System and method for biomagnetic data quality assessment
The classification and enhancement of biomagnetic data into high, medium, and low quality classes using TCNs addresses noise interference in biomagnetic sensors, ensuring reliable physiological information derivation, particularly in wearable devices.
Patent Information
- Application Number
- PCT/EP2026/058724
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-05
- Filing Date
- 2026-03-26
- Publication Date
- 2026-11-12
AI Technical Summary
Biomagnetic sensors, despite their increased sensitivity, are prone to recording magnetic noise that compromises the reliability of real-time applications by interfering with the interpretation of biomagnetic data, necessitating improved methods for noise reduction and quality assessment.
A method and system for classifying biomagnetic data into three classes: first (high quality), second (medium quality), and third (low quality), where high-quality data is used directly, medium-quality data is enhanced, and low-quality data is discarded, utilizing machine learning models like Temporal Convolutional Networks (TCNs) for classification and enhancement.
Ensures accurate and reliable biomagnetic data processing by filtering out low-quality data and enhancing medium-quality data, thereby improving the reliability of physiological information derivation from biomagnetic sensors, especially in wearable devices.
Smart Images

Figure EP2026058724_12112026_PF_FP_ABST
Abstract
Description
[0001] System and Method for Biomaqnetic Data Quality Assessment
[0002] Field of the invention
[0003] The present invention relates to a method and system for processing biomagnetic data, in particular, for assessing and improving the quality of biomagnetic data received from biomagnetic sensor(s).
[0004] Background to the invention
[0005] Biomagnetic sensors have emerged as an important tool for non-invasive physiological measurements, including but not limited to, signals originating from the heart, pulse, and other sources, recorded from a subject.
[0006] The increased sensitivity of biomagnetic sensors enables the detection of weak magnetic signals from biological sources. However, this increased sensitivity also results in the recording of additional sources of magnetic noise that can interfere with the interpretation of the biomagnetic data. Many real-time applications, such as heart rate monitoring and heart rate variability estimation, rely on clean, reliable and high-quality biomagnetic data where the effect of this noise on measurements is minimised or removed. When sensed biomagnetic signals are affected by external magnetic noise, the reliability of these applications is compromised. Present techniques may employ spectral filtering or matched filtering to reduce external magnetic noise or use gradiometry to reduce the effect of noise by measuring with two sensors and removing the common mode.
[0007] Embodiments of the present invention seek to provide an improved approach.
[0008] Summary of the invention
[0009] According to a first aspect of the present invention there is provided a method for (e.g. of) processing biomagnetic data obtained from a subject using one or more biomagnetic sensor(s), the method comprising:
[0010] receiving the biomagnetic data from the one or more biomagnetic sensor(s); andclassifying the biomagnetic data as being in a first, second or third class; wherein if:
[0011] (i) the biomagnetic data is in the first class, the method comprises making said biomagnetic data available for use in determining physiological information relating to the subject;
[0012] (ii) the biomagnetic data is in the second class, the method comprises enhancing said biomagnetic data;
[0013] (iii) the biomagnetic data is in the third class, the method comprises making said biomagnetic data unavailable for use in determining physiological information relating to the subject.
[0014] The first aspect extends to a computer program, a software tool and a non-transitory computer-readable medium, each comprising instructions that, when executed by a processing system, cause the processing system to perform the method according to the first aspect.
[0015] According to a second aspect of the present invention there is provided a system for processing biomagnetic data obtained from a subject using one or more biomagnetic sensor(s), the system comprising:
[0016] the one or more biomagnetic sensor(s);
[0017] wherein the system is configured to:
[0018] classify the biomagnetic data as being in a first, second or third class;
[0019] wherein if:
[0020] (i) the biomagnetic data is in the first class, the system makes said biomagnetic data available for use in determining physiological information relating to the subject;
[0021] (ii) the biomagnetic data is in the second class, the system enhances said biomagnetic data;
[0022] (iii) the biomagnetic data is in the third class, the system makes said biomagnetic data unavailable for use in determining physiological information relating to the subject.According to a third aspect of the present invention there is provided a wearable apparatus for processing biomagnetic data obtained from a subject using one or more biomagnetic sensor(s), the wearable apparatus comprising:
[0023] the one or more biomagnetic sensor(s);
[0024] wherein the wearable apparatus comprises a processing module configured to:
[0025] classify the biomagnetic data as being in a first, second or third class;
[0026] wherein if:
[0027] (i) the biomagnetic data is in the first class, the processing module makes said biomagnetic data available for use in determining physiological information relating to the subject;
[0028] (ii) the biomagnetic data is in the second class, the processing module enhances said biomagnetic data;
[0029] (iii) the biomagnetic data is in the third class, the processing module makes said biomagnetic data unavailable for use in determining physiological information relating to the subject.
[0030] Thus it will be seen that, in accordance with the invention, biomagnetic data obtained by one or more biomagnetic sensor(s) may be analysed to determine whether the data should be made available to be used in determining physiological information relating to the subject, whether it should be enhanced or whether it should be made unavailable for use (e.g. discarded). This is useful for biomagnetic measurements which make use of highly sensitive magnetic sensors particularly vulnerable to noise. Embodiments of the invention, therefore, may provide a robust mechanism for evaluating the quality of biomagnetic signals to ensure accurate downstream processing. The invention may be used to provide real-time feedback on signal quality assessment results.
[0031] The method may be a computer-implemented method. In a set of embodiments, the system or a processing module of the system receives the biomagnetic datafrom the one or more biomagnetic sensor(s) - e.g. before classifying the biomagnetic data as being in a first, second or third class.
[0032] In a set of embodiments, the system or a processing module of the system receives the biomagnetic data from the one or more biomagnetic sensor(s) - e.g. before classifying the biomagnetic data as being in a first, second or third class. In a set of embodiments, the processing module of the wearable apparatus receives the biomagnetic data from the one or more biomagnetic sensor(s) - e.g. before classifying the biomagnetic data as being in a first, second or third class.
[0033] The physiological information relating to the subject may be a physiological parameter of the subject. The physiological parameter of the subject may be one or more of the list comprising: pulse rate, heart rate, and heart rate variability.
[0034] The term 'subject' as used herein includes any human or non-human animal subject, including any human or non-human mammal, bird, fish, reptile, amphibian, etc. However, in a set of embodiments, the subject is a human. The one or more biomagnetic sensor(s) may be implantable inside a body of the subject or may be external to the body of the subject.
[0035] In a set of embodiments, the first class is indicative of high quality biomagnetic data, the second class is indicative of medium quality biomagnetic data, and the third class is indicative of low quality biomagnetic data. High quality biomagnetic data may be considered as data from which the (e.g. useful) physiological information (e.g. a physiological parameter) relating to the subject can be derived. Low quality biomagnetic data may be considered as data from which the (e.g. useful) physiological information (e.g. a physiological parameter) relating to the subject cannot be derived, even after enhancement. Medium quality biomagnetic data may be considered as data from which the (e.g. useful) physiological information (e.g. a physiological parameter) relating to the subject can be derived after enhancement of said biomagnetic data.The biomagnetic data is preferably time-series data, e.g. from the one or more biomagnetic sensor(s).
[0036] In a set of embodiments, the step of classifying the input data as being in a first, second, or third class uses a machine learning model. The machine learning model may be trained to receive as input biomagnetic (e.g. time-series) data and provide as output labelled biomagnetic data, wherein the label is (e.g. associated with the) first class, second class or third class.
[0037] In a set of embodiments, the machine learning model is based on a Temporal Convolutional Network (TCN). For example, the machine learning model may be a TCN or may comprise a TCN. Such a machine learning model would be capable of making signal quality assessments with low latency.
[0038] In a set of embodiments, the machine learning model has been trained on training data comprising labelled biomagnetic data, wherein the label corresponds to one of the first class, second class or third class.
[0039] The training data used to train the machine learning model is preferably obtained from biomagnetic sensor measurements from a diverse range of sources. For example, the training data may comprise labelled biomagnetic data from both human and non-human animal sources. The labelled biomagnetic data may be obtained from biomagnetic measurements on different areas on a subject’s body. The labelled biomagnetic data may be obtained in different environments (i.e. with varying levels of noise present) - e.g. magnetically shielded and unshielded environments. The training data may comprise (e.g. only) labelled TMR sensor data.
[0040] In a set of embodiments, the machine learning model has been trained on training data, wherein the training data comprises:
[0041] a. high quality biomagnetic data labelled as being in the first class; b. low quality biomagnetic data labelled as being in the third class; andc. biomagnetic data, which is not labelled as being in the first class or the third class, labelled as being in the second class.
[0042] In a set of embodiments, the machine learning model has been trained on training data, wherein the training data comprises:
[0043] a. biomagnetic data having a signal to noise ratio, SNR, of at least 10 db and / or having a noise level below (e.g. approximately) 25 PTRMS labelled as being in the first class;
[0044] b. biomagnetic data having an SNR of (e.g. approximately) up to 1 db and / or having a noise level of at least 25 PTRMS, labelled as being in the third class; and c. biomagnetic data which is not labelled as being in the first class or the third class, labelled as being in the second class.
[0045] Thus, high quality biomagnetic data may be biomagnetic data having a signal to noise ratio, SNR, of at least 10 db and / or having a noise level below (e.g. approximately) 25 PTRMS. Similarly, low quality biomagnetic data may be biomagnetic data having an SNR of (e.g. approximately) up to 1 db and / or having a noise level of at least 25 PTRMS. Medium quality biomagnetic data may be biomagnetic data not meeting the criteria of high quality or low quality.
[0046] In a set of embodiments, the biomagnetic data labelled as being in the first class is also in the frequency range of 1 Hz to 500 Hz.
[0047] In a set of embodiments, the machine learning model has been trained by active learning - e.g. pool-based sampling. This may comprise training the model on a fully labelled (e.g. small - e.g. having fewer than 50 examples) subset of a training dataset of biomagnetic data which yields class-membership probabilities for instances of biomagnetic data. Based on that, new samples about which the system is least confident, may be selected from a large pool of unlabelled biomagnetic data and passed to a labeller - e.g. a system operator, researcher, clinician - for labelling. These newly labelled samples may then be included in the training data for model retraining.The step of classifying the biomagnetic data as being in a first, second or third class will be referred to herein as the ‘classification step’. The step of enhancing the biomagnetic data classified as being in the second class will be referred to herein as the ‘enhancement step’.
[0048] The biomagnetic data may be re-sampled to a nominal or pre-determined sampling rate before the classification step. In embodiments wherein the classification step is performed using a machine learning model, the nominal or pre-determined sampling rate may be the sampling rate corresponding to data on which said machine learning model was trained.
[0049] In a set of embodiments, biomagnetic data is input (e.g. to the machine learning model both during training and at inference time) in the form of a window of timeseries data. Preferably, there are a plurality of (e.g. consecutive and overlapping) windows of time-series biomagnetic data. The plurality of windows may be input (e.g. to the machine learning model) one at a time. The plurality of windows may be classified one at a time.
[0050] Thus, in a set of embodiments, each window of time-series biomagnetic data, from a plurality of consecutive and overlapping windows of time-series biomagnetic data, is input to the classifying step one at a time.
[0051] In a set of embodiments, a sliding window mechanism is used, wherein a fixed-length window moves over a sequence of time-series biomagnetic data to extract overlapping sub-sequences for processing. In other words, consecutive windows of biomagnetic data may overlap. This allows the machine learning model to be more efficient at handling continuous time-series data, e.g. for longer term physiological monitoring. In a set of embodiments, the overlap between the windows of biomagnetic data is in the range of 25% to 75%, preferably 50%, of the length (i.e. in time) of the window.
[0052] Each window may have a length in the range of 200ms to 4 s, for example, in the range of 500 ms to 3 s. The applicant has found that the range of 500 ms to 3 s isparticularly suitable for applications wherein physiological parameters such as heart rate are measured. In a set of embodiments, each window has a length of between 500 ms and 1 s (e.g. approximately 750 ms). The Applicant has found this window length to be optimal for model precision.
[0053] As will be appreciated by those skilled in the art, biomagnetic sensors are magnetic sensors that can detect biomagnetic fields arising from a subject. For example, a biomagnetic sensor may be a tunnel magnetoresistance (TMR) sensor, a superconducting quantum interference device (SQUID), or an optically pumped magnetometer (OPM). The, or each biomagnetic sensor may be a magnetoresistance (MR) sensor . In a set of embodiments, the or each biomagnetic sensor is a tunnel magnetoresistance (TMR) sensor.
[0054] The applicant has found that TMR sensors are well-suited for wearable systems and devices because of the small sensor footprint (e.g. the
[0055] sensor package may be as small as 1.5 mm by 1.5 mm) that can be achieved without compromising on sensitivity. This small sensor footprint may allow biomagnetic measurements to be performed using a non-bulky portable and / or wearable device where the portable and or wearable device comprises one or more TMR sensor(s).
[0056] When TMR sensors are used, physiological (e.g. cardiac) monitoring may be performed in ambient environments without a magnetically shielded room and without sensor heating or cooling requirements. The ability to classify (e.g. by quality), and selectively enhance, biomagnetic data improves the performance of biomagnetic measurements when the one or more biomagnetic sensor(s) are incorporated into a wearable apparatus and / or system which can be comfortably worn under clothing, like a t-shirt. This may improve monitoring of biomagnetic activity in various settings, including everyday life and clinical environments, with no shielding and an uncontrolled environment.
[0057] In a set of embodiments, the, or each, biomagnetic sensor is configured to sense magnetic fields having a magnetic field strength below 100 pT, e.g. below 10 pT.As mentioned above, if the biomagnetic data is classified in the third class, the biomagnetic data is made unavailable for use in determining physiological information relating to the subject. This may comprise storing the (e.g. window of) third-class biomagnetic data with a flag indicating that said biomagnetic data is unusable and / or discarding the (e.g. window of) third-class biomagnetic data. In some embodiments, the third-class biomagnetic data may be stored (e.g. labelled as third-class) for use as training data. Thus, the amount of labelled third-class biomagnetic data available for retraining the model may be increased.
[0058] The biomagnetic data may be received continuously from the one or more biomagnetic sensors over several hours or days - for example, as part of a longterm monitoring system - or by way of a controlled measurement taking place over a short period of time - e.g. less than a minute.
[0059] It has been recognised by the applicant that to derive physiological information from biomagnetic data, there may be a minimum time period over which biomagnetic measurements must be taken. Therefore, the biomagnetic data may be collected in measurement cycles. Each complete measurement cycle preferably comprises a plurality of windows of biomagnetic data. It has been recognised by the applicant that approximately 30 seconds of biomagnetic data may provide sufficient time for useful physiological information to be derived for both consumer device health monitors and clinical applications, such as pulse rate, heart rate or heart rate variability. For non-clinical applications useful information may be obtained after as little as 3 seconds. Therefore, the biomagnetic data may be collected (e.g. for the purpose of determining physiological information relating to the subject) in one or more measurement cycles, each (e.g. complete) measurement cycle having a duration of at least 3 seconds, e.g. at least 4 seconds, e.g. at least 30 seconds. The physiological information relating to the subject may be determined from a measurement cycle. Where there is a plurality of measurement cycles, they may overlap - e.g. similarly to a sliding window mechanism. The invention (e.g. the method, system, or wearable apparatus) may be used to help ensure that the biomagnetic datacollected in each measurement cycle is of sufficient quality to determine accurate physiological information.
[0060] In one example, when a first measurement cycle is completed, the first-class data from said cycle can be used in determining physiological information relating to the subject. Subsequent completed measurement cycles may be used to update the physiological information - e.g. to provide an updated physiological parameter. Thus, the physiological information may be determined and updated in real time with low latency.
[0061] If, during a measurement cycle, there is not sufficient biomagnetic data classified as being in the first class then the measurement cycle may be terminated and a new measurement cycle may be initiated.
[0062] In a set of embodiments, after (e.g. a window of) biomagnetic data has been classified as being in the third class, the measurement cycle continues. In embodiments wherein the method processes a window of biomagnetic data at a time, there may be a plurality of consecutive windows being classified as third class (i.e. third-class data) wherein a new measurement cycle is initiated when a threshold criterion is reached. The threshold criterion may be a maximum number of consecutive windows being classified as third class. The maximum number of consecutive windows being classified as third class may be in the range of from 2 to 10. The threshold criterion may be a maximum amount of time wherein only third-class data has been identified. The maximum time duration may be between 200 ms and 3 s - e.g. between 200 ms and 2 s- e.g. between 200 ms and 1 s.
[0063] In some embodiments, during the measurement cycle, the overall quality of the biomagnetic data is determined based on the classification of the biomagnetic data obtained in said measurement cycle. If the biomagnetic data is determined to be of insufficient overall quality such that sufficient overall quality cannot be reached by the end of the measurement cycle, then the measurement cycle may be terminated (e.g. early).The new measurement cycle may be initiated automatically. Alternatively, a notification indicating that a new measurement cycle is required may be presented to a user. The user (e.g. the subject) may be prompted by a notification to manually restart the measurement cycle. The notification may comprise a prompt to the user to check and / or adjust one or more conditions before initiating the new measurement cycle. Such conditions may include one or more of the list comprising:
[0064] (i) a position of the one or more sensor(s) and / or the wearable apparatus (e.g. device);
[0065] (ii) an orientation of the one or more sensor(s) and / or the wearable apparatus (e.g. device),
[0066] (iii) temperature,
[0067] (iv) motion of the subject,
[0068] (v) proximity to sources of noise or artifacts (e.g. magnetic materials), (vi) other sources of noise or artifacts.
[0069] Noise in biomagnetic measurements may originate from a range of sources, e.g. environmental electromagnetic interference (EMI), motion artifacts, or physiological sources unrelated to biomagnetic signal being measured. The step of enhancing second-class biomagnetic data may comprise denoising. The step of enhancing may include filtering the biomagnetic data - e.g. using a frequency-selective filter. Such filtering may involve, for example, cancellation of 50Hz noise with a notch filter and / or removal of low frequency drift (0-5Hz) with a high pass filter.
[0070] Additionally or alternatively, methods based on independent component analysis or other subspace decomposition methods may be used to enhance the biomagnetic data classified as second class.
[0071] Techniques may be used to ‘latch’ on the signal of interest and remove undesired noise signals.In a set of embodiments, the biomagnetic data is obtained by performing a magnetocardiographic (MCG) measurement on the subject. Therefore, the one or more biomagnetic sensor(s) may be MCG sensor(s). Said MCG sensor(s) may be located around (e.g. positioned at) a chest portion of the subject and / or proximal to the heart of the subject. There may be a small gap of 1cm or less between the sensor(s) and the subject or the sensor(s) may be in contact with an external surface of the subject.
[0072] The step of enhancing may comprise processing (e.g. using a first enhancement machine learning model) the second class biomagnetic data to generate (e.g. determine or estimate):
[0073] (i) data relating to a primary or desired (e.g. MCG) signal of the subject; and (ii) data relating to one or more secondary or undesired sources (e.g. a source of environmental noise, a seismocardiographic (SCG) signal of the subject and / or a ballistocardiographic (BCG) signal of the subject).
[0074] In such a set of embodiments, the enhanced data may be the data relating to the primary or desired (e.g. MCG) signal of the subject. The data relating to the one or more secondary (e.g. undesired) sources may be discarded, stored or sent for further processing.
[0075] A second enhancement machine learning model may be used to classify the data output from the first enhancement machine learning model as data relating to a primary or desired (e.g. MCG) signal of the subject or data relating to one or more secondary or undesired sources. The second enhancement machine learning model may be trained to classify input data as primary or desired (e.g. MCG) signal data or data relating to one or more secondary or undesired sources. The second enhancement machine learning model may be trained using primary or desired (e.g. MCG) signal data and (e.g. signal) data relating to one or more secondary or undesired sources as training data. The second enhancement machine learning model may be based on a temporal convolutional network (TCN) or a linear classifier, e.g. linear discriminant analysis.In a set of embodiments, if the biomagnetic data is classified in the second class, after enhancing said biomagnetic data, the method comprises classifying the enhanced biomagnetic data as being in a first, second or third class; wherein if:
[0076] (i) the enhanced biomagnetic data is in the first class, said enhanced biomagnetic data is made available for use in determining physiological information relating to the subject;
[0077] (ii) the enhanced biomagnetic data is in the second class, the method comprises enhancing said enhanced biomagnetic data;
[0078] (iii) the enhanced biomagnetic data is in the third class, said enhanced biomagnetic data is made unavailable for use in determining physiological information relating to the subject.
[0079] In a set of embodiments, enhancement and classification steps can be repeated (e.g. as needed) up to a maximum number of iterations and / or for up to a maximum time duration. If the maximum number of iterations or maximum time duration is reached, and the enhanced biomagnetic data remains second class, then the enhanced biomagnetic data may be made unavailable for use in determining physiological information relating to the subject if the enhanced biomagnetic data is still classified in the second class. This helps to prevent second class data from being kept in an unduly long and potentially perpetual loop of classification and enhancement. Furthermore, this also helps to ensure that only second class biomagnetic data which can be successfully enhanced in a reasonable time is used in determination of physiological information relating to the subject.
[0080] The maximum number of iterations of enhancement and classification and / or the maximum time duration (e.g. for enhancement and classification of the same biomagnetic data) may be selected so that the biomagnetic data can be received, processed and used in determining physiological information relating to the subject in real time. It will be appreciated by those skilled in the art that each iteration may only take a few milliseconds.The maximum number of iterations may be 10, e.g. 3, e.g. 2. The maximum time duration may be between 200 ms and 1 s - e.g. between 200 ms and 800 ms -e.g. approximately 750 ms. In some embodiments, wherein a window of biomagnetic data is processed at a time, the maximum time duration may be approximately equal to the window length. This may provide energy savings and reduced processing overhead by preventing excessive attempts at enhancing data which cannot be enhanced to become ‘first-class’ biomagnetic data in a reasonable timeframe.
[0081] In a set of embodiments, a wearable apparatus (e.g. worn by the subject) comprises the one or more biomagnetic sensor(s). The system may thus comprise the wearable apparatus. The wearable apparatus may be attachable (e.g. in use attached) to the subject. One or more of the biomagnetic sensor(s) may be attachable (e.g. in use attached) to the subject. The attachment may be provided by one or more securing elements (e.g. straps or other suitable means) to secure the wearable apparatus and / or one or more of the biomagnetic sensor(s) to a body part - e.g. head, neck, arm, chest - of the subject.
[0082] The wearable apparatus may comprise one or more securing elements (e.g. one or more straps) arranged to attach the one or more biomagnetic sensor(s) to the subject. The or each securing element may be a strap. The or each securing element may be an adhesive - e.g. the wearable apparatus may be an adhesive patch.
[0083] The one or more biomagnetic sensor(s) may be provided within one or more sensor units. Each sensor unit may comprise a plurality of layers (in other words, each sensor unit may be arranged in layers); wherein at least one of the layers is a sensor layer.
[0084] The plurality of layers may comprise a motherboard layer and / or a battery layer. The sensor layer may comprise a substrate and one or more biomagnetic sensors. For example, the biomagnetic sensor layer may comprise a single biomagnetic sensor or an array of biomagnetic sensors. The plurality of layersmay be arranged so that in use the sensor layer is the layer most proximal to the subject.
[0085] The, or each, sensor unit may be arranged within a housing. Each sensor unit preferably has a respective housing. However, where there is a plurality of sensor units, the plurality of sensor units may be arranged within a common housing.
[0086] The housing may be a plastic housing. The housing may comprise an opening (e.g. an aperture or window) proximal to the one or more sensors in the, or each, sensor unit. The opening may be in use positioned between the, or each, sensor of each sensor unit and the body of the subject. In a set of embodiments, the system is arranged so that in use the, or each, opening is positioned between the, or each, sensor of each sensor unit and the subject.
[0087] The wearable apparatus may be a wearable device or garment, e.g. a smartwatch, a heart monitor (e.g. in the form of a chest strap or vest), a patch, or the like.
[0088] In a set of embodiments, the method comprises a step of determining physiological information (e.g. a physiological parameter) relating to the subject. Therefore, the system or wearable apparatus (e.g. the processing module) may be configured to determine physiological information (e.g. a physiological parameter) relating to the subject. In preferred embodiments, this determination step only uses data classified as being in the first class (e.g. which may include biomagnetic data which was initially classified as second class but after enhancement was classified as first class).
[0089] The physiological information relating to the subject (e.g. a physiological parameter) may be provided to a user or subject via an audio or visual output. The physiological information relating to the subject may be displayed via a display means (e.g. a screen).The system and / or wearable apparatus may, therefore, comprise a display means (e.g. a screen) for displaying the physiological information relating to the subject (e.g. a physiological parameter).
[0090] In a set of embodiments, the system is a health monitoring system.
[0091] As mentioned above, the wearable apparatus may comprise a processing module configured to carry out the method according to the first aspect. Similarly, the system may comprise a processing module or sub-system configured to perform the method according to the first aspect.
[0092] The processing module or sub-system may comprise a classification module configured to perform the classification step and a data enhancement module configured to perform the enhancement step. The processing module or subsystem may be configured to determine physiological information relating to the subject - e.g. a pulse rate or heart rate may be determined by determining timepoints in a time-series sequence of biomagnetic data wherein the (e.g. first-class) biomagnetic data is indicative of a pulse or heartbeat.
[0093] The wearable apparatus or system may determine the physiological information relating to the subject in the processing module of the wearable apparatus or system. Alternatively, to determine the physiological information, (e.g. first class) biomagnetic data may be transmitted via a wireless or wired connection to an external processing module or system (e.g. server, computer etc.) for determining physiological information relating to the subject.
[0094] The system and / or wearable apparatus (e.g. the processing module or subsystem) may comprise a digital signal processor (DSP), a field-programmable gate array (FPGA), and / or a microcontroller to execute real-time signal processing. For example, instructions for carrying out one or more (e.g. all) steps of the method may be comprised in a binary file uploaded to a memory of a microcontroller which at runtime may call the method as a subroutine or function in a single-thread program or a thread in multi-thread program.Embodiments of any aspect of the present invention may include one or more features of one or more of the other aspects of the present invention or its embodiments. Where reference is made to different embodiments or sets of embodiments, it should be understood that these are not necessarily distinct but may overlap.
[0095] Brief description of the drawings
[0096] Embodiments of the invention will now be described, by way of example, with reference to the drawings, in which:
[0097] Fig. 1 shows a flowchart depicting a method embodying the present invention;
[0098] Fig. 2 shows a block diagram of a system which may carry out the method of Fig. 1;
[0099] Fig. 3 shows an example flowchart of a method embodying the present invention;
[0100] Fig. 4 is a block diagram showing the training process of a machine learning model suitable for use in the method according to the invention;
[0101] Figs. 5A, 5B and 5C depict various wearables which may embody the invention;
[0102] Fig. 6 shows three traces of time-series biomagnetic data of varying quality;
[0103] Fig. 7 is an example output for an embodiment of the invention, when the biomagnetic data is in the first class, including displayed physiological information relating to the subject;
[0104] Fig. 8 shows second class biomagnetic data, before and after enhancement, and the displayed physiological information relating to the subject;
[0105] Fig. 9 is a graph indicating the precision according to input length for a machine learning model used in an embodiment of the invention;
[0106] Fig. 10 is a block diagram of an example of a Temporal Convolutional Network (TCN) suitable for use with the present invention;
[0107] Fig 11 shows performance metrics related to a trained machine learning model used in an embodiment of the present invention;
[0108] Fig. 12 is an exploded view of part of the wearable apparatus of Fig. 5A;Fig. 13 is a schematic diagram showing separation of desired and undesired signals as a form of biomagnetic data enhancement; and
[0109] Fig. 14 is a schematic diagram showing how enhanced data can be obtained using a machine learning model.
[0110] Detailed Description
[0111] The flowchart of Fig. 1 provides an overview of the method according to the present invention.
[0112] The method 100 relates to processing biomagnetic data obtained from a subject (e.g. a human) using one or more biomagnetic sensors, for example, a tunnel magnetoresistance (TMR) device.
[0113] In step 101 the biomagnetic data from the one or more biomagnetic sensor(s) is received. In step 102, the biomagnetic data is classified as being in a first, second or third class.
[0114] Depending on the outcome of step 102, the method can progress to one of three alternative steps 103, 104 or 105.
[0115] If at step 102, the biomagnetic data is classified in the first class, then the method progresses to step 103 and the biomagnetic data is made available for use in determining physiological information relating to the subject.
[0116] If at step 102, the biomagnetic data is classified in the second class, the method progresses to step 104 and the biomagnetic data is enhanced, for example, by passing the data to a denoising algorithm or model.
[0117] If at step 102, if the biomagnetic data is classified in the third class, the method progresses to step 105 and the biomagnetic data is made unavailable for use in determining physiological information relating to the subject, e.g. it can be discarded or stored with a flag indicating that the data should not be used.In some examples, there exists a feedback loop 106 between step 104 and step 102, wherein once the biomagnetic data is classified as second class and subsequently enhanced at step 104, it is then re-classified to see if the enhancement was sufficient to upgrade the biomagnetic data to first class data or if further enhancement is required.
[0118] The determination of physiological information relating to the subject is performed at step 107. This determination step preferably only uses first-class biomagnetic data. Between step 103 and step 107 the system may check that enough first-class biomagnetic data has been made available to determine the physiological information relating to the subject. For example, before the determination step, the system may check that a threshold criterion, e.g. a minimum proportion of first-class data for a given length of time, has been reached.
[0119] Fig. 2 shows a block diagram of a system 200 corresponding to the method of Fig. 1.
[0120] The system 200 may include a wearable apparatus (e.g. device) 200a having a radio antenna 201, a CPU 202, a memory 203, I / O devices and / or peripherals 204 including biomagnetic sensor(s) 205 - e.g. a TMR device - and a user interface (i.e. a touchscreen), a classification module 206, a data enhancement module 207 and a battery 208. In this example, the CPU 202, classification module 206 and data enhancement module 207 are separate. However, in some examples, the CPU 202 may also function as one or both of the classification module 206 and data enhancement module 207. The system 200 may include a processing module comprising the classification module 206, data enhancement module 207, CPU 202 and the memory 203.
[0121] Information obtained from the system 200 can be transmitted through wired or wireless means (e.g. radio antenna 201) to a server (not shown) and displayed via a mobile phone application or a digital health platform for remote monitoring. This may allow for real-time display and analysis of the physiological information relating to the subject (e.g. one or more physiological parameters), enablingremote monitoring of the condition of a subject (e.g. patient). This also allows a clinician to intervene promptly when necessary. For example, action or intervention may be necessary in the event of an arrhythmia, an impending stroke or increased anxiety in humans, or when identifying signs of distress, infection or delivery etc. in non-human animals.
[0122] Fig. 3 shows an example flowchart of a variant of the method embodying the invention, which will be described with reference to the operation of the system 200 and wearable device 200a depicted in Fig. 2.
[0123] A measurement cycle may be initiated for the wearable device 200a in any suitable way - e.g. by the user (i.e. the subject) of the device 200a initiating a measurement cycle, via the touchscreen 204. The user may select a heart rate estimation function via a software application on the device 200a which initiates a magnetocardiography (MCG) measurement.
[0124] During the measurement cycle, the sensor (e.g. TMR device) 301, 205 senses the biomagnetic field arising from the subject. In this example, the sensor 301, 205 is located on the chest of the subject 502 and the biomagnetic field sensed is indicative of cardiac activity. The sensor 301 , 205 provides the biomagnetic data in the form of a raw biomagnetic signal to a classifier 302 (i.e. classification module 206) which is a classification model based on a temporal convolutional network (TCN) architecture.
[0125] The raw biomagnetic signal data may be pre-processed before it is classified. For example, pre-processing may involve using an analog-to-digital converter, ADC, (not shown) to sample and digitise the signal if the biomagnetic signal is in analog form. Alternatively or additionally, if the classifier 302 is a machine learning model trained on biomagnetic data having a certain sampling rate, the biomagnetic data can be re-sampled to said sampling rate.The classification step involves classifying the biomagnetic data input to the classifier 302 into one of three classes - first class (e.g. high quality), second class (e.g. medium quality) or third class (e.g. low quality).
[0126] More specifically, the TCN classifier 302 takes as input a window of time-series biomagnetic data at a time and uses a sliding window mechanism to move over the sequence of biomagnetic data, such that each window consists of an overlapping sub-sequence of the time-series biomagnetic data. In this example, the window has an optimal length of 750 ms. The overlap between consecutive windows in this case is 50% of the window length. The output of the classifier is a labelled version of said window - the label corresponding to first class, second class or third class.
[0127] This means, therefore, that the classifier 302 classifies the biomagnetic data a window at a time during the measurement cycle.
[0128] If the window of biomagnetic data is classified as first class (i.e. high quality), then the window of biomagnetic data is made available for use in determining heart rate of the subject by recording 303 the window of biomagnetic data in the memory 203.
[0129] If the window of biomagnetic data is classified as second class (i.e. medium quality), then the window of biomagnetic data is passed to the data enhancement module 207 for enhancement 304 (e.g. denoising). Once enhanced the window of biomagnetic data may be re-classified to determine whether the enhanced window of biomagnetic data is sufficiently high quality to be classified in the first class. This loop of classification and enhancement may be repeated; however, a limit is placed on the duration and / or number of iterations of classification and enhancement for each window of biomagnetic data. For example, the feedback loop may take up to a maximum duration of approximately 750 ms after which the window of biomagnetic data is treated in the same way as the third-class data (e.g. discarded).If the window of biomagnetic data is classified as third class (i.e. low quality), then the window of biomagnetic data is made unavailable for use in determining heart rate of the subject by discarding or deleting said window of biomagnetic data 305.
[0130] If the number or duration of consecutive windows of biomagnetic data classified in the third class reaches a pre-determined maximum threshold, then a new measurement cycle may be initiated 306. This may be determined by an algorithm run by the CPU 202. The new measurement cycle may be automatically initiated or a notification may be provided to a user (e.g. the subject) to check and / or adjust certain conditions - such as the position or orientation of the wearable device 200a, temperature, motion of the subject, proximity to magnetic materials, or other sources of noise or artifacts - before initiating the new measurement cycle. Often, taking the wearable apparatus off and putting it on again or changing the environment, by moving to a new room, can work to reduce sources of noise.
[0131] If at step 102, the biomagnetic data is classified in the second class, the method progresses to step 104 and the biomagnetic data is enhanced, for example, by passing the data to a denoising algorithm or model (i.e. the data enhancement module 207).
[0132] If at step 102, if the biomagnetic data is classified in the third class, the method progresses to step 105 and the biomagnetic data is made unavailable for use in determining physiological information relating to the subject, e.g. it can be discarded or stored in the memory 203 with a flag indicating that the data should not be used.
[0133] Fig. 4 is a block diagram showing the training process of a machine learning model used in the classification step of the method according to the invention, such as the TCN already described.
[0134] The training method 400 involves supervised learning and may, in some cases, make use of active learning as shown in Fig. 4.Biomagnetic data from a diverse range of sources may be used for training the classification model 405. In the case of active learning, e.g. pool-based sampling, from a set of biomagnetic data, a subset of labelled biomagnetic data 401 is used to train the model 405. The subset of labelled data includes labelled first class biomagnetic data 402, labelled second class biomagnetic data 403 and labelled third class biomagnetic data 404. The labelled data 401, 402, 403, 404 may be in the form of labelled windows of time-series biomagnetic sensor data, wherein the windows all have the same length - e.g. approximately 750 ms.
[0135] The classification model 405 is first trained on a labelled subset 401 of the biomagnetic data which yields class-membership probabilities for instances of the biomagnetic data. Based on this, the model 405 issues a query 406 for new data. The query 406 causes new windows of time-series biomagnetic sensor data about which the system is least confident to be selected from a pool of unlabelled biomagnetic data 407 and passed to an Oracle 408 - e.g. a system operator, researcher, clinician - for labelling. The newly labelled windows of time-series biomagnetic sensor data are then included in the labelled biomagnetic data (i.e. the training data) 401 for model retraining.
[0136] The biomagnetic sensing is preferably performed using a wearable apparatus.
[0137] Figs. 5A, 5B and 5C depict various wearable devices 501, 505, 508 according to different examples of the invention.
[0138] Fig. 5A shows a heart monitor 501 comprising a sensor unit 503 and a strap 504 worn over a thin t-shirt on the chest 502 of a human subject.
[0139] Fig. 5B shows a smart watch 505 worn on the wrist 506 of a human subject. Fig.
[0140] 5C shows a collar monitor 508 worn on the neck 507 of a dog subject. Each of the smart watch 505 and the collar monitor 508 comprise a sensor unit attached to the subject via a strap.Turning to Fig. 6, an example from each class of biomagnetic data is provided. The uppermost trace 602 shows a high-quality heart signal which would be labelled as first-class data. The middle trace 604 shows a medium quality heart signal which would be labelled as second-class data. The lowermost trace 606 shows a low-quality heart signal which would be labelled as third-class data. When unlabelled these traces may be sources of input data for the classification step. When labelled, these traces may be sources of training data for the machine learning model.
[0141] Figs. 7 and 8 each show results of the method according to an embodiment of the invention where physiological information 706 relating to the subject includes heart rate and heart rate variability. As can be seen in Fig. 7, a clear cardiac trace 702 is obtained by a TMR sensor. Thus, the biomagnetic data measured by the sensor is classified in the first-class. This means that physiological information 706 can be readily determined without the need for enhancement. The physiological information 706 is determined by detecting the time points (e.g. 704) - indicated in Fig. 7 as vertical dashed lines - at which the pulses / beats occur. These time points are then used in determining heart rate and heart rate variability.
[0142] Fig. 8 shows an example where the biomagnetic data measured by the sensor(s) is in the second class, necessitating enhancement before the physiological information 806 is determined. The uppermost trace 808 of Fig. 8 is a medium quality signal trace 808 which is enhanced to provide a high-quality signal trace 810. In this example, the method of enhancement used was bandpass filtering, however, a similar result would be achieved using two or more iterations of single-channel independent component analysis, ICA. The time points (e.g. 804) - indicated in Fig. 8 as vertical dashed lines - at which the pulses / beats occur are determined to calculate the physiological information 806 relating to the subject - i.e. the heart rate and heart rate variability.
[0143] Fig. 9 is a graph obtained from 5-fold cross validation indicating the precision according to input length for a machine learning model 302, 405 used in anembodiment of the invention to classify the data as first, second or third class. The graph shows the precision of the classification model as a function of the length - in seconds - of the window of biomagnetic time-series data. The mean and standard deviation is also provided. As will be appreciated by those skilled in the art 5-fold cross validation involves randomly taking 80% of the data to train the model and using the last 20% to test the model with the unseen data. This is repeated 5 times to ensure that the results are not biased by the choice of the 80%. As can be seen in the graph, the best performance of the model is found between 500 ms and 4 s - with a peak at approximately 750 ms.
[0144] Fig. 10 is a block diagram of an example of a Temporal Convolutional Network (TCN) suitable for use with the present invention.
[0145] The TCN includes three consecutive feature extraction blocks 1012 (i.e. this is indicated by ‘x3’) and a classification section 1013. Each of the three feature extraction blocks 1012 includes: a first convolution layer 1001; a first chomp layer 1002; a first activation layer 1003; a first dropout layer 1004; a second convolutional layer 1005; a second chomp layer 1006; a second activation layer 1007; a second dropout layer 1008 and a batch normalisation layer 1009.
[0146] The input is time-series sensor data from the biomagnetic sensor(s) and the label is one of first, second and third class. The first causal convolution layer 1001 applies a convolution operation to the input 1012. The chomp layer 1002 removes padded elements from the output of the convolution layer 1001. Then the activation layer 1003 applies a non-linear activation function (e.g. ReLLI) to the output of the chomp layer 1002. The dropout layer 1004 is there to help prevent overfitting. This layer structure is repeated 1005 - 1008 before the batch normalisation layer 1009 which reduces internal covariate shift. Between the batch normalisation layer 1009 of the final block 1012 and the classification section 1013 there is a residual connection 1014 connecting the input to the output of the final block 1012.The layer sizes of the blocks are 16, 8 and 4 respectively. The kernel size is 17. The dilation grows, taking the layer number as the power of n, which is the block number. The dropout is set to 0.5 to avoid over-fitting.
[0147] The classification section 1013 consists of a linear layer 1010 followed by a softmax layer 1011. The output from the temporal convolutional feature extractor 1001- 1009, enhanced by dilated and causal convolutions, effectively concentrates most of the relevant information into the most recent time step. To mitigate overfitting due to the limited dataset size, only the final unit from each convolution channel is passed to the linear layer 1010. This linear layer 1010 maps the extracted features into three units, each corresponding to one of the three target classes - first class (i.e. high quality), second class (i.e. medium quality), and third class (i.e. low quality). Finally, the softmax layer 1011 is applied to convert these outputs into a probability distribution.
[0148] The TCN’s core components are dilated and causal convolutions. Causal convolutions link the output at time step t with prior layer elements before t, emphasizing time-related input data. Unlike traditional methods that rely on many layers and large kernels for broader reception fields, a TON adopts dilated convolutions. These convolutions use expanding kernel sizes in each layer, effectively increasing the reception field. This enhances feature recognition and performance without requiring a bigger network, preventing issues like overfitting, vanishing gradients, and excessive computation.
[0149] Fig. 11 demonstrates the performance of a classification model 405, 302 according to the present invention. The model is shown to perform well with a precision (i.e. the fraction of relevant instances among the retrieved instances) of approximately 0.85, a recall (i.e. the fraction of relevant instances that were retrieved) of approximately 0.83 and an F1 score (i.e. the harmonic mean of precision and recall) of approximately 0.83. These values indicate that the classification model performs well when a TCN architecture is used.Turning to Fig. 12, an exploded view of part of the wearable apparatus of Fig. 5A is shown. In this example, a single sensor unit 2 is shown. However, it is to be appreciated that the device may comprise further sensor units 2. The sensor unit 2 has a layered internal structure. The sensor unit 2 includes a sensor layer 4a, a motherboard layer 32 and lithium-ion battery layer 30 arranged in that order. The sensor layer 4a may be connected to the motherboard by pin connectors 34. The sensor layer 4a may comprise one or more TMR sensor(s) 4- e.g. an array of TMR sensors.
[0150] Each sensor unit 2 is contained in a plastic housing. The housing is substantially cuboidal and has a lid portion 38a and a base portion 38b, which may be connectable together via a friction fit or may click into place. The base portion 38b of the housing is arranged to be placed (e.g. pressed) against the body of the subject. The base portion 38b has a window or aperture 36 proximal to the sensor layer 4a to help allow biomagnetic signals to be sensed by the sensor layer 4a.
[0151] The base portion 38b has a contact surface (not shown) for contacting the subject 502. The window or aperture 36 may be an opening in the contact surface of the base portion 38b. Extending away from the contact surface and toward the lid portion 38a are one or more side walls. Each side wall of the base portion 38b has an attachment portion 40 for attaching a strap thereto. The attachment portions 40 may be used to attach the sensor unit 2 to one or more straps (shown in Fig. 5A) so that the sensor unit 2 can be worn by the subject. The attachment portions 40 may additionally or alternatively be used to attach the sensor unit 2 to one or more further sensor units 2 to build a multi-channel biomagnetic sensor wearable apparatus.
[0152] The data enhancement module may process second class biomagnetic data in any suitable way. In one example, frequency selective filtering is used.
[0153] In another example, depicted in Figs. 13-14, data relating to a desired or primary signal is extracted from a mixture of desired and undesired signals or primary andsecondary signals. In Figs, 13 and 14 the biomagnetic measurement is an MCG measurement and the wearable apparatus 501 is positioned on the chest of the subject 502 (as shown in Fig. 5A). The sensed biomagnetic data may also be influenced by secondary sources such as noise or undesired biomagnetic signals - e.g. environmental noise and / or by other biomagnetic signals such as SCG and BCG signals - sensed by the biomagnetic sensor(s). If the influence of these second sources is too strong, then the biomagnetic data may be classified as second or third class. In this case, the second class biomagnetic data may be passed to the data enhancement module for the signal of interest, in this case the MCG signal, to be ‘extracted’ to obtain a clear desired signal which may be considered the ‘enhanced’ biomagnetic data.
[0154] As shown in Fig. 13, a second-class biomagnetic data window from an MCG measurement cycle is provided from the classification module 206 via output 3 to the data enhancement module 207. The second-class biomagnetic data is a mix of the desired signal, e.g. the MCG or cardiac signal, and undesired or secondary signals from other sources - e.g. SCG and BCG signals.
[0155] The data enhancement module 207 processes the input data to unmix the second-class biomagnetic data window into its constituent sources - i.e. the signal of interest - i.e. data related to approximated MCG signals 7 (i.e. the enhanced biomagnetic data), data related to a signal from a first secondary (e.g. undesired) source 9 and data related to a signal from a further secondary (e.g. undesired) source 11. The enhanced biomagnetic data may thereafter be reclassified by the classification module 206.
[0156] Fig. 14 is a schematic diagram showing how data relating to these signals can be obtained using a trained machine learning model (i.e. a trained enhancement model) 48. As can be seen in Fig. 13, the classification module 206 takes data from the biomagnetic sensor(s) 4 and provides second class data, via output 3, to the data enhancement module 207. The data enhancement module 207 may use a trained machine learning model 48, to approximate the data relating to a primary or desired (e.g. MCG) signal 7 of the subject; and data relating to one ormore secondary or undesired sources 9, 11 (e.g. a source of environmental noise, SCG and / or BCG).
[0157] The enhancement model 48 may be continuously trained by a training module 17 which uses sensor data from biomagnetic sensor(s) 4 as training data 46. How the signals are separated into primary and secondary data will be explained in more detail below.
[0158] The trained enhancement model 48 may be trained using machine learning techniques for constrained source separation. In this example, a machine learning technique of ‘independent component analysis’, known as FastICA, is used. Independent component analysis is a method of recovering a version of original source signals by multiplying data from mixed statistically independent signals by an unmixing matrix. FastICA is a particularly efficient version which is preferable for use in low power applications.
[0159] A plurality of sources (i.e. in this case desired primary signals and undesired secondary signals) may be mixed in one or more different ways. Source separation software (e.g. FastICA) is used to separate the one or more signals into scaled and permuted versions of the original source signals.
[0160] FastICA uses a measure called negative entropy (i.e. negentropy) to obtain scaled and permuted versions of the original source signals. This measure of negative entropy may be known as Jmain. Typically, FastICA minimises Jmain to achieve separation of source signals.
[0161] In this case, different to blind source separation, some prior knowledge is incorporated into the separation process to extract only the signals of interest. For this, auxiliary cost functions are used. This approach has been applied to other types of signals and may be termed constrained ICA, or semi-blind source extraction.A set of auxiliary measures { JauXi, JauX2Jaux^^} are employed, each measure being respectively tuned to each one of the target sources (e.g. MCG, SCG and BCG signals).
[0162] A set of corresponding compound cost functions are also constructed - e.g. Ji = Jmain + a JauX1for extracting the first target source, J2= Jmain+ a JauX2for extracting the second target source and so on (wherein a is a constant). The auxiliary measures provide the constraints relevant to each target source and are determined based on the spatial or temporal characteristics (i.e. signatures) of each target source. For example, the spatial characteristics of a target source may be based on the spatial distribution of the target source with respect to the body of the subject; while temporal characteristics may be based on particular temporal features of signal traces of the target sources.
[0163] It will be appreciated by those skilled in the art that the invention has been illustrated by describing one or more specific embodiments thereof, but is not limited to these embodiments; many variations and modifications are possible, within the scope of the accompanying claims.
Claims
Claims1. A method for processing biomagnetic data obtained from a subject using one or more biomagnetic sensor(s), the method comprising:receiving the biomagnetic data from the one or more biomagnetic sensor(s); andclassifying the biomagnetic data as being in a first, second or third class; wherein if:(i) the biomagnetic data is in the first class, the method comprises making said biomagnetic data available for use in determining physiological information relating to the subject;(ii) the biomagnetic data is in the second class, the method comprises enhancing said biomagnetic data;(iii) the biomagnetic data is in the third class, the method comprises making said biomagnetic data unavailable for use in determining physiological information relating to the subject.
2. The method of any preceding claim, wherein the step of classifying the input data as being in a first, second, or third class uses a machine learning model.
3. The method of claim 1 or 2, wherein the machine learning model has been trained on training data, wherein the training data comprises:a. high quality biomagnetic data labelled as being in the first class; b. low quality biomagnetic data labelled as being in the third class; andc. biomagnetic data, which is not labelled as being in the first class or the third class, labelled as being in the second class.
4. The method of claim 2 or 3, wherein the machine learning model has been trained on training data, wherein the training data comprises:a. biomagnetic data, having a signal to noise ratio, SNR, of at least 10 db and / or a noise level below approximately 25 PTRMS, labelled as being in the first class;b. biomagnetic data, having an SNR of approximately up to 1db and / or having a noise level of at least 25 PTRMS, labelled as being in the third class; andc. biomagnetic data, which is not labelled as being in the first class or the third class, labelled as being in the second class.
5. The method of any one of claims 2 to 4 wherein the machine learning model is based on a Temporal Convolutional Network, TCN.
6. The method of any preceding claim, wherein the step of enhancing second-class biomagnetic data comprises denoising.
7. The method of any preceding claim, wherein the step of enhancing may comprise processing the second class biomagnetic data to generate:(i) data relating to a primary or desired signal of the subject; and(ii) data relating to one or more secondary or undesired sources.
8. The method of any preceding claim, wherein the, or each, biomagnetic sensor is a tunnel magnetoresistance, TMR, sensor.
9. The method of any preceding claim, wherein the physiological information relating to the subject is a physiological parameter of the subject.
10. The method of claim 9, wherein the physiological parameter of the subject is one or more of the list comprising: pulse rate, heart rate, and heart rate variability.
11. The method of any preceding claim, wherein the biomagnetic data is input to the classifying step in the form of a window of time-series data.
12. The method of claim 11, wherein each window of time-series biomagnetic data, from a plurality of consecutive and overlapping windows of time-series biomagnetic data, is input to the classifying step one at a time.
13. The method of claim 11 or 12, wherein each window has a length in the range of 200 ms to 4 s.
14. The method of any preceding claim, wherein each window has a length in the range of 500 ms and 1 s.
15. The method of any preceding claim, wherein the biomagnetic data is collected in one or more measurement cycles, each complete measurement cycle having a duration of at least 3 seconds.
16. The method of claim 15, wherein each complete measurement cycle has a duration of at least 30 seconds.
17. The method of claim 15 or 16, wherein if during a measurement cycle there is not sufficient biomagnetic data classified as being in the first class then the measurement cycle is terminated and a new measurement cycle is initiated.
18. The method of any preceding claim, wherein making biomagnetic data unavailable for use in determining physiological information relating to the subject comprises storing the third-class biomagnetic data with a flag indicating that the third-class biomagnetic data is unusable and / or discarding the third-class biomagnetic data.
19. The method of any preceding claim, wherein the biomagnetic data is resampled to a nominal or pre-determined sampling rate before the classification step.
20. The method of any preceding claim, wherein a wearable apparatus, worn by the subject, comprises the one or more biomagnetic sensor(s).
21. The method of any preceding claim, wherein the biomagnetic data is obtained by performing a magnetocardiographic (MCG) measurement on the subject.
22. A computer program comprising instructions which, when the program is executed by a processing system, cause the processing system to carry out the method of any preceding claim.
23. A non-transitory computer-readable medium comprising instructions that, when executed by a processing system, cause the processing system to perform the method of any preceding claim.
24. A system for processing biomagnetic data obtained from a subject using one or more biomagnetic sensor(s), the system comprising:the one or more biomagnetic sensor(s);wherein the system is configured to:classify the biomagnetic data as being in a first, second or third class;wherein if:(i) the biomagnetic data is in the first class, the system makes said biomagnetic data available for use in determining physiological information relating to the subject;(ii) the biomagnetic data is in the second class, the system enhances said biomagnetic data;(iii) the biomagnetic data is in the third class, the system makes said biomagnetic data unavailable for use in determining physiological information relating to the subject.
25. A wearable apparatus for processing biomagnetic data obtained from a subject using one or more biomagnetic sensor(s), the wearable apparatus comprising:the one or more biomagnetic sensor(s);wherein the wearable apparatus comprises a processing module configured to:classify the biomagnetic data as being in a first, second or third class;wherein if:(i) the biomagnetic data is in the first class, the processing module makes said biomagnetic data available for use in determining physiological information relating to the subject;(ii) the biomagnetic data is in the second class, the processing module enhances said biomagnetic data;(iii) the biomagnetic data is in the third class, the processing module makes said biomagnetic data unavailable for use in determining physiological information relating to the subject.
