Dietary monitoring

Abstract

There is provided a system for classifying food consumed by a user. The system comprises wearable apparatus comprising a sensor located so as to, in use when the apparatus is worn by a user consuming a food item, monitor an area of skin whose movement is correlated with a facial muscle associated with movement of the jaw. The sensor is configured to capture time-varying data describing movement of the area of skin in a plane defined by the area of skin. The system further comprises a processor comprising a machine learning system configured to process the captured time-varying data so as to classify the food item according to one or more of a predetermined plurality of texture qualities, each texture quality representing a physical characteristic of the food item. The food item may be classified as belonging to a predetermined food group.

Description

FIELD OF THE INVENTION This invention relates to monitoring dietary intake of a user. BACKGROUND OF THE INVENTION It is known that eating too quickly, eating too frequently, and eating unhealthy types of food can be linked to being overweight or obese. Many people are not aware of the speed at which they consume food, or how often they eat during the day. They may not track the types of food they are eating very carefully. Keeping track of these habits can help a person to eat more mindfully and lose weight. For example, eating slowly, taking smaller bites, and longer chewing, promote better satiety through prolonged exposure to food. Rapid eating, on the other hand, is associated with faster gastric emptying and reduced response to satiety-related hormones. Tracking what type of food is being eaten provides the individual with a more accurate picture of their eating habits and may prompt them to think more carefully about what they choose to eat. It can also be important to monitor food intake for reasons other than weight loss. For example, certain medications may require a person to not eat at certain times, or to not eat certain food types. Similarly, some individuals have intolerances to certain foods which are difficult to identify due to the long interval between consumption and gastrointestinal symptoms. Monitoring the eating habits of someone experiencing cognitive decline is especially important as malnutrition is commoner in the group affected by neurodegenerative diseases. There are various difficulties with accurately monitoring dietary intake of a person. Commonly used methods require the person to manually log their meals and snacks in a smartphone app or diary. These methods are inaccurate as they rely on the user remembering to log their food and logging it correctly. It is common for users to under-report what they have consumed. Other methods of determining what food someone has eaten involve chemical tests which are invasive and may be impractical for day-to-day monitoring. Emerging technologies such as wearable sensors offer the potential to improve eating assessment methods by passively measuring eating activity with minimal user interaction. Recent advances in wearable sensors for eating detection are disclosed in “Automatic, wearable-based, in-field eating detection approaches for public health research: a scoping review by NPJ Digit Med. 2020; 3: 38. Published online 2020 Mar 13. doi: 10.1038 / s41746-020-0246-2. Current methods of food analysis using wearable sensors involve use of a camera to capture photos of what the user is eating and perform image analysis on the photos to identify the food being eaten. The accuracy of these sensors in naturalistic settings is uncertain. There are technical, analytical, and multidisciplinary challenges of conducting in-the-wild measurements with wearable sensors. These challenges include sensor noise, sensor occlusion (in the case of cameras), and privacy concerns. The motion of the jaw during eating can provide useful insights into the food being eaten. In a laboratory, a mandibular kinesiograph has been used to track movement of the incisal point of a subject when eating three different food items, as detailed in “Predicting Foodstuff from jaw dynamics during masticatory crushing in man” by J.S Wang and C.S Stohler, Archives of Oral Biology Vol 36, Issue 3, pg 239-244,1991. This involved tracking movement of a magnetic device positioned on the teeth of the subject using a bulky experimental apparatus that is not suitable for day-to-day use. In “Adaptation of mastication mechanics and eating behaviour to small differences in food texture” by Benjamin Le Reverend et. al, Physiology and Behaviour, Vol 165, page 136-145, 2016, electromyography (EMG) and kinematics of jaw movements (KJM) techniques were used to follow mastication muscle activity and jaw motion during mastication of seven cereal products. The KJM technique involved tracking the motion of the jaw with a motion capture system using a marker stuck on the skin of the user. There is a need for a wearable apparatus that can accurately monitor dietary intake of a user day-to-day in a minimally invasive and unobtrusive way. SUMMARY OF THE INVENTION According to a first aspect of the present invention there is provided a system for classifying food consumed by a user, the system comprising: wearable apparatus comprising a sensor located so as to, in use when the apparatus is worn by a user consuming a food item, monitor an area of skin whose movement is correlated with a facial muscle associated with movement of the jaw, the sensor being configured to capture time-varying data describing movement of the area of skin in a plane defined by the area of skin; and a processor comprising a machine learning system configured to process the captured timevarying data so as to classify the food item according to one or more of a predetermined plurality of texture qualities, each texture quality representing a physical characteristic of the food item. The machine learning system may be configured to classify the food item by generating a texture classification comprising a classified texture quality for each of the one or more of the predetermined plurality of texture qualities. The system may comprise a camera configured to, in use when the user is consuming the food item, capture an image of the food item. The system may further comprise, at the processor, an image classification system configured to operate on the captured image so as to generate an image classification for the food item, the processor being configured to combine the texture classification for the food item with the image classification so as to identify the food item. The wearable apparatus may comprise the camera. The processor may be configured to transform the time-varying data into a frequency domain prior to classifying the food item. The machine learning system may be configured to process the time-varying data by determining one or more predetermined characteristics of the time-varying data. The machine learning system may be trained to correlate the one or more predetermined characteristics with the one or more of the predetermined plurality of texture qualities. The machine learning system may be configured to determine the one or more predetermined characteristics of the time-varying data by transforming the time-varying data into a frequency domain, the one or more characteristics comprising a frequency-domain characteristic. The frequency-domain characteristic may comprise one or more of a dominant frequency, spectral entropy, and energy in one or more frequency bands. The time-varying data captured during consumption of the food item may exhibit at least one cyclic movement of the area of skin. The one or more characteristics of the time-varying data may comprise one or more characteristics of the at least one cyclic movement. The one or more characteristics of the at least one cyclic movement may comprise one or more of: a maximum amplitude of the cyclic movement, a duration of the cyclic movement, an average velocity of the cyclic movement, a maximum velocity of the cyclic movement, and a rate of change of velocity of the cyclic movement. The time-varying data may exhibit a plurality of cyclic movements and the one or more characteristics of the plurality of cyclic movements comprises one or more of: an average frequency of the plurality of cyclic movements, a frequency variability of the plurality of cyclic movements, and a change in maximum amplitude of the plurality of cyclic movements overtime. The machine-learning system may be configured to use a neural network or a Random Forest or Gradient Boosting learning method. The machine learning system may be a neural network. The system may be further configured to receive a user-identified classification of the food item according to the one or more of the predetermined plurality of texture qualities, the system being configured to use the user-identified classification as training data for the machine learning system. The system may be configured to apply a clustering algorithm to a plurality of sets of time-varying data captured by the sensor during consumption of a plurality of respective food items so as to produce a clustered dataset, the system being configured to receive a user-identified classification of the clustered dataset with respect to the one or more of the predetermined plurality of texture qualities. The apparatus may comprise a pair of sensors. Each sensor in the pair may be located so as to, in use when the apparatus is worn by the user consuming the food item, monitor a respective area of skin whose movement is correlated with a corresponding muscle on either side of the face of the user. Each sensor in the pair of sensors may be configured to capture time-varying data describing movement of the respective area of skin in a plane defined by that area of skin. The machine learning system may be configured to process time-varying data captured by the pair of sensors so as to classify the food item according to the one or more of the predetermined plurality of texture qualities. The apparatus may comprise a lip sensor located so as to, in use when the apparatus is worn by the user, monitor an area of skin associated with a lip dilator and / or lip constrictor. The lip sensor may be configured to capture time-varying data describing movement of said area of skin. The machine learning system may be configured to process the time-varying data captured by the lip sensor in combination with the time-varying data captured by the sensor so as to classify the food item according to one or more of the predetermined plurality of texture qualities. The apparatus may further comprise a motion sensor and / or microphone. The machine learning system may be configured to process data captured by the motion sensor and / or microphone in combination with the time-varying data captured by the sensor so as to classify the food item according to one or more of the predetermined plurality of texture qualities. The one or more of the predetermined plurality of texture qualities of a food item may comprise a measure of hardness and / or a measure of cohesiveness and / or a measure of density of the food item. The sensor may be configured to capture time-varying data describing movement of the area of skin in two dimensions in the plane defined by the area of skin. The sensor may be an optical flow sensor configured to capture samples of the area of skin and compare the samples apart in time so as to detect movement of the area of skin relative to the optical sensor. The time-varying data may represent movement of the area of skin in two dimensions in the plane defined by the area of skin. The facial muscle may be one of a temporalis muscle, masseter muscle, a medial pterygoid, and a lateral pterygoid. The area of skin may overlie one or more of: a masticator muscle, a zygomaticus major and / or minor muscle, a levator labii superioris muscle, a levator labii superioris alaeque nasi muscle, a levator anguli oris muscle, a risorius muscle, a buccinator muscle, an orbicularis oris muscle. The processor may be configured to use the texture classification for the food item for monitoring dietary intake of the user. The processor may be configured to use the texture classification of the food item to estimate what food item was consumed. The processor may be configured to estimate what food item was consumed by comparing the texture classification of the food item with texture classifications of previously identified food items. The texture classification of the food item may comprise a coordinate of a texture-space, the texturespace having axes corresponding to each texture quality of the one or more of the predetermined plurality of texture qualities. The system may be configured to estimate what food item was consumed by comparing the coordinate with coordinates of previously identified food items. The system may be configured to receive blood glucose data of the user following consumption of the food item and use the texture classification of the food item and the blood glucose data to estimate what food item was consumed. The system may be configured to predict blood glucose data of the user following consumption of the food item using the texture classification of the food item and optionally a history of previously consumed food items. According to a second aspect, there is provided a method of classifying food consumed by a user, the method comprising: during consumption of a food item by a user, capturing time-varying data representing movement of an area of skin of the user in a plane defined by the area of skin, the movement of the area of skin being correlated with a facial muscle associated with movement of the jaw of the user; and at a machine learning system trained to correlate time-varying data with one or more of a predetermined plurality of texture qualities, processing the captured time-varying data so as to classify the food item according to the one or more of the predetermined plurality of texture qualities, each texture quality representing a physical characteristic of the food item. The method may comprise: forming training data by associating time-varying data with one or more of the predetermined plurality of texture qualities, the time-varying data representing movement of an area of skin of a user in a plane defined by the area of skin during consumption by the user of a food item having physical characteristics corresponding to the one or more of the predetermined plurality of texture qualities, and the movement of the area of skin being correlated with a facial muscle associated with movement of the jaw of the user; and training the machine learning system using the training data to correlate the time-varying data with the one or more of the predetermined plurality of texture qualities so as to enable the machine learning system to classify food according to the predetermined plurality of texture qualities. The forming training data may comprise associating each of a plurality of items of time-varying data with a respective set of one or more of the predetermined plurality of texture qualities. The method may comprise, prior to training the machine learning system, transforming the timevarying data into a frequency domain and training the machine learning system using the transformed time-varying data. Forming the training data may comprise determining one or more characteristics of the time-varying data, and associating the time-varying data with the one or more predetermined plurality of texture qualities comprises associating the one or more characteristics of the time-varying data with the one or more predetermined plurality of texture qualities. The time-varying data may be captured during consumption by the user of a plurality of food items, each having physical characteristics corresponding to the one or more of the predetermined plurality of texture qualities. The machine-learning system may be configured to use a neural network or a Random Forest or Gradient Boosting learning method. The machine-learning system may be a neural network. The one or more predetermined texture qualities of a food item may comprise a measure of hardness and / or a measure of cohesiveness and / or a measure of density of the food item. According to a third aspect, there is provided a system for classifying food consumed by a user, the system comprising: wearable apparatus comprising a sensor located so as to, in use when the apparatus is worn by a user consuming a food item, monitor an area of skin whose movement is correlated with a facial muscle associated with movement of the jaw, the sensor being configured to capture time-varying data describing movement of the area of skin in a plane defined by the area of skin; and a machine learning system trained to perform food classification based on a plurality of known food items belonging to a plurality of predetermined food groups, the machine learning system being configured to classify the food item being consumed as belonging to one of the predetermined plurality of food groups. BRIEF DESCRIPTION OF THE DRAWINGS The present invention will now be described by way of example with reference to the accompanying drawings. In the drawings: Figures 1a and 1b illustrate possible skin locations sampled by wearable apparatus shown in figure 2 and the locations of some facial muscles of interest. Figure 2 shows a pair of glasses as an example of the wearable apparatus. Figure 3 shows a schematic diagram of a system for detecting mouth activity. Figure 4 shows an example method for identifying a jaw activity. Figure 5 shows an example of steps that may occur to identify an eating episode. Figures 6a to 6d show an example of time-varying data captured by a temple sensor describing movement of an area of skin associated with the temporalis muscle in a plane defined by the area of skin during the beginning of an eating episode. Figures 7a to 7d show an example of time-varying data captured by a cheek sensor describing movement of an area of skin associated with the zygomaticus muscle in a plane defined by the area of skin during the beginning of an eating episode. Figure 8 shows an example of time-varying data captured by a temple sensor when the user is clenching their jaw. Figure 9 shows an example of time-varying data captured by a temple sensor when the user presses their tongue against the roof of the mouth. Figure 10 depicts an example method for monitoring a mouth activity. Figures 11a to 11c depict examples of time-varying data captured by a cheek sensor when the user is eating different types of food. Figure 12 shows an example texture space coordinate system. Figure 13 illustrates the different ways in which the mouth opens when eating different foods. Figure 14 is a schematic of an example processor. Figure 15 shows an example method for classifying food consumed by a user using a trained machine learning system. Figure 16 shows an example method fortraining a machine learning system. DETAILED DESCRIPTION OF THE INVENTION The following description is presented by way of example to enable a person skilled in the art to make and use the invention. The present invention is not limited to the embodiments described herein and various modifications to the disclosed embodiments will be apparent to those skilled in the art. It is desirable to detect, identify, and monitor mouth activities of a user accurately and automatically using a minimally invasive wearable device. In particular, it is desirable to detect, identify, and monitor jaw activities of a user. It is further desirable to detect, identify, and monitor eating activities of a user and provide feedback. Furthermore, it is desirable to use additional device data to understand the context of the user behaviour such as location, pose, posture, and physical activity. It is also desirable to detect the relationship between emotional responses and mouth activities of a user such as jaw clenching. A system for detecting a mouth activity of a user is described herein. More specifically, the system may be a system for detecting a jaw activity of a user. In particular, the system may be a system for detecting an eating activity of a user. Whilst the system may advantageously be used to monitor eating activities such as for weight loss applications, the system may also be used to detect bruxism (repeated clenching of the jaw), determine asymmetry in jaw motion, determine limitations in the range of jaw motion, and detect other mouth activities such as pushing the tongue against the teeth. Detection of mouth activities such as these may be important in identifying temporomandibular joint problems or looking for early warning signs of certain conditions. Identification of mouth activities can also be combined with information from other sensors that may be incorporated into a wearable device, for example regarding activation of other facial muscles that allow a contextual analysis of the mouth activity. Sensors worn on the head can be used to infer the location (e.g. using GPS), pose (e.g. sitting, standing, lying down), posture (e.g. the tilt of the head), emotional response and physical activity of the user. In combination with data from these sensors, a wearable device can provide more detailed understanding of eating behaviour, pathological jaw function / pain, or emotional, cognitive or physical strain. For example, some people clench their teeth, particularly when experiencing emotional or cognitive or physical strain. The system described herein can be used to infer the context of jaw clenching which may provide insights into the emotional state of the user. In some cases, mouth activities may also be used to control a computer or Al, e.g. for hands-free control of a device. For example, jaw motions can be used to indicate that the user wants to engage with a particular device. A pre-configured gesture could be used to "handshake" with the device and initiate further interactions. For example, the user may clench their jaw three times consecutively to activate an Al assistant. Head orientation and user location could be used to determine what the user is looking at. When the jaw motion is considered in the context of location data and head orientation, and potentially also a camera incorporated into the wearable apparatus, the Al assistant may be able to infer what actions the user wishes to take with the Al. For example, a user wanting to know the time might look down (in the same way that they would look at a wristwatch) and clench their jaw. This could tell the Al that the user wanted the time which could be given by audio or vision means. In another example, when navigating outdoors, the user might want to request navigation advice. When the user gets to a junction (which may be determined by GPS on the wearable apparatus) and is unsure on where to go, they might look left and right and clench the jaw to summon directions from an Al assistant. Such applications will be described in more detail later. Herein, a “mouth activity” refers to the activation of muscles associated with the mouth. In other words, a mouth activity is the activation of a muscle that can cause movement within the interior of the mouth, and / or movement of an exterior part of the mouth, and / or movement of the skin immediately surrounding the mouth. Mouth activities may include one or more of: opening the mouth, closing the mouth, mastication (chewing), drinking, swallowing, spitting, licking, speaking, singing, breathing, smiling, kissing, whistling, yawning, sneezing, coughing, blowing, movement of the tongue within the mouth (e.g. against the teeth), side-to-side jaw movement, front-to-back jaw movement, a jaw clench, and bruxism (repeated clenching of the jaw). Facial muscles associated with the mouth may include one or more of the muscles of mastication (the masseter, the temporalis, the medial pterygoid, and the lateral pterygoid), the tongue, the cheek muscles (the zygomaticus major and minor, the levator labii superioris, the levator labii superioris alaeque nasi, the levator anguli oris, buccinator), the risorius, the depressor anguli oris, the depressor labii inferioris, the orbicularis oris, and the mentalis. A “jaw activity” is a type of mouth activity that involves the activation of muscles associated with the jaw (i.e. jaw muscles). Jaw activity may include one or more of opening the mouth, closing the mouth, chewing, side-to-side jaw movement, front-to-back jaw movement, movement of the tongue against a part of the mouth, a jaw clench, and bruxism. Herein, references to mouth activities may include jaw activities. The muscles of mastication are muscles associated with the jaw. They control movement of the lower jaw (the mandible). The muscles of mastication may include the masseter, the temporalis, the medial pterygoid, and the lateral pterygoid. Each of these muscles may be called a masticator muscle herein. The activation of the jaw using any of the mastication muscles is a mouth activity, and more specifically, a jaw activity. Mastication (chewing) is a jaw activity because it involves the activation of the masticator muscles. Bruxism is a jaw activity because it involves activation of the temporalis muscle. Pressing the tongue against different parts of the mouth such as the front teeth or anterior palate may be considered a jaw activity because it involves (at least) the activation of one or more jaw muscles such as the masseter and / or temporalis. A “consuming” activity is a type of mouth activity. A consuming activity occurs when the user is consuming a substance such as food and / or drink and / or medication. Consuming activities may include one or more of opening the mouth, closing the mouth, chewing, sucking, swallowing, and more generally eating and drinking. An “eating activity” is a type of mouth activity, and, more specifically, a type of “consuming activity”. Eating activities may include one or more of opening the mouth, closing the mouth, chewing, and swallowing. Whilst the system described herein is predominantly described in the context of detecting jaw activities, it may be used for detecting mouth activities more generally such as those listed above. The inventors have recognized that mouth activities such as chewing can be accurately detected by measuring lateral movement of an area of skin that moves when the mouth moves. By measuring lateral movement of the skin (e.g. movement in the plane of the skin), rather than movement out of the plane of the skin, it can be possible to differentiate between different kinds of mouth activities, as well as determine metrics relating to these activities. In particular, the inventors have recognized that jaw activity can be identified by detecting one or more cyclic movements of the area of the skin in the plane of the skin. Features of the cyclic movements) can be used to identify the jaw activity. Figure 1a illustrates some of the muscles associated with the jaw and some example areas of skin that may be monitored by the system described herein. Not all the mouth muscles listed above have been labelled on figure 1a for clarity. Figure 1b shows the masticator muscles during a side-to-side grinding motion whilst chewing. One of the muscles associated with the jaw is the temporalis muscle 1. The temporalis muscle is near the temple and extends downwards in a direction towards the mouth. There is a temporalis muscle on each side of the face. The temporalis muscle controls movement of the lower jaw that can result in the opening and closing of the mouth. The cheek muscles such as zygomaticus major and minor are generally indicated at 2. Herein, references to the zygomaticus muscle mean either or both the zygomaticus major and minor. Also labelled at the risorius 7 and the buccinator 8. Figure 1 b depicts the other muscles of mastication muscles: the masseter 3, medial pterygoid 4 and the lateral pterygoid 5. Also shown in figure 1b is the temporomandibular joint 6. To detect a jaw activity, an area of skin whose movement can be correlated with the activity of a facial muscle associated with jaw movement is monitored. In other words, there is a relationship between movement of the area of skin being monitored and the activation of a facial muscle associated with movement of the jaw. For example, the area of skin being monitored may be correlated with (e.g. associated with) the masticator muscles. The area of skin being monitored need not overly a facial muscle directly associated with movement of the jaw. For example, the area of skin may overly a different facial muscle to one that is associated with movement of the jaw, provided that the movement of the area of skin can still be correlated with a facial muscle associated with movement of the jaw. For example, in figure 1a, the area of skin 22 may be monitored to detect jaw activity, even though it overlies the zygomaticus muscle 2 and not a masticator muscle. This is because the activation of the muscles associated with the jaw can cause movement of the area of skin 22 overlying the zygomaticus muscle. As another example, the area of skin being monitored 33 may be by the hinge of the jaw below the ear, as depicted in figure 1a. Whilst this area of skin is not directly overlying any facial muscle, activation of the muscles associated with the jaw can cause movement of this area of skin. In other examples, the area of skin may overly a facial muscle directly associated with movement of the jaw. For example, the area of skin 11 overlying the temporalis muscle 1 may be monitored. Figure 2 shows an example wearable apparatus 200 that is a part of a system 300 for determining a mouth activity of a user. In the example shown in figure 2, the wearable apparatus is a pair of glasses. The apparatus may alternatively be a headset, headband, hat, earphone, earpiece or any other wearable device or article that can act as a platform from which a sensor can monitor the relevant area of skin of the user. At least one sensor (e.g. a sensor 201) is positioned on the wearable apparatus. The sensor is located on the wearable apparatus so that when the wearable apparatus is worn by a user, the sensor can monitor an area of skin whose movement is correlated with a facial muscle associated with movement of the jaw, as described above. For example, one of the sensors 201 may monitor the area of skin 11 associated with temporalis muscle 1 when the glasses are worn by the user. As another example, one of the sensors 202 may monitor an area of skin 22 associated with the zygomaticus muscle 2. As described above, this is because movement of the area of skin 22 can be correlated with a facial muscle associated with jaw movement such as the temporalis muscle 11. In some examples, the area of skin may preferably overlie an insertion point of a facial muscle associated with movement of the jaw. By monitoring an area of skin overlying an insertion point of a facial muscle, a greater range of skin movement can be seen when that facial muscle is activated when compared with monitoring an area of skin positioned over the middle of the muscle. The area of skin being monitored may be in the range 0.1mm2 to 30mm2. The size of the area of skin monitored by the sensor may depend on the muscle or muscles associated with that area of skin, and / or the surface texture and shape of the skin in the area. The area of skin monitored by the sensor need not be circular or uniform in shape. For example, a sensor located so as to monitor an area of skin associated with the temporalis muscle may monitor a relatively large area of skin such as in the range 5mm2 to 20mm2. As the temporalis muscle acts over a relatively wide area, the sensor can monitor a relatively large area of skin without being in danger of picking up skin movement caused by a different muscle to the temporalis. In another example, a sensor located so as to monitor an area of skin associated with the levator anguli oris (near the edge of the lip) may monitor a relatively small area of skin in the range 0.1mm2 to 5mm2. This may avoid the sensor detecting movement of the area of skin that is caused by a different muscle to the levator anguli oris. By monitoring areas of skin that are associated with predominantly one muscle, data collected by the sensor can be more accurately attributed to a particular mouth activity or muscle activation. Alternatively, a sensor may monitor an area of skin on the face of the user that is associated with more than one muscle. For example, sensor 202 may monitor one area of skin that is associated with both the zygomaticus muscle and the orbicularis oris muscle (the muscle surrounding the lips). A sensor may monitor more than one area of skin associated with a respective muscle. For example, sensor 202 may monitor an area of skin associated with the zygomaticus muscle and separately an area of skin associated with the orbicularis oris muscle. As facial muscles tend to work together, it may be the case that one mouth activity will be caused by the movement of multiple related facial muscles. The mouth activity may be determined by monitoring one area of skin that is associated with multiple related facial muscles. It may not always be necessary to determine which muscle caused the movement of the area of skin as the ultimate goal may be only to determine the mouth activity, rather than specific muscle activity. The sensor (e.g. sensor 201) is configured to capture time-varying data describing movement of the area of skin in two dimensions in a plane defined by the area of skin. In other words, the sensor is configured to capture time-varying data describing lateral movement of the area of skin. That is, the direction of movement of the area of skin being measured by the sensor is along the surface of the skin. In some examples, this may be a direction transverse to the direction of data capture by the sensor. The time-varying data captured by the sensor may be one or more of: vector information (e.g. a coordinate or a pair of coordinates expressing a direction and magnitude of movement in the plane defined by the skin); a time series of images or other optical representations of the area of skin; or data in any form which captures time-varying contrast at the sampled area of skin (e.g. due to features on the skin, hairs, or its texture). For example, in figure 1a the area of skin 11 defines a plane in which a sensor (e.g. sensor 201) measures movement. The x-y axes in figure 1a depict two example directions in which movement may be measured in the plane 11. In this example, the area of skin 11 is shown as an x-y plane. In examples in which the sensor captures data along the z-axis, the x-y plane is transverse to the direction of data capture by the sensor. As the temple area is relatively flat, the area of skin monitored by the sensor may be relatively large so as to define a relatively large plane (e.g. 10mm2) in which movement may be measured. In examples in which the skin being monitored is particularly curved or bumpy, the area of skin may be smaller (e.g. 0.1mm2). In some examples, the data captured by the sensor may be substantially in the plane defined by the area the skin. In an example, the sensor is configured to capture time-varying data describing movement of the area of skin in two directions in the surface of the area of skin. It can be particularly advantageous to capture data describing movement of the area of skin in two directions (e.g. a first direction and a second direction) in the plane of the skin. The first and second directions do not need to be orthogonal, although they may be. In an example, the first and second directions may be labelled x and y directions. The x and y directions may be orthogonal. By capturing data in both the first and second directions in the plane of the skin, a richer source of information can be obtained and used to determine the user’s mouth activities, as will be explained below. In some examples, the sensor may be configured not to measure movement of the area of skin in a direction transverse to the plane defined by the area of skin. In other words, the sensor may be configured not to capture time-varying data describing movement of the area of skin in a direction out of the surface of the area of skin. In an example, the sensor is not a singular proximity sensor. The sensor may be any type of sensor that is adapted to capture time-varying data describing movement of the area of skin in two dimensions in a plane defined by the area of skin. In an example, the sensor may be an optical sensor. An optical sensor may be any sensor configured to use light (e.g. visible or infra-red light) to detect movement. The optical sensor may image the area of skin associated with a facial muscle of a user. In an example, the optical sensor is an optical flow sensor. An optical flow sensor is a sensor configured to capture (e.g. visually) samples of the area of skin and compare the captured samples apart in time so as to detect movement of the area of skin relative to the sensor. A surface sample captured by an optical flow sensor may be any kind of representation of the surface. This may be a representation of light modified by the surface such as a matrix, an interference pattern or speckle pattern created by the surface interacting with electromagnetic radiation such as a laser directed at the surface. Suitable optical flow algorithms may be used to track changes between samples which represent movement of the sampled surface. A suitable optical flow algorithm may be of the type used in computer mice to track relative movement of the mouse over a surface. The optical flow sensor outputs information representative of movement of the area of skin being monitored relative to the optical flow sensor- i.e., the “optical flow” of the skin movement. This may be in the form of a vector describing the magnitude and direction of the skin movement. For example, the sensor 201 shown in figure 3 is an optical flow sensor having a light sensor 206 and an optical flow processor 207 for processing the output of the light sensor so as to determine relative movement between the sensor 201 and the area of skin. The optical flow processor 207 provides an output representative of movement of the skin relative to the sensor. In an example, the optical flow processor 207 may be configured to capture lateral skin movement (i.e. skin movement in the plane of the skin) using suitable optical flow algorithms configured to track movement of a surface. The light sensor 206 may be configured to capture an array of pixels, e.g. a 16x16 array, 18x18 or 30x30 array of pixels. In alternative examples, the light sensor 206 may comprise a single pixel and the optical flow sensor may comprise a tens or other optica! element configured to randomise the light passing through it and onto the single pixel sensor (e.g. by means of a suitable translucent material or a randomising array of mirrors). The light sensor may be, for example, a Charge Coupled Device (CCD) or a Complementary Metal-Oxide-Semiconductor (CMOS) sensor. The optica! flow processor 207 and the light sensor 206 may be provided as a single integrated circuit. Alternatively, the optical flow processor 207 may be separate to the light sensor 206 and may be external to the sensor 201 at the wearable apparatus. For example, the optical flow sensor may transmit data obtained from the light sensor 206 to the external processor 207 for determining relative movement. As another example of an optical sensor, the sensor may be an event-based detector. For example, the sensor may be a Metavision ® sensor produced by Prophesee ®. When the event-based detector monitors an area of skin, rather than capturing frames at a predefined frequency, the sensor only registers changes (e.g. “events”) in the surface of the skin as they occur. In this way, less redundant data is collected and / or transmitted because the sensor is only taking a measurement when it detects motion or a change in the area being monitored. Sampling can also occur at a much higher frequency than when using conventional frame-based optical sensors. As another example, the optical sensor may be a camera combined with a motion detection algorithm. As another example, the optical sensor may be a millimeter wave detector. In some examples, the sensor may be configured to make use of expected directions of movement of the area of skin being monitored and ignore motion in other unexpected directions. Due to the underlying anatomy of the human face, areas of skin on the face can be expected to move in certain directions, depending on the muscle(s) being activated. For example, when a person smiles, there are defined flow directions (e.g. directions of movement) for parts of the face such as the skin near the lips or the cheeks. The sensor may ignore motion of the area of skin in directions that are not associated with the activation of the muscle(s) underling the area of skin being monitored. This processing may be performed at a processor external to the sensor, and / or it may be performed within the sensor. For example, the event-based sensor described above may incorporate a filter that filters out or ignores motion in directions that are not expected for the area of skin being monitored by that sensor. Preferably, the wearable apparatus may comprise a plurality of sensors. In particular, the apparatus may comprise a plurality of sensors of the type described above. That is, the apparatus may comprise a plurality of sensors each located so as to, in use when the apparatus is worn by a user, monitor a respective area of skin whose movement is correlated with a facial muscle associated with movement of the jaw of the user. Each sensor of the plurality of sensors may be configured to capture time-varying data describing movement of the respective area of skin in two dimensions in a plane defined by that area of skin. For example, the apparatus in figure 2 comprises four sensors 201 each located so as to monitor an area of skin associated with a temporalis muscle of the user, and three sensors 202 each located so as to monitor an area of skin associated with other facial muscles of the user such as the zygomaticus muscle. Sensors 201 and 202 are of the type described above. Sensors 203, 204 and 205 are a different type of sensor that are not part of said plurality of sensors and will be discussed later. Of said plurality of sensors, two or more may monitor respective areas of skin that are each associated with the same facial muscle of the user. For example, figure 2 shows two sensors 201 on the right-hand side of the glasses 200 from the point of view of the wearer (i.e. the left-hand side in the view in figure 2). Each of the sensors 201 is located on the glasses so as to monitor an area of skin associated with the right temporalis muscle of the user when worn (from the point of view of the user). As the sensors 201 measure activity of the temporalis muscle in this example, they will be referred to herein as “temple sensors” to distinguish them from other sensors described later. Preferably, the wearable apparatus comprises two or more sensors located on the wearable apparatus so as to, in use when the apparatus is worn by the user, monitor respective corresponding areas of skin on opposite sides of the face of the user. Suitably, the two or more sensors may be located so as to monitor respective areas of skin that are each associated with a muscle of a muscle pair on either side of the user’s face. Most facial muscles exist on both sides (i.e. the left and right side) of the human face as part of a pair of muscles. For example, there is a respective temporalis muscle on each side of a human face, one at each temple. Of the plurality of sensors, two or more may be positioned so as to monitor the same muscle pair on either side of the user’s face. In figure 2, for example, there are two sets of temple sensors 201, one on each arm 210 of the glasses. The two sets of sensors 201 each measure the same muscle pair (in this case, the left and right temporalis muscles). Of the plurality of sensors, two or more may monitor respective areas of skin that are each associated with different facial muscles. For example, in figure 2, the wearable apparatus comprises temple sensors 201 for monitoring respective areas of skin associated with the temporalis muscles, and cheek sensors 202 for monitoring respective areas of skin associated with cheek muscles such as the zygomaticus muscles. More generally, the apparatus may comprise a first sensor for measuring movement of an area of skin associated with a masticator muscle (e.g. the temporalis muscle), and a second sensor for measuring movement of an area of skin associated with a different facial muscle (e.g. the zygomaticus muscle). In an example, a third sensor may be located on the wearable apparatus so as to monitor an area of skin associated with a lip dilator and / or lip constrictor. A lip dilator may be one or more of: the zygomaticus major and minor, levator labii, levator labii, superioris alaeque nasi, risorius, depressor anguli oris, depressor labii inferioris, and platysma. A lip constrictor may be one or more of: the orbicularis oris and the levator anguli oris. The sensors may detect a range of movement from 1 pm to 10cm. In an example, a cheek sensor may measure skin movement of 5mm in a vertical direction (e.g. in the y-direction) relative to the user when the user opens their mouth wide. In an example, a temple sensor may measure skin movement of 0.7mm in the y-direction, and 1.5mm of skin movement in the x-direction in the plane of the skin when the user opens their mouth wide. The distance moved by the area of skin during a particular jaw activity will vary depending on the skin quality, age, and local anatomy. The apparatus may comprise othertypes of sensors which are not configured to capture data describing movement of the area of skin in a plane defined by the area of skin. For example, the apparatus may comprise ECG sensors. Figure 2 shows ECG sensors 203 located on the nosepiece of the glasses and on the temple tips of the arms of the glasses, where the glasses contact the skull. ECG sensors measure the electrical activity of the user’s heart using electrodes that contact the skin. In examples in which the wearable apparatus is not a pair of glasses, the ECG sensors may be located at any suitable position where contact with the user’s skin is likely. In an example in which the apparatus is an earpiece, the ECG sensors may be located so as to contact the outer ear of the user. As another example, the apparatus may comprise one or more motion sensors 204 such as an accelerometer or inertial measurement unit (IMU). An IMU may comprise an accelerometer and / or gyroscope and / or magnetometer, for example. The motion sensor measures movement of the wearable apparatus. Suitably, when the wearable apparatus is a pair of glasses, the motion sensor may be positioned towards the front of the glasses such as at the hinges or in the casing around the lenses. This may reduce the amount of damping of the movement of the glasses that may occur closer to the temple tips, where the glasses meet the ears of the user. Alternatively, the motion sensor may be positioned near the ear of the device to capture vibrations related to motion at the temporomandibular joints. Two or more motion sensors may be located at corresponding positions on either side of the apparatus so as to pick up on any asymmetry in the motion of the apparatus. The motion sensor(s) may be configured to act as a microphone if sampled at a high enough frequency. For example, the motion sensor may be sampled at 4000Hz in order to pick up vibrations in the audio range. Otherwise, the motion sensor may sample at 50Hz or adaptively change sampling frequency based on context such as location, or detected activity inferred by motion analysis algorithms. The motion sensor(s) may comprise an accelerometer. The motion sensor(s) may be an accelerometer. An IMU may comprise one or more of said motions sensors. The motion sensor(s) may be an IMU. The apparatus may comprise a motion sensor and / or an IMU which may comprise a (separate) motion sensor. Sensor 204 may alternatively or additionally comprise a microphone. The apparatus may further comprise a camera 205. Figure 3 shows a system 300 for detecting mouth activities. The system 300 comprises the wearable apparatus 200 and a processor 350. As described above, the apparatus 200 comprises at least one sensor 201 configured to capture time-varying data describing movement of an area of skin in two dimensions in a plane defined by that area of skin. Data captured by the at least one sensor is output to the processor 350. The processor processes the time-varying data received from the sensor so as to identify a mouth activity of the user. The at least one sensor may comprise the plurality of sensors described above having the same data capturing capability as the at least one sensor 201. As described above, there may be other sensors within the apparatus that do not have the same data capturing capability as the at least one sensor 201, such as the ECG sensors 203 and motion sensor 204. The processor 350 also receives data captured by these sensors. The outputs from the sensor(s) may be combined at the processor 350. Generally, processor 350 may be configured to perform various types of processing to the outputs from the sensors such as multiplexing, encoding, encrypting and / or compressing the outputs from the sensors. The processor 350 may be a part of the wearable apparatus. For example, the processor 350 may be incorporated into an arm 210 of the glasses 200. In an example, processor 350 may be configured to transmit processed data to an external computer system 380 such as a smartphone for further analysis. For example, after identifying a mouth activity of the user, the processor may provide feedback to a user via a smartphone app. Dashed line 312 indicates a possible boundary between system 300 and an external computer system 380. Alternatively, the processor 350 may be external to the wearable apparatus, such as in the computer system 380. The computer system could be any kind of system having suitable processing capabilities, such as a personal computer, server, smartphone, tablet, smartwatch, or dedicated processing unit. An application running at the smartphone may provide suitable processing capabilities in order to identify a mouth activity from the sensor data. In the example in which the processor is external to the apparatus, the wearable apparatus may comprise a data transmitter 310 for transmitting data from the sensor(s) to the external processor, e.g. wirelessly according to a Bluetooth protocol. In the case where the processor is incorporated onto the apparatus, the data transmitter 310 may be a part of the processor. Figure 4 shows an example method for identifying a jaw activity of a user using the system 300 described herein. At step S100, data describing movement of the area of skin in a plane defined by the area of skin is captured by at least one sensor. Preferably, the data describing movement of the area of skin is in two directions in the plane of the area of skin. This allows more detailed information to be determined regarding the jaw activities of the user when compared with sampling data only in one direction. Figures 6a to 6d show an example of the data that may be captured at this step using a temple sensor 201 configured to capture skin movement in two dimensions in a plane defined by the area of skin being monitored. The graphs shown in figures 6a to 6d depict the same data but from different perspectives along the same axes. The data is outputted by the sensor and received by the processor 350. At step S110, the time-varying data is processed (e.g. at processor 350) in order to detect at least one cyclic movement of the area of skin in the plane of the skin - e.g. in the two directions (e.g. axes). Figures 6a to 6d show example time-varying data from a temple sensor 201 in which the data is exhibiting multiple cyclic movements. This is most noticeable when the time-varying data is plotted on a 3-D graph of position in each of the two directions against time. Each of the directions may be an axis defined, for example, forthe purpose of processing the time-varying data captured by the sensor at the processor, and / or for the purpose of representing the time-varying data captured by the sensor monitoring the skin movement, and / or by the direction(s) of measurement of one or more of the sensors (e.g. the sensor may capture movement data along each of the axes). A cyclic movement may comprise the combination of an oscillating pattern overtime in each of the two directions individually. For example, as can be seen in figure 6c, the x-direction data exhibits an oscillating pattern over time. Similarly, as can be seen in figure 6d, the y-direction data exhibits an oscillating pattern overtime. Figures 6a and 6b show the cyclic movements in the x and y direction that are formed from the combination of an oscillation in the x direction overtime and an oscillation in the y direction overtime. As can be seen in figures 6a to 6d, the rate of change of the skin movement in each of the x and y directions overtime is not constant. The oscillations in each of the two directions may occur concurrently. The oscillations in each of the two directions may be out of phase with one another. The oscillations in each of the two directions may be sinusoidal or approximately sinusoidal. A cyclic movement does not necessarily require the data to exhibit a circular or oval pattern. A cyclic movement, or cyclic pattern, may comprise oscillating movements in the two directions over time. When plotted in three dimensions that comprise the x and y directions (as labelled herein for convenience) and time, a cyclic movement may represent a spiral or corkscrew shape - e.g. depending on the nature of the oscillations and their relative phases in each of the two directions. This shape may become apparent when the skin movement in two directions is plotted on a 3-D graph of each direction overtime - for example, as shown in Figures 6a to 6d. A cyclic movement need not be a perfectly closed loop in the x-y plane such that movement of the skin returns to exactly the same point as each cycle is completed. In some examples, a cyclic movement may be identified by detecting, in the time-varying data captured by the sensor, a combination of movements which represent the skin periodically returning to approximately the same point each cycle as detected by the sensor. In some examples, a cyclic movement may be identified by detecting a repetition of the gross movement of the skin. A cyclic movement may not be strictly periodic according to a fixed or varying period (although it could be). A cyclic movement may comprise movement of the area of skin in the plane which comprises components of movement in the two directions such that, after a single cycle, the area of skin returns approximately to its starting point. A plurality of cyclic movements may comprise movement of the area of skin in the plane which comprises approximate repetition of the components of movement in the two directions such that, on each cycle, the area of skin passes approximately through its starting point. The starting point of the area of skin may be any position of the area of skin relative to the sensor capturing the movement of that area of skin - the starting point may be position of the area of skin relative to the sensor when the jaw muscles (or more generally the facial muscles) are not activated such that face is at rest. The area of skin may be considered to have returned approximately to its starting point when the position of the area of skin along the two directions is within 25%, 20%, 15%, 10%, 5% or 2% of its starting point as determined by the sensor arranged to monitor the movement of the area of skin in the plane of the skin. To detect a cyclic movement, feature extraction methods or other data analysis techniques or machinelearning models may be used (e.g. by the processor 350) that look for a sequence of one or more cyclic movements. For example, information such as the size, cross-sectional area of the cyclic loop in the x-y plane, direction of movement, and turning points (e.g. where the gradient of the curve changes sign) of a potential cyclic movement may be extracted from the sensor data and analyzed to confirm the presence of a cyclic movement. In another example, the sensor data may be transformed into new data representations utilizing the 2D information from the sensors (e.g. movement measured in the x and y axis) at the same point in time. In an example, a simple 2D representation could be an image of the x-y plot, on top of which deep learning could be utilized to identify a cyclic movement. For example, a machine learning model could be trained to identify a cyclic movement in an image of the x-y plot based on images of sensor data exhibiting different cyclic movements having varying features such as their size, orientation and turning point. Multiple cyclic movements may be detected in the same way as individual cyclic movements. In many cases, a single jaw activity may be immediately preceded or followed by other jaw activities that are responsible for multiple cyclic movements, or a single jaw activity can be responsible for multiple cyclic movements. The processor may be configured to detect a plurality of cyclic movements at step S110. This may have the advantage that anomalies in individual cyclic movements within the plurality of cyclic movements can be ignored, whilst still picking up the majority of the plurality of cyclic movements. At step S120, the processor may determine one or more characteristics of the cyclic movement(s). The characteristics may be used to identify the jaw activity. Step S120 may be omitted, for example if the processor processes the time-varying data using machine learning techniques in which characteristics of the cyclic movement may not be explicitly identified, as described further below. Steps S120 and S110 may be performed simultaneously. The determination of characteristics of the cyclic movement(s) may be a part of detecting whether or not a cyclic movement or a plurality of cyclic movements are occurring. A characteristic of an individual cyclic movement may include any of the following: A maximum amplitude of the cyclic movement. The maximum amplitude may be in either or both of the two dimensions in which the sensor is configured to capture data. For example, the maximum amplitude may be in either or both of the two directions in which the skin movement is captured. An average velocity of the skin movement. The average velocity of the skin movement may be determined by taking two points in time and determining the distance travelled by the area of skin between those two points. A maximum velocity of the skin movement. The velocity of the skin movement at a particular point in time can be found from the gradient of the distance-time curve at that point in time. The maximum velocity of the skin movement is the gradient of the cyclic movement at its steepest point. An acceleration of the skin movement (e.g. the rate of change of the velocity of the skin movement). The acceleration of the skin movement can be found from the rate of change of the gradient of the distance-time curve. A duration of the cyclic movement (e.g. the time over which the cyclic movement is made). The orientation of the cyclic movement. The orientation may be determined by the angle made by the major axis of the cyclic movement with respect to a fixed axis. In examples in which the processor is configured to detect a plurality of cyclic movements, the characteristics of the plurality of cyclic movements may further include any of: The variance of the maximum amplitude of each cyclic movement in the plurality of cyclic movements (e.g. a change in maximum amplitude of each cyclic movement of the plurality of cyclic movements overtime). The average frequency of cyclic movements in the plurality (i.e. number of cyclic movements per unit time). The frequency variability of cyclic movements in the plurality (e.g. whether the frequency of the repeated cyclic movements is constant). In examples in which the wearable apparatus comprises a pair of sensors each located so as to monitor a respective area of skin associated with a corresponding muscle on either side of the face of the user, the processor may determine a measure of asymmetry between the cyclic movement or movements detected in the data from each sensor in the pair of sensors, and identify the jaw movement in dependence on the measure of asymmetry. Each of the above characteristics may be defined separately for the two directions in which the skin movement is captured (e.g. a first and a second direction). For example, the cyclic movement may have maximum amplitude in the first direction and a maximum amplitude in the second direction. The processor may determine some of the above characteristics for part (e.g. half) of a cyclic movement where applicable. For example, the processor may determine some of the above characteristics for an opening-phase of a cyclic movement and separately determine some of the above characteristics for a closing-phase of the cyclic movement. For example, an average velocity of the skin movement may be determined for the opening-phase of the mouth by determining the time taken to reach a maximum amplitude of the cyclic movement from a neutral position. The maximum amplitude of the cyclic movement is indicative of the mouth being at its most open point. This characteristic can be used to determine a jaw metric relating to the speed that the jaw opens. Similarly, the average velocity of the skin movement may be determined for the closing-phase by determining the time taken for the area of skin to return to the neutral position from the maximum amplitude. This provides information regarding how fast the jaw closes. When the user is chewing, the lower jaw partially opens and closes to chew the food. The processor may learn to distinguish the partial opening and closing of the jaw that occurs during chewing from a mouth-opening and mouth-closing event when the mouth is empty (e.g. during singing, talking, or when about to take a bite) based on characteristics of the cyclic moment such as its amplitude, and the context in which it occurs. Based on characteristics of the cyclic movement, the processor may optionally determine one or more jaw metrics of the user. A jaw metric quantifies the movement of the jaw that caused the cyclic movement. The following are examples of a jaw metric: how fast the jaw moved during the cyclic movement (jaw velocity), how wide the mouth opened (bite amplitude), the acceleration of the jaw during the cyclic movement (jaw acceleration), or how long it took for the jaw to open or close (mouthopening duration and mouth-closing duration). Table 1 shows a list of example jaw metrics and the characteristics of the cyclic movement that may be used to determine them. Jaw metric Characteristic of the cyclic movement used to determine jaw metric Bite amplitude Maximum amplitude of the cyclic movement Opening jaw velocity Average velocity of the skin movement during the opening-phase of the cyclic movement. Closing jaw velocity Average velocity of the skin movement during the closing-phase of the cyclic movement. Opening jaw acceleration Rate of change of the velocity of the skin movement during the opening-phase of the cyclic movement. Closing jaw acceleration Rate of change of the velocity of the skin movement during the closing-phase of the cyclic movement. Maximum jaw velocity The maximum velocity of the skin movement during the cyclic movement. Maximum jaw acceleration The maximum value of the rate of change of the velocity of the skin movement during the cyclic movement. Mouth-opening duration Time taken to reach a maximum amplitude of the cyclic movement from a neutral position. Mouth-closing duration Time taken for the area of skin to return to the neutral position from the maximum amplitude. Table 1 In examples in which the wearable apparatus comprises a pair of sensors each located so as to monitor a respective area of skin associated with a respective muscle of a pair of facial muscles, the processor may determine a jaw metric relating to the symmetry of the user’s jaw activity. A measure of jaw symmetry or asymmetry can be determined from a comparison between the data from the pair of sensors during a jaw activity. For example, each sensor in the pair of sensors may be configured to capture time-varying data describing movement of the respective area of skin in at least one dimension in a plane defined by that area of skin. The processor may process the data from the pair of sensors to identify a first cyclic movement of the area of skin monitored by a first one of the pair of sensors, and a second cyclic movement of the area of skin monitored by a second one of the pair of sensors. The first and second cyclic movements may then be compared so as to produce a measure of asymmetry relating to the jaw activity causing the first and second cyclic movements. In one example, base values of each jaw metric may be derived from images or videos captures of the individual’s mouth to calibrate the system. In another embodiment, the co-activation of facial muscles associated with lip dilation may further characterise the bite behaviour. At step S130, the processor identifies a jaw activity. The processor may identify a jaw activity based on the characteristics of the cyclic movement(s). Over time the values of certain characteristics can be associated with particular jaw activities. In some examples, the system may be trained for an individual by getting the user to log different jaw activities so that the processor can learn to associate a range of values for each characteristic with a jaw activity. In an example, the processor may be configured to identify a mouth-opening event in response to detecting a cyclic movement having a maximum amplitude in the dimensions in which the skin movement is captured that exceeds a predefined threshold. For example, the processor may be configured to determine that a mouth-opening event has occurred If the amplitude in the y-directlon exceeds a predefined threshold, and / or if the amplitude in the x-direction exceeds a (different) predefined threshold. The predefined threshold may be updated over time as the system learns the maximum dimensions of the user’s mouth such as their maximum bite amplitude. Based on the characteristics of the cyclic movement(s), the processor may be able to rule out certain jaw activities as not having occurred and may determine that other jaw activities are more likely to have occurred. For example, when talking, the jaw tends to move much more quickly and erratically than when eating. A short duration of a cyclic movement combined with a high average speed of the skin movement may suggest that the user is talking rather than eating. A high maximum amplitude suggests that the user may be taking a bite of food, singing or yawning. As another example, if the processor determines that the duration of the cyclic movement, and more specifically the duration of the opening and closing phase of the cyclic movement is unusually long, it may determine that the user is yawning or singing, rather than eating, for example. As will be described later, combining the data from mouth opening with data from other sensors such as a motion sensor and / or microphone may be used to identify and determine the jaw activity. An example of the latter is to distinguish pathological changes in jaw function such as trismus, or bruxism. Figure 8 shows an example of time-varying data captured by the temple sensor obtained whilst the user clenches their jaw repeatedly, exhibiting bruxism. As can be seen in figure 8, the area of skin being monitored exhibits a cyclic pattern in both directions in the plane of the area of skin. Although bruxism produces similar looking cyclic movements to that during chewing in this example, the maximum amplitude of the cyclic movements during bruxism is significantly (almost 10 times) smaller than the maximum amplitude of a cyclic movement during chewing. In addition, the maximum amplitude of the cyclic movements during bruxism does not reduce over time, unlike during chewing. Thus, the processor may be able to distinguish repeated jaw clenching from chewing based on the maximum amplitude of the cyclic movement(s). As explained further below, the processor may also be configured to use the context in which the cyclic movements occur to help identify the jaw activity. Whilst the system may determine a jaw activity from the detection of one individual cyclic movement of the area of skin, preferably, each cyclic movement is considered in the context of the data preceding and following that cyclic movement. For example, the processor may identify a jaw activity based on characteristics of the cyclic movement in combination with the position of the cyclic movement in a series of cyclic movements. For example, the processor may be configured to only identify a chewing activity if immediately preceding the chewing activity it detected an initial mouth-opening event and a mouth-closing event. This avoids the processor incorrectly identifying a chewing activity when instead the user is repeatedly clenching their jaw, for example. Thus, the processor may attribute a jaw activity to a cyclic movement in dependence on a position of the cyclic movement in a sequence of jaw activities. In addition to using multiple cyclic movements to determine the context in which a particular cyclic movement occurs, the processor may determine a jaw activity based on characteristics of a plurality of cyclic movements, as mentioned above. That is, instead of identifying a jaw activity based on one individual cyclic movement (even when considered contextually), one jaw activity may be responsible for a plurality of cyclic movements. Chewing is an example of this. The processor may be configured to determine chewing based on repeated cyclic movements. The repeated cyclic movements may at least initially have similar characteristics, such as similar duration per cycle, similar amplitudes, and similar maximum speeds. The repeated cyclic movements occur sequentially. Machine Learning The process of identifying a jaw activity from the time-varying data captured by the sensor(s) of the system may be performed using a machine learning algorithm or performed at a machine learning system. For example, a machine learning algorithm may be used in place of steps S110 and S120, or steps S110 and S120 may be performed implicitly (or explicitly) by a machine learning system. A machine learning system for learning jaw activities may be provided at a computer system to which wearable apparatus is connected, such as the external computer system 380. The machine learning system may provide a dataset for use by the processor so as to enable the processor to identify a jaw activity according to the scheme learned by the machine learning system. Alternatively, the machine learning algorithm and / or machine learning system may be provided at the processor 350. The wearable apparatus may be provided with a learning mode in which the output of the sensor is passed through to the machine learning system at the computer system, processor or system on chip or chiplet architecture. In one example, the time-varying data captured by the sensor is used as an input to a machine learning system such as a classifier. By causing a user to perform different jaw activities and informing the machine learning system which activities are being performed, the machine learning system can (using suitable algorithms) learn to identify which cyclic movements are associated with which jaw activities. In other words, the machine learning model may be provided with training data that includes preclassified jaw activities (e.g. already known jaw activities) and the labels associated with each known jaw activity. In this way, the machine learning model can be trained to label certain cyclic movements as corresponding to certain jaw activities. Preferably, the training data includes cyclic movements that correspond to a series of known jaw activities (e.g. each cyclic movement corresponding to a respective jaw activity) and cyclic movements that correspond to singular known jaw activity (e.g. a known jaw activity that corresponds to a plurality of cyclic movements). In this way, the machine learning model may be trained to distinguish between a jaw activity that causes multiple cyclic movements (e.g. chewing) and a jaw activity that causes one cyclic movement (e.g. mouth opening) but which may also be identified based on the context of the cyclic movement. The machine learning system for identifying a Jaw activity using time-varying data from the sensor(s) may be configured to perform some or all of the following steps: Preprocessing - the data from the sensor(s) may be segmented into time windows or frames. This can help in capturing the temporal dynamics of the activities occurring. Filtering - filtering techniques may be applied to remove noise or outliers from the data. Data augmentation - the data may be augmented with variations of the existing data to improve the generalization of the model. Techniques that may be employed in this step include time warping, jittering, or introducing random noise. Feature extraction - the relevant features from each segmented time window may be extracted. The relevant features can be used to represent the underlying patterns in the data. Relevant features may include: statistical features (e.g. mean, standard deviation, skewness, kurtosis); frequency domain features (e.g., Fast Fourier Transform (FFT) coefficients); time-domain features; autoregressive coefficients, or domain-specific features such as features that describe the cyclic patterns visible in the x-y data. Feature Selection - the most relevant features may be chosen to reduce dimensionality and computational complexity. Techniques like Recursive Feature Elimination (RFE) or feature importance from tree-based models can be used here. Model Training - In this step, a feature-based machine learning model is trained (e.g., Random Forest or any other algorithm based on feature input) using a training data set comprising examples of cyclic movement and associated data indicating the mouth activity represented by each cyclic movement, along with - typically - data which does not represent cyclic movement. The training data set may be formed at least in part using features extracted from each segmented time window in the manner described above. The training data may be processed using one or more of the steps described above prior to being input to the machine-learning model. In a more complex machine learning example, the processor may identify a jaw activity based on the cyclic movement of the area of skin using a neural network. For example, data from the sensor that is captured during different jaw activities in which a cyclic movement or a plurality of cyclic movements can be detected in the skin movement may be provided to the neural network as training data. The data from the sensors may be processed using some of the same steps described above, such as the preprocessing and filtering steps and the data augmentation. The neural network may be trained using the training data from the sensors (e.g. as discussed above) by optimizing the network for a specific classification task. As would be known by the skilled person, hyperparameters such as learning rate, batch size, and network architecture, can all be experimented with to achieve optimal performance. Thus, the neural network may be trained using the training data to classify a cyclic movement or a plurality of cyclic movements as corresponding to a jaw activity. The neural network may perform the classification based on the one or more characteristics of the cyclic movement as described above, but this step may not be explicitly performed. The neural network architecture used can be flexible, provided it is suitable for time-series data. Suitable neural network architectures include recurrent neural networks (RNNs), Long Short-Term Memory networks (LSTM), Convolutional Neural Networks (CNNs), Transformer, Informer, Residual networks, Inception networks, and similar. Any combination of layers based on existing deep-learning architectures may be used. In one example, the Spectro Temporal Residual Network (STResNet) can be used. STResNet extracts channel-specific spectro-temporal information. As spectral information, the networks calculate the spectrogram in decibels, i.e., Iog10 of the amplitude spectrogram, for each input window. For the temporal information, the network uses residual blocks that contain CNN layers with 1-dimensional filters. The output from all channel-specific layers is then merged by a fully connected layer. The fully connected layer is followed by a softmax layer, that outputs class probability estimates for each input segment. Once trained, the neural network can take as an input time-varying data from the sensor describing movement of an area of skin in a plane defined by the area of skin that exhibits a cyclic movement and classify the cyclic movement as corresponding to a jaw activity. The output of the neural network is the identified jaw activity. The neural network may be trained and / or implemented on a hardware accelerator in the processor or run as software on the processor. At step S140, optionally, feedback may be provided to a user about the jaw activity that has occurred. In an example in which the jaw activity is chewing, the processor may log the activity as an eating episode and inform the user of the time and duration of the eating episode, once it is over. The user may correct the logged activity if it has been incorrectly classified. This may be fed back to the machine learning algorithm or neural network so as to improve its accuracy. The sensor used to capture the data in the above-described method is not limited to a temple sensor. The sensor can be located so as to monitor any area of skin whose movement is correlated with a facial muscle associated with movement of the jaw. Figures 7a to 7d show a graph of time-varying data captured by a cheek sensor (e.g. sensor 202) configured to monitor an area of skin associated with the zygomaticus muscle. In this example, the cheek sensor is configured to capture data in two dimensions in a plane defined by the area of the skin, as described above in relation to figures 6a to 6d. It can be seen that similar cyclic movements exist in the data captured by the cheek sensor compared to the data captured by the temple sensor. Multiple sensors The above method is described with reference to the data captured by one sensor. Preferably, the processor is configured to identify a jaw activity based on data from a plurality of sensors of the type described above. The time-varying data captured by the plurality of sensors can be compared to identify a mouth activity. For example, steps S100 to S120 may be independently performed as above for data captured by each appropriate sensor in the wearable apparatus. Step S130 may be performed based on a combination of the data captured by the plurality of sensors. For example, the processor may compare the cyclic movement(s) detected from the data from each sensor so as to identify the jaw activity. By looking at data captured from multiple sensors, the accuracy and reliability of the system at identifying a jaw activity can be improved. Suitably, the apparatus comprises at least a first sensor configured to monitor an area of skin associated with the temporalis muscle, and a second sensor configured to monitor an area of skin associated with the cheek muscle such as the zygomaticus muscle. By monitoring an area of skin associated with the temporalis muscle and an area of skin associated with the zygomaticus muscle, the system may be more likely to correctly identify an eating activity because eating activities tend to cause movement of areas of skin around both the zygomaticus muscle and the temporalis muscle. By comparing data captured from both a temple sensor and a cheek sensor, the system is less likely to incorrectly classify a jaw activity such as bruxism (which may only be picked up from the temple sensor data) as an eating activity. In an example, the phase difference between the time-varying data from each sensor may be used to identify the mouth or jaw activity. In an example in which each sensor outputs data in the form of a timevarying vector describing movement of the respective area of skin, the phase difference between the two vectors may be used to help identify the mouth activity. This may be particularly effective when the apparatus comprises a pair of sensors each located on the apparatus so as to monitor respective corresponding areas of skin on either side of the user’s face. When chewing, for example, a user may have a dominant side of the mouth on which they preferentially chew their food, giving rise to a distinctive side-to-side grinding movement during chewing. This would be apparent when comparing the phase difference between the data of the two sensors (e.g. temple sensors 201) located on either side of the user’s face. The data may be compared to determine a measure of asymmetry in the mouth activity, as described above. The measure of asymmetry may be indicative of pathology of the joint, bone, muscles, teeth, gum, tongue or oral lining. When performing certain jaw activities, there is often a coordinated sequence of muscle activations that occur. For example, contraction of the muscle that facilitates mouth opening is usually coordinated with relaxation of the orbicularis oris muscle and differential contraction of the cheek muscles to retract the lips and expose the teeth. Similarly, during mastication the coordinated contraction and relaxation of the cheek muscles on both sides of the face facilitates the manipulation of the food bolus in collaboration with the tongue. This sequence of muscle activations can be detected with the system described herein comprising multiple sensors and used to identify or confirm the jaw activity occurring. For example, combining information from a lip dilator and / or lip constrictor sensor (e.g. a sensor located on the wearable apparatus so as to monitor an area of skin associated with a lip dilator and / or lip constrictor), with data obtained from the sensor(s) may improve the accuracy and reliability of the system in identifying mouth or jaw activities. A sensor or sensors for monitoring a lip dilator and / or lip constrictor may be particularly advantageous when determining a mouth-opening event, as the lip constrictors and lip dilators are most active during this event. As mentioned above, the wearable apparatus may comprise a motion sensor, which may be an inertial measurement unit (IMU). The motion sensor may be configured to measure tilting of the head of the user when the wearable apparatus is worn on the head of the user. This may allow the system to infer that a jaw activity (e.g. mouth-opening and mouth-closing) corresponds to consumption of a pill such as a food supplement or medication. When a user ingests a pill, they typically tilt their head backwards at the same time as swallowing the pill. The motion sensor can be used to detect the head tilting, which can be compared (e.g. at the processor) with information regarding the identified jaw activity (e.g. the opening and closing of the mouth) prior to the head tilt. The processor may infer that a pill has been ingested based on the comparison. In some examples, the wearable apparatus may comprise a detector for detecting a signal emitted from the pill to confirm ingestion of the pill. For example, the pill may emit acoustic or radio waves that may be detected at the detector in the wearable apparatus. This may allow for more accurate tracking of a user’s medication or supplements. In other examples, the wearable apparatus may comprise a detector for detecting a signal emitted from a medication container. For example, a unique acoustic signal from the medication container may be sensed by one or more microphones on the wearable devices to provide context that the subsequent head and jaw motions indicate medication ingestion from the container. Facial expression detection The system may further comprise sensors configured to monitor areas of skin that are not associated with or correlated with mouth muscles. For example, the apparatus may comprise sensors located so as to monitor from the upper facial muscles such as the contraction or passive stretching of corrugator supercillii, depressor supercillii, procerus, and orbicularis oculi. In the example in which the wearable apparatus comprises a pair of glasses, sensors for monitoring upper facial muscles such as the corrugator supercilli can easily be incorporated into the frame of the glasses. By monitoring the activation of upper facial muscles, the system may be configured to infer a facial expression of the user during the jaw activity. A method for inferring a facial expression of a user from sensor data that monitors movement of the skin around the upper facial muscles can be found in UK Patent No. GB2561537, which is incorporated by reference herein. When combined with information from the other sensors, the system may be able to determine an emotional reaction or preference of the user during a particular mouth activity. This may be provided as feedback to the user. For example, bruxism and / or individual jaw clenching are often triggered by emotional and / or physical stress, anxiety, pain, anger, or intense concentration. Jaw clenching may by itself indicate an emotional response such as the negative valence responses of anger, nervousness, tension or stress. Inferring the facial expression of the user whilst they are clenching their jaw and / or exhibiting bruxism provides the user with additional information regarding how they are responding to external factors. Generally, analysing the facial expression of the user whilst certain mouth activities are occurring may have a variety of applications, including psychological assessment of stress and anxiety, medical uses such as for detecting subtle signs of pain, monitoring progress in treatments such as for TMJ dysfunction or chronic neck pain, personalised pain management, research purposes such as understanding pain and muscle movements / contractions, and biofeedback therapy. The device may also be used for understanding workplace stress levels and employee well-being or assessing physical strain by assessment of motor overflow. In some examples, the processor may be configured to identify a mouth activity more generally, without necessarily requiring a cyclic movement to be determined. Whilst the sensors preferably capture data in two directions in the plane of the skin, in some cases the sensor may be configured to measure movement of the area of skin in only one direction in the plane defined by the area of skin. When measuring movement of the area of skin in only one direction, preferably that direction is the substantially vertical direction with respect to the user-i.e., in a direction substantially parallel to the sagittal plane of the user. This is because most areas of skin are likely to move the most in the vertical direction during jaw activities such as opening and closing the mouth than in other directions. If the sensor is configured to measure movement of the area of skin in only in one direction in the plane of the skin, an oscillating movement in that direction overtime (e.g. as shown in figures 6c and 6d) can be detected which may be used to identify a mouth or jaw activity. Tongue-pressing activity In an example, the apparatus may be configured to identify the user pressing their tongue against a part of their mouth. Pressing the tongue against a location in the mouth causes activation of at least one jaw muscle so as to stabilize the jaw in reaction to movement of the tongue. For example, the anterior part of the temporalis muscle may primarily be activated in response to the tongue being pressed against a part of the mouth. Thus, skin movement correlated to the activation of the anterior part of the temporalis muscle may be used to identify a tongue-pressing activity. When the user presses their tongue against a part of their mouth and then releases that pressure, an area of skin whose movement is correlated with activation of a jaw muscle may exhibit a cyclic movement which can be detected, e.g. by the sensors described herein. If the user repeatedly presses their tongue against a part of the mouth, multiple cyclic movements may be identified. Figure 9 shows data captured by a temple sensor on the apparatus described herein when the user repeatedly presses their tongue against their front teeth. In this example, when pushing the tongue against the front teeth, the anterior part of the temporalis muscle contracts. The maximum amplitude of the skin movement during this activity is, similarly to during bruxism, significantly smaller (e.g. 10 times smaller) than the maximum amplitude of the skin movement during other jaw activities such as chewing or opening the mouth. In an example, the amplitude of the skin movement can be used to identify a tongue-press, additionally or alternatively to detecting that the anterior part of the temporalis muscle is activated. Generally, the machine learning system can be trained to identify skin movement resulting from the user pressing their tongue against the mouth based on training data exhibiting this activity. Being able to identify the jaw activity of pressing the tongue against a part of the mouth can improve the robustness of the system in correctly identifying other jaw activities. The system described herein may further be used to indirectly infer the location of the tongue-press. Depending on what part of the mouth the tongue is pressing against, different jaw muscles may be activated and to differing degrees. In an example, when pushing the tongue against the front teeth or anterior palate, the anterior part of the temporalis muscle contracts. In an example, when pushing the tongue against the front teeth or anterior palate, the strength of activation of the temporalis muscle on the two sides of the face is approximately equal. In an example, when the tongue presses against the left teeth, the strength of activation of the jaw muscles (e.g. temporalis) on the right side of the face is generally greater than the strength of activation of the corresponding jaw muscles on the left side of the face. In an example, when the tongue presses against the right teeth, the strength of activation of the jaw muscles (e.g. temporalis) on the left side of the face is generally greater than the strength of activation of the corresponding jaw muscles on the right side of the face. The processor may therefore be configured to distinguish between a tongue press being performed by the user on the left and right teeth inside the mouth by comparing the strength of activation of one or more jaw muscles (e.g. degree of movement of the skin correlated with the respective muscle(s)) on the two sides of the face. The processor may be configured to identify a tongue press on the front teeth or anterior palate by detecting tongue activity in which the strength of activation of corresponding jaw muscles on the two sides of the face is substantially the same. The system can be trained to recognize different patterns of skin movement that correlate with different jaw muscle contractions each associated with the tongue pressing against a particular location within the mouth. This may be improved with calibration for a particular user. In particular, the system may infer the location of the tongue-press based on the balance of jaw muscle activation on either side of a user’s face. By monitoring the activity of the jaw muscles in the manner described herein it is therefore possible to determine where the tongue is being pressed inside the mouth - e.g. by comparing jaw muscle activity on the left and right side of the user’s face, and / or by looking at which section of the temporalis muscle is activated. This enables, for example, the tongue to be used as an input to a computer system without requiring an invasive sensors inside the mouth. Taking the temporalis muscle as an example, using data from a sensor located so as to monitor an area of skin associated with the temporalis muscle and configured to capture skin movement in two dimensions in a plane defined by the area of skin being monitored, the processor may determine pressing locations of the tongue within the mouth. The temporalis muscle is activated by several different nerves that cause different sections of the temporalis muscle to move. Generally, the anterior part of the temporalis muscle primarily activates in response to a tongue-press. When the user presses their tongue against different parts of the mouth, for example against the front teeth or the roof of the mouth, different sections of the temporalis muscle may be activated. The anterior part of the temporalis muscle may activate most strongly when the tongue is pressed against the front teeth or anterior palate. Depending on the section of the temporalis muscle that is activated, a different movement of the skin associated with (e.g. overlying) the temporalis muscle may be observed, enabling the movement of that skin to be used to determine which part of the temporalis has activated. Whilst movement of individual sections of the area of skin in the temple region may be difficult to discern using the naked eye, by monitoring the area of skin associated with the temporalis muscle with the sensor type described herein, the sensor may differentiate between movement of different parts of the skin within the area of skin being monitored. The processor may correlate particular movements of the area of skin being monitored with particular sections of the temporalis muscle being activated that are associated with certain tongue activities (e.g. using suitable training such as a machine learning algorithm or neural network as described above). The system may thus infer that particular skin movements of an area of skin associated with the temporalis correspond to particular tongue pressing activities. In a simple example, the area of skin being monitored by the sensor may be divided into quadrants. Each quadrant may be monitored independently by the same sensor. Movement of the skin within each quadrant may be associated with a particular tongue activity (e.g. once the system has been trained to recognise which quadrants correspond to which tongue activities). When movement is detected entirely or predominantly in a particular quadrant, it may be inferred that the tongue is performing the activity associated with movement of that quadrant. In more complex examples, the area of skin may be divided into a greater number of sections. Movement within each section or within a combination of sections may be attributed to a different tongue activity. The area of skin need not be divided into sections explicitly. The system may provide sensor data describing movement of different parts of the skin within the area of skin to the processor which may be configured to identify the tongue activity using a machine-learning algorithm or neural network. In another example, a comparison of the contraction of the jaw muscles on either side of a user’s face can be used to identify the tongue-pressing location within the mouth. For example, when pressing the tongue into the teeth on one side of the mouth, the activation of the masseter on that side of the face may be more pronounced than the activation of the masseter on the opposite side of the face. The same may apply to the activation of other jaw muscles on either side of the face, such as the temporalis muscle. When pressing the tongue at a central location in the mouth (e.g. in the middle of the roof of the mouth), the activation of the jaw muscles on either side of the face may be the same or similar. The activation of the jaw muscles is reflected in the amount of movement of an area of skin associated with those jaw muscles. The tongue-pressing location may be inferred from the balance of muscle activation on either side of the user’s face. This can be determined using sensors located on corresponding sides of the apparatus. The data from the sensors located so as to monitor an area of skin associated with a jaw muscle on either side of the user’s face may be used in combination with the technique described above in relation to monitoring sections of the temporalis muscle to infer the tongue-pressing location. Inferring the position of a tongue press within the mouth using the system herein may enable an alternative use of the system as an input device, e.g. for use by people with full body paralysis, or as input device for an augmented reality or virtual reality headset where it is advantageous for the user’s hands to be free to perform other tasks. The tongue pressing location may be used to control a device, for example as an alternative to a computer mouse or joystick. For example, four distinct tongue pressing locations may be used to operate a left-right-up-down controller or the like. Other mouth activities may also be used to remotely control a device, such as jaw clenching. A mouth activity can be used as an input method for certain types of questions or prompts to an Al assistant. For example, if a user is navigating a location or challenging task, rather than writing or speaking a message, explicit or implicit input can be used to facilitate communication. Implicit communications may include emotional responses such as facial, mouth, or jaw movements. Jaw movements indicative of stress can be used, such as jaw clenching. The intensity of the jaw clenching movements can be used to communicate the degree of stress. Explicit communication could be distinguishable from implicit by the duration or sequence of clenching. For example, a double clench, or triple clench may be pre-configured as an explicit communication. Integrating a jaw monitoring device that can read facial expressions into a human-computer interaction system could enhance each stage of the human-computer interaction process in several ways. These are described below with reference to an Al assistant. 1. Initiation: The Al system can detect when the user is puzzled or in need of information through their facial and / or mouth or jaw movements, potentially prompting the Al to offer assistance proactively. 2. Formulation: By reading the user’s facial and / or mouth or jaw movements, the Al might infer the user’s emotional state or level of understanding, allowing it to guide the formulation of their query more effectively. 3. Transmission: The Al could use visual cues to confirm whether the user’s request has been communicated successfully or if further clarification is needed. 4. Processing: Emotional cues such as stress indicators can help the Al understand the urgency or the emotional context of the request, influencing how it processes the information. 5. Response Formulation: There are ways that the system can provide value to the user. The Al can adjust its response based on the user’s emotional state - for instance, providing more detailed explanations if the user seems confused. The Al can modulate the delivery of its response, such as using a more empathetic tone if the user seems stressed. Based on past Interactions the Al can make suggestions inferred from their past emotional responses e.g. playing relaxing music. 6. User Interpretation: By continuously reading the user’s reactions, the Al can gauge whether the information is being understood and can offer further clarification as needed. 7. Feedback and Iteration: Real-time emotional feedback allows the Al to adjust its responses on the fly, improving the quality of the interaction and reducing the need for repeated queries. Eating episode example In a more specific example, the system described herein is used to monitor a user’s eating episodes. Figure 5 shows a series of mouth activities that occur during an eating episode. An eating episode occurs when the user consumes some food. For example, if the user eats a snack this is classified as an eating episode, even if it only involves one mouthful. As another example, when the user consumes a meal such as breakfast, lunch or dinner, the duration over which the meal is consumed is an eating episode. An eating episode starts when the user initiates consumption such as an initial bite of food and ends when the user stops consumption fora predetermined amount of time. The predetermined amount of time after which the eating episode is deemed to have finished may be updated by the system as it learns the user’s habits. For example, if the user is a particularly slow eater, the system may wait longer before it determines that the user has finished their eating episode. An eating episode can only start if a previous eating episode is not already in progress. Steps S510 to S540 are described below with reference to the graphs shown in figures 6a to 6d. An eating episode starts with an initial mouth-opening event at step S510. This can be identified using the techniques described above. In figures 6a to 6d, the initial mouth-opening event occurs within the first few seconds (e.g. at 0 <t <2 seconds). An example cyclic movement that may be associated with a mouth-opening event is marked out with a dashed box in figures 6b, 6c and 6d. In this case, the first cyclic movement has a relatively large maximum amplitude in both the x and y direction compared to the other cyclic movements, indicating that a mouth-opening event may be occurring. At step S520, subsequently to the mouth opening event, the user may close their mouth over some food that they are taking a bite from, e.g. an apple. The processor may determine that a mouth-closing event has occurred if, following the mouth-opening event, it detects a reduction in amplitude of the cyclic movement from the maximum amplitude to a value that is near the user’s closed-mouth position. The mouth-opening and mouth-closing event may be identified based on the same cyclic movement. In some examples, the processor may increase a sampling rate of the sensor after steps S510 and S520 occur. As the mouth opening and mouth closing event may indicate that an eating episode is starting, the sensor may be awakened from operating in a previous low-power mode in order to sample more frequently to detect the eating episode. This may save power by not operating the sensor at a high sampling rate when an eating episode Is unlikely to be occurring. After the mouth has closed at least partially, the user starts to chew the food. At step S540, in response to detecting repeated cyclic movements, a chewing activity is determined. As the repeated cyclic movements were immediately preceded by a mouth-opening and a mouth-closing event, it is expected that a chewing event may subsequently occur. Chewing may be confirmed by analysing the data for further features such as a reduction in amplitude of the cyclic movements overtime. For example, the slope of the dotted trapezium in figure 6d indicates the gradual reduction in maximum amplitude of the cyclic movements as the user continues to chew their food and break it down into smaller pieces. Had the graph in figure 6d continued for longer than 14 seconds, the amplitude would have decreased further. Assuming that the user is eating food, rather than chewing gum for example, there will eventually be a pause in the repeated cyclic movements whilst the user stops chewing to swallow their food (not shown in figures 6a to 6d). At step S550, a swallowing activity is inferred. The processor may be configured to infer a swallowing event in response to detecting an interruption in the repeated cyclic movements for predetermined interval of time. A reduction in the maximum amplitude of the cyclic movements immediately preceding the interruption may additionally be used to infer a swallowing event. If the apparatus comprises a motion sensor and / or microphone, a swallowing event can be confirmed based on a peak in the data from the motion sensor and / or microphone during the interruption in the repeated cyclic movements. The user may then either continue to chew the food or open their mouth to take another bite, and the process described above continues. Alternatively, after the swallowing activity, the eating episode ends. When the processor detects an eating episode (as evidenced by the sequence of mouth activities shown in figure 5), it may determine an eating metric of the user. An eating metric may include: The number of eating episodes per 24hrs The duration of each eating episode Eating episode chewing count (the number of chews per eating episode) Bolus duration (how long each mouthful is chewed for) Bolus chewing count (the number of chews per mouthful) Chewing rate (the number of chews per minute) Inter-bolus interval (gap between mouthfuls in the same eating episode) The eating metrics may be provided as feedback to the user, e.g. at step S140. The system may be pre-configured by the user or to learn using the neural network about the eating context. The eating context comprises the commonest eating locations, times or patterns of preceding sensor data (such as associated with food preparation). Sensor data may include the location of the user based on GPS, Wifi, Bluetooth or other local area network data. Sensor data may include motion data indicative of the pose, posture or the activity of the user. Sensor data may also include the time windows since the eating sensors last detected an eating episode. This may be used by the processor to predict the time of the next eating episode, as will be explained in the power-saving section later. Power saving features It is desirable to reduce the amount of time that the sensor or sensors in the wearable apparatus are using power where possible. The sensor(s) of the wearable apparatus are likely to be powered by a respective battery ora centralized single battery. It is advantageous to reduce the frequency with which the battery or batteries in the wearable apparatus need to be charged or replaced. The system may be configured to adapt the sampling rate (e.g. how often the sensor captures data) of each sensor in the wearable apparatus. Each sensor may have a variety of operating modes with varying sampling rates. For example, each sensor may have a stand-by mode in which the sensor does not capture any data, a power-saving mode in which the sensor captures a minimal number of samples, and a full-power mode in which the sensor takes a maximum number of samples. There may be more or fewer than three operating modes. The system may control each sensor to operate in one of the operating modes at a time. There may not be discrete operating modes for each sensor. For example, the sampling rate of each sensor may be varied by the system to any desired value, rather than controlling the sensor to operate in a particular mode. Some sensors in the system may be more power intensive than others. For example, if the wearable apparatus comprises a camera 205, the camera is likely to require more power to run than the temple sensor 201, which may be an optical flow sensor, for example. To save power, the system may only operate the high power sensor(s) during periods in which eating has been detected or is likely to be occurring. To detect this, the system may use low power sensor(s) to detect activity that indicates eating may be about to occur. The low power sensors may be controlled to sample data at a relatively low sampling rate until an eating activity is detected. For example, in response to detecting a mouth-opening event using a low power sensor, the system may increase the sampling rate of the low power sensor and / or the sampling rate of a high power sensor. If it is determined that the mouth-opening event did not lead to a chewing activity, and so an eating episode is not occurring, the sampling rate of the high power sensor and / or the low power sensor may be reduced. The low power sensor may continue to monitor at a relatively low sampling rate in case further jaw activities occur. The sampling rate of the motion sensor may be increased from 50Hz to 4000Hz when an eating episode is detected (e.g. using data captured by another sensor or from the motion sensor when sampling at a lower rate). When sampling at 4000Hz, the motion sensor can act like a microphone to pick up sound wave vibrations which may be used to confirm swallowing or the chewing of hard food. As sampling at 4000Hz will use more power than sampling at 50Hz, the sampling rate of the motion sensor may only be increased to this level when the system detects an eating episode. In a preferred example, the low-power sensor Is an accelerometer. An example power-saving method 600 for activating the sensors in the wearable apparatus will now be described, with reference to figure 10. In this example, the wearable apparatus is a head-worn device comprising a motion sensor (e.g. an accelerometer), in addition to one or more of the sensors described above (e.g. the temple sensor 201). The power-saving method may be used whilst monitoring any of mouth, jaw, consuming, and / or eating activities. Suitably, the power-saving method is used when the apparatus is being used to monitor one or more eating episodes. The inventors have recognized that mouth / jaw / consuming activities cause motion in the wearable apparatus that can be detected in a relatively low-power way using a motion sensor. Upon detection of movement of the wearable apparatus that is consistent with movement that might occur during a mouth, jaw, or consuming activity, the sampling rate of other more power-hungry sensors can be increased. If, following an increase in sampling rate of the other sensors, the potential mouth, jaw, or consuming activity is determined not to be occurring, or to have finished, the sampling rate of the power-hungry sensors may be decreased. At step S610, data is captured describing movement of the wearable apparatus using a motion sensor. Said data may be captured over a predetermined period of time. The motion sensor is configured to detect movement of the wearable apparatus relative to the user, when worn. The motion sensor may also be configured to detect gross movement of the wearable apparatus. Preferably, the motion sensor comprises an accelerometer. The sampling rate of the motion sensor during step S610 may be relatively low, e.g. 50Hz. The remaining sensors in the apparatus may be set to operate in a stand-by mode, e.g. where the sampling rate of the sensors is zero or very low. At step S620, the motion sensor data may be analyzed to determine whether the movement of the wearable apparatus meets certain criteria. For example, one criterion may be whether the relative movement of the wearable apparatus exceeds a predefined threshold, e.g. during the predetermined period. If the movement exceeds the predefined threshold, the method may proceed to step S630. Otherwise, the method returns to step S610. Other criteria may be whether the relative movement is below a different predetermined threshold. For example, if the apparatus moves too quickly with respect to the user, it may be determined that the user is not performing a mouth, jaw, or consuming activity. The filtering performed at step S620 may prevent unnecessary further analysis of the motion data in the case where the movement of the apparatus is too small or too big to be suggestive of a mouth activity occurring. Step S620 may be omitted. Preferably, step S620 is implemented directly on the motion sensor, e.g. as an in-built filtering operation. This avoids the need to communicate with the processor in the system. At step S630, the motion data is input to a machine learning model, e.g. at the processor 350. The machine learning model is trained to identify patterns in movement of the wearable apparatus as corresponding to a potential mouth activity. The model may be trained in the same way as the machine learning models described above, but where the training data is motion data obtained whilst the user is wearing the wearable apparatus and is performing known activities that include different mouth activities, non-eating activities and eating activities. The machine learning model does not need to identify the mouth activity itself, it only needs to identify whether it is likely that an mouth activity is occurring based on the motion data. If at step S635 the machine learning model identifies the motion data as corresponding to a potential eating activity (e.g. the answer to step S635 is yes), the method proceeds to step S640. Because steps S630 and S635 are preliminary steps that occur before one or more of the remaining sensors are activated to confirm and analyse the mouth activity, preferably the machine learning model is trained with a large degree of tolerance in what is considered to be a potential mouth activity. That is, the machine learning model is trained to be sensitive enough to detect all mouth activities and potential mouth activities (e.g. eating activities and potential eating activities), even if this includes some non-mouth activities (e.g. non-eating activities). To prevent frequent activations of the remaining sensors caused by false positives, one or more of the remaining sensors may only be triggered (e.g. awoken from stand-by mode) if the machine learning model identifies a potential mouth activity more than a predefined number of times within a predetermined period. For example, if the machine learning model outputs a flag when a potential mouth activity is detected, the method may only proceed to step S640 if the model outputs more than a threshold number of flags within a certain time frame (e.g. 10 seconds). This selective approach may ensure that the remaining sensors are activated only when substantial and consistent indications of a potential mouth activity are detected, thus enhancing accuracy while minimizing unnecessary activations. At step S640, the sampling rate of one or more sensors in the wearable apparatus is increased. The one or more sensors may include the motion sensor. Preferably, at least one of the one or more sensors is a sensor located so as to, in use when the apparatus is worn by a user, monitor an area of skin whose movement is correlated with a facial muscle associated with movement of the jaw. Said sensor is preferably configured to capture time-varying data describing movement of the area of skin in two dimensions in a plane defined by the area of skin (e.g. temple sensor 201 or cheek sensor 202). The increase in sampling rate may be an increase from zero, e.g. activating the sensor(s) from stand-by mode, or from a low-power mode sampling rate. The increase in sampling rate may be different for different sensors. Some sensors (e.g. very high-power sensors) may remain in standby mode until further checks are performed to confirm an eating activity. Once the sampling rate for at least one of the one or more sensors has been increased, the method may proceed to step S100 to determine more information regarding the potential mouth activity. The data capturing method and analysis of the data to determine the mouth activity may be as described above with reference to figure 4 and 5, although it may be performed in an alternative way. At step S650, a check is performed periodically to determine whether a mouth activity is detected. It may be the case that despite a potential mouth activity being identified in step S635, this was a false alarm and no such mouth activity was detected by the one or more sensors. In another case, a mouth activity may have initially been detected by the one or more sensors but it may have since ended. Alternatively, the mouth activity may still be ongoing, for example because the user is having a meal. When it is determined that no mouth activity has been detected for a predefined period, at step S660, the sampling rate of at least one of the one or more remaining sensors is reduced. Not all the sensors that were activated in step S640 need to have their sampling rate reduced at the same time, or reduced by the same amount. For example, some of the sensors may continue to sample at the same rate as I step S640, whilst some of the other sensors may be switched off or returned to standby mode. Each sensor may have its sampling rate decreased individually depending on its respective power requirements. During tests of the above method on eating episodes, on average, eating activities were detected within 15 seconds of the eating activity starting. In addition to or alternatively to the method described above, the system may be configured to control the sampling rate of one or more of the sensors in the wearable apparatus in dependence on a likelihood of a mouth activity occurring, or more specifically, an eating episode occurring. The system may determine the likelihood of a mouth activity occurring as a stand-alone feature, separately to controlling the sampling rate. The likelihood of a mouth activity occurring may be based on one or more of: the time of day, the times at which historic mouth activities occurred, the time elapsed since the most recent mouth activity, a user input, and learned user habits. The processor may be configured to predict the time of the next mouth activity (e.g. based on the likelihood of it occurring as described above) and increase the sampling rate of one or more sensors around the predicted time of the next mouth activity. For example, the system may set all sensors in the wearable apparatus to stand-by mode during certain times of the day in which the mouth activity being monitored (e.g. eating) is unlikely to occur, such as during the night. As another example, the user may provide the system with typical times during the day in which the mouth activity is likely to occur. In another example, the system may learn the times of day that the user is likely to perform that mouth activity. The system may increase the sampling rate of at least some of the sensors around those times. In a more specific example, the processor may be configured to predict when the next consumption episode will occur (e.g. drinking, eating, taking medication). The prediction may be based on the time since the last (e.g. most recent) consumption episode, for example as detected by the processor or as logged by the user. For example, the probability of an eating episode occurring increases non-linearly beyond a certain minimum threshold since the previous meal. The system may learn (e.g. using a neural network) patterns in the consumption habits of the userand base the prediction of the next consumption episode on the learned habits. The system may allow for the user to input their usual mealtimes into the system. Based on the predicted time of the next consumption episode, the system may increase the sampling rate of one or more sensors in the wearable apparatus around the times that the consumption episode is expected. This may preserve the battery life of the sensors by not sampling data when a consumption episode is unlikely to occur. For example, the system may increase the sampling rate of the sensors an hour or 30minutes before the expected time of the next consumption episode. In other words, the system may increase the sampling rate of one or more sensors during a time window that includes the predicted time of the next consumption episode. The duration of the window (e.g. how much earlier it begins before the expected time, and how much longer it continues after the expected time) may depend on the accuracy of previously predicted consumption episodes. For example, the user may tend to eat lunch at the same time every day, in which case the window may only begin 15 minutes before the expected time of the consumption episode. In another example, if the user tends to snack throughout the day such that the predicted time of the next consumption episode is relatively soon after the previous consumption episode, the window may span most of the time between successive consumption episodes. When predicting the time of the next consumption episode, the system may use data from one or more sensors arranged to detect information regarding the user’s external environment. Examples of external environmental information may be the location of the user (e.g. as determined by GPS) and what the user may be looking at (e.g. using the camera in combination with image recognition). For example, if the user is in a restaurant (e.g. as determined by the GPS) and is looking at a menu (e.g. as determined by image recognition from the camera data), the system may adjust the predicted time of the next consumption episode so that it falls within the next hour. To further save power, during an eating episode, the sensor may sample data periodically and for a predetermined amount of time. For example, the sensor may sample data for a period of 2 seconds, every 10 seconds. The gap between sampling periods may be determined by a minimum time for chewing a mouthful. In an extreme power saving mode, the system may prioritize determining the number of mouthfuls during the eating episode rather than dynamics of the jaw during chewing. So, the system may decide not to sample data multiple times during one mouthful. The gap between sampling periods may be personalized on the individual. For example, fora user who tends not to chew their food many times before swallowing, the gap between sampling periods may be relatively short so that the sensor can detect a mouth-opening event. In some cases, the gap between sampling periods may be determined based on a predicted time to chew a mouthful based on the bite amplitude during a mouthopening event. A large bite amplitude during an eating episode may suggest the user is taking a large bite of food which will take longer to chew than the average mouthful, which may be used to increase the gap between sampling periods for the sensor. Identifying food qualities The inventors have recognized that the data from the sensor(s) described above can be used to determine qualities (e.g. physical properties) of the food being eaten during an eating episode, in particular relating to the texture of the food. The food being eaten may be categorized into certain food types depending on its qualities. In some cases, the food being eaten can be identified. Thus, the system described herein may be used to track and analyse what is being consumed by the user. This can provide an unobtrusive and minimally Invasive way of monitoring dietary intake. The dynamics of the lips and jaw during an eating episode are affected by the qualities (e.g. physical characteristics) ofthe food being eaten. A quality of the food being eaten may be its texture, the size of the food being eaten (e.g. the size ofthe mouthful), orthe shape ofthe food being eaten (e.g. the shape ofthe mouthful). The way in which a user opens their mouth to consume the food will differ depending on the size and shape ofthe food, such as whether it is an apple or a crisp. Once in the mouth, the motion of the jaw will depend on the shape and texture of that food, such as whether it is crunchy or soft. The sensor or plurality of sensors on the wearable apparatus can provide information relating to the shape ofthe lips and mouth when initially taking a bite ofthe food, the dynamics of jaw movement when chewing the food, how long it takes to chew each mouthful, and other information regarding the eating episode which can be used to identify qualities ofthe food and even identify the food being eaten. Identifying qualities ofthe food and / or identifying the food being eaten based on jaw dynamics may be performed as a further step S560 in the method of identifying an eating episode described above with reference to figure 5. In a simple example, the bite amplitude ofthe user, or equivalently the maximum displacement ofthe area of skin from its equilibrium position, may be used to estimate the size and / or shape ofthe food item being eaten. Figure 13 shows three examples of how the mouth opens when eating foods of different sizes. As can be seen in the right image of figure 13, a person eating an apple is likely to open their mouth much wider than a person eating a biscuit, as shown on the left of figure 13, or a strawberry as shown in the middle image. Thus, the bite amplitude can be used to infer information regarding the size of the food item (e.g. mouthful) being eaten. As described above with reference to jaw metrics, the bite amplitude can be determined from the maximum amplitude of the cyclic movement corresponding to an initial mouth-opening event (e.g. as identified at step S510 of figure 5). The bite amplitude can be equivalently determined from the maximum displacement of the area of skin being monitored from an equilibrium position (e.g. a neutral closed mouth position) during an eating episode, without needing to identify a cyclic movement. As another example, one or more characteristics of a cyclic movement or plurality of cyclic movements detected during chewing can be used to determine a quality of the food being eaten. The characteristics of a cyclic movement and of repeated cyclic movements are described above. Said characteristics comprise, for example, the maximum amplitude of each cyclic movement, a change in the amplitude of the cyclic movements overtime, a duration of each cyclic movement, an average velocity of the area of skin during each cyclic movement, a maximum velocity of the area of skin during each cyclic movement, and a rate of change of velocity of the area of skin during each cyclic movement. These characteristics may be monitored for each cyclic movement in a plurality of cyclic movements. The changes in each characteristic over the plurality of cyclic movements may be used to determine one or more texture qualities ofthe food item being consumed. The characteristics may be determined for the closing and / or opening phase of each cyclic movement, in addition to or alternatively to identifying the characteristics for the overall cyclic movement. In particular, the characteristics of a cyclic movement in the closing-phase may provide information about the texture ofthe food being chewed, because the closing-phase is the part ofthe chewing cycle that crushes the food. For example, when taking a bite of something crunchy or hard, the mouth may close less than when taking a bite of something very soft because the hard food may be preventing the jaw from closing fully. The speed at which the jaw moves to close the mouth is also affected by the food type being eaten. The harder or chewier the food being eaten, the slower the jaw is likely to close when compared with eating something soft or crumbly. In an example, to determine a measure of hardness of a food item, the speed ofthe cyclic movement during the closing-phase and / or the rate of change of speed of the skin movement in the cyclic movement may be analyzed. Hard foods may require more force to chew during the closing phase which may result in sudden and fast movements ofthe jaw as the food gets broken down. This may be reflected by greater changes in speed ofthe skin movement in the closing phase. Once broken down, the jaw may close much fasterthan when chewing soft food, so there may be a relatively high maximum speed of the skin movement during the closing-phase of a cyclic movement when eating hard foods. Similarly, brittle foods such as nuts may show sharp changes in speed ofthe closing phase when the food is suddenly broken, which may correspond to high frequency cyclic movements with low amplitude. As another example, the change in amplitude of the cyclic movements during chewing can be used to determine a measure of hardness (e.g. in addition to or alternatively to the characteristics mentioned above). The amplitude of the cyclic movements during chewing should gradually decrease as the hard food is broken down. In contrast, the amplitude of the cyclic movements when chewing soft foods may not vary as much. For foods with a hard outer texture but a soft inner part, the amplitude of each cyclic movement may reduce rapidly over successive cycles, so the cyclic movements would exhibit a high amplitude variance. On the other hand, a food item with a firm consistency may have less variance in amplitude overtime. For example, figure 11a shows a graph of the time-varying data captured by a cheek sensor when the user chews hard food compared with soft food. It can be seen that the frequency of the cyclic movements (or the number of cycles per second) when chewing hard food is greater than when chewing softer food. Thus, the maximum velocity of the skin movement is greater when chewing hard food than soft food, in this example. It can also be seen from figure 11 a that the amplitude of the cyclic movements decreases faster during the initial few seconds ofchewing hard food when compared with the initial few seconds of chewing soft food. This reflects the greater amount by which the hard food is broken down compared to the softer food. The chewing of brittle and hard foods can also be detected on a microphone or motion sensor (e.g. accelerometer or IMU) which can pick up the vibrations in the mouth. Hard and brittle foods produce high-frequency and / or high amplitude vibrations in the mouth which not only can be picked up as sound on a microphone or motion sensor, but also may produce noise in the data of the sensor. A particularly noisy signal from the sensor whilst chewing may indicate the presence of very hard food being chewed. As another example, sticky or highly cohesive food may slow the velocity of the opening phase as the teeth may get stuck when trying to open the mouth. The average speed of the skin movement in the opening-phase of the cyclic movements may therefore be used to determine a relative value for the stickiness of the food being eaten. The closing-phase of each chew may also be slowed by the stickiness of the food. The average speed of the skin movement in the closing-phase of the cyclic movements may therefore also be used to determine a relative value for the stickiness of the food being eaten. Overall, the average speed of the skin movement during the cyclic movement may be lower when eating sticky food than the average speed of the skin movement when eating crumbly food. The amplitude of the cyclic movements overtime is likely to decrease more quickly with crumbly food as it is easily broken down. On the other hand, sticky cohesive food will take longer to break down and so the amplitude of the cyclic movements may take longer to reduce. Figure 11c shows the time-varying data captured by a cheek sensor when the user chews ephemeral (e.g. crumbly) food and persistent (i.e. chewy / sticky) food. It can be seen that the amplitude of the cyclic movements when chewing persistent / sticky food does not decrease as much over 6 seconds as the amplitude of the cyclic movements when chewing ephemeral foods. As another example, the density of the food item can be inferred from the velocity of the area of skin during chewing, and / or the time it takes to chew a mouthful before swallowing. Particularly dense / solid foods tend to take longer to chew and to break down. Airy foods, on the other hand, may be broken down more easily and quickly. For example, figure 11 b shows data captured by the cheek sensor when eating airy food compared with solid food. The airy food is broken down faster than the solid food, as evidenced by the pause of the cyclic movements from 5 seconds onwards in the top graph of figure 11 b, indicating that the food has been swallowed. The solid food, on the other hand, is still being chewed after 15 seconds. As another example, dry or overly moist foods can influence the ease of chewing and the duration of chewing per bolus. Particularly dry foods will take longer to chew before the user can swallow them compared to very moist foods. Often, very moist foods are likely to be easierto chew and have smoother jaw motion, so the rate of change of the speed of the jaw movement may be relatively constant. Table 2 shows a summary of examples of characteristics of a cyclic movement that may be used to identify values for different texture qualities. Example texture qualities Characteristics of the cyclic movement(s) Hard High amplitude of the cyclic movement, greater change in amplitude of the cyclic movements during chewing, variable average velocity of the area of skin during chewing, high maximum velocity of the area of skin, variable rate of change of the velocity of the area of skin (sudden changes in speed). Sticky Slow average velocity of area of skin during opening-phase and closing-phase of each chew, slow gradual change in amplitude of cyclic movements during chewing, long duration of each cyclic movement, low variability of velocity of area of skin. Dense Long duration of each cyclic movement, high number of chews per bolus (e.g. long amount of time to swallow a mouthful), slow maximum speed of skin movement, consistent speed of skin movement (e.g. low rate of change of velocity of skin movement). Dry High number of chews per bolus Table 2 The above examples in combination with the data shown in figures 11a to 11c demonstrate that the sensor data comprises valuable information that can be used to draw conclusions regarding the food item being consumed. Machine learning In preferred implementations, a machine learning system is used to process the sensor data and classify the food item being consumed. An example method S900 for classifying a food item in terms of its texture using a machine learning system will now be described, with reference to figure 14 to 16. As described above with reference to figure 3, the system 300 comprises the processor 350 and the wearable apparatus 200. When a user wears the wearable apparatus and consumes a food item, the sensor(s) on the wearable apparatus capture time-varying data describing movement of an area of skin whose movement is correlated with a facial muscle associated with movement of the jaw, as described above. The area of skin may overlie the temporalis muscle or zygomaticus muscle, for example. In this example, the processor comprises a machine learning system 700 trained to process the data captured by the sensor(s) of the wearable apparatus so as to classify the food item being eaten according to one or more of a predetermined plurality of texture qualities. Each texture quality represents a physical characteristic of the food item, such as a measure of the food’s hardness, a measure of the food’s stickiness (i.e. cohesiveness), ora measure of the food’s density. Figure 14 shows an example schematic of processor 350 comprising the machine learning system 700. The machine learning system may be the same as the machine learning system described above with respect to identifying a jaw activity. The machine learning system may operate at the same hardware as the machine learning system described above with respect to identifying a jaw activity. The processor 350 is configured to receive, as an input, the time-varying data x(t) from the sensor(s) on the wearable apparatus or a dataset derived from the time-varying data, captured whilst the user is consuming a food item (e.g. at step S901). Optionally, the processor may determine one or more characteristics Ci(x(t))...CM(x(t)) from the input data x(t) such as those described above which may also be provided as an input to the machine learning system. The machine learning system processes, at step S902, the received input data and optionally the one or more characteristics so as to classify the food item according to one or more of a predetermined plurality of texture qualities ... TN. In some examples, the data provided to the processor is derived from the time-varying data - for instance, the data may be pre-processed to improve performance of the machine learning system and / or enhance key characteristics of the data and / or adapt the data to the characteristics of the machine learning system. In some examples, the time-varying data is converted into the frequency domain, e.g. using a Fourier or Wavelet transformation. The machine learning system 700 may classify the food item being consumed by generating a texture classification for that food item. A texture classification for the food item may comprise a classified (i.e. determined) texture quality (CT,, CT2,... CTN) for each of the one or more of the predetermined texture qualities (TltT2... TN) according to which the machine learning system performed the classifying. For example, if the one or more of the predetermined plurality of texture qualities comprises a measure of hardness and a measure of cohesiveness, then the texture classification generated by the machine learning system according to those one or more predetermined texture qualities comprises a classified measure of hardness and a classified measure of cohesiveness for the food item. In other words, the one or more of the predetermined plurality of texture qualities represent the predefined metrics that are used by the machine learning system to produce a texture classification for a particular food item. In one example, a texture classification may comprise only one texture quality (i.e. N=1), such as hardness. In another example, a texture classification may be represented by an array of elements CTr, CT2,... CTn where each element in the array represents a classified texture quality (e.g. a determined value for the respective texture quality). For example, if the texture classification is made up of three different texture qualities where‘S represents a measure of hardness, T2 represents a measure of cohesiveness, and T3 represents a measure of density, the texture classification may be an array having three values, e.g. [CT,, CT2, CT3], where the position in the array dictates the texture quality to which the value relates. In the example in figure 14, the machine learning system outputs three classified texture qualities (i.e. N=3), but there may be more or fewer texture qualities than three. Each classified texture quality may have a relative value such as low, medium, or high, or it may be ranked with a number in a predefined range, e.g. from 0 to 10. In another example, the classified texture quality may be a binary number representing whether that food item was found to have the particular texture quality or not. For example, instead of representing the classified texture quality on a relative scale for the respective texture quality such as hardness, the classified texture quality may put the food item into a discrete category such as “hard” or “not hard” (i.e. soft). The format of the classified texture quality may depend on the format of the one or more of the predetermined texture qualities used in the training data when training the machine learning system, as will be explained later. The machine learning system may employ a feature-based learning technique or a deep-learning technique (e.g. using a neural network) with which to classify the food item. In both techniques, preprocessing may be applied to the data (e.g. at the processor) to remove noise and clean the data. For example, the techniques described above in relation to preprocessing, filtering, and data augmentation may apply to either implementation. In the feature-based learning technique, the machine learning system is trained to correlate predetermined characteristics Ci(x(t))...CM(x(t)) of the time-varying data with one or more of the predetermined plurality of texture qualities. Once trained, the machine learning system extracts the predetermined characteristics Ci(x(c))...Cm(x(c)) from captured time-varying data x(t) and uses the extracted characteristics to determine the classified texture qualities of the food item CT^ CT2,... CTN for the one or more of the predetermined texture qualities (TltT2... TN). The inventors have recognized that when the time-varying data from the sensor is transformed into a frequency domain (e.g. using a Fourier or Wavelet transformation), certain features of the data in the frequency domain can be used to differentiate between different textures. In particular, if the timevarying data is transformed into a frequency domain, one or more of a dominant frequency (e.g., the actual frequency and / or the dominant-frequency amplitude), spectral entropy, and energy in specific frequency bands of the transformed data may provide useful information regarding the hardness of the food being eaten. In another example, characteristics of the time-varying data in the time-domain may be used, additionally or alternatively to the frequency-domain characteristics. For example, the time-varying data is likely to exhibit at least one cyclic movement of the area of skin during chewing of the food item. The machine learning system may be configured to determine characteristics of the cyclic movement or plurality of cyclic movements and correlate those characteristics with the one or more predetermined plurality of texture qualities. The characteristics of the at least one cyclic movement and / or of the plurality of cyclic movements are described above. Said characteristics comprise, for example, the maximum amplitude of each cyclic movement, an average frequency of the plurality of cyclic movements, a frequency variability of the plurality of cyclic movements, a change in the amplitude of the cyclic movements overtime, a duration of each cyclic movement, an average velocity of the area of skin during each cyclic movement, a maximum velocity of the area of skin during each cyclic movement, and a rate of change of velocity of the area of skin during each cyclic movement. In the feature-based machine learning technique, a Random Forest algorithm, Gradient Boosting algorithm, or any other algorithm based on feature input may be used as is known in the art. In the deep learning technique, the machine learning system is a neural network such as a recurrent neural network. The neural network may be trained on time-varying data captured by the sensor whilst the user consumes a plurality of known food items each having physical characteristics corresponding to one or more of a predetermined plurality of texture qualities. Once trained, the neural network can take as an input time-varying data from the sensor captured during consumption of an unknown food item, and classify the food item according to the one or more of the predetermined texture qualities with which the network was trained. The output of the neural network may be the texture classification for the food item. As described above, the neural network may be trained and / or implemented on a hardware accelerator in the processor or run as software on the processor. The neural network architecture used can be flexible, provided it is suitable for time-series data. Suitable neural network architectures include recurrent neural networks (RNNs), Long Short-Term Memory networks (LSTM), Convolutional Neural Networks (CNNs), Transformer, Informer, Residual networks, Inception networks, and similar. Any combination of layers based on existing deep-learning architectures may be used. The neural network may be an end-to-end neural network - i.e. a neural network that operates on raw input data without requiring feature-extraction. In some examples, the machine learning system is a multimodal large language model (LLM). The multimodal LLM may comprise a neural network such as a transformer network which may be configured to process the sensor data in the same way as the neural networks described herein. Any suitable method for valuing the predetermined texture qualities of known food items may be used. In one example, a scale for each predetermined texture quality may be determined using specific foods. Preferably, the specific foods have consistent physical characteristics. For example, a raw carrot may define the maximum hardness level (i.e. the hardest food) and a piece of brioche may define the minimum hardness level (i.e. the softest food). Foods may be ranked by a user on a scale of hardness relative to the carrot and brioche. This method may be used for each predetermined texture quality (e.g. cohesiveness and density). The resulting rankings may be used as the values for the respective texture qualities of those foods. Table 3 shows a list of food items with relative values forthe respective texture qualities of hardness, cohesiveness, and density which may be determined in this way. Alternatively, a plurality of known food items may have their physical characteristics tested in a laboratory using a specified test for each of the predetermined texture qualities. For example, a texture analyzer such as an EZ-X series single-column universal test frame sold by Shimadzu ® or an MT Series Manual Test Stand sold by Ametek ® or a food texture analyser sold by PackTest ® may be used to determine values of each texture quality. To calibrate the system to a particular user, the user may be provided with a variety of known food items each having predetermined texture qualities and asked to consume each food item. Food Hardness Cohesiveness Density Apples (raw) Hard (due to cellulose) Moderate Moderate Almonds Hard Moderate High Carrots (raw) Hard (cellulose) Moderate Moderate to high Chicken breast (cooked) Firm Low Moderate Avocado Soft High Low Broccoli (steamed) Tender High Low Potato Chips Crispy (brittle) Low Low Fried Chicken Crispy (brittle) Low Moderate to high Bacon Crispy (brittle) Low Moderate to high Watermelon Juicy (soft) High Low Oranges Juicy (soft) High Low Tomatoes Juicy (soft) Moderate Low Croissants Soft Moderate Low Fish fillet Soft Moderate Moderate Puff Pastry Soft Moderate Low to moderate Bagels Chewy High Low Beef jerky Chewy High Moderate to high Caramel Chewy High Moderate to high Chocolate Cake Soft High Low to moderate Table 3 Generally, the food classification determined by the system for an unknown food item may be compared with food classifications of previously identified or known foods to estimate what food was being consumed by the user. The estimate may comprise a range (e.g. one or more) of suggested foods that are each associated with a food classification that is within a predetermined margin of the determined food classification. For example, the food classification for a raw carrot may be similar but not identical to the food classification for raw celery and raw radish. The system may identify a plurality of known food items (e.g. carrot or celery) each having a food classification that is similar to the determined food classification and provide the plurality of food items as the estimate for the unknown food item. In another example, the system may provide an estimate of the type of food being eaten (e.g. crunchy food, chewy food, soft food), rather than an estimate of what the food item being eaten is (or is similar to). The system may produce a likelihood of the unknown food item being one or more of a previously identified or known food item based on a comparison between the texture quality of the unknown food item and the texture quality(s) of the previously identified food item(s). The user may correct or confirm the estimate of the food item or type of food that the system produces, which may be used to improve the suggestion provided by the system in the future. Suitably, the food item being consumed may be placed on a spectrum for each texture quality, such as hardness (hard to soft), cohesiveness (ephemeral vs chewy), and density (airy vs dense). This may be visualized as a multi-dimensional space, where each dimension corresponds to a respective texture quality. The overall texture classification for a food item may be represented as coordinates in the multidimensional space. Figure 12 shows an example of this spectrum with three dimensions (hardness, cohesiveness, and density) forming a texture space. The texture space may comprise more dimensions than the three depicted in figure 12, corresponding to as many texture qualities as desired to describe the texture classification. The coordinates of the food in the texture space are a measure of the overall texture of the food. Overtime, the system can recognize that certain coordinates in the texture space correspond to certain foods. The more sensor data and corresponding texture data that the system gathers, the more accurate the system can be in identifying the food item being consumed. When the food being eaten is assigned coordinates in texture space, if those coordinates are sufficiently close to a previously identified food, it can be determined (e.g. to within a certain degree of accuracy) that the food being eaten is the same or similar. For example, if the coordinate of the unknown food item being eaten is within a predetermined distance to a known food item in each of the dimensions of the texture space, it may be estimated that the unknown food item is the same as that known food item. In another example, the system may base its estimate on the known food item that has a coordinate in texture space that is closest to the coordinate of the unknown food item. Generally, the system may form a likelihood (e.g. probability) of the unknown food item being a particular known food item based on a distance in texture space between the coordinate of the unknown food item and the coordinate of that known food item. The closer the coordinates of the unknown food item and the known food item are to each other, the higher the likelihood that the unknown food item is the same or similar (e.g. similar type) to that known food item. Thus, the system may map texture classifications to coordinates in texture space and determine an estimate of the unknown food item being eaten based on a comparison between the coordinate of the unknown food item and the coordinates of known food items. The food may be categorized into a food type relating to its texture classification. Examples of food types may be hard food, soft food, chewy food, crumbly food, solid food, and airy food. The system may be configured to receive a user-defined classification of the food item according to the one or more of the predetermined plurality of texture qualities. In an example, the system may apply unsupervised learning, for example a clustering algorithm such as K-means or DBSCAN, to the sensor data to group similar patterns together. Once the data is clustered, the system may ask the userto label the cluster according to the perceived texture qualities of the food items they consumed. This may help in reducing the burden on users by allowing themto label groups of similar food types instead of labelling each food type individually, increasing efficiency in labelling large datasets. The system may use the user-identified classifications as training data for the machine learning system, e.g. using semisupervised learning. The system may use the determined food classification to monitor dietary intake of the user. For example, the system may keep track of the estimated food items that were consumed by the user (e.g. based on the identified food classifications as described above) in a digital food diary or log. The system may record the time of day at which the food item was eaten. Such data would be readily available from the time-varying data associated with the food classification for that food item. Whilst the food items that were consumed by the user may not always be accurately identified by the system, or may only be identified by their food type (e.g. crunchy food), keeping a record of the estimated food item ortype of food item that was consumed may remind the user of what they have eaten. For example, the system may record in a food diary that a crunchy food item was consumed that was similarto carrot at 11am. The user can then correct or refine the record in the diary if it is inaccurate. This may help the user not to forget foods that they have eaten and allows them to maintain a more accurate food diary. In another example, the identified food may be used as input to existing nutrition services (e.g., MyFitnessPal ®) to provide more information regarding the macro- and micro- nutrients consumed in a specific meal. Generally, the accuracy of the system can be improved by additional data. For example, where contextual information is available (such as the food items in the kitchen or receipt information from electronic purchases), the range of food items to select from can be narrowed to improve the identification system. In an example, as described further below, the system described herein may be integrated with an image classification system 750 to identify a food item about to be eaten. Several image classification systems are available to recognise items of food based on image data such as a camera image. In one example, the system might capture one or more images of the food in front of the wearer to identify the food being eaten (e.g. by comparing differences in sequential images so that the specific food item that has been eaten can be identified). The system may combine the food identification from the image classification system with the texture analysis system. This has several benefits, particularly because most people have a relatively narrow dietary repertoire. Firstly, the system can be trained to associate the texture classification with the food identified in the images. This can help train the system to identify food being eaten based solely on sensor data. Furthermore, the appearance of the food may modify the textural characteristics for example by softening it, which may affect the predicted classification. Secondly, once the system is able to identify foods based on sensor data, the metabolic effects of a meal might be better predicted based on the order in which food is consumed. For example, eating vegetables first can reduce the spike in blood sugar after a meal. Companies such as January.AI ® are attempting to predict blood glucose levels based food photo logging. The identification of food types based on textures enables predictive estimation of post meal blood sugar. In another example, non-invasive blood glucose measurements can be combined with the sensor data to improve food classification. Non-invasive blood glucose monitors are commonly used by diabetics and increasingly by non-diabetics for wellness monitoring. The post-ingestion change in glucose level (or indeed any food metabolite) following ingestion of a quantity of food could be used to refine estimates and distinguish items with similar texture. For example, a chunk of cucumber may have a similar texture characteristic as an apple chunk but the sugar content of an apple is much higher than a cucumber, so the post-prandial rise in glucose levels will distinguish one from the other. Furthermore, detection of the food type and the sequence in which it’s eaten can be used refine predictions of blood glucose spikes. As is known in the art, eating certain foods such as vegetables prior to energy dense foods attenuates the subsequent glucose spike. The system described herein can be used to provide a history of the food items that have been consumed during an eating episode (e.g. including the order of consumption), which can be used to adjust or predict subsequent blood glucose levels of the user. For example, the history of the food items consumed may be provided to a blood glucose level prediction algorithm. An example of a blood glucose prediction algorithm which takes nutritional information as an input is January Al’s blood glucose level (BGL) prediction algorithm as described in the paper “Virtual blood glucose monitoring and prediction using machine learning” by Ashkan Dehghani Zahedani et. al, January Al which is hereby incorporated by reference. An example method S800 fortraining the machine learning system will now be described with reference to figure 15. At step S801, training data is formed by associating time-varying data with one or more of a predetermined plurality of texture qualities. The time-varying data is captured during consumption by a user of a food item having physical characteristics corresponding to the one or more of the predetermined plurality of texture qualities. The one or more texture qualities of the food item are known prior to performing the training. Preferably, the training data is formed from time-varying data captured during consumption of a plurality of different food items. Each food item may be consumed by the same user or by different users each wearing a wearable apparatus 200. The step of capturing the timevarying data (not shown) may be included as part of the training method S800. The time-varying data may comprise a plurality of sets of time-varying data, each set being captured during consumption of a respective food item having known texture qualities. As part of step S801, each set of time-varying data may be associated with a respective set of one or more of the predetermined plurality of texture qualities. For example, each set of time-varying data may be assigned one or more labels for the one or more texture qualities of the food item consumed whilst that set of time-varying data was captured. The time-varying data may also optionally be associated with the food item that was consumed during its capture, and one or more characteristics of the data as explained above with respect to the feature-based learning. For example, the training data may be represented as shown in table 4. The food item and texture qualities for each set of time-varying data in the training data are known. The below table may optionally comprise further columns for each of the one or more characteristics of the time-varying data that may be determined, e.g. if using a feature-based learning technique. Set number Time-varying data Food item Texture quality 1 Texture quality 2 1 X(tl) .. .X(tm) Carrot high high 2 X(tl) .. .X(tn) Chocolate low medium 3 X(tl) ...X(tk) Bread medium medium Table 4 Optionally, prior to step S801 at step S800a, the time-varying data is transformed into a frequency domain. This may help extract distinctive features of the time-varying data that can be used to train the machine learning system so that it can better associate the time-varying data with the one or more texture qualities. The subsequent training steps may be performed on the transformed data. Alternatively, this transforming step may be performed at step S801a as part of the feature-based learning technique. In step S801a, one or more of the above-mentioned characteristics (e.g. in the frequency domain and / or time domain) may be determined and associated with the texture qualities of the time-varying data. At step S802, the machine learning system is trained to correlate the time-varying data with the texture qualities. If using the feature-based learning technique, the machine learning system is trained to correlate the one or more characteristics of the time-varying data with the one or more of the predetermined plurality of texture qualities. In the deep-learning technique, the one or more characteristics may not be determined explicitly. Once trained, the machine learning system takes timevarying data captured whilst consuming an unknown food item, and outputs a classified texture quality for each of the one or more of the predetermined plurality of texture qualities. The wearable apparatus may comprise a camera, e.g. camera 205 shown in figure 2. The camera is located on the wearable apparatus so as to, in use when the apparatus is worn by the user consuming a food item, capture an image, video, or other data related to the food item. The inventors have recognized that the accuracy of the system’s food classification can be improved by combining the output of an image classifier (e.g. a neural network adapted and trained to perform image classification, as is known in the art) with the texture classification determined above. Thus, in an example, the processor 350 may further comprise an image classification system 750. The image classification system 750 is a machine learning system trained to operate on the captured image from the camera to generate an image classification I for the food item. Image classification is known in the art and any suitable image classification systems may be used to process the captured image so as to identify the food item. The processor may combine the image classification and the texture classification so to determine the food being eaten. In examples in which the wearable apparatus comprises a pair of sensors for monitoring corresponding muscles in a pair of muscles, the machine learning system may be configured to operate on the timevarying data captured by both sensors so as to classify the food item. This may be advantageous for identifying the cohesiveness of the food item. For example, when chewing particularly sticky foods it is likely that the food will remain on one side of the mouth, which will be reflected in a relatively high measure of asymmetry between skin movement captured by the pair of sensors during chewing. In examples in which the wearable apparatus comprises a lip sensor, the machine learning system may be configured to use data captured by the lip sensor in combination with the time-varying data captured by the sensor so as to classify the food item. This may be advantageous for determining the size of the food item being eaten which may be used to identify the food item being consumed. For example, when taking an initial bite, the user will alter their lips using specific facial muscle activations to anticipate and accommodate the food type and size. The way in which the mouth opens can be detected with a lip sensor positioned on the wearable apparatus so as to monitor an area of skin associated with a lip dilator and / or lip constrictor. In another example, when consuming liquid, the orbicularis oris is contracted to reduce spillage. Conversely, when trying to take a large bite from an apple, the lip dilators such as zygomaticus and depressor labii inferioris are contracted to help move the lips away from the teeth. So, the activation of the lip dilators and / or lip constrictors, as monitored by the lip sensor, can provide information relating to the size of the food being consumed and whether it is a solid or liquid. The above-described methods for identifying the food qualities may be performed by the processor in response to detecting that an eating episode has started. This may be in response to a user input (e.g. the user may inform the system that they are eating), or a signal from an external system, or it may be automatically detected by the system itself using the techniques described above with reference to detecting jaw activity.

Claims

1. A system for classifying food consumed by a user, the system comprising:wearable apparatus comprising a sensor located so as to, in use when the apparatus is worn by a user consuming a food item, monitor an area of skin whose movement is correlated with a facial muscle associated with movement of the jaw, the sensor being configured to capture time-varying data describing movement of the area of skin in a plane defined by the area of skin; anda processor comprising a machine learning system configured to process the captured timevarying data so as to classify the food item according to one or more of a predetermined plurality of texture qualities, each texture quality representing a physical characteristic of the food item.

2. The system as claimed in claim 1, wherein the machine learning system is configured to classify the food item by generating a texture classification comprising a classified texture quality for each of the one or more of the predetermined plurality of texture qualities.

3. The system as claimed claim 2, further comprising:a camera configured to, in use when the user is consuming the food item, capture an image of the food item;at the processor, an image classification system configured to operate on the captured image so as to generate an image classification for the food item, the processor being configured to combine the texture classification for the food item with the image classification so as to identify the food item.

4. The system as claimed in any preceding claim, wherein the processor is configured to transform the time-varying data into a frequency domain prior to classifying the food item.

5. The system as claimed in any preceding claim wherein the machine learning system is configured to process the time-varying data by determining one or more predetermined characteristics of the time-varying data, the machine learning system being trained to correlate the one or more predetermined characteristics with the one or more of the predetermined plurality of texture qualities.

6. The system as claimed in claim 5, wherein the machine learning system is configured to determine the one or more predetermined characteristics of the time-varying data by transforming the timevarying data into a frequency domain, the one or more characteristics comprising a frequency-domain characteristic.

7. The system as claimed in any claim 6, wherein the frequency-domain characteristic comprises one or more of a dominant frequency, spectral entropy, and energy in one or more frequency bands.

8. The system as claimed in any preceding claim, wherein the time-varying data captured during consumption of the food item exhibits at least one cyclic movement of the area of skin.

9. The system as claimed in claim 7 when dependent on claim 5, wherein the one or more characteristics of the time-varying data comprises one or more characteristics of the at least one cyclic movement.

10. The system as claimed in claim 9, wherein the one or more characteristics of the at least one cyclic movement comprises one or more of: a maximum amplitude of the cyclic movement, a duration of the cyclic movement, an average velocity of the cyclic movement, a maximum velocity of the cyclic movement, and a rate of change of velocity of the cyclic movement.

11. The system as claimed in claim 9 or 10, wherein the time-varying data exhibits a plurality of cyclic movements and the one or more characteristics of the plurality of cyclic movements comprises one or more of: an average frequency of the plurality of cyclic movements, a frequency variability of the plurality of cyclic movements, and a change in maximum amplitude of the plurality of cyclic movements overtime.

12. The system as claimed in any preceding claim, wherein the machine-learning system is configured to use a neural network or a Random Forest or Gradient Boosting learning method.

13. The system as claimed in any preceding claim, wherein the machine learning system is a neural network.

14. The system as claimed in claim 3 or any of claims 4 to 13 when dependent on claim 3, wherein the wearable apparatus comprises the camera.

15. The system as claimed in any preceding claim, wherein the system is further configured to receive a user-identified classification of the food item according to the one or more of the predetermined plurality of texture qualities, the system being configured to use the user-identified classification as training data for the machine learning system.

16. The system as claimed in any preceding claim, wherein the system is configured to apply a clustering algorithm to a plurality of sets of time-varying data captured by the sensor during consumption of a plurality of respective food items so as to produce a clustered dataset, the system being configured to receive a user-identified classification of the clustered dataset with respect to the one or more of the predetermined plurality of texture qualities.

17. The system as claimed in any preceding claim, wherein the apparatus comprises a pair of sensors, each sensor in the pair being located so as to, in use when the apparatus is worn by theuser consuming the food item, monitor a respective area of skin whose movement is correlated with a corresponding muscle on either side of the face of the user, each sensor in the pair of sensors being configured to capture time-varying data describing movement of the respective area of skin in a plane defined by that area of skin, the machine learning system being configured to process time-varying data captured by the pair of sensors so as to classify the food item according to the one or more of the predetermined plurality of texture qualities.

18. The system as claimed in any preceding claim, wherein the apparatus comprises a lip sensor located so as to, in use when the apparatus is worn by the user, monitor an area of skin associated with a lip dilator and / or lip constrictor, the lip sensor being configured to capture time-varying data describing movement of said area of skin, the machine learning system being configured to process the time-varying data captured by the lip sensor in combination with the time-varying data captured by the sensor so as to classify the food item according to one or more of the predetermined plurality of texture qualities.

19. The system as claimed in any preceding claim, wherein the apparatus further comprises a motion sensor and / or microphone, the machine learning system being configured to process data captured by the motion sensor and / or microphone in combination with the time-varying data captured by the sensor so as to classify the food item according to one or more of the predetermined plurality of texture qualities.

20. The system as claimed in any preceding claim, wherein the one or more of the predetermined plurality of texture qualities of a food item comprises a measure of hardness and / or a measure of cohesiveness and / or a measure of density of the food item.

21. The system as claimed in any preceding claim, wherein the sensor is configured to capture timevarying data describing movement of the area of skin in two dimensions in the plane defined by the area of skin.

22. The system as claimed in any preceding claim, wherein the sensor is an optical flow sensor configured to capture samples of the area of skin and compare the samples apart in time so as to detect movement of the area of skin relative to the optical sensor.

23. The system as claimed in any preceding claim, wherein the time-varying data represents movement of the area of skin in two dimensions in the plane defined by the area of skin.

24. The system as claimed in any preceding claim, wherein the facial muscle is one of a temporalis muscle, masseter muscle, a medial pterygoid, and a lateral pterygoid.

25. The system as claimed in any preceding claim, wherein the area of skin overlies one or more of: a masticator muscle, a zygomaticus major and / or minor muscle, a levator labii superioris muscle, a levator labii superioris alaeque nasi muscle, a levator anguli oris muscle, a risorius muscle, a buccinator muscle, an orbicularis oris muscle.

26. The system as claimed in any preceding claim, wherein the processor is configured to use the texture classification for the food item for monitoring dietary intake of the user.

27. The system as claimed in claim 26 or 27, wherein the processor is configured to use the texture classification of the food item to estimate what food item was consumed.

28. The system as claimed in claim 27 wherein the processor is configured to estimate what food item was consumed by comparing the texture classification of the food item with texture classifications of previously identified food items.

29. The system as claimed in any of claims 26 to 28, wherein the texture classification of the food item comprises a coordinate of a texture-space, the texture-space having axes corresponding to each texture quality of the one or more of the predetermined plurality of texture qualities.

30. The system as claimed in claim 29, wherein the system is configured to estimate what food item was consumed by comparing the coordinate with coordinates of previously identified food items.

31. A method of classifying food consumed by a user, the method comprising:during consumption of a food item by a user, capturing time-varying data representing movement of an area of skin of the user in a plane defined by the area of skin, the movement of the area of skin being correlated with a facial muscle associated with movement of the jaw of the user; andat a machine learning system trained to correlate time-varying data with one or more of a predetermined plurality of texture qualities, processing the captured time-varying data so as to classify the food item according to the one or more of the predetermined plurality of texture qualities, each texture quality representing a physical characteristic of the food item.

32. A method of training a machine learning system to classify food according to a predetermined plurality of texture qualities, the method comprising:forming training data by associating time-varying data with one or more of the predetermined plurality of texture qualities, the time-varying data representing movement of an area of skin of a user in a plane defined by the area of skin during consumption by the user of a food item having physical characteristics corresponding to the one or more of the predetermined plurality of texture qualities, and the movement of the area of skin being correlated with a facial muscle associated with movement of the jaw of the user; andtraining the machine learning system using the training data to correlate the time-varying data with the one or more of the predetermined plurality of texture qualities so as to enable the machine learning system to classify food according to the predetermined plurality of texture qualities.

33. The method as claimed in claim 32, wherein the forming training data comprises associating each of a plurality of items of time-varying data with a respective set of one or more of the predetermined plurality of texture qualities.

34. The method as claimed in claim 32 or 33 further comprising, prior to training the machine learning system, transforming the time-varying data into a frequency domain and training the machine learning system using the transformed time-varying data.

35. The method as claimed in any of claims 32 to 34, wherein forming the training data comprises determining one or more characteristics of the time-varying data, and associating the time-varying data with the one or more predetermined plurality of texture qualities comprises associating the one or more characteristics of the time-varying data with the one or more predetermined plurality of texture qualities.

36. The method as claimed in any of claims 32 to 34, wherein the time-varying data is captured during consumption by the user of a plurality of food items, each having physical characteristics corresponding to the one or more of the predetermined plurality of texture qualities.

37. The method as claimed in any of claims 32 to 35, wherein the machine-learning system is configured to use a neural network or a Random Forest or Gradient Boosting learning method.

38. The method as claimed in any of claims 32 to 37, wherein the machine-learning system is a neural network.

39. The method as claimed in any of claims 32 to 38, wherein the one or more predetermined texture qualities of a food item comprises a measure of hardness and / or a measure of cohesiveness and / or a measure of density of the food item.

40. The system as claimed in any of claims 27 to 30, wherein the system is configured to receive blood glucose data of the user following consumption of the food item and use the texture classification of the food item and the blood glucose data to estimate what food item was consumed.

41. The system as claimed in any of claims 27 to 30, wherein the system is configured to predict blood glucose data of the user following consumption of the food item using the texture classification of the food item and optionally a history of previously consumed food items.

42. A system for classifying food consumed by a user, the system comprising:wearable apparatus comprising a sensor located so as to, in use when the apparatus is worn by a user consuming a food item, monitor an area of skin whose movement is correlated with a facial5 muscle associated with movement of the jaw, the sensor being configured to capture time-varying data describing movement of the area of skin in a plane defined by the area of skin; anda machine learning system trained to perform food classification based on a plurality of known food items belonging to a plurality of predetermined food groups, the machine learning system being configured to classify the food item being consumed as belonging to one of the predetermined10 plurality of food groups.AMENDMENTS TO THE CLAIMS HAVE BEEN FILED AS FOLLOWS:-13 02 25CLAIMS1. A system for classifying food consumed by a user, the system comprising:wearable apparatus comprising a sensor located so as to, in use when the apparatus is worn by a user consuming a food item, monitor an area of skin whose movement is correlated with a facial muscle associated with movement of the jaw, the sensor being configured to capture time-varying data describing movement of the area of skin in a plane defined by the area of skin; anda processor comprising a machine learning system configured to process the captured timevarying data so as to classify the food item according to one or more of a predetermined plurality of texture qualities, each texture quality representing a physical characteristic of the food item.

2. The system as claimed in claim 1, wherein the machine learning system is configured to classify the food item by generating a texture classification comprising a classified texture quality for each of the one or more of the predetermined plurality of texture qualities.

3. The system as claimed claim 2, further comprising:a camera configured to, in use when the user is consuming the food item, capture an image of the food item;at the processor, an image classification system configured to operate on the captured image so as to generate an image classification for the food item, the processor being configured to combine the texture classification for the food item with the image classification so as to identify the food item.

4. The system as claimed in any preceding claim, wherein the processor is configured to transform the time-varying data into a frequency domain prior to classifying the food item.

5. The system as claimed in any preceding claim wherein the machine learning system is configured to process the time-varying data by determining one or more predetermined characteristics of the time-varying data, the machine learning system being trained to correlate the one or more predetermined characteristics with the one or more of the predetermined plurality of texture qualities.

6. The system as claimed in claim 5, wherein the machine learning system is configured to determine the one or more predetermined characteristics of the time-varying data by transforming the timevarying data into a frequency domain, the one or more characteristics comprising a frequency-domain characteristic.

7. The system as claimed in any claim 6, wherein the frequency-domain characteristic comprises one or more of a dominant frequency, spectral entropy, and energy in one or more frequency bands.13 02 258. The system as claimed in any preceding claim, wherein the time-varying data captured during consumption of the food item exhibits at least one cyclic movement of the area of skin.

9. The system as claimed in claim 7 when dependent on claim 5, wherein the one or more characteristics of the time-varying data comprises one or more characteristics of the at least one cyclic movement.

10. The system as claimed in claim 9, wherein the one or more characteristics of the at least one cyclic movement comprises one or more of: a maximum amplitude of the cyclic movement, a duration of the cyclic movement, an average velocity of the cyclic movement, a maximum velocity of the cyclic movement, and a rate of change of velocity of the cyclic movement.

11. The system as claimed in claim 9 or 10, wherein the time-varying data exhibits a plurality of cyclic movements and the one or more characteristics of the plurality of cyclic movements comprises one or more of: an average frequency of the plurality of cyclic movements, a frequency variability of the plurality of cyclic movements, and a change in maximum amplitude of the plurality of cyclic movements overtime.

12. The system as claimed in any preceding claim, wherein the machine-learning system is configured to use a neural network or a Random Forest or Gradient Boosting learning method.

13. The system as claimed in any preceding claim, wherein the machine learning system is a neural network.

14. The system as claimed in claim 3 or any of claims 4 to 13 when dependent on claim 3, wherein the wearable apparatus comprises the camera.

15. The system as claimed in any preceding claim, wherein the system is further configured to receive a user-identified classification of the food item according to the one or more of the predetermined plurality of texture qualities, the system being configured to use the user-identified classification as training data for the machine learning system.

16. The system as claimed in any preceding claim, wherein the system is configured to apply a clustering algorithm to a plurality of sets of time-varying data captured by the sensor during consumption of a plurality of respective food items so as to produce a clustered dataset, the system being configured to receive a user-identified classification of the clustered dataset with respect to the one or more of the predetermined plurality of texture qualities.

17. The system as claimed in any preceding claim, wherein the apparatus comprises a pair of sensors, each sensor in the pair being located so as to, in use when the apparatus is worn by the13 02 25user consuming the food item, monitor a respective area of skin whose movement is correlated with a corresponding muscle on either side of the face of the user, each sensor in the pair of sensors being configured to capture time-varying data describing movement of the respective area of skin in a plane defined by that area of skin, the machine learning system being configured to process time-varying data captured by the pair of sensors so as to classify the food item according to the one or more of the predetermined plurality of texture qualities.

18. The system as claimed in any preceding claim, wherein the apparatus comprises a lip sensor located so as to, in use when the apparatus is worn by the user, monitor an area of skin associated with a lip dilator and / or lip constrictor, the lip sensor being configured to capture time-varying data describing movement of said area of skin, the machine learning system being configured to process the time-varying data captured by the lip sensor in combination with the time-varying data captured by the sensor so as to classify the food item according to one or more of the predetermined plurality of texture qualities.

19. The system as claimed in any preceding claim, wherein the apparatus further comprises a motion sensor and / or microphone, the machine learning system being configured to process data captured by the motion sensor and / or microphone in combination with the time-varying data captured by the sensor so as to classify the food item according to one or more of the predetermined plurality of texture qualities.

20. The system as claimed in any preceding claim, wherein the one or more of the predetermined plurality of texture qualities of a food item comprises a measure of hardness and / or a measure of cohesiveness and / or a measure of density of the food item.

21. The system as claimed in any preceding claim, wherein the sensor is configured to capture timevarying data describing movement of the area of skin in two dimensions in the plane defined by the area of skin.

22. The system as claimed in any preceding claim, wherein the sensor is an optical flow sensor configured to capture samples of the area of skin and compare the samples apart in time so as to detect movement of the area of skin relative to the optical sensor.

23. The system as claimed in any preceding claim, wherein the time-varying data represents movement of the area of skin in two dimensions in the plane defined by the area of skin.

24. The system as claimed in any preceding claim, wherein the facial muscle is one of a temporalis muscle, masseter muscle, a medial pterygoid, and a lateral pterygoid.13 02 2525. The system as claimed in any preceding claim, wherein the area of skin overlies one or more of: a masticator muscle, a zygomaticus major and / or minor muscle, a levator labii superioris muscle, a levator labii superioris alaeque nasi muscle, a levator anguli oris muscle, a risorius muscle, a buccinator muscle, an orbicularis oris muscle.

26. The system as claimed in any preceding claim, wherein the processor is configured to use the texture classification for the food item for monitoring dietary intake of the user.

27. The system as claimed in claim 26 or 27, wherein the processor is configured to use the texture classification of the food item to estimate what food item was consumed.

28. The system as claimed in claim 27 wherein the processor is configured to estimate what food item was consumed by comparing the texture classification of the food item with texture classifications of previously identified food items.

29. The system as claimed in any of claims 26 to 28, wherein the texture classification of the food item comprises a coordinate of a texture-space, the texture-space having axes corresponding to each texture quality of the one or more of the predetermined plurality of texture qualities.

30. The system as claimed in claim 29, wherein the system is configured to estimate what food item was consumed by comparing the coordinate with coordinates of previously identified food items.

31. A method of classifying food consumed by a user, the method comprising:during consumption of a food item by a user, capturing time-varying data representing movement of an area of skin of the user in a plane defined by the area of skin, the movement of the area of skin being correlated with a facial muscle associated with movement of the jaw of the user; andat a machine learning system trained to correlate time-varying data with one or more of a predetermined plurality of texture qualities, processing the captured time-varying data so as to classify the food item according to the one or more of the predetermined plurality of texture qualities, each texture quality representing a physical characteristic of the food item.

32. A method of training a machine learning system to classify food according to a predetermined plurality of texture qualities, the method comprising:forming training data by associating time-varying data with one or more of the predetermined plurality of texture qualities, the time-varying data representing movement of an area of skin of a user in a plane defined by the area of skin during consumption by the user of a food item having physical characteristics corresponding to the one or more of the predetermined plurality of texture qualities, and the movement of the area of skin being correlated with a facial muscle associated with movement of the jaw of the user; and13 02 25training the machine learning system using the training data to correlate the time-varying data with the one or more of the predetermined plurality of texture qualities so as to enable the machine learning system to classify food according to the predetermined plurality of texture qualities.

33. The method as claimed in claim 32, wherein the forming training data comprises associating each of a plurality of items of time-varying data with a respective set of one or more of the predetermined plurality of texture qualities.

34. The method as claimed in claim 32 or 33 further comprising, prior to training the machine learning system, transforming the time-varying data into a frequency domain and training the machine learning system using the transformed time-varying data.

35. The method as claimed in any of claims 32 to 34, wherein forming the training data comprises determining one or more characteristics of the time-varying data, and associating the time-varying data with the one or more predetermined plurality of texture qualities comprises associating the one or more characteristics of the time-varying data with the one or more predetermined plurality of texture qualities.

36. The method as claimed in any of claims 32 to 34, wherein the time-varying data is captured during consumption by the user of a plurality of food items, each having physical characteristics corresponding to the one or more of the predetermined plurality of texture qualities.

37. The method as claimed in any of claims 32 to 35, wherein the machine-learning system is configured to use a neural network or a Random Forest or Gradient Boosting learning method.

38. The method as claimed in any of claims 32 to 37, wherein the machine-learning system is a neural network.

39. The method as claimed in any of claims 32 to 38, wherein the one or more predetermined texture qualities of a food item comprises a measure of hardness and / or a measure of cohesiveness and / or a measure of density of the food item.

40. The system as claimed in any of claims 27 to 30, wherein the system is configured to receive blood glucose data of the user following consumption of the food item and use the texture classification of the food item and the blood glucose data to estimate what food item was consumed.

41. The system as claimed in any of claims 27 to 30, wherein the system is configured to predict blood glucose data of the user following consumption of the food item using the texture classification of the food item and optionally a history of previously consumed food items.

42. A system for classifying food consumed by a user, the system comprising:wearable apparatus comprising a sensor located so as to, in use when the apparatus is worn by a user consuming a food item, monitor an area of skin whose movement is correlated with a facial muscle associated with movement of the jaw, the sensor being configured to capture time-varying data describing movement of the area of skin in a plane defined by the area of skin; anda machine learning system trained to perform food classification based on a plurality of known food items belonging to a plurality of predetermined food groups, the machine learning system being configured to process the time-varying data so as to classify the food item being consumed as belonging to one of the predetermined plurality of food groups.13 02 25

Images