Monitoring system

Abstract

A heart sensor device (10) comprises a sensor arrangement (14a, 14b) operatively connected to a controller (20), the controller (20) configured to obtain electro cardiac signals obtained from the sensor arrangement (14a, 14b), and further configured to obtain a sequence of electro cardiac signals comprising at least one heartbeat cycle; and to use a classification module to carry out an analysis of the electro cardiac signals of the at least one heartbeat cycle, to assign a classification of the signal, and to provide an output indicative of the classification; wherein the controller (20) and the classification module comprise a pre-set configuration allowing the analysis to be carried out without external data communication from the heart sensor device. The heart sensor device (10) can tolerate the presence of pacemaker signals and can be used during defibrillation.

Description

[0001] Monitoring system

[0002] Field of the Invention

[0003] The present invention relates to an assessment method and system for measuring electrocardiographic data of a user. More specifically, the present invention relates to a wearable device comprising a sensor arrangement and data evaluation module that are operable autonomously to provide an output of heart failure risk.

[0004] Despite recent advances in heart measurements and analysis of electrocardiographic (ECG) data, many tests require a patient to attend a clinical setting under supervision by clinical staff.

[0005] Recently, machine learning has been proposed as means for analysing heart rate and arrhythmic heart beat behaviour in an endeavour to discover early signs of congestive heart failure.

[0006] The present invention seeks to provide additional options for monitoring heart performance.

[0007] Summary of the Invention

[0008] In accordance with a first aspect of the invention, there is provided a heart sensor device as defined in claim 1 , comprising a sensor arrangement operatively connected to a controller, the controller configured to obtain electro cardiac signals obtained from the sensor arrangement, and further configured to obtain a sequence of electro cardiac signals comprising at least one heartbeat cycle; and to use a classification module to carry out an analysis of the electro cardiac signals of the at least one heartbeat cycle, to assign a classification to the signal; and to provide an output indicative of the classification; wherein the controller and the classification module comprise a pre-set configuration allowing the analysis to be carried out without external data communication from the heart sensor device.

[0009] The device may be a wearable, e.g. a device intended to be worn by a user for prolonged periods of time, if desired. The sensor arrangements will typically be electrodes, specifically passive electrodes, suitable for obtaining electrocardiographic measurements from a person. In some embodiments, the electrodes are gel electrodes. In some embodiments, the electrodes are dry electrodes. While dry electrodes may be susceptible to detachment and generate noisier data, they are believed to be more suitable for being worn for prolonged periods of time. For instance, the electrodes may be skin-conformal dry electrodes, which are electrodes designed to be worn for prolonged periods of time in contact with skin.

[0010] Version 2026-01-16Suitable electrodes for an ECG sensor system will be known to a skilled person and will not be described in detail. The electrodes may be disposable electrodes to be connected to the device for the duration of a measurement or to remain on place for a period of time that the user considers comfortable, e.g. for a day.

[0011] The controller is understood to comprise a processor and memory provided with instructions that can be executed by the processor for the processing of data obtained by the sensor arrangement.

[0012] Data obtained via the sensor arrangement are stored in a memory of the device. The device is configured to obtain ECG data, which is understood to be representable in waveform. An ECG waveform is understood to depict an ECG cycle of a heartbeat, herein referenced as a heartbeat cycle. A waveform is sometimes referenced as “peak”, whereas it is understood that a heartbeat cycle, or ECG waveform unit, comprises a group of peaks including a P wave, QRS complex, and T wave, wherein the R peak is usually visualised as characteristic peak of a beat. A heartbeat cycle may describe the interval between two R peaks. Alternatively, a heartbeat cycle may define an interval covering a successive P wave, QRS complex, and T wave. For a measurement, the device is configured to obtain ECG data for a period of time, continuously, sufficient to obtain at least one heartbeat cycle, e.g. of a group comprising a P wave, an immediately following QRS complex, and an immediately following T wave. For instance, the device may be configured to obtain ECG data comprising at least two, or more, heartbeat cycles. E.g. for a heart rate of about 60 bpm, a data set of ten seconds length is expected to contain about ten heartbeat units, and would therefore be expected to be comprise at least one (here: ten) of such heartbeat cycles.

[0013] In accordance with the first aspect, waveform data are analysed in a pre-configured classification module to determine whether a waveform belongs to a class of signal, and to provide an output indicative of the class of signal to which the waveform belongs.

[0014] The classification may be indicative of a heart condition, such as a risk of congestive heart failure, arrhythmic fibrillation, or others.

[0015] The classification module is pre-configured, with a pre-set configuration, such that the device is able to provide a classification output autonomously, without external data communication. By autonomously, it is meant that the device does not require access to external databases, such as via a network or cloud services, at the time of carrying out the analysis or classification, in order to analyse the waveform. For practical purposes, the autonomous classification may be implemented by an offline data processing, or so-called “airgapped” processing, without reliance on a connection to an external network during the analysis. Rather, a classification is performed within, or on, the device. For instance, the classification output may be provided in the form of a score calculated within the device.

[0016] Version 2026-01-16The ability to carry out a classification autonomously provides an improved level of data security, because the data can be analysed without having to transmit data externally.

[0017] In some embodiments, the classification module comprises a machine learning algorithm configured with, or using, a machine learning model.

[0018] As will be appreciated, a machine learning model may be one that has already been trained, to thereby provide a pre-configured classification module.

[0019] In some embodiments, at least one heartbeat cycle is processed by the classification module without modifying the pre-set configuration.

[0020] By “pre-set configuration”, the device configuration is meant that allows the device to carry out classification without external communication, i.e., offline, or autonomously. For instance, the pre-set configuration may comprise a classification model that is stored on the device as basis for classification. In that case, the device may operate such that data acquired during measurements is not used to further train the classification model while the device is offline.

[0021] In some embodiments, the sequence of electro cardiac signals is at least 1 , at least 2, at least 3, at least 4, at least 5, at least 6, at least 7, at least 8, at least 9, or at least 10 seconds long.

[0022] In some embodiments, the classification module is configured to accept into the analysis a sequence of electro cardiac signals that is no more than 20, 17, 15, 13, 11 , 10, or 9 seconds long.

[0023] The duration of time may be selected such that at least one heartbeat cycle can be obtained. The duration may be a fixed period of time. Without wishing to be bound by theory, it is believed that a consistent duration may improve compliance with measurement protocols by making it easier for a user to adhere to measurement protocols. A duration of around ten seconds is believed to provide a good balance between a measurement that is relatively short and thereby places lower demand on resources, and a period of time that is sufficiently long to capture usually more than one heartbeat cycle in most circumstances.

[0024] Depending on measurement conditions, a sequence may be several seconds long, sometimes in the region of a minute. However, in some measurement conditions, it can be difficult to ensure a continuous uninterrupted measurement without irregular data for more than a minute. An observation underlying the development of the invention was that good classification results can be obtained with sequences having a length in the region of 3 to 10 seconds length, or between 4 and 9 seconds length, e.g. 5 seconds or 7 seconds length, which is significantly shorter than measurements taking several minutes. As will be set out below, a classification based on relatively short duration measurements is more resilient against to irregular measurements that may be obtained from time to time.

[0025] Version 2026-01-16In some embodiments, the pre-set configuration comprises an algorithm to derive as an input for the classification module two or more descriptors obtained from analysing the at least one heartbeat cycle.

[0026] Herein, a descriptor of a heartbeat cycle, i.e. of a waveform within the heartbeat cycle, is understood to be an indicator describing one or more properties or features of the waveform within the heartbeat cycle. E.g., a descriptor may be a peak height, peak area, peak width at half of its maximum height, and others.

[0027] An appreciation underlying the embodiment was that the execution of a classification algorithm on an autonomous device may be improved, and may potentially be more robust, if descriptors are extracted from the waveform. In some embodiments, only a small number of descriptors may be used for analysis. Without wishing to be bound by theory, it is believed that analysis based on descriptors may be less susceptible to waveform noise than analysis of raw data.

[0028] In some embodiments, the two or more descriptors comprise at least two of: peak to average power ratio, Hjorth activity, Hjorth mobility, Hjorth complexity, skewness, and kurtosis.

[0029] As such, a descriptor may be considered a statistical parameter providing a representation of one or more morphological characteristics, and / or of a variation in morphology, of a series of ECG waveforms. For instance, in some embodiments, the classification model may use as few as two descriptors, e.g. Hjorth mobility and peak average power ratio. A surprising finding underlying the development of the present invention was that such descriptors may enable a reliable classification even if the sample size is relatively small, e.g. amounting to a few heartbeat units from measurement of a few seconds duration.

[0030] It will be appreciated that the number and selection of descriptors may depend on the classification model and on the number of groups distinguishable by the classification model.

[0031] In some embodiments, the device is configured to remove noise from the sequence of electro cardiac signals.

[0032] Noise may be removed by use of a bandpass filter, e.g. a 5th order Butterworth filter, or other suitable filter. Bandpass filtering allows movement noise and breathing noise to be removed. Furthermore, bandpass filtering was found to reduce and practically remove baseline “wandering”, to provide an even baseline.

[0033] In some embodiments, the device is configured to remove irregular data.

[0034] Version 2026-01-16The expression “irregular data” is understood to refer to data that is inaccurate despite noise removal and baseline straightening, and as such refers to data that is likely to fail classification and cannot be filtered by noise removal. Irregular data may be characterised by unexpected waveform shapes, and / or by waveform shapes indicating an erroneous measurement.

[0035] An example of irregular data may be a double-peak obtained instead of an expected single peak, due to an acquisition error or user movement. Irregular data may be detected by suitable means, such as applying cut-off values to statistical parameters, such as a peak amplitude-to-area ratio, or wave-to-wave interval. In some embodiments, a machine learning model may be used to classify irregular data. In some embodiments, which may or may not use machine learning models, the device may be pre-configured with reference data defining a range of expected data and optionally defining a range of known outlier data. In that case, data may be excluded as “irregular data” if it is outside the reference data, e.g. because it is outside the range of expected data and / or if it is outside the range of known outlier data. Irregular data are believed to be more likely when the device is used without supervision, and / or in a wider range of activities, and / or with dry electrodes. A configuration providing a capability to ignore irregular data, or to recognise data that is unlikely to contribute to a reliable analysis, is therefore believed to increase the robustness of the device for a wider range of use scenarios.

[0036] In some embodiments, the device is configured to exclude data comprising pacemaker signals.

[0037] An appreciation underlying the development of some embodiments was that pacemaker signals may be detected as regular, repeating signals, allowing pacemaker signals to be removed, and / or to be treated as irregular data, respectively.

[0038] In some embodiments, the device comprises a power source to provide power to the sensor arrangement and / or the controller.

[0039] As will be appreciated, the power source may be a battery which may be chargeable, or other chargeable power source. In some embodiments, the power source is a disposable battery.

[0040] In some embodiments, the device comprises, further, a communication interface operatively connected with the controller, and configured to provide the output upon a request signal received via the communication interface.

[0041] For instance, a communication interface may allow a user to use an external device, such as a mobile device, to transfer data from or to the device.

[0042] In some embodiments, the device is configured to obtain an electro cardiac signal upon receipt of a command signal received via the communication interface.

[0043] Version 2026-01-16In some embodiments, the classification of the signal determined by the classification module includes at least a classification indicative of congestive heart failure (CHF) and a classification indicative of normal sinus rhythm (NSR).

[0044] In some embodiments, the classification of the signal determined by the classification module includes one or more intermediary classes each indicating a different risk level between congestive heart failure (CHF) and normal sinus rhythm (NSR).

[0045] In some embodiments, the classification module is configured to assign a classification indicative of one or more conditions selected from atrial fibrillation, atrial tachycardia, arrhythmia, coronary artery disease, myocardial infarction, hyperkalaemia, hypokalaemia, hypercalcaemia, and hypocalcaemia.

[0046] In some embodiments, the heart sensor device forms part of a defibrillator system.

[0047] The device may be incorporated with a defibrillator. The device may be provided as a kit of parts comprising any one of the embodiments of the first aspect and a defibrillator.

[0048] In some embodiments, the heart sensor device is comprised in a wearable device and / or garment.

[0049] Embodiments of the first aspect may be incorporated with, or provided as, a wearable device or garment. This may facilitate the use of the device during exercise, treatment, outdoor pursuits, and / or other activities.

[0050] In accordance with another aspect of the invention, there is provided a defibrillator system comprising a heart sensor device according to any one of the embodiments of the first aspect.

[0051] In accordance with another aspect of the invention, there is provided a classification model trained to accept, as input for classification, descriptors selected from a sequence of two or more ECG waveforms, and to present, as an output, at least two different classification groups.

[0052] In some embodiments, the classification model is configured to determine whether the input data belongs to one of two or more classification groups.

[0053] In accordance with a further aspect of the invention, there is provided a method of training a classification model for use with a heart sensor device according to any one of the embodiments of the first aspect, the method comprising: providing ECG training data labelled with a reference classification; selecting, from the ECG training data, one or more ECG waveform segments of predetermined length, each of the segments representing a sequence of one or more heartbeat cycles; using the one or more ECG waveform segments as input for training a classification model, and providing, as an output, a classification model.

[0054] Version 2026-01-16In some embodiments, the method comprises removing artefacts from the one or more ECG waveform segments prior to using the one or more ECG waveform segments as an input for classification.

[0055] In some embodiments, the artefacts comprise pacemaker signals.

[0056] In some embodiments, the method comprises detecting irregular waveforms prior to using the one or more ECG waveform segments as an input for classification.

[0057] In some embodiments, the method comprises including irregular waveforms as an input for classification.

[0058] In that case, examples of known irregular waveforms by be used to enable a trained model to identify waveforms to be excluded from analysis.

[0059] In some embodiments, the method comprises extracting features for use as input for classification, the features comprising statistical descriptors of a series of waveform segments.

[0060] Such descriptors may comprise statistical parameters representing waveform morphology, and / or variation in morphology, of a series of ECG waveforms.

[0061] In some embodiments, the descriptors comprise at least two of: pe a k-to-a verage power ratio, Hjorth activity, Hjorth mobility, Hjorth complexity, skewness, and kurtosis.

[0062] In some embodiments, the method comprises removing artefacts from the one or more ECG waveform segments prior extracting features.

[0063] In some embodiments, the data representing a sequence of one or more heartbeat cycles are provided by descriptors defined by extracted features.

[0064] In some embodiments, providing ECG training data comprises obtaining ECG data from a sensor measurement.

[0065] In some embodiments, providing ECG training data comprises using ECG data from a database.

[0066] The database may comprise model data sets, rather than measured data. Likewise, the database may comprise data with erroneous or uncommon waveforms.

[0067] Features described in relation to any one or more of the embodiments of the first aspect may be combined with features of any one or more of the further aspects. One or more embodiments may be implemented in the form of software instructions. The software may be incorporated in a device

[0068] Version 2026-01-16according to any one or more of the embodiments of the first aspect. The device may comprise a processor and software instructions implemented by the processor to carry out the functionality of the other aspects.

[0069] Description of the Figures

[0070] Exemplary embodiments of the invention will now be described with reference to the Figures, in which:

[0071] Figure 1 is a schematic illustration of components of a wearable ECG sensor device;

[0072] Figure 2 is a schematic illustration of an interface;

[0073] Figures 3A and 3B show example waveforms and an example waveform series;

[0074] Figure 4A shows an example waveform exhibiting baseline wandering;

[0075] Figure 4B shows an example waveform exhibiting powerline interference;

[0076] Figure 5A shows an example waveform exhibiting muscle noise;

[0077] Figure 5B shows an example waveform showing electrode motion;

[0078] Figure 6 shows a sequence of exemplary steps of a classification method;

[0079] Figure 7 shows a sequence of exemplary steps of a training method,

[0080] Figures 8 and 9 are graphs showing classification performance of different-length segments, and Figures 10 and 11 are graphs illustrating performance indicators for different-length segments.

[0081] Referring to Figure 1 , a sensor device 10, here constituting a heart sensor device, comprises a main body 12 in the form of a housing comprising a connector arrangement 18 for connection of an arrangement of (here: two) electrodes 14a, 14b, such as passive ECG sensor pads, connected via connector lines 16a, 16b to the connector arrangement 18. The connector lines 16a, 16b and / or the connector arrangement 18 may be constituted by, or may comprise, connections such as sockets for detachable electrodes and corresponding connectors. The electrodes 14a, 14b and the connector lines 16a, 16b may be provided as an integrally formed component, for instance in the form of disposable sensors, or in the form of reusable sensors. The electrodes 14a, 14b may be dry sensors that can be worn by a user for several hours, although this is not necessarily a requirement of all embodiments. As such, the device 10 may provide a reuseable device that can be repeatedly used for prolonged periods of time, while the sensor electrodes may be exchanged more frequently, e.g. after single use or after a predetermined duration of use, or after a suitable number of uses.

[0082] The device 10 comprises a controller 20 operatively connected to, or constituted by, a processor 22 and memory 24. The device 10 comprises, further, a power source 26 such as a battery or chargeable power source, which may be chargeable via a charge port 28. The device 10 further comprises a communication interface 30. The communication interface 30 may be wireless, such as a Bluetooth (RTM) or NFC interface, and / or may comprise a wired connection via a communication

[0083] Version 2026-01-16port 32. Although illustrated as two separate ports 28, 32, a single port may be provided as charging port and as a communications port. The device comprises, further, a display 34 in the form of a module. The display 34 may be provided in the form of a screen, and / or in the form of indicator elements such as LEDs. For the purpose of the present disclosure, a skilled person is aware of a wide range of connector types, communication protocols and power supply protocols, including wired and wireless interfaces, which may be combined or may be used separately, and which are not described herein other than to provide context for the invention.

[0084] Referring to Figure 2, the communication interface 30 of the device 10 may permit communication with a user device 36, such as a handheld computer, or mobile phone, or bespoke reader, provided with an interface 38, here in the form of a touch screen interface, configured to allow a user to provide control commands to control the device 10. The interface 38 may display an output, such as a classification of a signal according to one of a plurality of classes, e.g. (as shown in Figure 2) one of five risk levels. The user device 36 is not necessarily used by the wearer of the electrodes. For instance, the user device 36 may be connected to a monitoring station of a care provider, to a relative, or other person. Likewise, the user device 36 may be used by a user to regularly retrieve data from the device 10 and / or provide software updates to the device 10. While Figure 2 illustrates a user device 36, the system may be set up such that user data can be accessed via an authorised account, such that the device 10 may be linked with one or more of several different user devices 36. Conversely, the functionality of the user device 36 may be provided by a station that may be connected to receive data from a plurality of different devices 10.

[0085] The illustration of five risk levels (here: “0” to “4”) in Figure 2 is understood to be exemplary. The device 10 may assign classifications that distinguish between different heart conditions and may apply any number of levels for each of the different heart conditions. Likewise, instead of, or in addition to, a risk level, the device may provide an output indicative of the number of occurrences of a risk level. It is envisaged that the device 10 may be programmed to differentiate between at least two or more conditions. For example, the classification model may be trained to classify ECG data according to the NYHA (New York Heart Association) Classification system, as follows:

[0086] Class No symptoms and no limitation in ordinary physical activity, e.g. shortness of breath when walking, climbing stairs etc.

[0087] Class II Mild symptoms (mild shortness of breath and / or angina) and slight limitation during ordinary activity.

[0088] Class III Marked limitation in activity due to symptoms, even during less-than-ordinary activity, e.g. walking short distances (20 m to 100 m). Comfortable only at rest.

[0089] Class IV Severe limitations. Experiences symptoms even while at rest. Mostly bedbound patients.

[0090] For instance, the device may be programmed to provide an output as one class selected from several classes including “Class I”, “Class II”, etc. Alternatively, the device may provide an output representative of a risk level, such as “low risk” in the case of class I or II, and “high risk” in the event

[0091] Version 2026-01-16of class III or IV. It will be understood that the NYHA Classification is an example of an established classification system, and that the invention is not necessarily limited to using a particular classification system. Embodiments of the invention may be configured to be compatible with already-established, and / or other classification systems. Embodiments of the invention may provide multiple outputs each corresponding to a different classification system.

[0092] The controller 20 is configured to obtain ECG data via the sensors 14a, 14b, when connected. ECG data will be understood to be measured in the form of a series of waveforms, each waveform constituting a waveform unit corresponding to a heartbeat cycle. With a typical heart rate of around 60 to 100 beats per minute, a measurement of around 10 seconds length can be expected to record between ten and 16 waveform units. The waveforms may be used in a classification to determine whether or not the waveforms are indicative of a heart failure risk.

[0093] In accordance with the invention, the analysis of the waveforms is carried out on the device 10, without reliance on a connection to an external processor. In the illustrated device, the analysis of the waveforms is carried out by the controller 20 on the device 10.

[0094] To illustrate different morphology that may be observed in ECG data, an exemplary clean waveform and four noise waveforms are illustrated below.

[0095] Figures 3A and 3B illustrate a “clean”, standard waveform 100 and a “clean”, standard series 110 of waveforms forming an ECG signal. A clean waveform 100 illustrates two heartbeat cycles, each heartbeat cycle expected to comprise a P-wave 102, a QRS complex 104, and a T-wave 106 in one heartbeat 101. Here, a first cycle comprises a first or preceding P-wave 102a, a first or preceding QRS complex 104a, and a first or preceding T-wave 106a, followed by a second or successive P-wave 102b, a second or successive QRS complex 104b, and a second or successive T-wave 106b. For instance, a heartbeat cycle may be defined as a period covering a P-wave, an immediately following QRS complex and an immediately following T-wave. Alternatively, another interval between like features may be taken as a heartbeat cycle, e.g. a R-peak to R-peak interval 108. An appreciation underlying the invention is that a time segment of a few seconds length, e.g. ten seconds, typically includes a few heartbeat cycles.

[0096] Figure 3B shows a clean series 110 of successive heartbeat cycles 112, 114, 116, 118, 120. As will be appreciated, using appropriate statistical or waveform analysis, it is possible to derive features indicative of periodicity or variation of successive cycles. For the purposes of illustrating the invention, Figure 3B shows an ECG data segment as an example of a sequence of electro cardiac signals, e.g. a ten-second segment, comprising at least one heartbeat cycle, here five successive heartbeat cycles 112, 114, 116, 118, 120.

[0097] Figure 4A illustrates a waveform 140 exhibiting baseline wandering, wherein the measured baseline signal 142 deviates from a straight (here: horizontal) line 144. Baseline wandering in an ECG signal

[0098] Version 2026-01-16refers to the low-frequency oscillations or drift observed in the baseline of the signal. It may manifest as a slow, periodic movement of the baseline away from its expected position along a horizontal axis, making it challenging to accurately assess the underlying cardiac activity. Several factors may contribute to baseline wandering, including motion artefacts, respiration, electrode contact issues, and variations in skin conductivity. Motion artefacts from patient movement or poor electrode-skin contact can introduce low-frequency variations in the baseline. Respiration-related baseline wandering may occur due to the expansion and contraction of the chest during breathing.

[0099] Figure 4B illustrates a waveform 150 exhibiting powerline interference, wherein the baseline comprises a superimposed high frequency noise 152 here presenting as a thicker band. Power line interference in an ECG signal refers to unwanted electrical noise or artefacts as may originate from an alternating current (AC) power supply. The interference typically occurs at the frequency of the power line, e.g. 50 Hz or 60 Hz. Additionally, power line interference may appear as periodic spikes or oscillations superimposed on the ECG waveform, making it challenging to distinguish it from actual cardiac signals.

[0100] Figure 5A illustrates a waveform 160 comprising muscular noise, indicating in a muscular noise region 162, affecting the baseline for a duration of several successive heartbeat cycles. Muscular noise in an ECG signal refers to unwanted electrical activity generated by the muscles surrounding the electrode placement area. This noise may be caused by contraction and relaxation of muscles, leading to interference in the recorded ECG signal. Muscular noise can obscure the underlying cardiac electrical activity, making it challenging to accurately interpret the ECG waveform.

[0101] Figure 5B illustrates a waveform 170 showing an electrode motion effect, indicated in a motion effect region 172, leading to a peak shift and / or peak broadening, here showing an irregularly broad peak 174. Motion artefacts in ECG signals can occur due to electrode motion, which results from movements of the electrodes on the skin surface. These movements can lead to fluctuations in the electrical contact between the electrodes and the skin, causing disturbances in the ECG waveform. Conventionally, ensuring proper electrode placement and securing electrodes firmly on the skin can help minimize motion artefacts and improve the quality of ECG recordings. The suggestion made in this disclosure is to enable a classification method to tolerate the presence of such artefacts by being able to exclude them from classification.

[0102] As will be appreciated, the artefacts illustrated in Figures 4A to 5B, an others, may present simultaneously and affect signal classification as well as extraction of descriptors, such as peak features derived from peak height, peak-to-peak distance, and others.

[0103] An appreciation contributing to the development of embodiments was that some artefacts, such as muscular noise due to movement or breathing, may not be present in every measurement segment and / or may not render every measurement unsuitable for classification. By obtaining measurements in short, individual segments, the device is more robust to occurrence of occasional artefacts,

[0104] Version 2026-01-16because the presence of an artefact may be less likely in individual, shorter, segments than in a longer, continuous data set. It is believed that increasing the robustness of the classification algorithm to such artefacts increases the utility of the device with dry electrodes and / or for use during activities that may otherwise be more likely to interfere with electrocardiographic measurements.

[0105] During the development of the invention, it was further appreciated that the device 10 may be used during otherwise potentially interfering procedures, such as by users wearing pacemakers. To this end, the device may be trained to tolerate by ignoring, and / or to recognise, within one or more classification models used as part of a pre-set configuration, artificially created signal components such as pacemaker signals, or other artificially created signal components that may alter a frequency or shape of a natural heartbeat cycle. In some embodiments, the device may be configured to exclude data segments comprising pacemaker signals as irregular waveforms.

[0106] For instance, the device may be trained to recognise an artificially created signal, such as a pacemaker pulse signal, and exclude data comprising a pacemaker pulse signal from a turning count algorithm. As such, a pacemaker pulse signal (or other artificially created signal) may be recognised and excluded from further analysis. In this manner, the device increases the likelihood that only “natural” heartbeat cycles, whether indicative of disease or not, are used for classification.

[0107] In some embodiments, the device may be configured to detect pacemaker signatures as signal, for instance in the form of classification modules trained to recognise the presence of a pacemaker signature (e.g. a pacemaker pulse pattern).

[0108] Figure 6 illustrates a sequence of steps of a method 40 that may be carried out using the device 10. In step 42, ECG sensors are provided. The sensors may be consumables such as the sensors 14a, 14b illustrated in Figure 1. In an optional step 44, the controller 20 may await a connection between the communications interface 30 and an external device, such a device 36, to await a user command before commencing a measurement. However, this is not necessarily a requirement of all embodiments. In other embodiments, the device 10 may obtain sensor measurements continuously or in periodical intervals.

[0109] In step 46, the device 10 records ECG raw data of a pre-determined length, e.g. ten seconds length. Using the sensors 14a, 14b, the device may obtain several data sets or data segments of predetermined length, e.g. ten-second segments, for analysis. The data segments may have length of between 2 to 15 seconds, e.g., 3 seconds, 4 seconds, 5 seconds, 6 seconds, 7 seconds, 8 seconds, 9 seconds, or 10 seconds. Without wishing to be bound by theory, it is believed that shorter segments are less prone to irregular data within the segment. To provide an illustrative example, a ten-second segment may contain an irregular feature, potentially causing the entire ten-second segment potentially to be excluded from analysis. If, instead, two or three data segments of three second length are measured, an irregular pattern may affect only one of the 3-second data segments. The device 10 may be configured to determine if the sum of durations of individual data

[0110] Version 2026-01-16segments amounts to a minimum cumulative duration. For instance, the device 10 may comprise a configuration to acquire a minimum duration of measurements, e.g. at least 30 seconds of analytical data. In that case, the device 10 may obtain at least three ten-second segments. For instance, provided all three ten-second segments are used for classification, the minimum cumulative 30 seconds of analytical data have been acquired. Alternatively, if one or more of the ten-second segments is excluded, e.g. after having been classified as containing an irregular waveform or too many irregular waveforms, the device 10 may obtain additional data segments until a minimum cumulative duration is reach, e.g. until a total of the required cumulative 30 seconds of analytical data have been obtained. In this manner, a data set comprising 30 seconds of data of good quality may be acquired relatively quickly. As may be imagined, even if, for instance, only every second data set is of sufficient quality, it is still possible (in this example) to obtain 30 seconds of good quality data within a minute (assuming a first, third, and fifth ten-second data set are excluded and analysis is based on a second, fourth, and sixth data set). The decision whether or not a waveform is irregular may be based on a classification of a waveform shape, or on a classification of descriptors derived from a waveform.

[0111] In step 48, the ECG raw data is processed to remove artefacts. To provide illustrative examples, a bandpass filter such as a Butterworth bandpass filter (order 5) and a notch filter may be applied to remove artefacts such as baseline wandering or power line interferences.

[0112] The processing may be carried out on each individual segment, e.g. if the device 10 was configured to obtain a series of three ten-second waveform segments, the processing for artefact removal may be applied to each of the ten-second waveform segments individually.

[0113] In step 50a, the classification model uses, as input data, artefact free data in the form of a series of ECG waveforms. As an alternative or in addition, in step 50b, the classification model uses, as input data, descriptors derived from a series of ECG waveforms.

[0114] In step 52, the controller 20 uses the processor to carry out a classification of the artefact-free data, using a classification model that has been pre-programmed. The classification model can, therefore, operate without requiring a connection to an external network and without need to upload the ECG data.

[0115] Step 52 may involve extraction of one or more descriptors, for each one of the waveform segments. In an optional step 54, the data may be processed to identify data suitable for classification, and data that is too irregular and would fail classification. As will be appreciated, the separation of irregular segments from suitable segments may be carried out contemporaneously during one or more of steps 48, 50 and / or 52, when extracting descriptors, and / or on extracted descriptors, as the case may be.

[0116] Version 2026-01-16In optional step 56, the analysis is repeated for subsequent data segments. As will be understood, only data that has been retained after filtering and removal of irregular data is used for classification, and some data segments may not be used. Therefore, the process may be repeated, e.g. from step 46, to obtain and / or process further data segments, for a predetermined number of cycles, e.g. to obtain six classifications, or twelve classifications, in the manner of a turning count algorithm.

[0117] For instance, the algorithm may determine that a minimum number of repeated classifications need to be obtained, e.g., at least five out of six consecutive classifications need to be provided for a consistent classification level, such as “healthy”, before an output is provided of the classification. The device may be configured to prevent a classification output unless a minimum number of consistent classification results, or a minimum number of consensus classification results, has been obtained. The device may be configured to prevent a classification output unless a minimum cumulative duration of measurements has been classified.

[0118] An appreciation underlying the development of embodiments of the invention was that the use of relatively short segments, e.g. ten-second segments, firstly facilitates the shortening of the measurement time and associated resources. Furthermore, the use of relatively short segments makes it practical to employ a turning count algorithm, wherein a series of successive measurements can be made that are spaced over a period of time to collect a series of discrete measurements to check for consistency of the classification output. In this manner, a classification output can be based on a predetermined minimum number of readings, e.g. at least six readings, or at least twelve readings.

[0119] In 58, the classification model is used to determine a risk level based on the waveforms received in step 50a, and / or based on the descriptors received in 50b. It will be appreciated that the risk level so determined corresponds to the raw data, or a data segment thereof, obtained in step 46.

[0120] In step 60, the controller 20 generates an output indicative of the risk level. The output may be provided via the communication interface 30, as a notification or message to a user device 36, and / or via the display 34. In an optional step 62, the controller 20 further outputs the raw data, artefact free data, any descriptors derived, and / or stores data in the memory 24, and / or deletes the raw data. For practical purposes, the device 10 may employ a ring buffer, which may be part of the memory 24, to allow the device 10 to retain a pre-determined amount of data on the device, e.g. a typical amount corresponding to a week of measurements, to one or more days of measurements, and / or to a predetermined number of measurements, such as 1000 (one-thousand) ten-second segments. It will be appreciated that these values are only exemplary and dependent on memory configuration, power supply configuration, and data connectivity, and others, which may be selected differently for different user groups or application scenarios.

[0121] The risk level may be a two-tier risk level, to distinguish between congestive heart failure (CHF) or normal sinus rhythm (NSR). The risk level may be one of more than two tiers, e.g., a three-tier, four-

[0122] Version 2026-01-16tier, five-tier or higher-tier risk level. For instance, the output may be provided in the form of a score from 0 to 4, 0 being lowest risk (NSR) and 4 being highest risk (CHF). The lowest level (0) may correspond to a healthy scan, and other levels (1 to 4) may correspond to pre-determined classes I to IV, for instance.

[0123] In addition, or as an alternative, the algorithm may be configured to determine the presence of one of several heart conditions, such as arrythmia. Exemplary conditions that may be analysed from a waveform include atrial fibrillation, atrial tachycardia, arrhythmia, coronary artery disease, myocardial infarction, hyperkalaemia, hypokalaemia, hypercalcaemia, and hypocalcaemia.

[0124] Such conditions may present as changes in heart rhythm and waveform. Atrial Fibrillation, for instance, is a heart condition characterized by irregular and often rapid heartbeats, affecting the upper chambers (atria). Atrial Tachycardia is an abnormal rapid heartbeat originating in the atria, disrupting the heart’s normal rhythm. Arrhythmia, generally, is a broad term referring to irregular heart rhythms, which can include conditions like atrial fibrillation and atrial tachycardia. Coronary artery disease and myocardial infarction are conditions in which blood supply to the heart leads to reduced heart performance that may present in the heartbeat waveform.

[0125] Furthermore, the invention is believed to allow the distinction of elevated or insufficient levels of potassium (hyperkalaemia and hypokalaemia, respectively) and calcium (hypercalcaemia and hypocalcaemia, respectively).

[0126] In an endeavour to reduce processing resources required on the device 10, the invention suggests that in several embodiments, an analysis is carried out using descriptors (see step 54 in Figure 6), thereby reducing a need to process raw ECG data.

[0127] Figure 7 illustrates an exemplary method 70 of training, or creating, a classification model that can be implemented (e.g., installed) to operate without requirement for an external connection, which may be described as operating autonomously, offline, or airgapped. The method 70 comprises a step 72 of providing ECG training data. The ECG training data may have been measured from test users and / or from ECG data bases. In an optional step 74, the ECG training data is classified using a reference method, such as a classification standard, gold standard, or consensus opinion. For some heart conditions, training data sets may be available that are pre-labelled with a risk score.

[0128] In step 76, one or more segments of the ECG training data are selected that correspond to a predetermined segment length to be used later by the device. For instance, the device 10 may be configured to obtain measurement segments of a predetermined duration, such as 3, 4, 5, 6, 7, 8, 9, 10, 11 , 12, 13, 14, 15, or 20 seconds duration. As such, the segments fortraining may be selected to have corresponding duration. A device may comprise a pre-set configuration for multiple different segment lengths. In step 78, the ECG training data are subjected to an artefact-removal procedure, such as peak normalisation or baseline alignment.

[0129] Version 2026-01-16In step 80, a feature extraction procedure is carried out to extract statistical parameters describing the waveform in time domain, e.g., the morphology of the waveform and its variation throughout a pre-determined period of time. As will be appreciated, a segment of, e.g., ten seconds length may comprise around ten wavelength units. It was found by the inventor that a relative low number of peaks may suffice to extract statistical descriptors indicative of variations or, conversely, consistent behaviour. As such, the data segments used for classification may be relatively short, in the region of a few seconds, and so the measurement duration and corresponding power requirement, memory requirement, and data processing, place relatively small demands on the device 10. For instance, the data segment may have a duration of 3 seconds, 5 seconds, 7 seconds, or another suitable duration. Example descriptors include Hjorth activity, Hjorth mobility, Hjorth complexity, skewness, kurtosis, and pe a k-to-a verage power ratio, and others. The descriptors may be defined as follows:

[0130] Peak-to-averaqe power ratio (PAPR) is the ratio between the power of a signal’s peak and the average power of the entire signal, here the duration of a data segment.

[0131]

[0132] where, N is the length of a segment (e.g., ten seconds).

[0133] Hjorth activity indicates the signal’s variance, which is calculated by squaring the standard deviation as given below:

[0134] >

[0135]

[0136] where, ps is the mean value of s(n):

[0137]

[0138] The variance measures the signal’s variability about its average value. A relatively large variance indicates a wide spread of the signal’s amplitude around the mean value, while the opposite is indicated by a small variance.

[0139] Hjorth mobility is the reciprocal ratio between the signal’s standard deviation and that of the signal’s first derivative.

[0140] Mobility[s

[0141]

[0142] where, s'(n) is the first derivative of s(n):

[0143] s'(n) = fs[s(n+ 1) - s(n)]

[0144] Version 2026-01-16and fs is the sampling frequency at which the ECG signal is recorded.

[0145] The Hjorth mobility parameter gives an indication of the standard deviation of the signal’s power spectrum over the entire frequency range. That is to say, a signal with focused frequency content in a narrow range has a lower mobility than that of another signal with a widely spread power spectrum.

[0146] Hjorth complexity is calculated by dividing the Hjorth mobility of the signal’s first derivative by that of the signal as follows [1 , 2]:

[0147] „ i r r M mobility [s' (n)]

[0148]

[0149] <r

[0150] Complexity [s(n)J = - — —r, ... = - - — = -s" x o- 5 -s

[0151] mobility [s(n)] as< / aso ,

[0152] where, s "(n) is the second derivative of s(n):

[0153] s"(n) = fs[s'(n + 1) - s'(n)]

[0154] The Hjorth complexity is a dimensionless parameter which indicates the similarity between the signal’s morphology and that of a sinusoidal waveform. The complexity value converges to 1 as the signal’s shape gets more like a pure sine wave. In other words, the minimum signal’s Hjorth complexity value is one which is obtained in the case of a pure sine wave.

[0155] Skewness: Skewness indicates a degree of asymmetry of a distribution of waveforms. Zero skewness indicates a symmetric distribution, while negative and positive skewness indicate left and right skewed distributions, respectively, and the magnitude reflects the degree of skewness. The skewness of s(n) is calculated as shown below:

[0156] Skewness

[0157]

[0158] Kurtosis measures the existence of extreme values or outliers in a data set. High kurtosis indicates that the distribution of the data has heavy tails, while low kurtosis indicates that the distribution has light tails. The kurtosis of s(n) is calculated as follows:

[0159]

[0160] It will be appreciated that the statistical parameters described above are only examples of descriptors that may be used for classification, and that other parameters, as well as combinations of different parameters, may be used as alternative and / or in addition. The present parameters are examples of time-domain morphology of ECG data. Other descriptors may be based on frequency-domain data.

[0161] Version 2026-01-16Embodiments of the invention allow one or more suitable descriptors for classification of different heart conditions and risk levels thereof to be determined.

[0162] In step 82, a feature selection process is carried out by way of a statistical correlation to establish descriptors with high predictive power for different risk levels. The process may be repeated by returning to step 80 and using a different selection of descriptors, until in step 84, a selection of descriptors is identified as suitable input parameters for a classification model. To provide an illustrative example, in initial prototypes, a support vector machine (SVM) model was used as linear classification algorithm, and was found to provide a high predictive score of CHF or NSR conditions using only two predictors (Hjorth mobility and peak-to-average power ratio). However, while these two predictors were found to provide reliable classification results in initial tests, the invention is not necessarily intended to be limited to two specific descriptors. Different correlations may show different discriminatory power for different heart conditions.

[0163] In an optional step 86, data to be excluded may be identified and / or flagged as irregular data, for instance when the process of extracting descriptors in step 84 fails. While Figure 7 illustrates step 86 as following step 84, in some embodiments step 86 may be carried out at one or more different stages of the method, e.g. before the feature extraction procedure of step 80, and / or after step 80 and before the feature selection procedure of step 82, and / or after the descriptor identification step 82.

[0164] In step 88, a classification model is trained. In the illustrated example, the classification model is trained using the descriptors as input, rather than raw or artefact-free waveform data. The method 70 may return to step 80, or 78, or 72, to further improve the classification model. As will be appreciated, the classification model may be provided by, or may use, a machine learning model.

[0165] In step 90, the classification model is provided, for instance in the form of a classification module, for use with the device 10 for use by the controller 20. As will be appreciated, in use, the classification module may receive data from the sensors 14a, 14b purely for classification, without necessarily further modifying the classification model. However, in some embodiments, data obtained via the sensor arrangement may be exported, via the communication interface 30, to be made available for subsequent training and or review of future iterations of a classification model.

[0166] The classification process may be carried out by the processor, in the form of software instructions implemented by the processor under the control of the controller.

[0167] While the device 10 is illustrated only schematically in Figure 1 , it is contemplated for the device 10 to be provided in relatively small form, even in patch form, for instance of few square centimetres size, and in the region of one to a few millimetres in thickness, so as to be suitable in day-to-day use as a wearable patch device. The device may be part of a wearable item and / or a garment.

[0168] Version 2026-01-16Figures 8 to 11 are graphs provided to illustrate the variation in performance based on the duration of data segments used in test classification. In each of the graphs, the data analysis used a two-class classification, to separate the data segments into either normal sinus rhythm (NSR) or congestive heart failure (CHF). It will be appreciated that other classifications may use a different number of classes, and or may use different granularity, such as multiple classes for different stages of CHF.

[0169] The analysis was carried out using different classification models, here SVM-RBF (Support Vector Machine with Radial Basis Function), Logistic Regression (LR), and KNN (k-Nearest Neighbours algorithm). The analysis of the results was carried out on test data using 3, 5, 7, or 10 second segment length, and compared to a refence classification. The data segments were obtained from publicly available reference data (Physionet, Goldberger, A., Amaral, L., Glass, L., Hausdorff, J., Ivanov, P.C., Mark, R., Mietus, J.E., Moody, G.B., Peng, C.K. and Stanley, H.E., 2000. PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation [Online]. 101 (23), pp. e215-e220. RRID:SCR_007345.) The data included the BIDMC Congestive Heart Failure Database (Published: Oct. 14, 2000. Version: 1.0.0) and the MIT-BIH Normal Sinus Rhythm Database (Published: Aug. 3, 1999. Version: 1.0.0) and included 15 CHF patients (11 males, 4 females; age range: 22-71 years) with New York Heart Association (NYHA) class 3-4 severity, characterized by moderate to severe physical activity restrictions. The MIT-BIH NSR dataset comprises 18 healthy individuals (5 males, 13 females; age range: 20-50 years) (see Table 1). A total of 3,300 10-second single-lead ECG segments were extracted: 1 ,500 from CHF patients (BIDMC) and 1 ,800 from healthy controls (MIT-BIH). The performance of the classification was compared against the classification of the CHF and NSR data.

[0170] Figure 8 shows a graph 200, illustrating the sensitivity (y axis) of the analysis for different segment lengths of 3, 5, 7 and 10 length duration. Line 202 illustrates sensitivity results using SVM-RBF. Line 2024 illustrates sensitivity results using Logistic Regression. Line 206 illustrates sensitivity results using KNN. It is pointed out that the y axis of graph 200 spans a range from 0.970 to 1.000 to better visualise differences in sensitivity.

[0171] Figure 9 shows a graph 210, illustrating the specificity (y axis) of the analysis for different segment lengths of 3, 5, 7 and 10 length duration, using the same data sets as in Figure 8. Line 212 illustrates specificity results using SVM-RBF. Line 214 illustrates specificity results using Logistic Regression. Line 216 illustrates specificity results using KNN. It is pointed out that the y axis of graph 210 spans a range from 0.990 to 1.000 to better visualise differences in specificity.

[0172] Table 1 shows the sensitivity and specificity values corresponding to the 3, 5, 7, and 10-second segments shown in Figures 8 and 9.

[0173] Version 2026-01-16Segment Duration Model Sensitivity Specificity

[0174] 3 sec SVM (RBF) 0.9919 0.9991

[0175] Logistic Reg. 0.9855 0.9954

[0176] KNN 0.9793 0.9909

[0177] 5 sec SVM (RBF) 0.9991 0.9973

[0178] Logistic Reg. 0.9973 0.9910

[0179] KNN 0.9927 0.9900

[0180] 7 sec SVM (RBF) 0.9973 0.9973

[0181] Logistic Reg. 0.9946 0.9910

[0182] KNN 0.9891 0.9910

[0183] 10 sec SVM (RBF) 0.9919 0.9991

[0184] Logistic Reg. 0.9919 0.9936

[0185] KNN 0.9891 0.9910

[0186] Table 1 : Sensitivity and specificity, respectively, for 2-tier classification (CHF-NSR) based on different classification models.

[0187] Comparing Figures 8 and 9, it can be seen that for the given test data, the sensitivity 202 and specificity 212 of the SVM models performed higher than the corresponding sensitivity 204 and specificity 214 of the Logistic Regression, which performed higher than the corresponding sensitivity 206 and specificity 216 of the KNN model. Based on the trained models, the was highest among all models for the 5-second segments, and second highest with 7-second segments. Conversely, specificity was lower at 5-second and 7-second segments. Nevertheless, all length segments had relatively high sensitivity of typically at least 0.98 and relatively high specificity of typically at least 0.99.

[0188] The sensitivity results of Figure 8 suggest that the method disclosed herein can be used for a reliable detection of CHF, via classification of different ECG segment durations, including relatively short time windows, indicating suitability for rapid and low-latency analysis.

[0189] The specificity results of Figure 9 suggest that the method disclosed herein can be used for an accurate identification of NSR signals, via classification of different ECG segment durations, reducing false positive CHF detections.

[0190] The results of Figures 8 and 9 together support a finding that the method is robust for CHF-NSR discrimination, via classification of segments of few seconds lengths, even if the segments are as short as 3, 5, 7, or 10 seconds.

[0191] Figures 10 and 11 show relative deployment efficiency 220 (normalised to 1 for the highest efficiency), wherein a higher deployment efficiency score is better, and relative computational load

[0192] Version 2026-01-16230 (normalised to 1 for the highest computational load) for segments of 3, 5, 7 and 10 seconds length, respectively, wherein a lower computational load score is better. A first deployment efficiency score 221 for 3-second segments was around 1. A second deployment efficiency score 222 for 5-second segments was around 0.8. A third deployment efficiency score 223 for 7-second segments was around 0.6. A fourth deployment efficiency score 224 for 10-second segments was around 0.4. A first computational load score 231 for 3-second segments was around 1. A second computational load score 232 for 5-second segments was around 0.9. A third computational load score 233 for 7-second segments was around 0.7. A fourth computational load score 234 for 10-second segments was around 0.5.

[0193] As might be expected, the shortest segments (here, the 3-second segment) have the highest deployment efficiency within the tested data. Likewise, as might also be expected, the shortest segments require the lowest computational load within the tested data.

[0194] The findings support that relatively short data segments, of 3, 5, or 7 seconds length, perform no worse than longer segments, while also offering better performance with respect to deployment efficiency and computational load. Figures 10 and 11 support that the classification and analysis can be carried out autonomously on a sensor device with relatively low demand on computing resources and power supply.

[0195] The device may be a low-power-consumption device. In some embodiments, the device 10 may be passive, to be powered by a user device such as an NFC powered device, to activate for a measurement when in NFC proximity, and to remain otherwise passive. By reducing the input parameters to relatively few descriptors, it is believed that even relatively small batteries may allow a device to be active for several days without need for recharging.

[0196] Embodiments of the invention may be sufficiently inert to be used during defibrillation procedures. The device may remain in place, with its sensor electrodes attached to a user, if necessary, during defibrillation. Alternatively, and / or in addition, the device may be used as a monitoring device during defibrillation, to allow a response of a patient to be monitored while subjected to a defibrillation procedure. As will be appreciated, the device may integrate a protection arrangement to shield components from effects of a defibrillator, such as a hardware-based protection circuit (such as a gas discharge tube) to safeguard the heart sensor device. Suitable protection circuit arrangements and surge protectors will be known to a skilled person. As such, it is envisaged that some embodiments may comprise electrical shielding components to allow them to be used with a defibrillator. Embodiments may be part of a kit, comprising a defibrillator. The device may be used during and / or after use of a defibrillator. In that case, the classification module of the device may analyse electro cardiac signals, and may be configured to provide an output indicative of whether or not electro cardiac signals have changed (e.g. improved) after use of a defibrillator.

[0197] Version 2026-01-16A classification module may be able to determine whether or not an electro cardiac signal should be assigned a class indicative of a condition potentially treatable by defibrillation, e.g. tachycardia or fibrillation. A classification module may be able to determine the absence of a electro cardiac signal indicative of a condition treatable by defibrillation and / or whether or not defibrillation should be avoided or discontinued For instance, an output may indicate if a heart rate stabilizes within a normal expected range (e.g., between 60 and 100 bpm), or if rhythms are regular based on an assessment of regularity and / or morphology of QRS complexes, or if P-waves are present, where the presence of P-waves may indicate sinus node recovery, or if QRS complexes have a morphology indicative of successful rhythm restoration. To this end, the device may provide an output indicative of a class such as “heart rate within normal range” “heart rate outside normal range”, “QRS rhythm regular”, “P-wave present”, and / or “QRS morphology within normal range”, and / or others. In this manner, the device may be used for pre-defibrillation and / or for post-defibrillation signal monitoring.

[0198] It is believed that the device may be configured to detect, via use of suitable classification models, the presence of one or more of the following conditions to the extent that the present a characteristic ECG waveform or heartbeat cycle, and / or characteristic variations of successive heartbeat cycles, and / or characteristic artefacts or statistical behaviour:

[0199] Atrial fibrillation: a heart condition characterized by irregular and often rapid heartbeats, affecting the upper chambers (atria).

[0200] Atrial tachycardia: an abnormal rapid heartbeat originating in the atria, disrupting the heart's normal rhythm.

[0201] Arrhythmia: a broad term referring to irregular heart rhythms, which can include conditions like atrial fibrillation and atrial tachycardia.

[0202] Coronary artery disease: coronary artery disease is a condition where the blood vessels supplying the heart muscle become narrowed or blocked, leading to reduced blood flow and potential complications.

[0203] Myocardial infarction: myocardial infarction, commonly known as a heart attack, occurs when blood flow to a part of the heart is blocked, resulting in damage to the heart muscle.

[0204] Hyperkalaemia: elevated levels of potassium in the blood, which can impact heart function.

[0205] Hypokalaemia: insufficient levels of potassium in the blood, potentially leading to abnormal heart rhythms.

[0206] Hypercalcaemia: elevated levels of calcium in the blood, potentially affecting the heart's electrical system.

[0207] Hypocalcaemia: lower-than-normal levels of calcium in the blood, which can also impact the heart's electrical activity.

[0208] Whilst the principle of the invention has been illustrated using exemplary embodiments, it will be understood that the invention is not so limited and that the invention may be embodied by other variants defined within the scope of the appended claims.

[0209] Version 2026-01-16

Claims

CLAIMS:

1. A heart sensor device comprising a sensor arrangement operatively connected to a controller, the controller configured to obtain electro cardiac signals obtained from the sensor arrangement, and further configured:to obtain a sequence of electro cardiac signals comprising at least one heartbeat cycle; and to use a classification module to carry out an analysis of the electro cardiac signals of the at least one heartbeat cycle, to assign a classification of the signal, and to provide an output indicative of the classification;wherein the controller and the classification module comprise a pre-set configuration allowing the analysis to be carried out without external data communication from the heart sensor device.

2. The heart sensor device according to claim 1 , wherein the classification module comprises a machine learning algorithm configured with, or using, a machine learning model.

3. The heart sensor device according to claim 1 or 2, wherein at least one heartbeat cycle is processed by the classification module without modifying the pre-set configuration.

4. The heart sensor device according to any one of the preceding claims, wherein the sequence of electro cardiac signals is at least 1 , 2, 3, 4, 5, 6, 7, 8, 9, or at least 10 seconds long.

5. The heart sensor device according to any one of the preceding claims, wherein the classification module is configured to accept into the analysis a sequence of electro cardiac signals that is no more than 20, 17, 15, 13, 11 , 10, or 9 seconds long.

6. The heart sensor device according to any one of the preceding claims, wherein the pre-set configuration comprises an algorithm to derive as an input for the classification module two or more descriptors obtained from analysing the at least one heartbeat cycle.

7. The heart sensor device according to claim 6, wherein the two or more descriptors comprise at least two of: peak to average power ratio, Hjorth activity, Hjorth mobility, Hjorth complexity, skewness, and kurtosis.

8. The heart sensor device according to any one of the preceding claims, configured to remove noise data from the sequence of electro cardiac signals.

9. The heart sensor device according to any one of the preceding claims, configured to remove irregular data.Version 2026-01-1610. The heart sensor device according to any one of the preceding claims, configured to exclude data comprising pacemaker signals.

11. The heart sensor device according to any one of the preceding claims, comprising a power source to provide power to the sensor arrangement and / or the controller.

12. The heart sensor device according to any one of the preceding claims, further comprising a communication interface operatively connected with the controller, and configured to provide the output upon a request signal received via the communication interface, wherein, optionally, the device is configured to obtain an electro cardiac signal upon receipt of a command signal received via the communication interface.

13. The heart sensor device according to any one of the preceding claims, wherein the classification of the signal determined by the classification module includes at least a classification indicative of congestive heart failure (CHF) and a classification indicative of normal sinus rhythm (NSR).

14. The heart sensor device according to claim 13, wherein the classification of the signal determined by the classification module includes one or more intermediary classes each indicating a different risk level between congestive heart failure (CHF) and normal sinus rhythm (NSR).

15. The heart sensor device according to any one of the preceding claims, wherein the classification module is configured to assign a classification indicative of one or more conditions selected from atrial fibrillation, atrial tachycardia, arrhythmia, coronary artery disease, myocardial infarction, hyperkalaemia, hypokalaemia, hypercalcaemia, and hypocalcaemia.

16. The heart sensor device according to any preceding claim, forming part of a defibrillator system.

17. The heart sensor device according to any preceding claim, comprised in a wearable device and / or garment.

18. A defibrillator system comprising a heart sensor device according to any one of claims 1 to 15.

19. A classification model trained to accept, as input for classification, descriptors selected from a sequence of two or more ECG waveforms, and to present, as an output, at least two different classification groups.

20. The classification model according to claim 19, configured to determine whether the input data belongs to one of two or more classification groups.Version 2026-01-1621. A method of training a classification model for use with a heart sensor device according to any one of claims 1 to 17, the method comprising:providing ECG training data labelled with a reference classification;selecting, from the ECG training data, one or more ECG waveform segments of predetermined length, each of the segments representing a sequence of one or more heartbeat cycles; andusing the one or more ECG waveform segments as input fortraining a classification model, based on the reference classification, andproviding, as an output, a classification model.

22. The method according to claim 21 , comprising removing artefacts from the one or more ECG waveform segments prior to using the one or more ECG waveform segments as an input for classification.

23. The method according to claim 21 or 22, comprising detecting irregular waveforms prior to using the one or more ECG waveform segments as an input for classification.

24. The method according to claim 23, comprising including irregular waveforms as an input for classification.

25. The method according to any one of claims 21 to 24, comprising extracting features for use as input for classification, the features comprising statistical descriptors of a series of waveform segments.

26. The method according to claim 25, the descriptors comprise at least two of: peak-to-average power ratio, Hjorth activity, Hjorth mobility, Hjorth complexity, skewness, and kurtosis.

27. The method according to claim 25 or 26, comprising removing artefacts from the one or more ECG waveform segments prior extracting features.

28. The method according to claim 25 or 27, wherein the artefacts comprise pacemaker signals.

29. The method according to any one of claims 25 to 28, wherein the data representing a sequence of one or more heartbeat cycles are provided by descriptors defined by extracted features.

30. The method according to any one of claims 25 to 29, wherein providing ECG training data comprises obtaining ECG data from a sensor measurement and / or from a database.Version 2026-01-16

Images