Neurological movement disorders

Abstract

The present invention relates to methods for monitoring a subject who has been diagnosed as having or likely to have neurological motor disease or disorder, the method comprising: obtaining a pose derived from movement data comprising joint location data for a plurality of joints of the subject while performing a balance test; and determining the cosine similarity between the pose derived from the movement data and a reference pose derived from reference movement data, wherein said cosine similarity is indicative of the presence and / or severity of a neurological motor disease or disorder Related systems and clinical methods are also described.

Description

[0001] NEUROLOGICAL MOVEMENT DISORDERS

[0002] FIELD OF THE INVENTION

[0003] The present invention relates to methods for monitoring, diagnosing or prognosing a subject with a neurological movement disorder. The present invention relates in particular to methods for monitoring, diagnosing or prognosing a subject with a neurological movement disorder by detecting the presence and / or extent of dystonia in a subject by comparing movement data from the subject with control movement data. Related methods and products are also described.

[0004] BACKGROUND TO THE INVENTION

[0005] Clinical monitoring of patients with neurological movement disorders such as Huntington’s disease and Parkinsons’s disease typically uses complex clinical scales. For example, the current state of the art for assessment of patients with Huntington Disease in clinical trials includes the UHDRS (Unified Huntington’s Disease Rating Scale; Huntington Study Group 1996). The UHDRS is a clinical rating scale to assess four domains of clinical performance and capacity in HD: motor function, cognitive function, behavioral abnormalities and functional capacity. Each domain is associated with a series of tests and / or checklist questions, the outcomes of which are scored and added up to produce a score. For example, the motor assessment part of the UHDRS alone includes 15 different tests. These types of clinical scales must be assessed by trained professionals, are time consuming and burdensome for the patient, and are inherently subjective even when applied with the required level of expertise. This means that patient status is impractical to assess frequently, and must be assessed in a clinical context (i.e. excluding regular monitoring in flexible settings).

[0006] Therefore, there remains a need for improved methods for monitoring and / or diagnosing subjects with neurological movement disorders.

[0007] SUMMARY OF THE INVENTION

[0008] The inventors postulated that movement sensors data, particularly imaging based movement sensors obtained from balance tests, may be used to differentiate between patients with neurological movement disorders involving symptoms that affect posture and balance, such as dystonia, provided that an appropriate signal could be extracted from such data. They observed that at a high level shows different trends for healthy subjects and patients with Huntington’s disease (HD), and postulated that the presence and / or extent of a posture or balance disorder could be quantified by comparing the pose of subjects during balance tests to a reference pose corresponding to healthy subjects. They designed a sensitive method to perform this comparison which uses the cosine similarity between the pose of a subject and a reference pose. Using data from a HD cohort that verified the hypothesis that when quantified in this way, the distance between the pose of dystonic patients and a reference pose, is larger than the distance between the pose of a healthy patient and the reference pose. Thus, they showed that the new method can detect the presence and extent of dystonia in patients, as well as differentiate between HD and healthy subjects, and is predictive of existing clinical metrics. Thus, the new method provides a method for monitoring patients with neurological movement disorders that is automated (and hence objective), fast, minimally burdensome and with low equipment requirement.

[0009] Thus, according to a first aspect, the disclosure provides a method of monitoring a subject who has been diagnosed as having or likely to have neurological motor disease or disorder, the method comprising: obtaining, by a processor, a pose derived from movement data comprising joint location data for a plurality of joints of the subject while performing a balance test; and determining, by said processor, the cosine similarity between the pose derived from the movement data and a reference pose derived from reference movement data, wherein said cosine similarity is indicative of the presence and / or severity of a neurological motor disease or disorder.

[0010] The method may have any one or more of the following optional features.

[0011] A reference pose can be a pose derived from movement data for a plurality of reference subjects comprising joint location data for a plurality of joints of the respective subjects while performing a balance test. The method can further comprise obtaining, by said processor, the reference pose from said reference movement data, wherein the reference pose comprises summarized joint location data over a plurality of time points for the plurality of reference subjects. The reference subjects can be healthy subjects. The plurality of reference subjects can comprise at least 10, 20 or 30 reference subjects.

[0012] The method can comprise obtaining, by said processor, movement data comprising joint location data for a plurality of joints of the subject while performing a balance test, and obtaining, by said processor, a pose derived from said movement data, wherein a pose comprises summarized joint location data over a plurality of time points. The method can further comprise said processor applying one or more denoising algorithms to the movement data. The one or more denoising algorithms can be applied prior to obtaining a pose. The one or more denoising algorithms can selected from: (i) a smoothing algorithm, optionally a low pass filter, optionally a Savitsky-Golay filter, (ii) a filter that removes any frame of the movement data where the coordinates of a subset of the plurality of joints satisfy one or more predetermined criteria that apply to angles between said subset of joints, optionally wherein frames where the angle between [Right Shoulder, Spine Shoulder, Left Shoulder] is below a threshold selected between 135° and 155°, optionally 145°, are removed, and / or wherein frames where the angle between [Elbow, Shoulder, Spine Shoulder] is above a threshold selected between 120° and 140°, optionally 130° are removed, and (iii) a filter that removes movement data corresponding to a single balance test, where the movement data comprises fewer than a predetermined number of frames, optionally fewer than 40 frames.

[0013] A frame refers to a particular time point of movement data, associated with coordinates for all measured joints at the time point. Thus, movement data comprises joint location data for a plurality of joints of the subject at a plurality of time points. Each time point may be referred to as a frame. The movement data together may be referred to as a time series, which comprises a plurality of time series each made up of coordinates for a joint at a plurality of time points. When a smoothing algorithm is applied, this may be applied individually to each such time series.

[0014] The movement data can comprise joint location data for a plurality of joints of the subject for a plurality of time points. A balance test can comprise the subject standing at ease for a predetermined period of time. The predetermined period of time can be at least 10 seconds, at least 20 seconds or about 30 seconds. The movement data can comprise joint location data acquired at a sampling rate of at least 10 Hz. The movement data can comprise joint location data for a plurality of joints of the subject while performing a plurality of separate balance tests.

[0015] A pose can comprise average joint locations over a plurality of time points. Joint location data can comprise 2D or 3D coordinates of the plurality of joints. All coordinates can be 3D coordinates. The movement data may have been acquired using a 3D camera. The movement data can have been acquired using a movement sensor selected from a 2D camera and a 3D camera. The movement data can have been acquired using a stereocamera. The neurological disease or disorder can be dystonia or a neurological disease associated with dystonia, posture and / or balance impairment. The monitoring can comprise detecting or assessing dystonia, posture and / or balance impairment. The subject can be a subject who has been diagnosed as having or being likely to have Huntington’s disease, Parkinson’s disease, or having suffered a stroke. The dystonia can be generalized dystonia, multifocal dystonia, segmental dystonia or hemidystonia.

[0016] The plurality of joints can include a plurality of locations on the subject’s torso. The locations on the subject’s torso can comprise: head, base of the neck (shoulder centre), left shoulder, right shoulder, spine at navel level (spine), pelvis (hip centre), left hip, right hip. The plurality of joints can include a plurality of knee-up locations, wherein knee-up locations are joints located from the knee of the subject to the head of the subject. The plurality of joints can not include (i.e. can exclude) joints on the hands and wrists. The knee-up locations can comprise or consist of: head, base of the neck (shoulder centre), left shoulder, right shoulder, spine at navel level (spine), pelvis (hip centre), left hip, right hip, left knee, right knee, and optionally left elbow and right elbow.

[0017] The method can further comprise aligning the subject’s pose to a reference coordinate, wherein aligning the pose to a reference comprises modifying the coordinates of the plurality of joints while maintaining all relative distances between the joints, such that a predetermined joint is at a predetermined position. The method can further comprise normalising the movement data or pose, wherein normalising movement data or a pose derived therefrom comprises adjusting the coordinates of the plurality of joints in each frame of the movement data or in the pose based on the distance between a predetermined pair of joints in the frame.

[0018] According to a second aspect, there is provided a method of selecting a subject for participating in a clinical trial, the method comprising monitoring the subject using the method of any embodiment of the first aspect and selecting the subject for participating in a clinical trial when the cosine similarity or a value derived therefrom satisfies a predetermined criterion or the subject is classified in one of a plurality of classes using a machine learning model trained to classify subjects using values of a plurality of biomarkers comprising said cosine similarity or a value derived therefrom.

[0019] Also described herein according to a third aspect is a method of determining the effect of a therapeutic compound or composition for treating a neurological movement disease or disorder, or recommending a therapeutic compound or composition for treating a neurological movement disease or disorder in the subject, the method comprising monitoring the subject using the method of any embodiment of the first aspect and determining that the therapeutic compound or composition is effective in treating the neurological movement disease or disorder, or recommending the subject for treatment with the therapeutic compound or composition when the cosine similarity or a value derived therefrom satisfies a predetermined criterion or the subject is classified in one of a plurality of classes using a machine learning model trained to classify subjects using values of a plurality of biomarkers comprising said cosine similarity or a value derived therefrom.

[0020] According to a fourth aspect, there is provided a method of treating a subject with a neurological movement disease or disorder, the method comprising: monitoring the subject using the method of any embodiment of the first aspect; selecting the subject for treatment with a therapeutic compound or composition for treating a neurological movement disease or disorder when the cosine similarity or a value derived therefrom satisfies a predetermined criterion or the subject is classified in one of a plurality of classes using a machine learning model trained to classify subjects using values of a plurality of biomarkers comprising said cosine similarity or a value derived therefrom; and administering the therapeutic compound or composition in an effective amount. The therapeutic compound or composition can be a therapeutic compound or composition recommended for treatment of subjects with dystonia, or with symptomatic HD or PD. The dystonia can be generalized dystonia, multifocal dystonia, segmental dystonia or hemidystonia.

[0021] According to any embodiment of any aspect, the method may further comprise providing to a user, through a user interface, one or more of the value of the cosine similarity, an indication of a classification derived from said cosine similarity, a recommended therapy associated with said cosine similarity or classification, and a report comprising any one or more of the above.

[0022] According to a further aspect, there is provided a system comprising: a processor; and a computer readable medium comprising instructions that, when executed by the processor, cause the processor to perform the (computer-implemented) steps of the method of any preceding aspect. The system can further comprise a movement sensing device operably coupled to the process for acquisition of the movement data. The system can further comprise one or more cameras. The one or more cameras can each be 2D or 3D cameras. A 3D camera can be a stereocamera. According to a further aspect, there is provided a non-transitory computer readable medium or media comprising instructions that, when executed by at least one processor, cause the at least one processor to perform the method of any embodiment of any aspect described herein.

[0023] According to a further aspect, there is provided a computer program comprising code which, when the code is executed on a computer, causes the computer to perform the method of any embodiment of any aspect described herein.

[0024] DESCRIPTION OF FIGURES

[0025] Figure 1 shows an embodiment of a system for monitoring a subject, according to the disclosure.

[0026] Figure 2 is a flow diagram showing, in schematic form, a subject monitoring method according to the disclosure.

[0027] Figure 3 shows joint location data for subjects with HD (blue) compared to a reference (red), in a balance test. From left to right the data shows: data for a subject with the largest dissimilarity to control amongst the cohort of HD patients tested (manifest group), data for a subject that has a median dissimilarity to control in the cohort (premanifest group), and data for a subject with the lowest dissimilarity to control in the cohort (premanifest group). The manifest group included the subset of HD patients in the cohort that showed symptoms of HD (i.e. clinically diagnosed as having HD), and the premanifest group included the subset of HD patients in the cohort that carry a HD mutation but are not clinically diagnosed as having HD. Axes show distances in meters to a reference coordinate (spine joint).

[0028] Figure 4 shows dissimilarity scores between poses in a balance test of every 2 pairs of subjects in a study comprising healthy, premanifest and early to moderate HD patients. The dissimilarity between poses was calculated as described on Figures 5-6 except that pairs of subjects were used instead of a subject and a reference pose. The resulting values (i.e. for each subject, a vector of dissimilarity scores to all other subjects) were then analysed by principal component analysis and the plot shows the values along the first two principal components, for individual subjects (points) and curves fitted to groups of subjects (lines and shaded areas - lines are fitted linear models and shaded areas are 95% confidence interval around fitted model). The points and curves are coloured according to the subject group: group 1 (blue)=healthy, group 2 (orange)=premanifest, group 3 (green)=manifest severity stages l-ll, group 4 (red)=manifest severity stage III.

[0029] Figure 5 illustrates schematically the concept of cosine similarity used in the disclosure.

[0030] Figure 6 illustrates schematically a process for quantifying a biomarker of dystonia used in embodiments of the disclosure.

[0031] Figure 7 illustrates sets of joints that are used in embodiments of the present disclosure. A. Torso 18 joints set. B. Knee-up 1 12 joints set. C. no hands and feet 1 16 joints set. D. Full joint set (20 joints).

[0032] Figure 8 shows a correlation matrix (Pearson correlation coefficients) showing the correlation between the similarity metric calculated according to Figure 6 and clinical metrics, for a cohort of HD patients. A. 8 joints, B. 12 jointssim_score=score calculated as explained on Figure 6, CAG=number of CAG repeats in the HTT gene, group=disease group as on Figure 4 (i.e. healthy, premanifest, severity stages l-ll, severity III), dystonia=sum of dystonia scores in UHDRS (i.e. maximal dystonia trunk, maximal dystonia right upper extremity, maximal dystonia left upper extremity, maximal dystonia right lower extremity, maximal dystonia left lower extremity), tfctot=total functional capacity, tmstot=total motor score, age=age of the patients, dclscore=diagnostic confidence level. Disease groups were assigned by clinicians during recruitment screening, using the definitions of the stages of HD as described in Shoulson, 1981. Diagnostic confidence levels is a measure of motor abnormality assessed by clinicians as: 0 = normal (no abnormalities), 1 = non-specific motor abnormalities (less than 50% confidence), 2 = motor abnormalities that may be signs of HD (50-89% confidence), 3 = motor abnormalities that are likely signs of HD (90-98% confidence), 4 = motor abnormalities that are unequivocal signs of HD (>99% confidence).

[0033] Figure 9 shows the distributions of similarity scores calculated according to Figure 6 for a cohort of HD patients and healthy controls, separated by patient group. A, B: 8 joints. C, D. 12 joints. A, C. Distribution of dissimilarity scores (1 - cosine similarity between patient and reference skeleton) in healthy patients (group 1.0, left), premanifest HD patients (group 2.0, middle) and manifest HD patients (group 3.0, right). Each boxplot shows the median and interquartile range (IQR, extends between Q1 and Q3, respectively the 25thand 75th percentiles), whiskers extend to min(min observed value, Q1-1.5*IQR) and max(max observed value, Q3+1.5*IQR). B, D. Same data as in A, C but showing only the healthy (left) and HD manifest (right) groups. In C (12 joints): premanifest HD patients vs manifest HD patients (2.0 v.s. 3.0): t-test independent samples with Bonferroni correction, p- value=3.111*10'6, t statistic=-5.534; healthy control vs premanifest HD patients (1.0 v.s. 2.0): t-test independent samples with Bonferroni correction, p-value=3.468*10'4, t statistic=4.356; healthy control vs manifest HD patients (1.0 v.s. 3.0): t-test independent samples with Bonferroni correction, t statistic=-0.7130, non-significant. In D (12 joints), healthy control vs manifest HD patients (1.0 v.s. 3.0): t-test independent samples with Bonferroni correction, p-value=9.347*1 O'4, t statistic=-3.456. In A (8 joints): premanifest HD patients vs manifest HD patients (2.0 v.s. 3.0): t-test independent samples with Bonferroni correction, p-value=4.484*1 O'2, t statistic=2.517; healthy control vs premanifest HD patients (1.0 v.s. 2.0): t-test independent samples with Bonferroni correction, p-value=1 .678*1 O'2, t statistic=2.959; healthy control vs manifest HD patients (1.0 v.s. 3.0): t-test independent samples with Bonferroni correction, t statistics .112, non-significant. In B (8 joints), healthy control vs manifest HD patients (1.0 v.s. 3.0): t-test independent samples with Bonferroni correction, p-value=1.409*10'1, t statistic=1 .489.

[0034] Figure 10 shows a scatterplot of dissimilarity scores calculated according to Figure 6 (y axis) as a function of age (x axis) for a cohort of HD patients (Group 3=manifest HD patients, Group 2=premanifest HD patients) and healthy controls (Group 1 ). This indicates that the dissimilarity score increases with age within the manifest patient group, but not in the control or premanifest group. This confirms that dystonia as quantified herein is associated with disease severity and not age. Indeed, the age of onset for Huntington Disease patients is between 30 and 50 years old. Once the patients start to show the symptom, their motor functions decrease as they age because of disease progression. The progression rate of individuals is dependent on the number of CAG repeats in their HTT gene (as is the age of onset), which explains at least part of the variability in the scores within the patient population. Nevertheless, the metrics described herein are able to capture disease severity as evidenced by increasing values as the patient population ages (and their disease worsens).

[0035] DETAILED DESCRIPTION

[0036] In describing the present invention, the following terms will be employed, and are intended to be defined as indicated below. In this specification and the appended claims, the singular forms "a", "an" and "the" include plural referents unless the context clearly dictates otherwise. The terms "a" (or "an"), as well as the terms "one or more," and "at least one" can be used interchangeably herein. Furthermore, "and / or" where used herein is to be taken as specific disclosure of each of the two specified features or components with or without the other. Thus, the term "and / or" as used in a phrase such as "A and / or B" herein is intended to include "A and B," "A or B," "A" (alone), and "B" (alone). Likewise, the term "and / or" as used in a phrase such as "A, B, and / or 0" is intended to encompass each of the following aspects: A, B, and C; A, B, or C; A or C; A or B; B or C; A and C; A and B; Band C; A (alone); B (alone); and C (alone).

[0037] Wherever aspects are described herein with the language "comprising," otherwise analogous aspects described in terms of "consisting of' and / or "consisting essentially of' are also provided. The term "about" as used in connection with a numerical value throughout the specification and the claims denotes an interval of accuracy, familiar and acceptable to a person skilled in the art. In general, such interval of accuracy is ± 15 %. Units, prefixes, and symbols are denoted in their Systeme International des Unites (SI) accepted form. Numeric ranges are inclusive of the numbers defining the range. The headings provided herein are not limitations of the various aspects or aspects of the disclosure, which can be had by reference to the specification as a whole.

[0038] As used herein, the terms “computer system” of “computer device” includes the hardware, software and data storage devices for embodying a system or carrying out a computer implemented method, such as e.g. for printing a 3D object as part of a 3D printing system. For example, a computer system may comprise one or more processing units such as a central processing unit (CPU) and / or a graphical processing unit (GPU), input means, output means and data storage, which may be embodied as one or more connected computing devices. Preferably the computer system has a display or comprises a computing device that has a display to provide a visual output display (for example in the design of the business process). The data storage may comprise RAM, disk drives or other computer readable media. The computer system may include a plurality of computing devices connected by a network and able to communicate with each other over that network. For example, a computer system may be implemented as a cloud computer. The term “computer readable media” includes, without limitation, any non-transitory medium or media which can be read and accessed directly by a computer or computer system. The media can include, but are not limited to, magnetic storage media such as floppy discs, hard disc storage media and magnetic tape; optical storage media such as optical discs or CD-ROMs; electrical storage media such as memory, including RAM, ROM and flash memory; and hybrids and combinations of the above such as magnetic / optical storage media.

[0039] The methods described herein are computer implemented unless context indicates otherwise. Indeed, the methods described herein relate to the use of a custom designed video game for the treatment of symptoms of motor neuron disease, and as such by definition requires the present of a gaming engine, graphics processor, etc.

[0040] The methods described herein may be provided as computer programs or as computer program products or computer readable media carrying a computer program which is arranged, when run on a computer, to perform the method(s) described herein. As used herein, the term “computer readable media” includes, without limitation, any non-transitory medium or media which can be read and accessed directly by a computer or computer system. The media can include, but are not limited to, magnetic storage media such as floppy discs, hard disc storage media and magnetic tape; optical storage media such as optical discs or CD-ROMs; electrical storage media such as memory, including RAM, ROM and flash memory; and hybrids and combinations of the above such as magnetic / optical storage media.

[0041] Systems

[0042] Figure 1 shows an embodiment of a system that may be used according to embodiments of the present disclosure. For example, the system may be used for analysing motion data from a subject 1 , for diagnosing dystonia in a subject 1 , for providing a prognosis for a subject 1 , for treating a subject 1 , for selecting a subject 1 for participating in a clinical trial, and / or for monitoring a subject 1. The system comprises a computing device 4, which comprises a processor 401 and computer readable memory 402. In the embodiment shown, the computing device 1 also comprises a user interface 403, which is illustrated as a screen but may include any other means of conveying information to a user such as e.g. through audible or visual signals, by producing a report, etc. The computing device 4 is communicably connected, such as e.g. through a network 3, to one or more motion sensors 204, here illustrated as a camera, and / or to a computing device 2 (comprising one or more processors 201 and one or more memories 202) or database 203 associated with the motion sensors 204. The one or more motion sensors (e.g. cameras) 204 are used to record movements of the subject 1 , and the recorded movements are analysed by the processor 201 and / or the processor 401 using instructions stored on memories 402 and / or 202. For example, raw data from the sensor 204 may be analysed by processor 201 (e.g. using instructions stored on memory 202) to produce joint coordinates data which may be stored in memory 202, database 203 and / or directly provided to computing device 4. The joint coordinates data may be analysed by processor 401 (e.g. using instructions stored on memory 402) to produce a diagnosis and / or prognosis for the subject 1 using the joint coordinates data. Any other combination of locations of processing steps and storing of instructions may be used, such as e.g. all processing being done at computing device 2 or computing device 4. The computing device 4 may be a server, smartphone, tablet, personal computer or other computing device. The computing device 4 is configured to implement methods as described herein, and may be referred to as “remote computing device” because it may not be physically located near the sensor 204, and may process motion data at a location and / or time remote from data acquisition. Communication between the computing device 2 and the remote computing device 4 and / or database 203 may be through a wired or wireless connection, and may occur over a local or public network 3 such as e.g. over the public internet. The motion sensors 204 are typically in wired connection with the processor 201 , but they may instead or in addition be in wired or wireless connection with computing device 4 and / or in wireless connection with processor 201 . Any of the steps of any method described herein may be implemented by processor 201 and / or processor 401 , executing instructions stored on memory 202 and / or memory 402. Computing device 4 may be configured to store data (such as e.g. any output of any step of any method described herein, movement data, a report comprising a prognosis or diagnosis obtained using a method as described herein) on memory 402 and / or to provide such data to a further computing device (not shown) such as e.g. a computing device associated with a user, such as a healthcare professional.

[0043] As used herein, the term “motion sensor” refers to a sensor that is capable of measuring the 2D or 3D position of one or more features of a subject (also referred to herein as “joints”). A motion sensor or set of motion sensors may also be referred to as a motion capture system. A motion capture system may comprise one or more imaging based sensors and / or one or more inertial measurement units (I Mils, also referred to as wearable motion sensors). An IMU is a device that measures linear acceleration and rotational rate (angular velocity). IMUs typically comprise a gyroscope and an accelerometer. A motion capture system is a system that comprises only imaging based sensors (e.g. a single motion sensor in the form of a camera or set of cameras). Indeed, systems that rely on wearable motion sensors are much more cumbersome to set up as they require positioning of a plurality of sensors on the patient. By contrast, imaging-based sensors can be set up without patient manipulation (i.e. requiring no physical interaction with the patient or special expertise) and with little to no equipment other than the sensor itself. The motion sensor is preferably a markerless motion sensor. A markerless motion sensor is a sensor that does not rely on the presence of a specific marker (e.g. reflective marker) on the subject. The motion sensor may be configured to acquire measurement in a continuous manner. As the skilled person understands, continuous measurement refers to the acquisition of measurements at a predetermined sampling frequency. In other words, continuous measurements refer to the acquisition of time series comprising data at each of a plurality of time frames. The time frames are defined by the sampling frequency of the sensor. The predetermined sampling frequency may also be referred to as the nominal sampling frequency of sensor. For example, a sensor may have a nominal sampling frequency is 30 Hz, which should result in measurements recorded every 33ms. In practice the sampling rate of a sensor may not be perfectly consistent, and some variation around the nominal sampling frequency may be expected. For example, using a sensor with a nominal sampling frequency of 30 Hz may result in time series comprising time frames that are usually separated by 33ms, as well as time frames separated by longer periods such as e.g. 100 ms, or shorter periods. The motion sensor may comprise a camera. A camera may be a 2D camera or a 3D camera. A 2D camera may be for example a Webcam connected to a computing device. A 2D camera may be a device configured to obtain 2 dimensional images of a subject. A motion sensor comprising a 2D camera may be configured to measure the 2D position of one or more features (locations) of a subject and derive 3D coordinates from these. 2D cameras are advantageously cheap and widely available. Further, a plurality of 2D cameras may be used and their signals may be combined to derive 3D coordinates. A camera may be a 3D camera, such as a stereo camera. A 3D camera may be a device configured to obtain information comprising 3D images. A 3D camera may be used to obtain 3D position information for one or more features (locations) of a subject, for example derived from information comprising the 2D coordinates (e.g. x, y coordinates) and depth (e.g. z coordinate) of one or more features of a subject. The features for which a position may be measured may comprise one or more joints of a subject, as further described below. A motion sensor comprising a 3D camera may be selected from: Microsoft Kinect, Intel Realsense, Orbbec, and OAK-D. The Microsoft Kinect motion sensor (more information at: azure.microsoft.com / en-us / services / kinect-dk / ) may be able to measure the 3D coordinates of 32 joints. The Intel Realsense sensor (more information at: www.intelrealsense.com) may be able to measure the 3D coordinates of 18 joints. The Orbbec sensor (more information at: orbbec3d.com) may be able to measure the 3D coordinates of 19 joints. In embodiments, the motion sensor is a Microsoft Kinect sensor. The present inventors found this sensor to provide very high accuracy and detailed information about a subject’s movements. Any movement sensor that can record the 2D or 3D localisation (e.g. coordinates in a 2D or 3D reference system) of a plurality of joints in real time (i.e. at a sampling rate of at least 10 Hz, such as e.g. between 10 and 30 Hz) may be used in the context of the present disclosure. It is advantageous for the sensor to be a sensor that can record the 3D localization of a plurality of joints as a time series (i.e. at a plurality of time points). Indeed, the present inventors found 3D location data to lead to more accurate results in the methods described, particularly for subjects that have unstable balance on the z (depth) axis. However, 2D location data is still likely to be informative in many circumstances.

[0044] Detecting dystonia, posture and balance disorders

[0045] The present disclosure provides methods of detecting the presence and / or severity of dystonia, posture and balance disorders, using movement data from said subject acquired during one or more balance tests.

[0046] Dystonia is a movement disorder characterized by involuntary muscle contractions that cause slow repetitive movements or abnormal posture. The most common type of dystonia is focal dystonia that affects the neck in particular, in which case it is referred to as “cervical dystonia” (or spasmodic torticollis). The methods described herein are particularly concerned with dystonia that is not limited to the neck (i.e. not cervical dystonia). This may be generalized dystonia (affecting most or all of the body), focal dystonia (localized to a specific part of the body) that is not cervical dystonia, multifocal dystonia (involving two or more parts of the body), segmental dystonia (affects two or more adjacent parts of the body), or hemidystonia (involving the arm and leg on the same side of the body). Thus, the term “dystonia” as used herein may refer to generalized, multifocal, segmental or hemidystonia. Dystonia is very frequent in clinically symptomatic HD patients. For example, Louis et al. (1999) identified the prevalence of dystonia of any severity in a HD cohort to be 95.2%. In this cohort, the dystonia was present in several body regions and involved a variety of pathological movements and postures. A variety of methods have been proposed to study cervical dystonia (CD). Each of these quantified angles of rotation that specify head pose. For example, Zhang et al, 2022 proposed a method that uses a deep neural network to estimate the 3D projections of facial landmarks from 2D video recordings, from which the three angles of rotation that specify head pose (roll - corresponding to laterocollis tilt, pitch - corresponding to anterocollis I retrocollis, and yaw - corresponding to torticollis I rotation) are inferred with a generalized direct least-square method. These were shown to correlate with the Toronto Western Spasmodic Torticollis Rating Scale (TWSTRS). The method is strictly limited to CD, is computationally intensive, and does not rely on a standard movement test. Nakamura et al. 2019 used data from the Kinect sensor to measure neck angles (angles of the yaw axis (rotation), roll axis (lateral tilting), and pitch axis (sagittal flexion and extension) of the neck by tracking the position of the subject’s face, shoulders, and trunk) and semi-automatically calculate the TWSTRS severity scale score (some items of the scale being automatically calculated from the measured angles and others being manually assessed by an expert). The method essentially replaces some of the items of the TWSTRS with automated versions that perform the same assessment. It is again strictly limited to CD. A similar approach is proposed by Ye et al. (2022), who used a combination of a frontal Kinect 3D camera and a lateral 2D camera to calculate the yaw axis angles (rotation), roll axis angles (lateral tilting), and pitch axis angles (sagittal flexion and extension), which are then used to score the rotation (torticaput / torticollis), laterocollis (latercaput), and antecollis / retrocollis (antecaput / retrocaput) subscales of the TWSTRS severity scale. Again, the emphasis is on using cameras to automate the same assessments that are made manually in the TWSTRS. Approaches based on inertial sensors have also been proposed. For example, den Hartog et al. (2022) proposed a method of measuring dystonia in patients with cerebral palsy using machine learning analysis of data from 4 IMUs attached to the wrists and ankles, respectively, of subjects. The approach is cumbersome in that it requires the use of sensors worn by the subject, computationally intensive since 11 signals are obtained for each sensor at each time point of a time series (4 accelerations, 4 angular velocities and 3 Euler angles), and not interpretable as the data is analyzed by extracting time domain and frequency features by fast Fourier transform, which are the input of classification models (i.e. it is difficult to trace back the physical manifestations that caused the movement data to display features that eventually led to a patient being classified in a particular severity class). Park et al. (2019) proposed an approached for evaluation of CD based on IMUs that is more similar to the imaging-based ones above, where IMUs were attached to the head and neck of subjects and used to quantify involuntary movements of the neck by the rotation angle (RA) and magnitude of angular velocity (MAV). The mean and peak RA and MAV are then correlated with TWSTRS scores. This suffers from similar drawbacks as the imaging-based methods for CD evaluation mentioned above, trading simplicity of analysis of the data for a less practical set up due to the use of wearable sensors. In the context of posture assessment, Zhang et al. (2021 ) proposed a method to assess patients with Parkinson’s disease using Kinect and machine learning. The method comprises processing 3D data from the Kinect to obtain eight quantified coronal and sagittal features of the trunk (i.e. 2D features) chosen to map to predetermined postural angles that are known to be affected in postural abnormalities. These were quantified through a series of steps (stand at ease for 5s, actively correct their posture for 5s, turn left 90° then stand at ease for 5s, then actively correct the posture for another 5s), then used as input to a machine learning model trained to predict the corresponding doctors’ MDS-UPDRS-III 3.13 (the 13th item of the third part of Movement Disorder Society-Sponsored Revision of the Unified Parkinson’s Disease Rating Scale) scores. Because of the choice to map 3D data to specific predetermined 2D angles, the data processing is complex and the eventual result is heavily impacted by the prior knowledge involved in selecting the particular features that are quantified. Further, an 8 features vector for a patient is not directly interpretable, and the data collection process relies on a non-standard test requiring expert monitoring. By contrast with any of these approaches, the present disclosure provides methods that measure body dystonia (i.e. not CD) based on a new metric of posture abnormality assessment during a balance test. This approach is applicable to set ups comprising even a single camera (either 2D or 3D, preferably 3D), and involves quantification of a metric that is new in this context but intuitively interpretable as a score between 0 and 1 quantifying departure from a healthy pose. Further, the approach is extremely computationally efficient as it uses the 3D or 3D joint coordinates from the motion sensors without complex image processing, machine learning etc., and uses data from an extremely simple and non- burdensome test that is widely practiced in the clinic.

[0047] The principles described herein are expected to apply equally to any balance test, and particularly to standing balance tests. As used herein, a balance test is a standard clinical test in which a subject is asked to stand for a predetermined period of time. In embodiments, the subject is asked to stand at ease, with their arms by their side. The subject may be asked to stand as still as possible. The predetermined period of time is typically 30 seconds but can be any time between 10 seconds and 60 seconds or more. Thus, movement data recorded during a balance test may be any movement data acquired while a subject is standing at ease with their arms by their side for at least a predetermined period of time, whether the subject was formally asked to do so or not. In some versions of the balance test, subjects may be asked to stand with their arms crossed. The methods described herein are equally applicable to such contexts. Such balance tests are also known as Berg balance tests. They are commonly performed as part of the Berg balance scale (Berg et al. 1992), to assess elderly patients with balance impairment and patients with acute stroke. Any of the standing balance tests in the Berg Balance scale may be used in the context of the present disclosure, such as e.g. “standing unsupported” (in which a subject is asked to stand for up to two minutes without holding on), “standing unsupported with eyes closed” (in which a subject is asked to close their eyes and stand still for 10 seconds), “standing unsupported with feet together” (in which a subject is asked to stand for up to 1 minute with their feet together without holding on), “standing unsupported one foot in front” (in which a subject is asked to stand for up to 30 seconds with one foot directly in front of the other).

[0048] The methods described herein are able to distinguish between dystonia and chorea. Chorea is a movement disorder characterized by repetitive, brief, irregular involuntary movements, typically involving the face, mouth, trunk and limbs. By contrast, dystonia is slower although it can be episodic (i.e. a patient may experience dystonia in a nonpermanent manner). This is because the methods described herein combine signal that has been acquired over the duration of a balance test, where the effect of sudden movements can be distinguished from prolonged postural abnormalities such as those associated with chorea. Thus, it is particularly advantageous for the movement data used in the present methods to be summarized (e.g. averaged) over a period of time, such as a part of or an entire balance test or even a plurality of balance tests for the same subject.

[0049] The term “movement data” refers to data comprising coordinates of a plurality of joints as a time series (i.e. data from one or more motion sensors as described above). The term “joint” as used herein refers to a location on the body of a subject, the spatial position of which can be measured using a motion sensor as described herein. A joint is typically but not necessarily located on a physical joint of the subject. A joint may be selected from a location that aligns with a subject’s: head, base of the neck (which can also be referred to as “shoulder center”), left shoulder, right shoulder, spine at navel level (which can also be referred to as “spine”), pelvis (which can also be referred to as “hip centre”), left hip, right hip, left knee, right knee, left ankle, right ankle, left foot, right foot, left elbow, right elbow, left wrist, right wrist, left hand, right hand. An example of such a set of joints as measured by the Kinect platform is shown on Figure 7D. In the methods of the present disclosure, the plurality of joints may include one or more locations on each limb and the torso of the subject. The plurality of joints may exclude joints associated with feet, hands and fingers (e.g. finger-tip joints). These are typically measured with higher levels of noise and are less likely to be informative of posture. For example, the plurality of joints can include locations of the subject’s: head, base of the neck (shoulder centre), left shoulder, right shoulder, spine at navel level (spine), pelvis (hip centre), left hip, right hip, left knee, right knee, left ankle, right ankle, left elbow, right elbow, left wrist, right wrist. Such as set of joints is referred to herein as “no hands / feet” of “16 joints” set. An example of such a set of joints is illustrated on Figure 7C. The plurality of joints may additionally exclude joints associated with wrists and ankles. The present inventors have found these joints to be particularly noisy in the context of the biomarker metrics described herein, and have therefore found exclusion of these joints ot lead to biomarkers that better correlate with clinical variables. Thus, the plurality of joints can include a subset of the above locations, such as a plurality of knee-up locations, wherein knee-up locations are joints located from the knee of the subject to the head of the subject, or a plurality of torso locations. The plurality of torso locations can comprise: head, base of the neck (shoulder centre), left shoulder, right shoulder, spine at navel level (spine), pelvis (hip centre), left hip, right hip. Such a set of joints can be referred to herein as “torso” or “8 joints” set. An example of such a set of joints is illustrated on Figure 7A. The plurality of knee-up locations can comprise the torso locations and the left and right knee locations. The plurality of knee-up locations can further comprise the left and right elbow locations. Such a set of joints can be referred to herein as “knee-up” or “12 joints” set. An example of such a set of joints is illustrated on Figure 7B. Thus, the plurality of knee-up locations can exclude all extremities locations from the wrists and ankles, i.e. hands, wrists, ankle and feet locations. The present inventors found that the hands and wrist locations did not add signal that was not already captured by the elbow locations for the purpose of dystonia detection, and as such not including them improved the simplicity and signal to noise ratio of the method. Any set of joints that is intermediate between the above may be used, such as e.g. a set including the torso joints and any one or more of: the left and right elbow joints, the left and right knee joints, the left and right wrist joints, the left and right ankle joints, the left and right feet joints and the left and right hand joints may be used. In embodiments, additional or alternative joints may be used such as e.g. a spine at chest level joint (instead or in addition to the spine at navel level joint), a left and right clavicle joint (instead or in addition to the left and right shoulder joints), and / or a left hand tip and right hand tip joint (instead or in addition to the left and right hand joints). A set of joints is preferably bilateral, i.e. including both left and right joints for any bilateral joint. As will be described further below, movement data for a subject is analysed by obtaining summarized coordinates for each of a plurality of joints over a time series of coordinates. In other words, for each joint, a time series of 2D or 3D coordinates is obtained, and a summarized set of coordinates (comprising 2 coordinates in the case of 2D coordinates and 3 coordinates in the case of 3D coordinates) is obtained for each joint. A summarized set of coordinates may be an average or median, or trimmed versions thereof, of a set of coordinates. In embodiments where a smoothing and / or filtering is applied to the data prior to summarization, the average or median can be used. Indeed, the effect of outlier coordinates due to e.g. jerky movements would already have been attenuated by the smoothing and / or filtering process. A plurality of sets of coordinates (each associated with a joint) may together be referred to as a “pose” or “subject’s pose”. Thus, movement data may also be seen as comprising a time series of poses, and a summarized version thereof may also be referred to as a “pose”.

[0050] The methods described herein make use of a reference pose. A reference pose comprises coordinates of a plurality of joints. A reference pose can be compared with coordinates for corresponding joints in a subject, in order to assess the subject as described herein. A reference pose may be derived or may have been derived from movement data for a plurality of reference subjects. The movement data may have the same characteristics as movement data obtained to assess a subject. For example, it may have been collected during one or more balance tests, it may have been preprocessed in the same manner, and / or it may have been acquired using the same type of motion sensors. The plurality of reference subjects may comprise or consist of healthy subjects. Healthy subjects are subjects that are not known to have a neurological movement disorder. The plurality of subjects may comprise at least 10, at least 20, at least 30 or about 40 subjects. The plurality of subjects may be within the same age range as the subject to be assessed. In other words, movement data for a subject to be assessed may be compared with a reference pose derived from movement data for a plurality of subjects within a predetermined age range that encompasses the age of the subject. Alternatively, the same reference pose may be used regardless of the age of the subject. Indeed, dystonia is not believed to be associated with aging, i.e. there is no correlation between age and dystonia for healthy subjects, and for disease subjects (e.g. HD patients) the severity of dystonia typically correlates with the number of CAG repeats that the individual carries (see e.g. data on Figure 8). In embodiments, a similarity score comparing a subject and reference pose may be assessed in an age-specific manner, such as e.g. by comparison to a predetermined threshold that is dependent on the age of the subject. For example, when considering a subject or study population with a wide range of ages or large differences between the subject I study group and the subjects from which the reference pose has been obtained (e.g. when comparing subjects with a more than 20 years difference in age, such as e.g. a reference pose obtained for subjects in the 40-60 years old age range and a subject / study population in the 20-40, 40-80 or 80-100 years old range), a threshold on similarity score indicating an abnormal posture may depend on the age of the subject whose posture is being assessed.

[0051] The methods described herein comprise obtaining a score that quantifies the similarity between a subject’s pose and a reference pose. The similarity is calculated as a cosine similarity between the two poses to be compared, or a distance derived therefrom as (1 - cosine similarity). Given a pose for a subject s: ps= {ji , js jn} where n is the number of joints, and ji = {Xi, y,, z} are the Cartesian coordinates of the joint i for patient s, and a reference pose pR= { ki, k2 kn} where ki = {ki, ki, k} are the Cartesian coordinates of the joint i for the reference pose PR, the cosine similarity is calculated as: Cos(ps, pR) = ps. pR / , ||.|| is the norm of the vector. This can also be expressed as cos(ps, The cosine similarity is a score between 0 and 1 , where poses that are more similar to each other have higher scores. In order to obtain a metric that is more intuitively interpretable as severity, the cosine similarity can be transformed into an equivalent distance metric = 1 - Cos(ps, pR). Reference to a “score”, “similarity”, or “similarity score” encompass both the cosine similarity and its distance complement, unless context indicates otherwise.

[0052] Figure 2 is a flow diagram showing, in schematic form, a method according to the disclosure. The method may comprise at step 200, receiving movement data comprising coordinates of a plurality of joints of the subject, at a plurality of time points (i.e. a time series of joint coordinates). The movement data may be received as it is being acquired, or may have been previously acquired and may be received from a computing device, user interface or memory. Receiving movement data comprising coordinates of a plurality of joints of the subject at a plurality of time points may comprise receiving data from one or more cameras (directly or indirectly, such as e.g. through a computing device associated with the one or more cameras). The movement data is movement data that has been recorded during a balance test. In other words, the plurality of time points may be time points during which the subject was standing at ease with their arms by their side. At step 210, the data may be pre-processed by applying one or more denoising algorithms. For example, the data may be filtered to remove any frame that satisfies one or more predetermined criteria that apply to the angle between predetermined joints, such as the angle between predetermined joints being above or below a predetermined threshold. The predetermined threshold and the predetermined joints may be selected to exclude noisy frames and / or frames where the individual performed an unexpected movement (e.g. where the joint coordinates are indicative of a departure from a range of expected poses considering the balance test that is being performed). In embodiments, frames where the angle between [Right Shoulder, Spine Shoulder, Left Shoulder] was below 145° were removed. In embodiments, frames where the angle between [Elbow, Shoulder, Spine Shoulder] was above 130° were removed. Other sets of body joints and threshold angles may be defined that achieve the same effect of removing noisy I uninformative frames. Further, the one or more denoising algorithm may comprise applying a low pass filter, optionally after applying filtering to remove selected frames. Any digital low pass filters known in the art may be used, such as e.g. a Savitzky-Golay filter, a low pass Butterworth filter, low pass Chebyshev filter, etc. In embodiment, movement data comprising fewer than a predetermined number of frames after filtering may be excluded from further processing. At step 220, the data may be normalized and or aligned, as will be explained further below. At step 230, the data is used to obtain a pose for the subject. The pose is obtained by summarizing the set of coordinates for each of the plurality of joints over the plurality of time points or a subset of the plurality of time points. For example, a pose may be obtained as the average coordinates for each joint over the plurality of time points. Steps 200-220 are optional because the method may also start from a previously determined pose. At step 240, a reference pose is obtained. The reference pose may have been previously determined using movement data for a plurality of reference subjects. Alternatively, the reference pose may be determined at step 240 using movement data associated with a plurality of reference subjects. This movement data may be pre- processed as explained by reference to step 210, and a pose may be obtained which summarized the movement data for the plurality of reference subjects. This may comprise obtaining respective poses for the respective reference subjects, including normalising and / or aligning these data, as explained by reference to step 220, prior to obtaining a summarized reference pose. At step 220, the movement data may be normalized and / or aligned to a reference. Aligning movement data to a reference comprises modifying the coordinates of the joints while maintaining all relative distances between the joints, such that a predetermined joint is at a predetermined position. The predetermined position may be the origin of the coordinates system used, or any other position provided that the same position is used for as the location of the predetermined joint in the reference pose. The predetermined joint may be a joint that is near the central axis of a subject, such as a spine joint (e.g. spine joint at navel level, or pelvis / hip centre). Normalising movement data is optional as the biomarker metrics described herein are based on angles and are therefore expected to be mostly independent of the subject’s height and width. Normalising movement data can be performed in embodiments to account for differences in distance between the subject and the sensor (e.g. camera). Normalising movement data may comprise adjusting the coordinates of all joints in each frame of the movement data by dividing the limb lengths (distance between pairs of joints in the frame) by the distance between a predetermined pair of joints. The predetermined pair of joints is advantageously chosen such that the distance between the joints in the pair is expected to have the least variation (compared to all possible pairs of joints in the data) between subjects. . The predetermined pair of joints may be the head joint and the shoulder centre joint (i.e. head and base of the neck). This may advantageously decrease the effect of sensor distance differences between subjects. In other embodiments, the movement data may be normalized by scaling each limb length to a reference limb length, such as an average limb length over a reference population. The length of the neck (captured here as the distance between the head joint and the shoulder centre joint, which corresponds to the base of the neck) is believed to have the least variation between people. Therefore, this distance is particularly useful to normalize limb lengths between individuals. At step 250, a metric based on the cosine similarity between the reference pose and the subject pose is obtained. The metric can be the cosine similarity itself, or a distance derived therefrom. The metric is indicative of the presence and / or severity of dystonia in the subject. At optional step 260, the metric obtained at step 250 is used to make a determination, such as a diagnosis, prognosis, treatment recommendation, treatment evaluation, clinical trial participation decision, etc. as will be described further below.

[0053] Monitoring, diagnosis and prognosis of neurological movement disorders

[0054] Referring back to Figure 2, the metrics described herein may be used at step 260 to diagnose a subject as having dystonia, to diagnose a subject as being at high risk of fall, or to diagnose a subject as having a particular subtype of neurological movement disorder, such as a severity group or prognosis group. Further, the methods described herein can be used to monitor a subject with a neurological movement disorder, for example to monitor disease progression, to assess the effect of one or more disease modifying therapeutic, to diagnose a subject has having a neuromuscular movement disorder in a particular category associated with a particular clinical metric of motor function, or to select a subject for participating in a clinical trial. For example, the methods described herein can also be used to provide a prognostic for a subject or classify the subjects between a plurality of disease subtypes, by classifying the subject between at least a first class that has a more severe disease and a second class that has a less severe disease, using the metric obtained at step 250. For example, the metric may be compared to a predetermined threshold, and the subject may be classified in the first class when the dissimilarity (distance derived from the cosine similarity metric) obtained at step 250 is above the predetermined threshold. Respective thresholds may be used for each of a plurality of classes. Further, the metric as described herein may be used in combination with one or more additional clinical scores and / or biomarkers to classify a subject, using a plurality of criteria that apply to each of such additional clinical scores and / or biomarkers (for example in the form of a lookup table), or a machine learning model trained to classify subjects between said plurality of classes using training data comprising values of the metric as described herein and any additional clinical scores and / or biomarkers for a plurality of subjects with known class labels. For example, a first class may comprise manifest HD patients and a second class may comprise healthy patients and / or premanifest HD patients. As another example, a first class may comprise manifest HD patients, a second class may comprise premanifest HD patients, and a third class may comprise healthy patients.

[0055] Similarly, the methods described herein can also be used to determine whether a patient is at high risk of fall, by classifying the subject between at least a first class that has a high risk of fall and a second class that has a lower risk of fall, using the metric obtained at step 250. For example, the metric may be compared to a predetermined threshold, and the subject may be classified in the first class when the dissimilarity (distance derived from the cosine similarity metric) obtained at step 250 is above the predetermined threshold. Respective thresholds may be used for each of a plurality of classes associated with different levels of risk of fall. Further, the metric as described herein may be used in combination with one or more additional clinical scores and / or biomarkers to classify a subject, using a plurality of criteria that apply to each of such additional clinical scores and / or biomarkers (for example in the form of a lookup table), or a machine learning model trained to classify subjects between said plurality of classes using training data comprising values of the metric as described herein and any additional clinical scores and / or biomarkers for a plurality of subjects with known class labels (i.e. subjects assigned to the predetermined classes of risk of falls). Further, the methods described herein can also be used to determine whether a patient has dystonia, by classifying the subject between at least a first class that has dystonia or more severe dystonia and a second class that does not have dystonia or has a less severe form of dystonia, using the metric obtained at step 250. For example, the metric may be compared to a predetermined threshold, and the subject may be classified in the first class when the dissimilarity (distance derived from the cosine similarity metric) obtained at step 250 is above the predetermined threshold. Respective thresholds may be used for each of a plurality of classes associated with different severity of dystonia. Further, the metric as described herein may be used in combination with one or more additional clinical scores and / or biomarkers to classify a subject, using a plurality of criteria that apply to each of such additional clinical scores and / or biomarkers (for example in the form of a lookup table), or a machine learning model trained to classify subjects between said plurality of classes using training data comprising values of the metric as described herein and any additional clinical scores and / or biomarkers for a plurality of subjects with known class labels (i.e. subjects assigned to the predetermined classes of risk of dystonia I severity of dystonia).

[0056] The methods described herein may also be used to monitor a subject, such as e.g. to determine the presence of and / or rate of disease progression in the subject. Such methods may comprise determining a first value of a (dis)similarity metric as described herein at a first time point and at a second value of a (dis)similarity metric as described herein at a second time point, wherein the difference between the first and second values is indicative of the presence and / or rate of disease progression in the subject. For example, the difference being above a predetermined threshold may be indicative of the subject having progressive disease. Conversely, the difference being at or below the predetermined threshold may be indicative of the subject having stable disease. Further, the magnitude of the difference may be indicative of the rate of progression of the subject’s disease.

[0057] The metrics described herein may be predictive of one or more clinical metrics of neurological movement disease severity (also referred to herein as “clinical disease severity metric” or “disease severity metric”). A clinical disease severity metric may be any metric used in clinical assessment of motor symptoms of patients with neurological movement disorders. A clinical disease severity metric may be a dystonia score. A clinical disease severity metric may be a motor score. A clinical disease severity metric may be a metric that combines one or more scores of the UHDRS. For example, a clinical disease severity metric may be selected from: a dystonia score of the UHDRS, the total functional capacity score of the UHDRS, or the total motor score of the UHDRS. A dystonia score of the UHDRS may be the maximal dystonia score of the UHDRS. This is a score between 0 and 4 assessed by a trained practitioner in relation to dystonia of the trunk and extremities, with 0=absent, 1 =slight / intermittent, 3=moderate / common, and 4=marked / prolonged (Unified Huntington’s Disease Rating Scale; Huntington Study Group 1996). A maximal dystonia score of the UHDRS may be quantified with respect to a specific part of the body, such as the trunk, right or left upper extremity, right or left lower extremity. A dystonia score of the UHDRS may be the sum of a plurality of dystonia scores. For example, a dystonia score of the UHDRS may be the sum of the maximal dystonia scores for one or more or all of: trunk, right upper extremity, left upper extremity, right lower extremity, left lower extremity. The total motor score of the UHDRS is the sum of respective scores from 0 to 4 for each of 15 motor function assessments (ocular pursuit, saccade initiation, saccade velocity, dysarthria, tongue protrusion, maximal dystonia, maximal chorea, retropropulsion pull test, finger taps, pronate / supinate-hands, luria, rigidity-arms, bradykinesia-body, gait, tandem walking), with increasing scores indicating increasing severity of the respective symptom (e.g. for dysarthria, dystonia, chorea, rigidity-arms, bradykinesia-body, luria) I decreasing performance of the respective test (for ocular pursuit, saccade initiation, saccade velocity, tongue protrusion, retropulsion pull test, finger taps, pronate / supinate hands, gait and tandem walking). The total functional capacity score of the UHDRS is the sum of scores obtained by answering a predetermined set of yes / no questions (yes=1 point, no=0 point) related to common daily tasks, a score on a scale of 0 to 100 related to independence with guidance by 10 points intervals, and respective scores from 0 to 2 or 3 for predetermined functional capacity items related to daily activity, which higher scores indicating higher independence for the independence scale and higher functional capacity for the functional capacity criteria (see Huntington Study Group 1996, Annex 2). A clinical disease severity metric may be a diagnostic confidence level, such as a HD clinical manifestation diagnostic confidence level that quantifies the severity of motor abnormality associated with HD, as assessed by clinicians. For example, a diagnostic confidence level may be a score selected from: a first level (e.g. 0) for a subject with no motor function abnormalities (normal), a second level (e.g. 1 ) for a subject with non-specific motor abnormalities (such as e.g. presence of motor abnormalities that cannot be confidently assigned to HD, i.e. motor function abnormalities that are not HD-specific), a third level (e.g. 2 or 3) for a subject with motor abnormalities that can be signs of HD (such as e.g. presence of motor abnormalities that can be present in HD but can also be present in other conditions, where in embodiments two different sublevels, e.g. respectively 2 and 3, can be used depending on the level of confidence as to whether the subject motor abnormalities are signs of HD, a lower sublevel being used when the confidence is lower), and a fourth level (e.g. 4) for a subject with motor abnormalities that are unequivocal signs of HD. While not strictly speaking a metric of severity of HD, a diagnostic confidence level correlates with the severity of symptoms at the time of diagnosis (and also correlates very strongly with HD disease stage). Similar scales exist in other diseases and similar scores on such scales can be used. For example, selected scores of the Unified Parkinson’s Disease Rating Scale (UPDRS) or its updated version MDS-UPDRS (Goetz et al. 2008) may be used. In particular, one or more scores of part III or part IV-A of the UPDRS or MDS-UPDRS may be used, such as the rigidity score (item 3.3), gait (item 3.10), freezing of gait (item 3.11 ), postural stability (item 3.12), posture (item 3.13), time spent with dyskinesias (item 4.1 ), or functional impact of dyskinesias (item 4.2), all of which are scores between 0 and 4, 0 being normal and 4 being severe. As another example, selected scores of the Berg balance scale (or the total score of the Berg Balance scale) can be used. In particular, one or more scores associated with standing balance tests of the Berg Balance scale (“standing unsupported”, “standing unsupported with eyes closed”, “standing unsupported with feet together”, “standing unsupported one foot in front”) or any other item or combination of items of the Berg balance scale may be used. All items of the Berg balance scale are scores between 0 and 4, with 4 being normal and 0 being severe (needing assistance to perform the task).

[0058] Thus, also described herein are methods of monitoring a subject, the method comprising determining a similarity or dissimilarity metric as described herein, wherein the metric is indicative of a clinical disease severity metric. The method may further comprise determining the value of said clinical disease severity metric using the (dis)similarity metric. For example, this may comprise determining that the value of said clinical disease severity metric is in a first range when the (dis)similarity metric is above a predetermined threshold, and in a second range when the (dis)similarity metric is at or below the predetermined threshold. Respective thresholds may be used for each of a plurality of ranges of the clinical disease severity metric. Further, the metric as described herein may be used in combination with one or more biomarkers to predict a clinical disease severity metric for a subject, using a plurality of criteria that apply to each of such additional biomarkers (for example in the form of a lookup table), or a machine learning model trained to classify subjects between a plurality of classes associated with different ranges of the clinical disease severity metric (or trained to predict the value of said clinical disease severity metric) using training data comprising values of the metric as described herein and any additional biomarkers for a plurality of subjects with known values for the clinical disease severity metric. As used herein, the term "subject" is a human subject. The subject can be a subject who has or is at risk of developing a neurological movement disorder. A neurological movement disorder (also referred to as neuromuscular disease) may be any disease or disorder that is associated with dystonia and / or posture or balance defects. A neurological movement disorder may be Huntington’s disease, Parkinson’s disease, or a stroke. The methods of the present disclosure are particularly advantageous in the context of these diseases because severity of dystonia for these patients is associated with a higher risk of fall with potentially severe consequences. Thus, the subject may be an elderly subject, such as a subject of at least 60, 65, 70, or 75 years old. The subject may be a subject who has been or is being treated with one or more therapeutics. A therapeutic may be any compound or composition for treating a neuromuscular disease or disorder. In embodiments, the neurological movement disorder is Huntington’s disease.

[0059] For example, the methods described herein may be used in the context of monitoring subjects treated with one or more therapeutics under clinical trial. In such contexts the methods described herein may be used to determine the effect of the one or more therapeutics (e.g. based on the presence of and / or rate of progression of the subject’s disease determined by quantifying the (dis)similarity metric as described herein at a first and second time points, the second time points being a time point after commencing of treatment of the subject with the one or more therapeutics), and / or to select or exclude subjects from participating in a clinical trial (e.g. based on the presence of and / or rate of progression of the subject’s disease determined as described above by quantifying the (dis)similarity metric as described herein at a first and second time points, the second time points being a time point after commencing of treatment of the subject with the one or more therapeutics that are the object of the clinical trial, or based on the classification of the subject in one or more classes as described above where classification between one of such classes is a predetermined inclusion or exclusion criterion of the clinical trial). Similarly, the methods described herein may be used to provide a treatment recommendation for a subject. For example, the methods may comprise classifying the subject between one or more classes as described above and recommending a treatment indicated for the respective class. For example, a subject may be classified in a first class having severe disease, and may be recommended for treatment with a therapeutic indicate for treatment of severe disease. As another example, a subject may be classified in a first class having dystonia or more severe dystonia, and may be recommended for treatment with a therapeutic indicated for treatment of dystonia. As another example, a subject may be classified in a first class having a high risk of fall, and may be recommended for treatment with a therapeutic indicated for treatment of subjects with more severe disease. Any such methods may also comprise selecting the patient for treatment with the recommended therapeutic and / or administering the recommended therapeutic to the patient.

[0060] Related methods of assessing the effect of one or more therapeutics on subjects (e.g. the effect on dystonia) are also described, the methods comprising comparing the values of (dis)similarity metrics as described herein (or values derived therefrom such as differences in these value or rate of change of these values over time) between a first cohort of subjects receiving the therapeutic and a second cohort of control subjects.

[0061] In general, a method described herein may comprise predicting a status of the subject using the value of a (dis)similarity metric as described herein, using one or more predetermined relationships between the value of said metric (and optionally one or more additional biomarkers) and the status of a subject. Said one or more predetermined relationships may be in the form of a lookup table, a predetermined threshold or a trained machine learning model. The predetermined relationships may have been obtained using training data comprising the values of said metric (and optionally one or more additional biomarkers) for a plurality of subjects with known status. The method of any aspect may comprise predicting a status of the subject using the value of said metric (and optionally one or more additional biomarkers) as input to a machine learning model that has been trained to predict said status using training data comprising the values of said metric for a plurality of subjects with known status. The plurality of subjects with known status may comprise at least 10, 20, 30 or 40 subjects. The predicted I known status may be selected from: a binary category, a multiclass category, and the value of a continuous variable. A binary category may be a disease vs healthy status, or a first category corresponding to a first range of values of one or more clinical metrics and a second category corresponding to a second range of values of the one or more clinical metrics. A multiclass category may be one of a plurality of categories each corresponding to a different disease status (e.g. healthy vs one of a plurality of diseases, a plurality of diseases with different severity, etc.), or a plurality of categories each corresponding to a different range of values of one or more clinical metrics. A continuous variable may be a clinical disease severity metric. Thus, predicting a status of the subject may comprise predicting the value of one or more clinical disease severity metrics (as specific values or ranges thereof). EXAMPLES

[0062] The invention is further illustrated by the following examples. It will be appreciated that the examples are for illustrative purposes only and are not intended to limit the invention as described above. Modification of detail may be made without departing from the scope of the invention.

[0063] Introduction

[0064] A balance test, where patients are asked to stand still and keep their balance for a period of time (typically 30 seconds) in a comfortable position, is a commonly used test in clinical settings. The present inventors analysed movement data in patients with Huntington’s disease while performing such tests, and noticed that there was a distinction in this data between healthy patients and HD patients with clinical symptoms. They therefore set out to devise an approach to quantify this difference, with applications in movement disorders including but not limited to HD. The approach compares the pose of subjects to a reference pose, as illustrated on Figure 3. They validated the approach as a diagnostic and prognostic tool by showing that it can be used to differentiate between healthy and diseased patients, as well as predict clinical metrics of disease severity.

[0065] Methods

[0066] Cohort. A cohort of 120 patients was used comprising 40 healthy relatives of HD patients (22 males, 18 females), 40 HD premanifest patients (18 males, 22 females), and 40 manifest patients (21 males, 19 females). The characteristics of the cohort are shown in Table 1 below.

[0067] Table 1. Demographics of patient cohort used.

[0068] Patients were assigned to disease groups by clinicians during recruitment screening, using the definitions of the stages of HD as described in Shoulson, 1981. In particular, patients were classified as preclinical stage (no motor symptoms, also referred to here as “premanifest”), Stage 1 (early stage; 0-8 years from disease onset, motor symptoms are not debilitating and most individuals are fully functional at home and at work, with predisease levels of independence), Stage 2 (early intermediate stage; 3-13 years from disease onset, patients begin to experience impairments in day-to-day living, assistance may be required in some aspects of daily tasks), Stage 3 (late intermediate stage; 5-16 years from disease onset, patients are no longer able to work or manage household responsibilities, cognitive abilities may become impaired requiring substantial help for daily living activities), or Stage 4 (early advanced stage; 9-21 years from disease onset, need for full assistance in daily living, patient aware of daily activities that have to be done but require major assistance to do them). There were no Stage 5 (Advanced stage) patients. The following inclusion criteria were used: all participants are male or female, 18-75 years of age inclusive, are capable of providing informed consent, are capable of complying with study procedures, are participating in the full study and are ambulatory. For the healthy control group, the following inclusion criteria were used: have no known family history of HD or have a known family history of HD but have been tested for the huntingtin gene glutamine codon (CAG) expansion and are not at genetic risk for HD (CAG<36). For the premanifest HD group the following criteria were used: do not have clinical diagnostic motor features of HD, defined as UHDRS Diagnostic Confidence Score <4, and have CAG expansion >40. For the early to moderate HD group, participants were selected with the following criteria: have clinical diagnostic motor features of HD, defined as UHDRS Diagnostic Confidence Score=4, and have CAG expansion >36, and have stage I, II or III HD, defined as UHDRS Total Functional Capacity (TFC) scores between 4 and 13 inclusive. All participants had not used investigational drugs or participated in a clinical drug trial within 30 days prior to sampling visit. The primary objective of the study is tolerability and feasibility of conducting smartphone and smartwatch based remote patient monitoring in HD. This was assessed over a longitudinal study with 3 assessment visits: equipment issue, interim (12 months after equipment issue), and return visits (18 months after equipment issue or earlier on request by participant). At equipment issue visit, the subject is provided with the devices and clinically established scales are assessed including the UH DS, SDMT, Stroop Word reading. The UHDRS (Unified Huntington’s Disease Rating Scale), developed by the Huntington Study Group to provide a uniform assessment of the clinical features and course of HD has undergone reliability and validity testing that support its use in longitudinal studies. The scale assesses 4 domains associated with HD: motor function, cognitive function, behavioural abnormalities and functional capacity. The UHDRS total functional capacity (TFC) represents the investigator’s assessment of the patient’s capacity to perform a wide range of activities of daily living including working, chores, managing finances, eating, dressing and bathing. It is based on a brief interview with the patient and the trial partner. Scores range from 0 to 13, and higher scores represent better functioning. The UHDRS independence scale (IS) is the investigator’s assessment of the patient’s degree of independence. The scale consists of 19 discrete levels ranging from 10 to 100 (by 5) where no special care needed corresponds to a scale of 100 and tube fed and total bed care corresponds to a scale of 10. The UHDRS total motor score (TMS) is the sum of the individual motor ratings obtained during administration of the motor assessment portion of the UHDRS. Scores range from 0 to 124, and higher scores represent more severe impairment. The Symbol Digit Modality Test (SDMT) is used to assess attention, visuoperceptual processing, working memory and psychomotor speed. It has been shown to have strong reliability and validity. The patient must pair abstract symbols with specific numbers according to a translation key. The test measures the number of items correctly paired (maximum of 110) in 90 seconds. The Stroop Word Reading (SWR) test is a measure of processing and psychomotor speed. Patients are presented with a page of colour names printed in black ink and are asked to read aloud as many words as possible within a given amount of time. The number of words read correctly is counted.

[0069] At each onsite visit, participants are asked to conduct “Active Tests” tasks under the supervision of a person trained on the digital biomarker approach. The active tests are a group of preselected and piloted tests adapted from well validated clinical assessments deployed via the use of the smartphone. They include patient reported outcomes (2 to 4 mood / health daily questions depending on the schedule day [How are you feeling physically right now?; Overall, how is your mood right now?; In the past 7 days, how often did your movements (chorea)interfere with your ability to get dressed?; In the last 7 days, I had trouble finishing thing because of my movements (e.g. chorea)?], the EQ-5D-5L and the WHODAS), cognitive tests (the Symbol Digit Modality Test, and the Word Reading Test), upper body motor tests, and stability and gait tests (such as the Balance test, the U-turn test, and the Walk test). A subset of the Active Tests and the onsite-only assessments are recorded using a Microsoft Kinect sensor. Microsoft Kinect is a marker-free motion capture system which can be used for sensitive three-dimensional motion tracking. It detects the locations of 20 body points, and composes a skeletal model of the user at a frequency up to 30 Hz. The sensor data collected from Kinect will be used in this study to accurately analyse the whole body posture, sway, gait and chorea. The EQ-5D-5L Questionnaire is a generic, preference-based health utility measure with questions about mobility, self-care, usual activities, pain / discomfort and anxiety / depression that can be expressed as a composite score of the patient’s health status. The WHO Disability Assessment Schedules 2.0 (WHODAS) is a generic assessment instrument for health and disability that takes about 5 to 20 minutes to be self-completed electronically by the participants. It consists of 12 questions and 3 visual analogue scales. The electronic Balance test assesses the patient’s static balance function by recording physical movements while the patient stands as still as possible while wearing the smartphone and wrist-worn wearable. Sensor-based approaches for measuring static balance have been shown to be sensitive to differences in symptoms in early HD. The test is also part of established scales used in HD, such as the Berg Balance Scale. The electronic U-Turn test is designed to assess gait and lower-body bradykinesia. The patient walks and turns safely at least 5 times between 2 points which are at least 4 steps apart (while wearing the smartphone and wrist-worn wearable). The electronic Walk test captures elements of gait, body bradykinesia and tandem walking abnormalities. The patient walks as fast as is safely possible for 200 meters or 2 minutes. Ideally, the test is performed in a straight path with no obstacles (e.g., in a park). Sensorbased approaches for measuring gait have been shown to be sensitive to differences in symptoms in early HD. The Berg Balance Scale is a test used to objectively determine the ability to safely balance during a series of predetermined tasks. We will use the following tasks: standing unsupported with eyes closed; standing unsupported with feet together; standing unsupported with one foot in front; standing on one leg. Each is scored from 0 to 4, “0” indicating the lowest level of function and “4” the highest level of function. The Timed Up and Go Test is a test used to determine fall risk and measure the progress of balance, sit to stand, and walking. The participant begins sat back in a standard arm chair, stands, walks up to a line on the floor 3 meters away from the chair, turns, and walks back to the chair and sits down. The time in taken to perform this task is recorded. Data acquisition. All participants were asked to stand still with their hands by their side and keep their balance for 30 seconds in a comfortable position in front of a Kinect v2 camera (part of the Berg Balance Scale assessment performed at onsite visits as explained above). A single test was done per patient per session. Repeated tests are done over multiple sessions to demonstrate progression or therapeutic effect. Data comprising joint coordinates for all 27 joints tracked by the Kinect software as respective time courses were obtained. The data was denoised using 3 separate filters: (1 ) frames where the angle between [Right Shoulder, Spine Shoulder, Left Shoulder] was below 145° were removed (such angles indicate that one of these data points are mislocated and noisy); (2) frames where the angle between [Elbow, Shoulder, Spine Shoulder] was above 130° were removed (this angle would be expected to be roughly between 95° and 130°, when standing with hands relaxed beside the body); (3) the remaining data was smoothed using Savitzky- Golay filtering (Savitzky & Golay, 1964; Savgol implementation in scipy, see docs.scipy.org / doc / scipy / reference / generated / scipy.signal.savgol_filter.html). Note that other criteria that can filter abnormal I unexpected frames could be used, for example other configurations of body joints which are relevant to the test and are expected be maintained during the test, could be used to remove noisy data. Further, any low pass filter could be used to smooth the data, such as e.g. a low pass Butterworth filter, low pass Chebyshev filter, etc. Sequences with fewer than a predetermined number of frames after filtering were excluded. In the data analysed in these examples, removing the faulty frames in some cases makes the sequence very short (about 20-30 frames left), and those very short videos were excluded from further analysis. Additionally, for all statistical analyses, because the group sizes differ, 18 videos were randomly sampled from each group. This process was repeated 10 times to ensure that each patient is represented at least once. 18 was the number of videos post filtering in the smallest group (group 3). Other numbers <18 could have been used instead.

[0070] An average joint position across the 30 seconds time course was obtained for each joint for each patient, leading to a single averaged coordinate value per joint (illustrated on Figure 3, in 2D). All data was aligned by aligning the coordinates for all patients at the spine joint (also referred to as “spine mid”) joint (see Figure 3 which shows aligned average poses). In other words, time courses of coordinates for the joints in each individual recording are modified such that the location of the spine joint in the chosen reference coordinate system is the same (see Figure 3 which shows aligned average poses). The data was normalised prior to averaging to account for differences in body height as measured by the sensor (which is influenced by the distance between the subject and the sensor) by normalising all coordinates at each time point (i.e. for each frame) based on the distance between the head joint and the shoulder centre joint. In other words, for each frame, all coordinates (in a system aligned to the spine joint) were multiplied by the distance between the head joint and the shoulder centre joint in the same frame. The resulting normalized coordinates were then averaged to obtain a pose. Analysis was performed using 16 joints (all Kinect joints excluding hand tips, thumbs, neck, hand and feet, as illustrated on Figure 7C, with the complete set of Kinect joints available in this study - 20 joints- illustrated on Figure 7D), or subsets of 12 joints (as per 16 joints but additionally excluding ankles and wrists, as illustrated on Figure 7B - referred to as the “knee-up” set of joints) or 8 joints (as per the knee-up set but additionally excluding elbows and knees, as illustrated on Figure 7A - referred to as the “torso” set of joints). All data was acquired in the same clinical context using the same setting.

[0071] Reference pose. A reference pose was obtained as the average across averaged coordinates (as above) for a plurality of healthy subjects (10 subjects in the present examples). This is illustrated on Figure 3 as the red pose, with poses of subjects to be compared to this pose illustrated in blue. In the present examples the 10 subjects used were manually selected as subjects that maintained their balance throughout the test. However, the inventors verified that the same approach could be applied with any number of reference subjects (even a single reference subject). The use of a plurality of subjects (here 10 subjects) helps to reduce noise in the data.

[0072] Dissimilarity metric. The calculation of a dissimilarity metric is illustrated on Figure 6. The body joint coordinates (after denoising, alignment, normalisation and averaging - although note that normalization is optional since the metric is based on angles) are then used to computationally compare a patient with a reference pose, to find the differences between two poses. This comparison is calculated by computing the cosine similarity distance between each body joint of the patients and the reference. In another words, given list of joints of patient p1= {ji, j2 jn} where n can be 20 (all joints), 16, 12 or 8 in the present examples, and ji = {Xi, y,, z} are the Cartesian coordinates of the joint i for patient pi, and the reference pR= { ki, k2 kn} where ki = {ki, ki, k} are the Cartesian coordinates of the joint i for the reference pose PR, the cosine similarity distance is calculated as: Cos(p1, pR) = 1- Cos(p1, pR). Note that either a similarity metric or a dissimilarity metric (i.e. cos or 1- cos) can be used. The similarity output is a number between 0 and 1 , where the closer to 0 indicates the smaller distance, thus higher similarity, and the closer to 1 , is an indicator of larger difference between two poses (or vice versa if a similarity is used instead of a dissimilarity).

[0073] Clinical scores. A variety of clinical scores were quantified for each patient by qualified clinicians. These included: dystonia score, Total Functional Capacity, Total Motor score, Diagnostic Confidence Level (DCL). The total motor score was determined according to the UHDRS (Unified Huntington’s Disease Rating Scale; Huntington Study Group 1996). The UHDRS is a clinical rating scales to assess four domains of clinical performance and capacity in HD: motor function, cognitive function, behavioral abnormalities and functional capacity. The total motor score is the sum of respective scores from 0 to 4 for each of 15 motor function tests (ocular pursuit, saccade initiation, saccade velocity, dysarthria, tongue protrusion, maximal dystonia, maximal chorea, retropropulsion pull test, finger taps, pronate / supinate-hands, luria, rigidity-arms, bradykinesia-body, gait, tandem walking), with increasing scores indicating increasing disease severity. The dystonia score used was the sum of the maximal dystonia score from the total motor score, for the trunk and each extremity (i.e. upper left, upper right, lower left, lower right). The total functional capacity score is the sum of scores obtained by answering a predetermined set of yes / no questions (yes=1 point, no=0 point) related to common daily tasks, a score on a scale of 0 to 100 related to independence with guidance by 10 points intervals, and respective scores from 0 to 2 or 3 for predetermined functional capacity items related to daily activity, which higher scores indicating higher independence for the independence scale and higher functional capacity for the functional capacity criteria (see Huntington Study Group 1996, Annex 2). The diagnostic confidence level (DCL) is a standard measure used for clinical diagnosis in at-risk individuals and is based solely on the motor evaluation. It is assessed by a clinician based on the following scale: 0 = normal (no abnormalities), 1 = non-specific motor abnormalities (less than 50% confidence), 2 = motor abnormalities that may be signs of HD (50-89% confidence), 3 = motor abnormalities that are likely signs of HD (90-98% confidence), 4 = motor abnormalities that are unequivocal signs of HD (>99% confidence).

[0074] Results

[0075] Figure 3 shows aligned joint location data for subjects with HD (blue) compared to a reference (red), in a balance test. From left to right the data shows: data for a subject with the largest dissimilarity to control amongst the cohort of HD patients tested (manifest group), data for a subject that has a median dissimilarity to control in the cohort (premanifest group), and data for a subject with the lowest dissimilarity to control in the cohort (premanifest group). This data shows that there is likely to be some clinically relevant information in pose data from a balance test, provided that a biomarker metric could be developed to capture this. This is what the present inventors set out to do. Based on the data on Figure 3, the inventors postulated that an angle-based metric may be informative. They therefore calculated the cosine similarity between aligned and normalized poses for each pair of subjects in the data, thereby obtaining for each subject a vector of cosine similarities between the subject’s pose and that of every other subject in the set. They found that some subjects are dissimilar to every other subject in the data. In order to better visualize this data, they performed a principal component analysis (PCA), the results of which are shown on Figure 4. Figure 4 shows values along the first two principal components of PCA on vectors of pairwise cosine dissimilarity between poses in a balance test (where cosine similarity is computed as described below). The plot shows data, for individual subjects (points) and curves fitted to groups of subjects (lines and shaded areas). The points and curves are coloured according to the subject group: group 1 (blue)=healthy, group 2 (orange)=premanifest, group 3 (green)=manifest severity 3 (corresponding to UHDRS severity I, II), group 4 (red)=manifest severity 4 (corresponding to UHDRS severity III). The data shows that there is a clear distinction between the healthy and manifest group (compare the blue and red+green data series). However, it is more difficult from this data to distinguish between healthy and premanifest groups (blue and orange data series), and between the different manifest groups (severity groups 3 and 4, green and red data series).

[0076] Based on this data, the inventors reasoned that it would be possible to design a metric that quantifies the difference between healthy subjects and subjects with dystonia, from the movement data collected during a balance test. They designed a method to do this based on cosine similarity with a reference pose. Figure 5 illustrates schematically the concept of cosine similarity. Cosine similarity is the cosine of the angle between two vectors and indicates whether two vectors are pointing in roughly the same direction. The cosine similarity metric finds the cosine of angles between two poses. In the context of diagnosing dystonia, this was found to be more informative than the Euclidean distance, which is based on the magnitude of the joint locations. Figure 6 illustrates the process to obtain the similarity score used. The metric was quantified using 3 different subsets of the joints for which data was available, which are illustrated on Figure 7A-C (with the complete set of joints illustrated on Figure 7D). This indicated that the 12 joints version (knee-up joints, i.e. all joints from the knee upwards, excluding hands, neck and wrists) struck the best balance between simplicity and accuracy, as additional joints on the hands, wrists, ankles and feet did not contribute much information to the detection and were not located as reliably (see e.g. Figure 8 which shows that correlations between the similarity metric and other disease severity metrics were significant when using 8 or 12 joints, whereas the correlations were lower when using 16 joints (data not shown), and Figure 9 which shows that differences in the similarity metric between patient groups were more pronounced with 12 joints than with 8 joints).

[0077] The similarity scores calculated as described on Figure 6 were then quantified for all patients using the 12 joints illustrated on Figure 7B. These were used to verify the hypothesis that given a reference pose PR = {ki,k2,k3 kn} the distance between the pose of dystonic patient PD = {ji 2J3 jn} and the reference, is larger than the pose of a healthy patient PH = {ji J2J3 jn} and the reference, i.e. D(PR , PD) > D(PR , PH). Figure 9 shows the distributions of similarity scores for a cohort of HD patients and healthy controls, separated by patient group (manifest and premanifest patients, where the premanifest patients are not expected to have dystonia whereas at least some of the manifest patients are expected to have dystonia), when calculated using 8 joints (Figure 9A, B) or 12 joints (FigureC, D). The data on Figure 9C shows that the distribution of similarity scores for manifest patients is significantly different from the distribution of similarity scores for healthy patients, with the expected D(PR ,PD) being significantly higher than the expected D(PR ,PH). In other words, this data verifies the hypothesis that D(PR ,PD) > D(PR ,PH). This difference did not reach significance when using 8 joints (Figure 9A, B), although a difference is still clearly visible and reached significance between the premanifest and manifest HD groups. Clinically speaking, the premanifest group usually have little or no sign of motor function impairment. Thus, the data show that there is a clear difference between healthy I premanifest and manifest groups, although there is some population variability (which would be expected in any population but especially in this case as not all patients will experience dystonia and not all patients that experience dystonia will experience dystonia at the particular time that the balance test was performed). Given this inherent population variability, the marked difference that the new metric was able to capture indicates that the metric is very sensitive to detect dystonia and other pathologies affecting posture in movement disorders. This indicates that the metric would be even more sensitive if calculated using movement data acquired over a plurality of balance tests performed at different times, as this increases the chances of at least one of these tests showing dystonia symptoms, which the metric would capture with high sensitivity even amongst other tests where the symptoms are not present. Further, the inventors also showed that the averaging over time for a balance test pronounces the differences between patients with and without dystonia, and enables to distinguish between chorea and dystonia. Finally, the use of 12 joints was also shown to be more informative than 8, although 8 is sufficient to gather some useful information.

[0078] Next, the inventors sought to determine whether the new metric could be predictive of existing clinical metrics. They therefore calculated the Pearson correlation coefficient between the similarity score calculated for each patient and a plurality of clinical variables associated with these patients. Figure 8 shows correlation matrices showing the correlation between the similarity metric calculated according to Figure 6 (using 8 or 12 joints, respectively in Figure8A and Figure 8B) and clinical metrics, for the cohort of subjects studied. The data on Figure 8B show that, when using 12 joints (which best separates the subject groups between healthy, premanifest and manifest), the score calculated as explained on Figure 6 (sim_score) is significantly correlated to the disease group (healthy=0, premanifest=1 , manifest=2), the number of CAG repeats (which is known to correlate with disease severity), the sum of dystonia scores (dystonia), the total motor score (tmstot), the age of the patients (which also correlates with disease severity in the diseased group as the disease has a typical range of age of onset and is degenerative), and the diagnostic confidence level (dclscore - which is indicative of motor symptoms). The DCL correlates with clinical symptom severity in practice (since it captures motor symptoms as well as confidence of assignment of these symptoms to HD specifically). Table 2 below shows the correlation and correlation p-values of these metrics on Figure 8B. No correlation was found with the total functional capacity (tfctot) when using 12 joints, which is not surprising as this does not directly relate to dystonia and instead captures severity of disease associated with many different symptoms that may not include dystonia. The similarity metric correlated with the total motor function score (which includes the dystonia score as well as scores that quantify many other motor symptoms) and even more strongly with the dystonia score. The similarity metric also correlated with age of the subjects, which is not unexpected as HD disease severity increases with ages (as a result of disease progression). This is also illustrated on Figure 10, which shows a trend towards higher dissimilarity scores with increasing age in the HD patient groups but not in the control group. When using 8 joints, the metric still correlated with CAG, disease group and age, indicating that the approach is relatively robust to the use of smaller set of joints. When using 16 joints the correlations were lowered, indicating that in this particular dataset some of these additional joints introduced noise. Whether this is the case or not may depend on the choice of joint, filtering approach and size of the dataset.

[0079] Figure 8B).

[0080] As explained above, for all statistical analyses (all data shown on Figures 8 and 9 and Table 2), because the group sizes differ, 18 videos were randomly sampled from each group. The inventors also tested whether the metric could differentiate between groups using larger numbers of more heterogenous subjects. They therefore combined group 3 and 4 (all manifest patients) and healthy and pre-manifest (control) and selected 30 samples instead of 18, from each of these two groups. The dissimilarity metric was also significantly different between these two groups, i.e. t-test independent samples t-statistic=2.8237, p- value=0.0065 (2-sided) and 0.0032 (1 -sided (manifest-healthy)>0). The manifest group had an average dissimilarity score of 0.164 (standard deviation 0.0234) while the healthy / premanifest group had an average of 0.150 (standard deviation 0.0148). This data confirms that the metric described herein is a robust indicator of the presence of dystonia symptoms and HD status.

[0081] Conclusions

[0082] The data in this example shows that a metric derived from balance test motion data as described herein can differentiate between healthy and HD patients with dystonia symptoms (and can also distinguish between patients with dystonia and patients with chorea symptoms), and is predictive at least to some extent of clinical motor function metrics (without requiring the time and qualified labour intensive assessment of such metrics, and without the inherent subjectivity associated with such metrics). To the best of the inventors knowledge, this is the first time that a posture similarity metric such as cosine similarity was used to monitor patients with movement disorders, and the first time that trunk / whole body dystonia was studied using an automated method in patients with movement disorders.

[0083] References

[0084] Huntington Study Group. Unified Huntington’s Disease Rating Scale: Reliability and Consistency. Movement Disorders. Vol. 11 , No. 2, 1996, pp. 136-142.

[0085] Louis ED, Lee P, Quinn L, Marder K. Dystonia in Huntington's disease: prevalence and clinical characteristics. Mov Disord. 1999 Jan;14(1 ):95-101.

[0086] Zhang Z, Cisneros E, Lee HY, Vu JP, Chen Q, Benadof CN, Whitehill J, Rouzbehani R, Sy DT, Huang JS, Sejnowski TJ, Jankovic J, Factor s, Goetz CG, Barbano RL, Perlmutter JS, Jinnah HA, Berman BD, Richardson SP, Stebbins GT, Cornelia CL, Peterson DA. Hold that pose: capturing cervical dystonia's head deviation severity from video. Ann Clin Transl Neurol. 2022 May;9(5):684-694.

[0087] Nakamura T, Sekimoto S, Oyama G, Shimo Y, Hattori N, Kajimoto H (2019) Pilot feasibility study of a semi-automated three-dimensional scoring system for cervical dystonia. PLoS ONE 14(8): e0219758.

[0088] Zhang, Z., Hong, R., Lin, A. et al. Automated and accurate assessment for postural abnormalities in patients with Parkinson’s disease based on Kinect and machine learning. J NeuroEngineering Rehabil 18, 169 (2021 ).

[0089] Ye C, Xiao Y, Li R, Gu H, Wang X, Lu T, Jin L. Pilot Feasibility Study of a Multi-View Vision Based Scoring Method for Cervical Dystonia. Sensors (Basel). 2022 Jun 20;22(12):4642. den Hartog D, van der Krogt MM, van der Burg S, Aleo I, Gijsbers J, Bonouvrie LA, Harlaar J, Buizer Al, Haberfehlner H. Home-Based Measurements of Dystonia in Cerebral Palsy Using Smartphone-Coupled Inertial Sensor Technology and Machine Learning: A Proof-of- Concept Study. Sensors. 2022; 22(12):4386.

[0090] Park, J., Yang, K.Y., Lee, J. et al. Objective Evaluation of Cervical Dystonia Using an Inertial Sensor-Based System. J. Med. Biol. Eng. 39, 305-314 (2019).

[0091] Goetz, C.G. et al. Movement Disorder Society-Sponsored Revision of the Unified Parkinson’s Disease Rating Scale (MDS-UPDRS): Scale Presentation and Clinimetric Testing Results. Movement Disorders. Vol. 23, No. 15, 2008, pp. 2129-2170.

[0092] Shoulson I. Huntington disease: functional capacities in patients treated with neuroleptic and antidepressant drugs. Neurology. 1981 Oct;31 (10):1333-5.

[0093] Savitzky, A. & Golay, M. J. E. Smoothing and Differentiation of Data by Simplified Least Squares Procedures. Anal Chem, 36(8): 1627-1639, 1964. Berg KO, Wood-Dauphinee SL, Williams JI, Maki B. Measuring balance in the elderly: validation of an instrument. Can J Public Health. 1992 Jul-Aug;83 Suppl 2:S7-11 .

[0094] All publications mentioned herein are incorporated by reference in their entirety.

Claims

CLAIMS:

1. A computer-implemented method of monitoring a subject who has been diagnosed as having or likely to have neurological motor disease or disorder, the method comprising: obtaining a pose derived from movement data comprising joint location data for a plurality of joints of the subject while performing a balance test; and determining the cosine similarity between the pose derived from the movement data and a reference pose derived from reference movement data, wherein said cosine similarity is indicative of the presence and / or severity of a neurological motor disease or disorder.

2. The method of claim 1 , wherein the a reference pose is a pose derived from movement data for a plurality of reference subjects comprising joint location data for a plurality of joints of the respective subjects while performing a balance test, optionally wherein the method further comprises obtaining the reference pose from said reference movement data, wherein the reference pose comprises summarized joint location data over a plurality of time points for the plurality of reference subjects.

3. The method of claim 2, wherein the reference subjects are healthy subjects, and / or wherein the plurality of reference subjects comprises at least 10, 20 or 30 reference subjects.

4. The method of any preceding claim, wherein the method comprises obtaining movement data comprising joint location data for a plurality of joints of the subject while performing a balance test, and obtaining a pose derived from said movement data, wherein a pose comprises summarized joint location data over a plurality of time points.

5. The method of any preceding claim, wherein the method further comprises applying one or more denoising algorithms to the movement data.

6. The method of claim 5, wherein the one or more denoising algorithms are applied prior to obtaining a pose, and / or wherein the one or more denoising algorithms are selected from:(i) a smoothing algorithm, optionally a low pass filter, optionally a Savitsky-Golay filter,(ii) a filter that removes any frame of the movement data where the coordinates of a subset of the plurality of joints satisfy one or more predetermined criteria that apply to angles between said subset of joints, optionally wherein frames where the angle between [Right Shoulder, Spine Shoulder, Left Shoulder] is below a threshold selected between 135° and 155°, optionally 145°, are removed, and / or wherein frames where the angle between [Elbow, Shoulder, Spine Shoulder] is above a threshold selected between 120° and 140°, optionally 130° are removed, and(iii) a filter that removes movement data corresponding to a single balance test, where the movement data comprises fewer than a predetermined number of frames, optionally fewer than 40 frames.

7. The method of any preceding claim, wherein the movement data comprises joint location data for a plurality of joints of the subject for a plurality of time points, and / or wherein a balance test comprises the subject standing at ease for a predetermined period of time, optionally wherein the predetermined period of time is at least 10 seconds, at least 20 seconds or about 30 seconds, and / or wherein the movement data comprises joint location data acquired at a sampling rate of at least 10 Hz, and / or wherein the movement data comprises joint location data for a plurality of joints of the subject while performing a plurality of separate balance tests.

8. The method of any preceding claim, wherein a pose comprises average joint locations over a plurality of time points, and / or wherein joint location data comprises 2D or 3D coordinates of the plurality of joints.

9. The method of any preceding claim, wherein the movement data has been acquired using a movement sensor selected from a 2D camera and a 3D camera, optionally a stereocamera.

10. The method of any preceding claim, wherein the neurological disease or disorder is dystonia or a neurological disease associated with dystonia, posture and / or balance impairment, and / or wherein the monitoring comprises detecting orassessing dystonia, posture and / or balance impairment and / or wherein the subject is a subject who has been diagnosed as having or being likely to have Huntington’s disease, Parkinson’s disease, or having suffered a stroke.

11. The method of claim 10, wherein the dystonia is generalized dystonia, multifocal dystonia, segmental dystonia or hemidystonia.

12. The method of any preceding claim, wherein the plurality of joints include a plurality of knee-up locations, wherein knee-up locations are joints located from the knee of the subject to the head of the subject, optionally wherein the plurality of joints do not include joints on the hands and wrists, and / or wherein the knee-up locations comprise or consist of: head, base of the neck (shoulder centre), left shoulder, right shoulder, spine at navel level (spine), pelvis (hip centre), left hip, right hip, left knee, right knee, and optionally left elbow and right elbow.

13. The method of any preceding claim, wherein the method further comprises: (i) aligning the subject’s pose to a reference coordinate, wherein aligning the pose to a reference comprises modifying the coordinates of the plurality of joints while maintaining all relative distances between the joints, such that a predetermined joint is at a predetermined position; and / or (ii) normalising the movement data or pose, wherein normalising movement data or a pose derived therefrom comprises adjusting the coordinates of the plurality of joints in each frame of the movement data or in the pose based on the distance between a predetermined pair of joints in the frame.

14. A method of selecting a subject for participating in a clinical trial, determining the effect of a therapeutic compound or composition for treating a neurological movement disease or disorder, or recommending a therapeutic compound or composition for treating a neurological movement disease or disorder in the subject, the method comprising monitoring the subject using the method of any preceding claim and selecting the subject for participating in a clinical trial, determining that the therapeutic compound or composition is effective in treating the neurological movement disease or disorder, or recommending the subject for treatment with the therapeutic compound or composition when the cosine similarity or a value derived therefrom satisfies a predetermined criterion or thesubject is classified in one of a plurality of classes using a machine learning model trained to classify subjects using values of a plurality of biomarkers comprising said cosine similarity or a value derived therefrom.

15. A system comprising: a processor and a memory storing instructions that, when executed by the processor, cause the processor to implement the method of any of claims 1 to 14.

Images