step 2220 and typically operates in a similar manner to the step 2040.

[0213] The method 2200 continues to noise reduction step 2240. At step 2240 the application 333 performs signal averaging of the normalised waveform signals generated at step 2230. The signal averaging is performed over a predetermined interval or window length to reduce background noise, typically in a similar manner to the step 2050. In the arrangements described the interval is 12 seconds. However different intervals can be used dependent on factors such as noise present, patient pulse rate, patient breathing cycle and the like. The averaging window determines the accuracy and resolution of the blood pressure measurements extracted from the heart sounds. The averaging should be long enough to comprise several heart beats but not so long as to comprise significant blood pressure changes. Depending on variable such as noise and patient pulse rates, the interval should typically be under two minutes. The steps 2230 and 2240 in combination can be considered to generate averaged waveforms.

[0214] The method 2200 continues from step 2240 to an obtaining step 2250. The step 2250 operates to obtain the element (or elements) 2170 determined at step 2160 for the subject. The one or more elements 2170 may be stored in a memory of the processor device 130 (such as memory 306) or on a different device in communication with the processor device 130 via a communication network, such as a cloud server, a separate clinical device, or the like. Using the example of Fig. 19A for example, the step 250 may obtain element Pi3_Pi5.

[0215] The method 2200 continues from step 2250 to a determining step 2260. The step 2260 executes to determine ratio values of the element(s) obtained at step 2250 over an interval. The interval may be predetermined, or may relate to time elapsed since measurements started at step 2210 or a last iteration of the method 2200. For example, referring to Figs. 18A and 18B, if the element obtained at step 2250 was Pn_N22, the step 2260 would execute to determine ratios of the absolute magnitude of the peak P11 of Fig. 18A to the absolute magnitude of the notch N22 of Fig. 18B over the measurement interval.

[0216] The method 2200 continues from step 2260 to a relative blood pressure determining step 2270. At step 2270 the ratio values determined at step 2260 are provided to the trained model 2080 output at step 2070 for determining relative changes in blood pressure. The model 2080 can return a result indicating relative changes in blood pressure. For example, the model 2080 can return one of “Increasing”, “Decreasing” or “Constant”. If more than one element 2170 is selected at step 2160 and used at step 2270, the output may be based on one or more values of “Increasing”, “Decreasing” or “Constant”. For example, an average or most common value may be used, or a value based on an element marked as having a highest score may be used.

[0217] The method 2200 can be repeated as long as signals are received at step 2210. The method 2200 can be used to determine continuous relative changes in blood pressure, or at specific interval, depending on the needs of the patient or subject.

[0218] Fig. 23 shows a method 2300 of calibrating or tuning a model to measure blood pressure for a subject. The method 2300 can be implemented by the software 333, as stored on the memory 309 and controlled by execution of the processor 305.

[0219] The method 2300 starts at a training sound capture step 2310 to capture or generate a sound recording of operation of the subject’s heart according to a training protocol. The training protocol requires examples of patterns of the subject’s heart to be captured during both exercise and recovery periods for a period of time sufficiently long to capture trends in heart sound (S1 , S2, S3 and S4) waveform shape. The training protocol may be similar to inventor experiments described in relation to Fig. 2 in some implementations.

[0220] The method used to generate training samples can be similar to the step 2010. For example, in one implementation, the training protocol relates to assessment of handgrip Maximal Voluntary Contraction (MVC). The participant or subject is in a seated position, holding the dynamometer in their dominant arm maintaining the elbow joint at 90 degrees. The subject squeezes the dynamometer as hard as possible for one second and then relaxes. The exercise is triplicated with 30 seconds of rest between repetitions. The mean of the three readings is the MVC.

[0221] The subject is then asked to remove upper-body clothing for the placement of the ECG leads on their chest and abdomen. The participant is connected to a blood pressure monitor (for example 270), a sound sensor (the sound sensor 110), an ECG (such as 250) and finally a force-sensing resistor (FSR, e.g. 260) on the arm under the cuff which produces a signal marking the start and end of a blood pressure recording. Heart sounds and ECG are continuously recorded during the experiment. The participant is required to sit down upright on a chair with their back supported. The subject’s non-dominant arm will rest on a table, and the subject will rest for 10 minutes before the first measurement. After the 10-minute rest period ends, blood pressure is measured three times separated by 5 minutes between recordings. The ECG data and heart sounds are also recorded and saved simultaneously.

[0222] The subject starts the exercise by squeezing the dynamometer at 30% MVC for a period of 90 seconds maintaining the elbow joint at 90 degrees. The blood pressure readings are taken during the last 20 seconds by starting the automated BP monitor 270 around 55 seconds into the exercise. This is because the HEM907 takes around 35 seconds to provide a single reading. After 90 seconds, the subject is asked to stop the exercise. During the recovery period, 3 additional BP measures will be taken with 1 , 3, 5 minutes post exercise. At ten minutes post recovery, the second 90-sec exercise starts (similar protocol).

[0223] The above protocol provides an example implementation only. Other exercise forms may be used to implement the exercise and recovery periods. For example, the subject may be placed on a treadmill, similarly to in a stress test, or perform some other form of exercise such as lifting a weight. The type of exercise may be varied based on considerations such equipment available, location and physical abilities of the subject. The MVC example described above can be particularly useful if the subject has limited mobility or energy.

[0224] The period of time over which the training measurements are used can also be varied. The period of time for each or exercise and recovery periods should be sufficiently long to allow heart rate in increase and decrease in an observable manner and can vary based on the exercise method used and the condition of the subject. For example, a subject with relatively high blood pressure or a known heart condition, the exercise period may be shortened if the subject’s heart rate could be expected to increase overly quickly. In measuring relative blood pressure changes, a minimum training set could be one set of exercise and recovery, over a total time of 5 to 7 minutes but will depend on factors such as the cardiac health, age and medical conditions of each individual subject. Alternatively, individual training may be substituted by using data from a pool of individuals. If data from a relatively large pool of individuals shows at least one element which has a consistently high correlation with changes in blood pressure, that at least one element and the corresponding behaviour with blood pressure may be used in place of training in some implementations.

[0225] At step 2310 the system 100 is used to capture phonocardiogram readings during the exercise and recovery periods. The sensor 110 is held against skin of the subject in a location suitable for capturing sounds indicating the subject’s heartbeat. For example, the sensor can be placed on the subject’s chest close to the heart. The sound sensor 110 starts records sounds at step 2310. The step 2310 outputs a training set comprising ECG readings 2305 and associated blood pressure readings 2307. In the example of Fig. 20, the ground truth data can relate to blood pressure measurements and / or whether blood pressure is increasing, decreasing or relatively continuous. In contrast, in the method 2300, the ground truth data relates to the actual blood pressure value measurements 2307.

[0226] The method 2300 continues from step 2310 to a communicating step 2315. At step 2315, the sounds captured at step 2310 (for example sounds 2305) are communicated to the processor device 130. In an example arrangement, the sounds are a phonocardiogram communicated directly from the stethoscope 110 to the processor device 130 via the network 120. In other arrangements, the sounds may be transmitted to a further storage device, for example a cloud server. Each portion of the recorded sounds 2305 may tagged with being forone of exercise and recovery periods. However, being tagged with the blood pressure measurements 2307 is typically sufficient. In some arrangements, the simultaneously recorded blood pressure readings (for example for blood pressure monitor 270) are associated with corresponding times of the recorded sounds.

[0227] The method 2300 continues from step 2315 to a waveform extraction step 2320. The step 2320 executes to identify (extract) heart sound waveforms (S1 and S2 waveforms and, in some implementations, S3 and / or S4 waveforms) from the reference phonocardiograms 2305 captured during the training protocol at step 2310. The step 2320 operates in a similar manner to the steps 2020 and 2220. Each waveform corresponds to operation of the heart and has a plurality of peaks and notches, each of the peaks and notches allocated a label (number). The heart sound waveforms can be identified using a model. Examples of suitable models for identifying S1 and S2 waveforms are a Deep Neural Network (DNN) or a Hidden Markov Model (HMM) or any other type of model that can be trained to extract complex patterns. In some implementations, only S1 and S2 waveforms can be extracted (identified). In other implementations, as in the example described, all of S1 , S2, S3 and S4 waveforms are extracted. Similar processes are used to extract S3 and S4 waveforms as S1 and S2 waveforms.

[0228] The method 2300 continues from step 2320 to a synchronisation step 2330. The synchronisation step 2330 operates in a similar manner to the step 2030 to synchronise the heart sound waveforms extracted at step 2320 with the blood pressure readings taken at step 2310 (such as 2307) and the associated tagging or ground truth data of the blood pressure measurement taken using traditional methods at step 2310. The synchronisation allows features of each of the extracted heart waveform sounds to be correlated with the blood pressure measurements taken during the training protocol. Synchronisation can be achieved using multiple methods, for example based on timestamps, recorded start time, use of a synchronising system such as the BioPac system 235 or the like. The synchronised waveforms are used to extract individual feature data from the heart sound waveforms obtained from execution of the step 2320, as described in the following steps.

[0229] The method 2300 continues from step 2330 to a normalisation step 2340. The step 2340 executes in a similar manner to the step 2040 to normalise the synchronised heart sound waveforms using techniques such as min-max ratio, peak normalisation, square of min-max ratio and the like. The normalisation step typically operates to remove outliers.

[0230] The method 2300 continues from step 2340 to a noise removal step 2350. At step 2350 the application 333 performs signal averaging of the generated and synchronised heart sound waveform over a predetermined interval or window length to reduce background noise. The step 2350 operates in a similar manner to the step 2050. Any suitable signal averaging technique that can increase strength of each heart sound waveform’s signal relative to noise can be used. In the arrangements described, the interval is 12 seconds. However different intervals can be used dependent on factors such as noise present, patient pulse rate, patient breathing cycle and the like. For example, breathing data may be used to estimate the averaging interval and adapt the interval over time to further individualise use of heart sound waveforms. For example, a longer breathing cycle may result in a longer averaging interval and vice versa.

[0231] The method 2300 continues from step 2350 to a feature determining step 2360. The feature determining step 2360 executes to extract or determine at least one feature (also referred to as an element) of the heart waveforms determined at step 2320 which show behaviour that corresponds with fluctuations in blood pressure. The extracted feature or element can also be considered to provide a ratio in form of Xap Ys* as described above. Fig. 24 shows an example method 2400, described hereafter, as implemented at the step 2360.

[0232] The method 2300 continues from step 2360 to a training step 2370. The step 2370 operates to select the one or more extracted features (elements) and the associated synchronised blood pressure readings and associated training data. In some arrangements, the step 2370 executes to train a deep regression model used to determine blood pressure for the subject. The step 2370 can also operate to train a deep neural network such as a multi-layer one-dimensional convolutional neural network or a variational autoencoder using the blood pressure 2360 (such as 2307) from step 2310 and the synchronised features extracted at step 960. In some instances, the model may be partially trained. For example, initial layers of the deep neural network may be pre-trained in relation to change in peaks and notches of heart sound waveforms relative to blood pressure, for example based on training samples obtained using the step 2010 for multiple individuals. The training at step 2370 may relate to training one or more final layers of a convolutional neural network. The training at step 2370 be performed based on the selected one or more elements of step 2360, for example by providing Pi3_Pi5 data for training in the example of Fig. 19A. Training can be performed using standard machine learning regression and training techniques. For example, regression techniques can include linear, logistic, ridge, lasso, polynomial and Bayesian regression techniques. Example supervised learning techniques can include convolutional neural network or a variational autoencoders, or other similar techniques. Similarly to the step 2070, the step 2370 can execute to obtain a pre-existing model. The previously trained model would typically be developed overtime using the previous iterations of steps 2310 to 2370 and / or experimentation. The previously trained model and may relate to training data from one or more individuals having one or more similar characteristics, such as age range, cardiac characteristics, medical conditions and the like. The step 2370 in other implementations requires training to be performed for each individual.

[0233] The method 2300 outputs a trained model 2380, whether generated or obtained from a pre-existing model. The model 2380 is trained to generate a blood pressure measurement corresponding to the particular subject using data from an identified element for the subject.

[0234] Fig. 24 shows the method 2400 of determining an element of a subject’s heart sounds, providing an example implementation of the step 2360. The method 2400 can be implemented by the software 333, as stored on the memory 309 and controlled by execution of the processor 305.

[0235] The method 2400 starts at an identify waveform step 2410. The step 2410 operates to receive the noise removed waveforms generated at step 2350, to identify a start and end of a waveform using the waveforms over the interval window. The step 2410 can operate in a similar manner to the step 2020 and use models to identify waveforms such as a Deep Neural Network (DNN) or a Hidden Markov Model (HMM) or any other type of model that can be trained to identify complex patterns may be used to determine the start and end of each waveform. The step 2410 can be implemented for each heart sound, for example for each of S1, S2, S3 and S4.

[0236] The method 2400 continues under execution of the processor 333 from step 2410 to an identifying step 2420. The step 2420 executes to detect peaks and notches in the waveform identified in step 2410. The peaks and notches can be determined using known peak detection algorithms and may be based on factors such as changing slope, absolute magnitude and the like. A threshold minimum absolute magnitude may be used to in identifying peaks and notches in some implementations.

[0237] The step 2420 executes to find peaks and notches for each of the heart sound waveforms for the subject in a similar manner to the step 2120 and as described for example in relation to Figs. 18A and 18B. Each peak and notch is allocated a number, for example as shown in Figs. 18A and 18B. For example, the step 2420 can execute to find peaks and notches for an S1 waveform, peaks and notches for an S2 waveform, peaks and notches for an S3 waveform and peaks and notches for an S4 waveform if all 4 waveforms have beneextracted. In the example of Figs. 18A and 18B, where only S1 and S2 waveforms are detected, the step 2420 executes to identify (i) the peaks P11 , P12, P13 and P14 and the notches N11 , N12, N13 and N14 for the S1 waveform 1810, and (ii) the peaks P21 , P22, P23, and P24 and the notches N21, N22, N23 and N24 for the S2 waveform 1830.

[0238] The method 2400 continues from step 2420 to a ratio determining step 2430. The step 2430 executes to determine potential element ratios present in the heart sound waveforms using the peaks and notches identified in operation of the step 2420. The step 2430 effectively operates, for each peak or notch, to determine a set of ratios with each of the other peaks and notches identified. Each ratio will comprise the selected peak or notch and one other peak or notch. Duplicate ratios, for example inverse ratios, can be discarded.

[0239] The step 2430 operates in a similar manner to the step 2130. Using the example of Figs. 18A and 18B and using the peak P11 as a selected peak, the ratios determined would reflect those in Table 1 described hereinbefore.

[0240] The step 2430 operates to determine ratio combinations for each park or notch, generating data similar to that of Table 1 for each of P11, P12, P13 and P14 and the notches N11, N12, N13 and N14 for the S1 waveform 180, and (ii) the peaks P21, P22, P23, and P24 and the notches N21 , N22, N23 and N24 in the example of Figs. 18A and 18B.

[0241] The method 2400 continues from step 2430 to a ratio determining step 2440. The step 2440 executes to calculate values for each ratio determined in step 2430 across the set of samples collected in training, for example across exercise and recovery periods. The ratios are determined as the magnitude of magnitudes for Xap Ys*, where each of X and Y can be a peak or notch in any of the derived sounds, the first subindex relates to the number of S1 to S4 and the second subindex is the peak / notch number or label, as described above. The step 2440 operates in a similar manner to the step 2140.

[0242] The method 2400 continues under execution of the processor 333 from step 2440 to a score determining step 2450. The step 2450 compares fluctuation in the set of values determined for each ratio in step 2430 with fluctuation in the blood pressure measurements taken at step 2310 (i.e. the measurements 2307). The step 2450 operates to determine a correlation score for each ratio based on changes in the ratio values compared to changes in synchronised measured blood pressure. The step 2450 operates in a similar manner to the step 2150. The correlation may be determined using correlation or pattern matching algorithms such as linear or non-linear regression algorithms with consideration of noise and outliers.

[0243] The method 2400 continues under execution of the processor 333 from step 2450 to an element determining step 2460. The step 2460 executes to select at least one element 2470 based on the correlation scores determined at the step 2450. The step 2460 operates in a similar manner to the step 2160. The step 2460 may select the score based on any of a number of criteria. Typically, the one or more elements are determined to be a highest score. The one or more elements may also be based on a highest score across a longest period of time and / or using a threshold. As shown in Figs. 19A and 19B, at least one element was found which has suitably high correlation with changes in blood pressure for most individuals in the study. For each individual, at least one element can be found that has a suitably high level of correlation. Further, at least one common element can typically also be found in larger sets of data measured for several individuals, the at least one common element having a suitably high correlation with changes in blood pressure for most individuals in that data set. The one or more elements selected at step 2460 can therefore relate to one or more elements with highest scores for the individual or based on a relatively large pool of individuals where the same element has been observed to have a high correlation with changes blood pressure.

[0244] The method 2400 ends after execution of the step 2460. In some implementations, the method 2400 may return a single determined element as 2470, for example Pi3_Ni5 of Fig. 19A. In other implementations, the method 1000 may be configured to return more than one element as output 2470. As shown in Figs. 19A to 19D, more than one element may have a suitable correlation to changes in blood pressure for an individual. While use of a single element for the output 2470 can be sufficient, use of more than one element may provide benefits of redundancy or double checking or to account for effects of noise.

[0245] The methods 2300 and 2400 may be implemented for one or both of systolic and diastolic blood pressure.

[0246] Fig. 25 shows a method 2500 of determining relative blood pressure in a subject. The method 2500 can be implemented by the software 333, as stored on the memory 309 and controlled by execution of the processor 305. The method 2500 may be implemented for one or both of systolic and diastolic blood pressure, corresponding to whether the methods 2300 and 2400 were implemented for one or both of systolic and diastolic blood pressure for the individual patient.

[0247] The method 2500 starts at a sound capturing step 2510. At the step 2510, the sound sensor 110 is held against the subject skin close to the subject’s heart and captures sounds made by operation of the subject’s heart. The step 2500 can include the sound sensortransmitting the sounds to the processor device 130, similarly to operation of step 2315 in implementations where the sound sensor 110 and the processor device 130 are separate devices. The sounds can be communicated at discrete intervals or continuously.

[0248] The method 2500 continues under execution of the processor 305 from step 2500 to an extraction step 2520. At step 2520 the software 333 executes to extract or identify a plurality of waveforms for heart sound signals from the sounds captured at step 2510. The step 2520 operates in a similar manner to the steps 2020 and 2320. Each waveform corresponds to operation of the heart and has a plurality of peaks and notches. Each of the peaks and notches is allocated a number, as described in steps 2120 or 2420 and described in relation to Figs. 18A and 18B. The extracted (identified) waveform signals can include at least S1 and S2 waveforms. In some implementations, the extracted waveform signals can also include S3 and / or S4 waveforms. The heart sounds waveforms can be identified or extracted using a model at step 2520. Examples of suitable models are a Deep Neural Network (DNN), a Hidden Markov Model (HMM), Convolutional Neural Network, or any other type of model that can be trained to extract complex patterns. In the example discussed below, both S1 and S2 waveforms are used to determine blood pressure measurements. However, in other implementations, one or more of S1, S2, S3 and S4 waveforms can be used. The waveforms identified at step 2520 include at least the waveforms extracted at training, that is the waveforms identified at step 2320.

[0249] The method 2500 continues from step 2520 to a normalising step 2530. The step 2530 operates similarly to the steps 2040, 2240 and 2340 and executes to normalise the heart sound waveforms extracted at step 2520 using techniques such as min-max ratio, peak normalisation, square of min-max ratio and the like. The step 2530 operates to remove or reduce outliers in the heart sound waveforms extracted at step 2520 and typically operates in a similar manner to the step 2040.

[0250] The method 2500 continues to noise reduction step 2540. At step 2540 the application 333 performs signal averaging of the normalised waveform signals generated at step 2530. The signal averaging is performed over a predetermined interval or window length to reduce background noise, typically in a similar manner to the step 2350. In the arrangements described the interval is 12 seconds. However different intervals can be used dependent on factors such as noise present, patient pulse rate, patient breathing cycle and the like. The averaging window determines the accuracy and resolution of the blood pressure measurements extracted from the heart sounds. The averaging should be long enough to comprise several heart beats but not so long as to comprise significant blood pressure changes. Depending on variable such as noise and patient pulse rates, the interval should typically be under two minutes. The steps 2530 and 2540 in combination can be considered to generate averaged waveforms.

[0251] The method 2500 continues from step 2540 to an obtaining step 2550. The step 2550 operates to obtain the element (or elements) determined at step 2460 for the subject. Using the example of Fig. 19A for example, the step 2550 may obtain element Pi3_Pi5.

[0252] The method 2500 continues from step 2550 to a determining step 2560. The step 2560 executes to determine the ratio values of the element(s) obtained at step 2550 over an interval. The interval may be predetermined, or may relate to time elapsed since measurements started at step 2510 or a last iteration of the method 2500. For example, referring to Figs. 18A and 18B, if the element obtained at step 2550 was Pn_N22, the step 2560 would execute to determine ratios of the absolute magnitude of the peak P11 to the absolute magnitude of the notch N22 over the measurement interval.

[0253] The method 2500 continues from step 2560 to a relative blood pressure determining step 2570. At step 2570 the ratios determined at step 2560 are provided to the model 2480 generated at step 2470 for determining blood pressure measurements. The model 2480 can return a result indicating a determined blood pressure reading. For example, the model 2480 can return values determined based on standard units and range of blood pressure measurement, for example 85 mmHg.

[0254] The method 2500 can be repeated as long as signals are received at step 2510. The method 2500 can be used to determine continuous relative changes in blood pressure, or at specific interval, depending on the needs of the patient or subject.

[0255] The example arrangements described collect reference blood pressure measurements (such as 1307 and 1517) using auscultatory methods, such as using a mercury sphygmomanometer or cuff device. Other known methods can be used to collect reference bold pressure readings, such as oscillometric methods, ultrasound, finger cuffs and the like.Industrial Applicability

[0256] The arrangements described are applicable to the medical, fitness and wellbeing industries and particularly for medical services where cardiac health and blood pressure are relevant.

[0257] The arrangements described allow a blood pressure measurement to be made using a sound sensor such as a phonocardiogram. Accordingly, blood pressure measurements may be made in a way that is non-invasive to the subject. Additionally, sounds sensors such asphonocardiograms can be implemented as wearable devices and are more suitable for continuous monitoring than previous solutions. Use of a sound sensor allows factors such as mechanical properties of the arteries to be minimised compared to solutions using pulse transit time.

[0258] In calibrating the model to each individual’s S1 / S2 waveform patterns in Fig. 15, the arrangements described allow the blood pressure measurement to account for individual variances in heart operation. Further, the calibration step 1530 can be repeated as a subject’s circumstances change, allowing for ongoing fine tuning. Use of an interval at step 1740 further allows anomalies in readings to be smoothed out.

[0259] In the alternative embodiment of Fig. 12, use of sound measurements allows relative changes in blood pressure to be non-invasively and quickly identified.

[0260] In addition to decreasing invasiveness and patient discomfort, using phonocardiograms to measure blood pressure can simplify the steps needed to take multiple blood pressure measurements. Standard equipment used in medical settings, an electronic stethoscope is sufficient once calibration has been performed.

[0261] Further, issues such as sweat and changes in skin features will not affect measurements.

[0262] In determining at least one element that correlates with changes in blood pressure the arrangements described can be implemented with reduced computation, for example on an edge or wearable device.

[0263] As the arrangements described are based on captured heart sounds after training is completed, the arrangement described are more suitable for continuous use. Further, the requirement to use a cuff is removed. Use of a sound sensor can be easier for the subject or patient to implement and may in some cases be implemented by the patient in their home after training has been completed. Using the arrangements described, at least one element with the structure described that is appropriate for a given person or subject based on data collected for the subject at a given time may identified.

[0264] The arrangements described can also be adapted as a patient’s circumstances or health changes by re-training the patient-specific model. Additionally, to improve accuracy over time additional training could be performed for the trained models generated at execution of Figs. 21and 24 at different points in time. As described above, for each individual, at least one element can be found that has a suitably high level of correlation. Further, at least one common element can typically also be found in larger sets of data measured for several individuals, the at least one common element having a suitably high correlation with changes in blood pressure for most individuals in that data set.

[0265] The foregoing describes only some embodiments of the present invention, and modifications and / or changes can be made thereto without departing from the scope and spirit of the invention, the embodiments being illustrative and not restrictive.

[0266] (Australia Only) In the context of this specification, the word “comprising” means “including principally but not necessarily solely” or “having” or “including”, and not “consisting only of”. Variations of the word "comprising", such as “comprise” and “comprises” have correspondingly varied meanings.

Claims

CLAIMS:

1. A computer-implemented method of measuring blood pressure behaviour of a subject, the method comprising: extracting at least one waveform from a sound recording of operation of the subject’s heart, each extracted waveform representing one of S1 and S2 sounds generated by operation of the subject’s heart; averaging each of the extracted waveforms over a predetermined interval; extracting features of the each of the averaged waveforms based on a shape of each of the averaged waveforms; and providing the extracted features to a model trained for the subject to determine a measurement of blood pressure behaviour of the subject.

2. The method according to claim 1, wherein the sound recording is a phonocardiogram.

3. The method according to claim 1, wherein the determined measurement of blood pressure is one of (i) a measurement value, or (ii) a relative measurement.

4. The method according to claim 3, wherein the relative measurement is one of a relative increase a relative decrease or remaining at a constant level.

5. The method according to claim 3, wherein the model is a trained deep neural network.

6. The method of claim 1, wherein extracting features comprises matching a portion of one of the extracted waveforms corresponding to a single heartbeat to one of a plurality of different waveform shapes and selecting a pattern related to notches and peaks of the waveform as the extracted features.

7. The method according to claim 1, wherein extracting features comprises selecting a maximum amplitude of the noise reduced waveform over the predetermined interval.

8. The method according to claim 1, wherein extracting the features comprises performing a KLT analysis on the waveforms and extracting first coefficient values of the KLT analysis.

9. The method according to claim 5, further comprising receiving a reference sound recording and corresponding blood pressure measurements obtained using a cuff for the subject, and calibrating a model for determining blood pressure, the calibrating comprising: extracting at least one of a reference S1 waveform and an S2 waveform from the reference sound recording, synchronising the at least one reference waveform to the measurements obtained using the cuff, averaging each of the synchronised reference waveforms over a predetermined interval; extracting features of the each of the averaged reference waveforms based on amplitude , and tuning a model using the extracted features, wherein the tuned model is used to determine the measurement of blood pressure behaviour.

10. The method according to claim 9, further comprising determining if placement on the subject of a sensor generating the recording has changed, determining a reference measurement, wherein the blood pressure measurement is determined relative to the blood pressure measurement.

11. The method according to claim 9, wherein the refence sound recording is associated with identifying a location on the subject’s body a sound sensor was placed to generate the recording.

12. A non-transitory computer-readable medium storing instructions executable on a processor, the instructions configured to implement a method of measuring blood pressure of a subject, the method comprising: extracting at least one waveform from a sound recording of operation of the subject’s heart, each waveform representing one of S1 and S2 sounds generated by operation of the subject’s heart; averaging each of the extracted waveforms over a predetermined interval; extracting features of the each of the averaged waveforms based on a shape of each of the averaged waveforms; and providing the extracted features to a model trained for the subject to determine a measurement of blood pressure behaviour of the subject.

13. A system, comprising: a sound sensor configured to measure a sound recording of operation of a subject’s heart; a processor; and a memory, the memory storing instructions executable on the processor to: extract at least one waveform from the sound recording, each waveform representing one of S1 and S2 sounds generated by operation of the subject’s heart; average each of the extracted waveforms over a predetermined interval; extract features of the each of the averaged waveforms based on a shape of each of the averaged waveforms; and provide the extracted features to a model trained for the subject to determine a measurement of blood pressure behaviour of the subject.

14. A method of measuring blood pressure of a subject, comprising: capturing sounds made by the subject’s heart; identifying a plurality of waveforms from the captured sounds, each waveform corresponding to operation of the heart and having a plurality of peaks and notches, each of the peaks and notches allocated a label; determining an element for the subject, the element being derived at a training stage as showing correlation with variations in blood pressure, wherein the element is a ratio of absolute magnitudes of form Xap Ys*, where each of X and Y is a peak or notch in any of the identified waveforms, each of a and 6 identifies one of the plurality of waveforms and each of 0 and (|) relates to the allocated label of the corresponding peak or notch in the corresponding waveform; and obtaining a blood pressure measurement by providing values of the determined element in the identified waveforms to a trained model.

15. The method according to claim 14, wherein the identified plurality of waveforms comprises waveforms of S1 and S2 sounds.

16. The method according to claim 14, wherein the identified plurality of waveforms comprises waveforms of S1 , S2, S3 and S4 sounds.

17. The method according to claim 14, wherein the blood pressure measurement indicates a relative change in blood pressure of the subject.

18. The method according to claim 14, wherein the blood pressure measurement provides a blood pressure measurement value.

19. The method according to claim 14, wherein the element derived at the training stage as showed a correlation of 0.40 or above with variations in blood pressure.

20. The method according to claim 14, wherein the captured sounds are a phonocardiogram captured by an electronic stethoscope.

21. A system for measuring blood pressure of a subject, comprising: a sound sensor configured to capture sounds made by the subject’s heart; a processor; and a memory, the memory storing instructions executable on the processor to: identify a plurality of waveforms from the captured sounds, each waveform corresponding to operation of the heart and having a plurality of peaks and notches, each of the peaks and notches allocated a label; determine an element for the subject, the element being derived at a training stage as showing correlation with variations in blood pressure, wherein the element is a ratio of absolute magnitudes of form Xap Ys*, where each of X and Y is a peak or notch in any of the identified waveforms, each of a and 6 identifies one of the plurality of waveforms and each of 0 and (|) relates to the allocated label of the corresponding peak or notch in the corresponding waveform; and obtain a blood pressure measurement by providing values of the determined element in the identified waveforms to a trained model.

22. The system of claim 21, wherein the sound sensor is an electronic stethoscope.

23. The system of claim 21, wherein the sound sensor is an electronic stethoscope and the processor and the memory are integral to the electronic stethoscope.

24. A non-transitory computer-readable medium storing instructions executable on a processor, the instructions configured to implement a method of measuring blood pressure of a subject, the method comprising:receiving captured sounds made by the subject’s heart; identifying a plurality of waveforms from the captured sounds, each waveform corresponding to operation of the heart and having a plurality of peaks and notches, each of the peaks and notches allocated a label; determining an element for the subject, the element being derived at a training stage as showing correlation with variations in blood pressure, wherein the element is a ratio of absolute magnitudes of form Xap Ys*, where each of X and Y is a peak or notch in any of the identified waveforms, each of a and 6 identifies one of the plurality of waveforms and each of 0 and (|) relates to the allocated label of the corresponding peak or notch in the corresponding waveform; and obtaining a blood pressure measurement by providing values of the determined element in the identified waveforms to a trained model.

Images