Medical devices and systems

Abstract

The present invention provides a wearable device for detecting movements of a wearer, device having: a substrate which is flexible and elastically deformable; a plurality of electrodes formed on the substrate; and a plurality of strain sensors formed on the substrate and configured to detect movement of the substrate, wherein: each strain sensor is formed at the intersection of a pair of said electrodes and comprises an insulating layer separating the pair of electrodes and one or more graphene connections conductively linking the electrodes in a plurality of orientations, such that relative movement between the pair of electrodes causes a change in the resistance of the graphene connection or of each of one or more of the graphene connections The wearable device is particularly suited for the detection of the breathing patterns of the wearer during sleep. A system for detecting and classifying the breathing patterns of a wearer during sleep is also provided.

Description

[0001] MEDICAL DEVICES AND SYSTEMS

[0002] The present invention relates to medical devices and systems. The invention is of particular, but not exclusive, relevance to wearable devices, for example for the monitoring and detection of sleep patterns.

[0003] Sleep occupies about one-third of a person's daily life, and its quality is crucial to overall human well-being. Statistics indicate that more than 60% of adults suffer from poor sleep quality, which results in the loss of approximately 44 to 54 working days per year and contributes to an estimated annual global GDP reduction of between 0.64% and 1.31% [1, 2, 3], Among the culprits are sub-healthy or high-risk sleep patterns such as mouth breathing, snoring, teeth grinding (“bruxism”), and sleep apnea, all of which significantly contribute to the degradation of sleep quality [4, 5, 6, 7], Characteristics of these patterns are set out below.

[0004] Nasal breathing is the natural and preferred mode of respiration during sleep. This involves inhaling and exhaling through the nostrils. It facilitates air filtration, humidification, and temperature regulation, promoting efficient gas exchange in the lungs.

[0005] In contrast to nasal breathing, mouth breathing occurs when airflow bypasses the nasal passages, leading to inhalation and exhalation through the mouth. It may result from nasal congestion, structural abnormalities, or habitual behaviours, potentially affecting respiratory efficiency and oral health.

[0006] Snoring is characterized by turbulent airflow causing vibrations of the soft tissues in the upper airway during sleep. It often arises due to partial obstruction of the airway, leading to disrupted sleep architecture and daytime fatigue.

[0007] Bruxism refers to the involuntary grinding, clenching, or tapping of teeth during sleep, often associated with stress, anxiety, or sleep disorders. It can lead to dental problems, temporomandibular joint dysfunction, and disrupted sleep.

[0008] Central Sleep Apnea (CSA) involves periodic cessation of airflow during sleep due to temporary failure of the brain to send signals to the respiratory muscles, yet without blockage in the airway. It disrupts normal breathing patterns and may be associated with neurological conditions or cardiac dysfunction.

[0009] Obstructive Sleep Apnea (OSA) is characterized by recurrent episodes of complete or partial upper airway obstruction during sleep, leading to breathing pauses and oxygen desaturation. It is commonly associated with obesity, anatomical factors, and increased risk of cardiovascular morbidity.

[0010] These sleep patterns manifest through vibrations and disturbances in physiological structures such as the velum, oropharynx, tongue, and epiglottis. Vibrations generated by airflow turbulence, muscle contractions, or neurological abnormalities propagate through the extrinsic laryngeal muscles, leading to audible sounds and physiological responses.

[0011] These unhealthy sleep states have been extensively studied and identified as risk factors for various chronic diseases, including cardiovascular disease, diabetes, and emotional disorders [8, 9, 10], To improve people's sleep quality and to provide early warnings for related chronic diseases, monitoring and identifying these unhealthy sleep states have become a focal point in modern health management.

[0012] Similar issues apply in relation to veterinary health monitoring associated with detecting breathing patterns.

[0013] For example similar pathological conditions exist in companion animals including, but not limited to, brachycephalic dog breeds suffering from brachycephalic obstructive airway syndrome (BOAS). Cardiac and respiratory conditions in horses can include atrial fibrillation, equine recurrent airway obstruction (asthma), and exercise-induced pulmonary haemorrhage. Neurological disorders in both dogs and horses including, but not limited to epilepsy, paroxysmal dyskinesia, degenerative myelopathy, sleep disorders, neuropathic pain, equine protozoal myeloencephalitis, wobbler syndrome and equine dysautonomia may also be monitored through breathing patterns.

[0014] Current veterinary diagnostics rely on in-clinic endoscopy under sedation, all of which are invasive, costly and unsuitable for longitudinal assessment or home screening. Consequently, there remains an unmet need for a comfortable, high-fidelity wearable device that can continuously capture animal breathing, such as canine and equine respiratory effort and vibration signatures, in a real-world setting, enabling early detection, therapy optimisation and remote tele-veterinary care.

[0015] Traditional sleep monitoring methods, such as polysomnography (PSG), although comprehensive and accurate, require specific facilities and involve complex, costly equipment, making them unsuitable for long-term or home use

[0011] ,

[0016] In recent years, the rapid development of smart wearable devices has provided a promising approach to portable sleep health monitoring [12, 13, 14], The most common monitoring platforms integrate photoplethysmography (PPG) sensors and motion sensors within watches or wristbands, in both academia and industry [15, 16, 17], Although these platforms offer convenience as a scaled-down version of PSG, they inherently lack the capability to collect sufficient human physiological data required for analysing different sleep states

[0018] ,

[0017] Some designs integrate physical sensors such as humidity, mechanical, and acoustic sensors in mask-like setups placed near the nose to monitor airflow and sounds during sleep [19, 20, 21, 22, 23], While these devices cover a broader range of physiological information, achieving comprehensive sleep pattern monitoring with such setups requires integrating multiple sensors, either within the same area of the body or across different body locations, which adds bulkiness and reduces system energy efficiency, hindering long-term continuous wear [19, 21, 22],

[0018] An alternative approach takes inspirations from the modalities of PSG, integrating electrophysiological sensors like electroencephalograms (EEG), electrooculogram (EOG), and electromyogram (EMG) into facial or ear areas using user-friendly technologies like electronic skin and miniaturized integration [24, 25, 26], Such methods collect essential indicators related to sleep health. However, despite the improvement in the level of comfort and information richness, the monitoring accuracy of non-invasive electrophysiological sensors is still hindered by inherent artefact noise, leaving room for improvement in accuracy [27, 28],

[0019] Accordingly, there is still a lack of versatile technology with both good comfort and high precision for continuous monitoring and recognition of various sleep patterns. The present invention aims to solve one or more of the above problems.

[0020] An aim of the present invention is to provide more portable, cost-effective, and / or user-friendly solutions to the detection and recognition of sleep patterns.

[0021] A further aim of the present invention is to provide a signal modality that encompasses the information necessary for analysing various sleep states and desirably can be captured by wearable devices in a manner that provides a high signal-to-noise ratio and comfort.

[0022] A further aim of the present invention is to provide wearable devices which are comfortable and capable of detecting wearer data which can be processed for analysing and determining sleep states.

[0023] A first aspect of the present invention provides a wearable device for detecting movements of a wearer, device having: a substrate which is flexible and elastically deformable; a plurality of electrodes formed on the substrate; and a plurality of strain sensors formed on the substrate and configured to detect movement of the substrate, wherein: each strain sensor is formed at the intersection of a pair of said electrodes and comprises an insulating layer separating the pair of electrodes and one or more graphene connections conductively linking the electrodes in a plurality of orientations, such that relative movement between the pair of electrodes causes a change in the resistance of the graphene connection or each of one or more of the graphene connections.

[0024] The wearable device of this aspect is able to collect mixed-mode signals generated by vibrations from various sleep activities such as breathing, snoring, teeth grinding, and sleep apnea, which are transmitted to the extrinsic laryngeal muscles from multiple anatomical locations including the velum, oropharynx, tongue, and epiglottis.

[0025] References to “wearer” and “user” will be understood to apply to both human and other animal subjects. Applications in veterinary health monitoring and diagnostics may include cardiac and respiratory conditions in horses, such as atrial fibrillation, equine recurrent airway obstruction (asthma), and exercise-induced pulmonary haemorrhage, or neurological disorders in both dogs and horses including, but not limited to, epilepsy, paroxysmal dyskinesia, degenerative myelopathy, sleep disorders, neuropathic pain, equine protozoal myeloencephalitis, wobbler syndrome, equine dysautonomia.

[0026] In certain embodiments the device is configured to be worn around the neck of the wearer so as to detect vibrations from these muscles.

[0027] In embodiments intended for veterinary applications, the wearable device may be integrated into a canine or equine collar sized to encircle the neck area of the wearing animal. The array can detect micro-vibrations associated with inspiratory and expiratory effort.

[0028] Preferably the wearable device has at least four strain sensors and these strain sensors form an array. The array may be regularly or irregularly arranged, for example the array of sensors may be arranged in a polygon configuration such as a rectangle, or a square. Providing an array of sensors can mean that accurate positioning of the wearable device on the user is not required in order to obtain usable signals. This can make the device suitable for use in a non-clinical setting (e.g. the home) and / or without input from a medical professional.

[0029] In certain embodiments the electrodes are arranged in a grid pattern such that there are a plurality of electrodes running in a first direction in the plane of the device and a plurality of electrodes running in a second direction, substantially perpendicular to the first direction, in the plane of the device.

[0030] In some embodiments the one or more graphene connections in the strain sensors may be cylindrical, i.e. forming a quarter ring in each quadrant between the electrodes. In other embodiments the one or more graphene connections may be in a square / diamond configuration, i.e. forming a straight connection in each quadrant between the electrodes.

[0031] Preferably the one or more graphene connections within each strain sensor are substantially identical so that they form a network of identical parallel resistances between the electrodes.

[0032] In certain embodiments the graphene connection(s) comprise(s) ordered cracks. Ordered cracks can provide for a highly sensitive strain sensor. Graphene with ordered cracks can provide a reliable linear response to small uniaxial strains (e.g. from 0.1% to 5%) with a gauge factor over 100. In certain embodiments the device further includes a starching layer formed between the substrate and the strain sensors. The starching can provide a smooth and uniform surface on the substrate for the application of subsequent elements. Alternatively or additionally the starching can improve the adhesion of the subsequent elements (such as the electrode and graphene) to the substrate which may reduce the likelihood of delamination under mechanical stress.

[0033] The starching layer can provide high stiffness which can modify the rigidity of selected areas on the textile, redistributing strain caused by large-scale body movements such as rolling or turning during sleep which are unrelated to the wearer’s breathing patterns which are desired to be detected.

[0034] In certain embodiments the starching layer is only formed in the vicinity of each strain sensor. This can allow the remainder of the substrate layer to retain its natural properties which may assist in terms of, for example, stretchability and breathability.

[0035] Preferably the wearable device is washable and / or breathable. This can allow the device to be comfortable to the wearer and to be reused on a regular basis.

[0036] Washability of the device can be understood to mean that the performance of the device is not significantly affected by multiple washes. For example a device with performance degradation of less than 20%, preferably less than 10%, more preferably less than 5% after a large number of washes (e.g. more than 500 or more than 1000) could be considered washable.

[0037] Breathability is often considered in terms of the moisture vapour transmission rate (MVTR). Generally materials with an MVTR of over 5,000 g / m2 / 24h are considered highly breathable. Preferably the wearable device of this aspect has an MVTR of over 10,000 g / m2 / 24h and preferably around 20,000 g / m2 / 24h or higher.

[0038] Preferably the device is configured to detect the breathing patterns of the wearer. More preferably the device is configured to detect breathing patterns of the wearer whilst the wearer is sleeping. Such breathing patterns may include nasal breathing, mouth breathing, snoring, bruxism, central sleep apnoea and obstructive sleep apnoea. The device may further include a wired or wireless connector for transmitting signals from the strain sensors. The signals from the strain sensors may be transmitted to processing electronics (which may include a processor and / or memory) which is remote from the device. Alternatively or additionally the wearable device may have processing capability itself.

[0039] In certain embodiments the electrodes and / or strain sensors or both are printed onto the substrate, for example by screen printing or other additive manufacturing methods.

[0040] In certain embodiments the substrate is a textile material. This can allow easy integration of the device into standard clothing items, for example as part of the collar of a shirt or t-shirt. This integration can be achieved by attaching the wearable device to an existing garment design using known techniques, or the components of the wearable device may be formed on a garment during its production.

[0041] Direct printing on garments allows for the integration of multi-functional electronic elements directly onto the garment with scalability and design flexibility.

[0042] As described in the embodiments below, examples of multi-channel graphene textile strain sensors according to this aspect can provide very high sensitivity (gauge factor >100), scalability (±20% conductivity fluctuation), and durability (stable over 10,000 cycles of stretching tests) to continuously monitor subtle vibrations of the extrinsic laryngeal muscles while ensuring user comfort.

[0043] The wearable device of this aspect may thus enhance diagnostic precision in sleep monitoring. It may also contribute to therapeutic interventions for sleep-related disorders.

[0044] The wearable device of this aspect may provide for personalized medicine strategies in relation to detected patterns in the user’ s breathing, including but not limited to breathing during sleep.

[0045] The wearable device of this aspect may also be used for personal health tracking, for example in conjunction with and communication with a user’s portable device (such as a smartphone, smart watch, etc.) and software running on that device. The device of the above aspect may include some, all or none of the above-described optional and preferred features in any combination.

[0046] A further aspect of the present invention provides a system for determining breathing patterns of a subject, the system including: a device according to the above first aspect, including some, all or none of the optional and preferred features of that aspect; and a processor, wherein the processor is configured to process signals produced by the strain sensors when the device is worn by the subject in order to determine breathing patterns of the subject whilst the subject is sleeping.

[0047] In certain embodiments the processor applies a trained neural network to the signals produced by the strain sensors to determine the breathing patterns.

[0048] The present inventors have developed a suitable neural network model. Explainable artificial intelligence visualizations confirm that this model comprehensively understands the sleep patterns, avoiding biases towards noisy regions, thus demonstrating its robustness. Moreover, transfer learning tests show that after few-shot learning (with only 15 samples per class), the model can achieve up to 95% classification accuracy on new users, showcasing the system's powerful generalization capabilities.

[0049] The processor may be configured to select a signal from a single one of the strain sensors and to use the selected signal (and potentially only the selected signal) to determine the breathing pattern. The inventors have found that the strain sensors on the wearable device generally detect the same patterns of movement, but at different intensities, depending on the position of each strain sensor. Therefore selection of one of the signals is sufficient for accurate determination, whilst reducing the total amount of input data to be processed.

[0050] This can also mean that the wearable device does not need to be accurately or consistently positioned on the wearer in order to produce useful outputs. This can make the wearable device suitable for repeated use in a non-clinical setting without any input from a medical professional.

[0051] In certain embodiments the processor is configured to determine and select the most powerful signal from the signals produced by each of the strain sensors and to use that selected signal. Preferably the processor is configured to determine and distinguish between a plurality of predefined breathing patterns. More preferably the processor is configured to determine and distinguish between each of nasal breathing, mouth breathing, snoring, bruxism, central sleep apnoea and obstructive sleep apnoea.

[0052] The processor may be configured to categorise the subject’s breathing pattern into one or more of the plurality of predefined breathing patterns. The categorisation may be performed for selected segments of the subject’s breathing and segments may be categorised differently.

[0053] The system may be configured to display or otherwise output the results of the determination and / or categorisation.

[0054] The system of the above aspect may include some, all or none of the above-described optional and preferred features in any combination.

[0055] A further aspect of the present invention provides a method of manufacturing a device according to the above-described first aspect (including some, all or none of the optional and preferred features of that device), in which the various components are formed on the textile substrate by an additive manufacturing process such as screen printing.

[0056] Unless indicated otherwise, any of the features (including the optional or preferred features) described in relation to one of the above aspects are equally applicable in combination with the devices, systems and methods of any of the other above-described aspects.

[0057] The invention is described below, by way of example, with reference to the accompanying figures in which:

[0058] Figure la shows an overview of a wearable device according to an embodiment of the present invention;

[0059] Figure lb shows the variability in the vibration responses produced by different sleep patterns;

[0060] Figure 2 shows the detailed composition of a strain sensor array forming part of a wearable device according to an embodiment of the present invention; Figure 3 shows, schematically, the steps in the manufacture of the strain sensor array of Figure 2;

[0061] Figure 4 shows a sensing simulation for the strain sensor array of Figure 2;

[0062] Figure 5 shows the formation of ordered cracks in the graphene elements of the strain sensor array of Figure 2;

[0063] Figure 6 shows how starching a textile layer can affect the penetration depth of graphene ink;

[0064] Figure 7 shows the resistance response to different cyclic tensile strains during testing of the strain sensor array of Figure 2;

[0065] Figure 8 shows the dynamic response to cyclic tensile strain of the strain sensor array of Figure 2 at different frequencies;

[0066] Figure 9 shows the results of a durability test for the strain sensors in the strain sensor array of Figure 2 over 10,000 cycles;

[0067] Figure 10 shows the strain response of the stretchable silver electrodes in the strain sensor array of Figure 2;

[0068] Figure 11 shows the resistance and gauge factor for 50 different strain sensors such as those used in embodiments of the present invention;

[0069] Figure 12 shows the results of a washability test on the strain sensor array of Figure 2;

[0070] Figure 13 shows various characteristics of a strain sensor array forming part of a wearable device according to an embodiment of the present invention;

[0071] Figure 14 shows the effects of positioning the strain sensor array of Figure 13 in different locations on the wearer’s throat; Figure 15 shows, schematically, the components of a sleep behaviour recognition model used in a system according to an embodiment of the present invention;

[0072] Figure 16 is a confusion matrix showing the classification of six sleep patterns using the model shown in Figure 15;

[0073] Figure 17 compares the model shown in Figure 15 with known neural network backbones;

[0074] Figure 18 is a process visualisation of random hyperparameter optimisation in the model shown in Figure 15;

[0075] Figure 19 shows the receiver operating characteristic (ROC) curves for classification of sleep patterns using the model shown in Figure 15;

[0076] Figure 20 shows SmoothGrad distributions of the contributions that signals make to the classification output in the model shown in Figure 15;

[0077] Figure 21 shows t-SNE visualisations comparing the distribution of raw data to features extracted from the model shown in Figure 15;

[0078] Figure 22 shows a flowchart of a generalisation process using the model shown in Figure 15 and the results of that process; and

[0079] Figure 23 is a relative intensity plot of breathing data obtained from a dog wearing a collar adapted to incorporate a strain-sensor array as illustrated in Figures 1 and 2.

[0080] Figure 1 provides an overview of a wearable device 1 according to an embodiment of the present invention as integrated into a system according to an embodiment of the present invention and the processing performed on signals from the wearable device 1.

[0081] Figure la shows the wearable device 1 as worn around the neck of a user. The device is a textile garment with a six-channel strain sensor 10 incorporated into the collar area. Whilst the device in Figure la is shown as being positioned around the neck of a human user, it will be appreciated that the wearable device may be formed with the same overall structure whilst being configured to be worn around the neck of another animal such as a dog or horse.

[0082] The positioning of the garment and strain sensor around the neck and the configuration of the device allow the device to detect subtle vibrations at the extrinsic laryngeal muscle, which are induced by physiological vibrations emanating from various anatomical locations such as the velum, oropharynx, tongue, and epiglottis. These vibrations are captured by the strain sensor array 10. The signals from the channel with the strongest response (highest power density) are processed by a deep learning neural network, which is designed to recognize and analyze sleep patterns.

[0083] Figure lb provides a visualisation of the signals of six different sleep patterns as detected by the strain sensor array (single, strongest channel response), demonstrating that these patterns exhibit distinctive behaviours across epochs.

[0084] Figure 2a illustrates the strain sensor array 10 in more detail. The strain sensor array 10 consists of six individual strain sensors 20 arranged in a regular 3x2 rectangular grid configuration. It will be appreciated that other configurations of the strain sensors 20 can be used to make up the array 10. For example, a square or rectangle with 4, 8, 9, 10, 12 or more strain sensors could be used. Alternatively other regular or irregular arrangements (including other polygonal configurations could be used.

[0085] The strain sensor array 10 is screen-printed onto a high-neck top made of elastic knitted fabric 12 using a multi-layer screen printing process. The fabrication process is illustrated schematically in Figure 3 and the individual steps are described further below. The number of prints refers to the number of cycles the substrate is printed for. During each cycle, the silk frame is lowered to the substrate, the flood blade travels backward to spread the ink, the squeegee travels to print, and the silk frame is raised to leave the substrate.

[0086] The textile substrate was first washed with detergent, thoroughly dried, and then treated with UV-ozone for 5 minutes to clean the surface. Screen printing was performed using a 165T polyester silk screen on a semi-automatic printer (Kippax & Sons Ltd.) set with a squeegee angle of 45 degrees, a spacer of 2mm, a coating speed of lOmm / s, and a printing speed of 40mm / s. Printing pressure was pneumatically controlled, with higher pressure applied for the viscous starching agent, and moderate pressure for the thinner graphene ink and silver ink to reduce penetration. After each printing pass, the textile was blown to dry. After printing, the sensor was washed with water to remove CMC Na and dried at 80 °C overnight. A biaxial strain of around 10% was then applied to induce the formation of ordered cracks.

[0087] Overall, the fabrication process is scalable and compatible with industrial garment printing processes, making it feasible for mass production of smart garments.

[0088] A starching layer 14 consisting of sodium carboxymethyl cellulose (CMC Na) and polyurethane acrylate (PUA) is printed onto the interconnect and sensing areas to customize the rigidity and printability of the substrate. Both cellulose derivatives and acrylates are common starching agents in garment printing industry

[0029] , CMC Na is a water-soluble polymer derived from cellulose known for its film-forming ability, which provides a smooth and uniform surface

[0030] , CMC Na also improves the adhesion of the printed graphene ink to the substrate, reducing the likelihood of delamination under mechanical stress. The PUA provides high stiffness and excellent adhesion by forming a robust, cross-linked network upon UV exposure

[0031] , Introducing a rigid PUA layer in the textile strain sensor array modifies the rigidity of selected area on the textile, redistributing strain caused by body movements during sleep.

[0089] The isolated area remains inert to large-scale uniaxial stretching, ensuring that local strain is accurately measured without interference, as shown in Figure 4. Figure 4a shows the setup of a sensing simulation for the PUA-isolated textile strain sensing array, with the Young's modulus of PUA set at 1 GPa. A I N force is applied by a ball to the centre of the array to simulate skin contact. Figure 4b shows the results of the simulation which reveals 40% uniaxial strain isolated from the array, with significant localized strain of 4.8% at the centre where the ball contacts, while other sensing areas show a 2.6% strain. This means that various interfering sources such as physiological or external artefacts, such as heart rate or macroscopic movements of the wearer or the device during sleep, resulting in uniaxial strain to the device, are not detected or do not affect the performance of the individual strain sensors 20.

[0090] The UV-curable nature of the starching materials also prevents clogging of the screen during the printing process, extending the lifespan of the printing screen and maintaining consistent printing quality. After starching treatment, crossbar silver electrodes 14 separated by an insulating layer 15 are printed, followed by the graphene sensing layer 16. The graphene sensing layer comprises exfoliated graphene flakes bound with ethyl cellulose (EC). The graphene flake is selected to be under 1 pm before ink formulation to ensure stable dispersion and avoid mesh clogging. During the screen-printing process, the ink is squeezed through the mesh by the squeegee, depositing a thin film onto the substrate. The stress concentration at the boundaries of the textile structural units induces the formation of regular cracks as shown in Figure 5 which shows this structure schematically (top) and with a SEM image (bottom). This process does not require complex pre-stretching or pre-treatment steps, making it compatible with garment processing.

[0091] However, capillary forces can cause the ink to spread, while air pockets can block ink deposition, leading to variability and poor printing quality. Excessive penetration of graphene ink into the textile can create a graphene / textile composite, which is insensitive to strain and acts as an extra conductive pathway to the surface cracking layer.

[0092] To inhibit ink penetration and air trapping during the printing process, another starching treatment with CMC Na was used to create a controlled surface for ink deposition, preventing deep penetration and ensuring the graphene forms a brittle surface layer that cracks under strain.

[0093] As shown in Figure 6, the penetration depth of graphene ink in starched textile is significantly lower compared to untreated textile. Figure 6a is an SEM image of textile directly printed with graphene ink, while Figure 6b is an SEM of textile treated with CMC Na solution and printed with graphene ink. Dashed boxes label the ink penetration depth in each sample.

[0094] The graphene / EC coating with ordered cracks shows a reliable linear response to small uniaxial strains from 0.1% to 5% with a gauge factor over 100, as shown in Figure 7 which shows the resistance response to cyclic tensile strains of 5%, 1%, 0.5%, and 0.1%; the inset shows a zoomed-in view of the response at 0.1%.

[0095] Moreover, the strain sensors 12 demonstrate a rapid response to straining cycles with frequencies ranging from 1Hz to 10 Hz, enabling the real-time monitoring of fast and subtle vibrations produced by the throat during sleep. This is shown in Figure 8 which shows the results of a dynamic response test under uniaxial cyclic tensile at 1% strain with different frequencies: 10 Hz, 5 Hz, 2 Hz, and 1 Hz. The inset shows the gauge factor at each frequency.

[0096] The durability of the graphene sensor was tested through tensile tests at 1% uniaxial strain and 1 Hz frequency, showing consistent performance as demonstrated in Figure 9 which shows a durability test of the graphene strain sensors under 10,000 cycles of 1% strain.

[0097] In contrast to the high response of the strain-sensing layer, the stretchable silver electrodes exhibit high conductivity with negligible strain response, ensuring stable electrical connections as shown in Figure 10 which shows the strain response of the silver electrodes under 1% strain.

[0098] To ensure consistency and reliability, the performance distribution of two-terminal resistance and gauge factor was studied across 50 separate strain sensor units. A resistance variation of less than 12.69% and a gauge factor variation of less than 9.16% were achieved by controlling the printing conditions and applying the starching treatments, as shown in Figure 11.

[0099] The washability of the sensors was tested by immersing the devices in water under 500 rpm stirring for 5 to 60 minutes, drying at room temperature. The performance degradation after washing was less than 10%. Similar results were observed when detergent was added. Figure 12 shows the results of the washability test under room temperature (RT) and 500 rpm stirring with a magnetic stir bar. Grey dashed lines show the maximum and minimum values measured during the experiment.

[0100] The textile substrate is known for its breathability, and after being integrated into a device, our smart garment still maintains excellent breathability. A comparison was made of pure textile and the textile substrate printed with the strain sensor array in the moisture vapor transmission rate (MVTR) test (20 °C , 59% RH). There was no decrease of the breathability after printing the sensor array on the textile substrate. Besides, the breathability of the printed textile strain sensor array (-20,000 g / m2 / 24h) is much better than that of Tegaderm (a commonly used medical dressing manufactured by 3M), which is usually 800-2000 g / m2 / 24h [9],

[0101] A biocompatibility test was also carried out which involved comparison of wearing the textile sensor and an Ecoflex substrate for 8 hours (a typical sleep period). No sign of irritation or side effect was observed in the textile sensor area, whilst skin irritation appeared in the Ecoflex area.

[0102] Monitoring of sleep behaviour with the strain sensor array

[0103] The wearable device 10 of the above embodiment was tested on volunteers. The six-channel strain sensor array 12 was integrated into the collars of textile garments which were positioned around the participants' necks. This setup was designed to collect minute vibrational signals from the extrinsic laryngeal muscles, which vary according to different sleep behaviours, as shown schematically in Figure 13a.

[0104] A multiplexer was used to read the responses from each circular six-channel piezoresistive strain sensor arranged in a cross-bar structure for further analysis. In the design of the strain sensor array 12 as discussed above, four circular sensing channels are formed by the graphene quarter-ring connections between the crossing electrodes. This configuration can be equivalently viewed as a parallel connection of four variable resistors, as shown in Figure 13b. This design provides two-dimensional sensitivity to both horizontal and vertical strain.

[0105] Figure 13c illustrates the response signals during a 10-second epoch of nasal breathing, captured by the six channels at standard wearing positions. Correlation analysis revealed that although the intensity of the strain responses varied across different channel locations covering the throat area, the signal characteristics were highly correlated (Pearson correlation coefficients greater than 0.9 between any two channels). This pattern persisted even when the sensor array was worn askew, not in the standard position (see Figure 14, in which the triangle in the configurations at the top shows the position of the centre of the throat, which shows the variation between the four tested sites).

[0106] These findings demonstrate that the wearable device according to embodiments of the invention only actually requires a single strain sensor, as increasing the number of sensors does not change the information available. However, providing an array of strain sensors, such as (but not limited to) the array of six sensors shown in the above embodiment and illustrated and tested in Figures 13 and 14, which utilizes a relatively large coverage area of the strain sensor array, ensures that the region with the strongest response falls within the device sensitive area. As a result, the wearable device according to embodiments of the invention does not require precise positioning, which gives a level of resilience to positional variances that augments the practicality of this monitoring system in real-world settings.

[0107] Furthermore, due to the high correlation between channels, subsequent pattern recognition (for example neural network-based pattern recognition as described below) only needs to consider the channel with the strongest response as being representative, which improves the inference efficiency of the pattern recognition algorithm while ensuring accuracy.

[0108] To verify the reliability of the proposed approach, a comprehensive dataset was obtained encompassing six distinct sleep behaviour patterns: nasal breathing, mouth breathing, snoring, bruxism, CSA, and OSA. The dataset was collected from seven healthy subjects and spans a health spectrum from healthy and sub-healthy to high-risk categories. As detailed further below, for the three conditions (bruxism, OSA, and CSA) which are rare among healthy individuals, the data were simulated under the guidance of medical experts. In contrast, data for mouth breathing, nasal breathing, and snoring were collected from actual sleep sessions.

[0109] As visualized in Figure 13d, the temporal and spectral characteristics of these sleep behaviours were meticulously analyzed focusing on the signals emanating from the channel exhibiting the strongest response. It was observed that the effective vibrational signals originating from the extrinsic laryngeal muscles are predominantly found within the low-frequency domain, specifically below 10 Hz. This frequency band captures the physiological nuances of each sleep pattern, enabling a precise delineation of the behaviours. Nasal and mouth breathing exhibit fundamental differences in their spectral signatures, reflecting variations in airflow mechanics and potential diagnostic markers for respiratory efficiency. The irregular and prominent vibrational patterns associated with snoring suggest disrupted airflow dynamics and may serve as indicators of upper airway resistance. Notably, the episodic high-amplitude signals of bruxism provide clear evidence of nocturnal teeth grinding, which could be associated with high stress levels or sleep disturbances. The absence of vibrational activity during respiratory pauses in CSA and the erratic signal fluctuations indicative of breathing efforts against obstruction in OSA are consistent with the clinical understanding of these two conditions.

[0110] Sleep patterns recognition with a deep learning model After simple pre-processing, which involved labeling, segmentation (10s signals were segmented into one sample), and Z-score normalization, signals from the channel with the strongest response are fed into a specially designed deep learning model for sleep pattern recognition. The pipeline for this approach is illustrated in Figure 15.

[0111] The neural network is composed of three core components: First, the learnable positional encoder adopts a residual bidirectional LSTM (BiLSTM) framework to understand the sequential nature of the input data. This approach surpasses traditional sinusoidal positional encoding by dynamically learning positional information, which is particularly advantageous in sleep pattern analysis where the temporal relationship between events can signify different breath cycles or disturbances

[0033] , Second, the multi-head self-attention module, derived from Transformer architecture, enables nuanced discrimination of significance across the data sequence. By effectively utilizing long-range dependencies, the model ensures a comprehensive understanding of the sequential data, reflecting the true complexity of sleep behaviours over time

[0034] , Lastly, the one-dimensional Residual Network (ResNet) layers function as a hierarchical feature extractor. With their ability to skip connections, they prevent the vanishing gradient problem and enhance the flow of information, thus allowing the model to learn more complex patterns effectively. The ID convolution within these layers ensures that the model is tailored to handle time-series data, offering a nuanced analysis of temporal patterns

[0035] ,

[0112] Figure 16 displays the confusion matrix for the model's classification of sleep patterns. As can be seen from Figure 16, the model achieved strong classification results: the accuracy for each category was above 95%, with an overall accuracy of 98.6%.

[0113] Figure 17 shows a comparison of model efficiency (measured in FLOPs), accuracy, and number of parameters with known state-of-the-art neural network backbones. As shown in Figure 17, the model surpassed the performance of its individual component architectures, such as the Transformer and ID ResNet, as well as other state-of-the-art models in terms of accuracy. An additional data point is provided in Figure 17 which represents the same model described above, but with a pruning of the least important 50% of the nodes in the ID-ResNet module of the model and re-training to obtain a pruned model. As shown in Figure 17, not only did the FLOPs (a critical indicator of model inference speed and energy efficiency) decrease by 30% in this pruned model, but there was also a slight increase in accuracy. This can be attributed to the phenomenon known as "pruning-induced efficiency", where removing redundant or less important nodes can lead to a more streamlined and efficient network. The re-training phase helps the model to re-allocate its resources towards the most salient features, potentially improving generalization and thus accuracy

[0036] ,

[0114] Figure 18 illustrates the process of hyperparameter optimization for the model. It can be observed that in the majority of hyperparameter combinations, the model exhibits high accuracy (greater than 90%). This indicates that the model's performance is not overly sensitive to the specific values of its hyperparameters, demonstrating the model's robustness. Moreover, such robustness might also imply that the fundamental features learned by the model are strong predictors of sleep patterns, allowing for decent performance despite variations in model configuration.

[0115] Figure 19 shows the ROC curves for the model's classification of each sleep pattern type. The AUC values are almost equal to 1 for each classification task, indicating that the model is effective and has achieved satisfactory classification performance.

[0116] SmoothGrad visualizations in Figure 20 provide a clear depiction of how different segments of the signal contribute to the model's classification decisions

[0037] , The shading intensity represents the degree to which each point in the time series influences the output, with darker shades indicating higher importance. For instance, the “Nasal Breath” and “Mouth Breath” classifications exhibit consistent contribution patterns throughout the breathing cycle, reflecting the model's reliance on rhythmic features for these classes. Conversely, “Snoring” shows a more variable contribution pattern, which likely corresponds to the erratic nature of snoring events. The “Bruxism” class demonstrates pronounced contributions at peaks, which may correspond to teeth grinding instances. In the cases of “CSA” and “OSA”, the model identifies critical contributions at pause cycle and obstruction cycle. This distribution of contributions is consistent with established physiological patterns, which means that the model gives appropriate weight to relevant features across different classes, reflecting a balanced understanding of the data rather than an overreliance on certain input aspects that could lead to skewed predictions (e.g. noise). Furthermore, this underscores the model's capacity to not only recognize but also assign appropriate significance to the distinct temporal features within the complex landscape of sleep-related signals, enhancing the interpretability of its predictions. Figure 21 compares the distribution of raw data with the features extracted by the model in the t-SNE plane, illustrating the model's ability to discern and delineate the complex structure within the data. The t-SNE visualization of extracted features shows distinct clusters corresponding to different sleep patterns, which are not as discernible in the raw data. This indicates the effectiveness of the model's feature extraction in capturing the underlying relationships and patterns necessary for accurate classification.

[0117] To assess the model's generalization ability, transfer learning tests were conducted, applying the model trained on six participants to the dataset of a new participant. As shown in Figure 22, with only 15 samples per category for few-shot learning, the model achieved an accuracy of up to 95% on the new subject. In contrast, training only on the new participant's dataset without transfer yielded a few-shot learning accuracy of merely 80%. This reflects the model's ability to adapt and maintain performance across different individuals, showcasing its robust transferability and capacity to leverage previously learned patterns to quickly adapt to new, unseen data.

[0118] Decoding human sleep patterns is important yet complex. Despite the encouraging development of wearable devices for sleep health monitoring in recent years, creating a system that combines versatility, comfort, and accuracy remains a significant challenge. In this work, we have designed a smart garment integrated with a six-channel strain sensor array. This ultrasensitive strain sensor array, characterized by its excellent robustness and durability, can collect subtle vibrations from the extrinsic laryngeal muscles associated with various sleep patterns, and its multi-channel design eliminates the need for positioning due to its spatial resolution. Despite utilizing only a single modality of strain response signals, our smart garment, equipped with a customized deep learning neural network, can comprehensively analyze and recognize subtle vibrations originating from various physiological sites and transmitted to the extrinsic laryngeal muscles. It accurately classifies six sleep patterns: nasal breath, mouth breath, snoring, bruxism, CSA, and OSA. Additionally, it can efficiently and effectively adapt to new users, maintaining high accuracy in its classifications. We believe our smart garment offers a promising solution for versatile sleep monitoring in wearable devices, not only suitable for the consumer electronics market to provide ongoing sleep monitoring for general users but also as a convenient, low-cost alternative for clinical sleep monitoring. In another embodiment, which is not illustrated in the figures, a veterinary smart collar is provided for canine respiratory monitoring. An adjustable dog collar (inner circumference 330-440 mm; width 45 mm) was manufactured which incorporates a strain-sensor array as described with reference to Figures 1 and 2 above. The array is centred ventrally on the subject dog so that three channels span the junction of the sternohyoid and sternothyroid muscles. Raw resistance data (100 Hz) are streamed to a laptop, where neural-network pipeline of Figure 15 is executed. Figure 23 shows the relative intensity plot of the dog’s detected breathing over a short period.

[0119] Materials

[0120] TIMREX KS 25 Graphite (particle size of 25pm) was sourced from IMERYS. Stretchable conductive silver ink was obtained from Dycotec Materials Ltd. Ethyl cellulose and sodium carboxymethyl cellulose were purchased from SIGMA-ALDRICH. Flexible UV Resin Clear was acquired from Photocentric Ltd. The textile substrate, composed of 95% Polyester and 5% spandex, was procured from Jelly Fabrics Ltd.

[0121] Ink Formulation

[0122] The graphene ink for screen printing was prepared following a reported method. Briefly, 100g of graphite powder and 2g of ethyl cellulose (EC) were mixed in IL of isopropyl alcohol (IP A) and stirred at 3000 rpm for 30 minutes. The mixture was then added into a high-pressure homogenizer (PSI-40) at 2000 bar pressure for 50 cycles to obtain graphene dispersion. The graphene dispersion was centrifuged at 5000g for 30 min to remove unexfoliated graphite. To prepare the CMC Na starching solution, CMC Na was dissolved in water at 20% wt. concentration.

[0123] Characterization of Structure and Performance

[0124] The size distribution of graphene flakes was analyzed using a Bruker Icon Atomic Force Microscope (AFM) in an area approximately 20 pm x 20 pm. Scanning Electron Microscopy (SEM) images were taken with a Magellan 400, after sputtering the textile samples with a 5 nm layer of gold to enhance conductivity. Optical images were captured using an Olympus microscope. Experimental setup of data acquisition

[0125] For data acquisition, a potentiostat (EmStat4X, PalmSens) and a multiplexer (MUX8-R2, PalmSens) were used as the primary readout modules. These modules consistently supplied a IV voltage, with the resulting output being the current passing through the strain sensors. A sampling frequency of 100Hz was selected and the data segmented into 10-second epochs for detailed analysis. The data collection process was specifically crafted to reflect real-world conditions, accommodating variations in the positioning and tightness of the collar with each use.

[0126] Throughout the data collection process from various participants, avoided strict calibration of the collar’s position or tightness was deliberately avoided. Participants were simply advised to wear the smart garment comfortably and put the collar around their necks, ideally positioned at the mid- to upper-throat level. This method ensured that the dataset obtained represented a wide range of real-life scenarios, capturing the inherent variability in the collar’s positioning and tightness across different users and experimental setups.

[0127] Sleep behaviour dataset collection

[0128] All participants were healthy students (7 participants, average age 25, 4 males and 3 females). As a result, they did not exhibit significant symptoms of Bruxism, CSA, or OSA. Therefore, these three types of sleep behaviour were simulated following training under the guidance of medical experts. For the collection of bruxism, simulated instances of grinding, clenching, and tapping were included; for CSA, participants were instructed in voluntary end-expiratory central apnea during breathing; for the more challenging simulation of OSA, participants were trained to utilise the Muller manoeuvre to maintain a lower intrathoracic pressure

[0038] and followed the characteristic descriptions of reports from the American Association of Sleep Medicine (AASM), introducing clinical SpO2 as an auxiliary simulation tool, marking a segment as a valid OSA pattern only when SpO2 continuously fell below 90% within the epoch, or a continuous decrease of more than 4% from baseline

[0039] , These simulations were carefully developed with clinicians to ensure the primary mechanics of airway obstruction and breath manoeuvre conform realistic situations. The data for the other three behaviours (nasal breath, mouth breath, and snoring) were collected during actual sleep states, with the breathing mode identified and checked using video footage and simultaneous audio recording. All data collection was performed in a supine position.

[0129] In total, 2119 epochs were collected, including 728 epochs of nasal breath, 701 epochs of mouth breath, 262 epochs of snoring, 180 epochs of bruxism, 102 epochs of CSA, and 146 epochs of OSA.

[0130] The forgoing description is exemplary in nature only, and the skilled person will understand that changes and variations on the disclosed embodiments are possible within the scope of the claims. The claims define the invention.

[0131] REFERENCES

[0132] [1] Hyun, S., Hahm, H. C., Wong, G. T. F., Zhang, E. & Liu, C. H. Psychological correlates of poor sleep quality among U.S. young adults during the COVID-19 pandemic. Sleep Medicine 78, 51-56 (2021).

[0133] [2] Wang, P. et al. Prevalence and associated factors of poor sleep quality among Chinese older adults living in a rural area: a population-based study. Aging Clinical and Experimental Research 32, 125-131 (2019).

[0134] [3] Berhanu, H., Mossie, A., Tadesse, S. & Geleta, D. Prevalence and Associated Factors of Sleep Quality among Adults in Jimma Town, Southwest Ethiopia: A Community-Based Cross-Sectional Study. Sleep Disorders https: / / www.hindawi.com / joumals / sd / 2018 / 8342328 / (2018).

[0135] [4] Jefferson, Y. Mouth breathing: adverse effects on facial growth, health, academics, and behaviour. Gen Dent 58, 18-25 (2010).

[0136] [5] Pevemagie, D., Aarts, R. M. & Meyer, M. D. The acoustics of snoring. Sleep Medicine Reviews 14, 131-144 (2010).

[0137] [6] Manfredini, D., Ahlberg, J., Winocur, E. & Lobbezoo, F. Management of sleep bruxism in adults: a qualitative systematic literature review. Journal of Oral Rehabilitation 42, 862- 874 (2015).

[0138] [7] Malhotra, A. et al. Metrics of sleep apnea severity: beyond the apnea-hypopnea index. Sleep 44, zsab030 (2021).

[0139] [8] Bussadori, S. K., Motta, L. J., Horliana, A. C. R. T., Santos, E. M. & Martimbianco, A. L. C. The Current Trend in Management of Bruxism and Chronic Pain: An Overview of Systematic Reviews. Journal of Pain Research 13, 2413-2421 (2020).

[0140] [9] Parati, G., Lombardi, C., Castagna, F. et al. Heart failure and sleep disorders. Nat Rev Cardiol 13, 389-403 (2016).

[0141]

[0010] Lee, S. W. H., Ng, K. Y. & Chin, W. K. The impact of sleep amount and sleep quality on glycemic control in type 2 diabetes: A systematic review and meta-analysis. Sleep Medicine Reviews 31, 91-101 (2017).

[0142]

[0011] Fino, E. & Mazzetti, M. Monitoring healthy and disturbed sleep through smartphone applications: a review of experimental evidence. Sleep and Breathing 23, 13-24 (2018).

[0143]

[0012] Kwon, S., Kim, H. & Yeo, W.-H. Recent Advances in Wearable Sensors and Portable Electronics for Sleep Monitoring, i Science 24, 102461 (2021).

[0144]

[0013] Guillodo, E. et al. Clinical Applications of Mobile Health Wearable-Based Sleep Monitoring: Systematic Review. JMIR mHealth and uHealth 8, el0733 (2020).

[0145]

[0014] Tang, C. et al. A roadmap for the development of human body digital twins. Nature Reviews Electrical Engineering 1, 199-207 (2024).

[0146]

[0015] Bosi, M. et al. Qualitative Phenotyping of Obstructive Sleep Apnea and Its Clinical Usefulness for the Sleep Specialist. International Journal of Environmental Research and Public Health 17, 2058 (2020).

[0016] Shen, Q. et al. Multitask Residual Shrinkage Convolutional Neural Network for Sleep Apnea Detection Based on Wearable Bracelet Photoplethysmography. IEEE Internet of Things Journal 9, 25207-25222 (2022).

[0147]

[0017] Asgari Mehrabadi, M. et al. Sleep Tracking of a Commercially Available Smart Ring and Smartwatch Against Medical-Grade Actigraphy in Everyday Settings: Instrument Validation Study. JMIR mHealth and uHealth 8, e20465 (2020).

[0148]

[0018] Pan, Q., Brulin, D. & Campo, E. Current Status and Future Challenges of Sleep Monitoring Systems: Systematic Review. JMIR Biomedical Engineering 5, e20921 (2020).

[0149]

[0019] Yue, O. et al. Skin-inspired wearable self-powered electronic skin with tunable sensitivity for real-time monitoring of sleep quality. Nano Energy 91, 106682 (2022).

[0150]

[0020] Honda, S., Hara, H., Arie, T., Akita, S. & Takei, K. A wearable, flexible sensor for realtime, home monitoring of sleep apnea. iScience 25, 104163 (2022).

[0151]

[0021] Sun, J. et al. Multifunctional wearable humidity and pressure sensors based on biocompatible graphene / bacterial cellulose bioaerogel for wireless monitoring and early warning of sleep apnea syndrome. Nano Energy 108, 108215-108215 (2023).

[0152]

[0022] Tarim, E. A., Erimez, B., Degirmenci, M. & Tekin, H. C. A Wearable Device Integrated with Deep Learning-Based Algorithms for the Analysis of Breath Patterns. Advanced intelligent systems 5, 2300174 (2023).

[0153]

[0023] O’Hare, E., Cogan, J. A., Dillon, F., Lowery, M. & O’Cearbhaill, E. D. An Intraoral Non-Occlusal MEMS Sensor for Bruxism Detection. IEEE Sensors Journal 22, 153-161 (2021).

[0154]

[0024] Kwon, S. et al. At-home wireless sleep monitoring patches for the clinical assessment of sleep quality and sleep apnea. Science Advances 9, eadg9671 (2023).

[0155]

[0025] Tabar, Y. R. et al. At-home sleep monitoring using generic ear-EEG. Frontiers in Neuroscience 17, 987578 (2023).

[0156]

[0026] da Silva Souto, Carlos F., et al. Pre-gelled Electrode Grid for Self-Applied EEG Sleep Monitoring at Home. Frontiers in neuroscience 16, 883966 (2022).

[0157]

[0027] Yin, J., Wang, S., Tat, T. & Chen, J. Motion artefact management for soft bioelectronics. Nature Reviews Bioengineering 1-18 (2024) doi:https: / / doi.org / 10.1038 / s44222-024-00175- 4.

[0158]

[0028] Kotte, S. & Kumar Dabbakuti, J. R. K. Methods for removal of artifacts from EEG signal: A review. Journal of Physics: Conference Series 1706, 012093 (2020).

[0159]

[0029] Tian, Y. et al. Applications of adhesives in textiles: A review. European Polymer Journal 167, 111089 (2022).

[0160]

[0030] Rahman, M. S. et al. Recent Developments of Carboxymethyl Cellulose. Polymers 13, 1345 (2021).

[0161]

[0031] Tabasum, Shazia, et al. Properties of the modified cellulosic fabrics using polyurethane acrylate copolymers. Carbohydrate polymers 94, 866-873 (2013).

[0032] Tang, C., Xu, M., Yi, W. et al. Ultrasensitive textile strain sensors redefine wearable silent speech interfaces with high machine learning efficiency, npj Flex Electron 8, 27 (2024).

[0162]

[0033] Weld, H., Huang, X., Long, S., Poon, J. and Han, S.C. A survey of joint intent detection and slot filling models in natural language understanding. ACM Computing Surveys 55, 1-38 (2022).

[0163]

[0034] Vaswani, A. et al. Attention is all you need. In Proc. 31st International Conference on Neural Information Processing Systems 6000-6010 (Curran Associates Inc., 2017).

[0164]

[0035] He, K., Zhang, X., Ren, S. and Sun, J. Deep Residual Learning for Image Recognition. openaccess.thecvf.com. (2016).

[0165]

[0036] Zhu, M. and Gupta, S. To prune, or not to prune: exploring the efficacy of pruning for model compression. arXiv: 1710.01878. (2017).

[0166]

[0037] Smilkov, D., Thorat, N., Kim, B., Viegas, F. and Wattenberg, M. SmoothGrad: removing noise by adding noise. arXiv: 1706.03825. (2017).

[0167]

[0038] Gaisl, T. et al. Simulated Obstructive Sleep Apnea Increases P-Wave Duration and P- Wave Dispersion. PLoS ONE 11, eO 152994 (2016).

[0168]

[0039] Berry, Richard B., et al. Rules for scoring respiratory events in sleep: update of the 2007 AASM manual for the scoring of sleep and associated events: deliberations of the sleep apnea definitions task force of the American Academy of Sleep Medicine. Journal of clinical sleep medicine 8, 597-619 (2012).

[0169] All of the above references are hereby incorporated by reference in their entirety.

Claims

CLAIMS1. A wearable device for detecting movements of a wearer, device having: a substrate which is flexible and elastically deformable; a plurality of electrodes formed on the substrate; and a plurality of strain sensors formed on the substrate and configured to detect movement of the substrate, wherein: each strain sensor is formed at the intersection of a pair of said electrodes and comprises an insulating layer separating the pair of electrodes and one or more graphene connections conductively linking the electrodes in a plurality of orientations, such that relative movement between the pair of electrodes causes a change in the resistance of the graphene connection or of each of one or more of the graphene connections.

2. The wearable device of claim 1 wherein there are at least four strain sensors forming an array.

3. The wearable device of claim 2 wherein the electrodes are arranged in a grid pattern such that there are a plurality of electrodes running in a first direction in the plane of the device and a plurality of electrodes running in a second direction, substantially perpendicular to the first direction, in the plane of the device.

4. The wearable device of any preceding claim wherein the one or more graphene connections form a cylindrical shape.

5. The wearable device of any preceding claim wherein the graphene connection comprises ordered cracks.

6. The wearable device of any preceding claim further including a starching layer formed between the substrate and the strain sensors.

7. The wearable device of claim 6 wherein the starching layer is only formed in the vicinity of each strain sensor.

8. The wearable device of any preceding claim wherein the device is washable and / or breathable.

9. The wearable device of any preceding claim wherein the device is configured to be worn around the neck of the wearer.

10. The wearable device of any preceding claim wherein the device is configured to detect the breathing patterns of the wearer.

11. The wearable device of any preceding claim, further including a wired or wireless connector for transmitting signals from the strain sensors.

12. The wearable device of any preceding claim wherein the electrodes and / or strain sensors are printed onto the substrate.

13. The wearable device of any preceding claim wherein the substrate is a textile material.

14. A system for determining the breathing patterns of a subject, the system including: a device according to any one of the preceding claims; and a processor, wherein the processor is configured to process signals produced by the strain sensors when the device is worn by the subject in order to determine breathing patterns of the subject whilst the subject is sleeping.

15. The system of claim 14 wherein the processor applies a trained neural network to the signals produced by the strain sensors to determine the breathing patterns.

16. The system of claim 14 or claim 15 wherein the processor selects a signal from a single one of the strain sensors and uses the selected signal to determine the breathing pattern.

17. The system of any of claims 14-16 wherein the processor is configured to determine and distinguish between each of nasal breathing, mouth breathing, snoring, bruxism, central sleep apnoea and obstructive sleep apnoea.

18. A method of manufacturing a wearable device, the method including the steps of:screen printing a plurality of electrodes onto a flexible and elastically deformable substrate; forming a plurality of strain sensors, each at the intersection of a pair of said electrodes, by forming an insulating layer separating the pair of electrodes at the intersection and one or more graphene connections conductively linking the electrodes in a plurality of orientations.

19. The method of claim 18 further including the step of applying a starching treatment to the substrate prior to the formation of the electrodes and strain sensors.

Images