WO2007029012A1

WO2007029012A1 - Categorising movement data

Info

Publication number: WO2007029012A1
Application number: PCT/GB2006/003336
Authority: WO
Inventors: Lars Adde; Øyvind STAVDAHL; Gabriela Espinosa Porragas; Pál Renton BERGE
Original assignee: Ntnu Technology Transfer As; Jackson, Robert
Priority date: 2005-09-09
Filing date: 2006-09-11
Publication date: 2007-03-15
Also published as: GB0518480D0

Abstract

A method and apparatus for categorising data derived from the movements of a living subject in which the movement data is processed to extract information and the extracted information is then classified into one of a plurality of categories using a classification model that has been trained using training data derived from the movements of other subjects whose category is known. The extracted information relates to patterns of movement which may, or may not, be readily recognisable to a human observer. In this way, the present invention can take into account movement phenomena that are otherwise incomprehensible to humans (or at least not readily recognisable or describable), for example because they involve complex inter-relationships of the movements of a plurality of limbs. The method and apparatus can be used to categorise movement data of infants according to whether or not they have, or are susceptible to, cerebral palsy.

Description

Categorising Movement Data

The present invention relates to categorising data relating to the movement of a living subject. More specifically, the invention may be applied to categorising movement data indicative of medical conditions, such as Cerebral Palsy (CP).

CP is a non-progressive motor impairment syndrome secondary to lesions or anomalies of the brain arising in the early ages of development. The type of motor impairment is divided into different categories according to which functions or body parts that are affected. The seriousness of CP differs from almost invisible disability to a serious handicap. Cerebral palsy affects approximately 1 in 500 infants. The risk of CP is highest in extremely premature infants (birth weight less than 1 kg and/or gestational age less than 28 weeks).

For the last two decades, technical advances and improvements in obstetric and neonatal care have led to a decrease in perinatal mortality. Especially among the extremely premature infants, the chance of survival has greatly increased. Despite the continued decrease in mortality rates, the incidence of neurosensory and developmental handicaps has remained constant. This means that a larger number of infants survive without major sequelae, but there are also a larger number of survivors born with a high risk of major handicaps such as cerebral palsy (CP), mental retardation, deafness and blindness. There is also a risk of minor developmental problems such as clumsiness, attention deficit hyperactivity disorder (ADHD) and/or learning problems. Infants with birth asphyxia or other perinatal brain injuries (e.g. cerebral infarction, Central Nervous System infection or neonatal sepsis) also constitute a population at high risk of neurological developmental disorders of which cerebral palsy is a major one.

As a result of the complexity of neonatal brain development and the different risk factors in different developmental stages, it is difficult to predict the neurological outcome in young infants. A diagnosis of CP is often not established until the age of 12-18 months and some of the mildest forms may still not be diagnosed before the age of four.

Early diagnosis of neurological development disorders is important because children with pathological development may benefit from early intervention training programmes, and healthy children can be dismissed from extensive assessment programmes at an early stage.

A number of techniques have been used to assess the brain at an early age. The techniques vary from clinically based methods requiring no equipment, such as various forms of neurological assessments tests, to sophisticated technical assessments, such as brain imaging (ultrasound, computer tomography and magnetic resonance imaging) and neurophysiologic tests, including electroencephalograms (EEG) and visual or sensory evoked potentials (VEP and SEP).

The introduction of ultrasound techniques (US) and magnetic resonance imaging (MRI), have contributed to a better and earlier diagnosis of brain defects in neonates. However, normal cerebral MRI and US can be found in infants who later develop abnormally and vice versa. The accuracy of the different assessment techniques to predict the neurological outcome of newborn babies at risk shows a large variation. In addition to a lack of accuracy in predicting the outcome, some of the methods need advanced technological equipment.

It has been recognised that the 'spontaneous movement' of an infant is dependent on its motor coordination and can be related to whether or not the infant has CP. A special type of spontaneous movement has been studied that has been termed 'general movements' (GM). Observation of an infant's GM by a clinician has been used to estimate whether or not the infant has CP. This is termed the 'general movement assessment technique' (GMA) and has shown promising scientific results. Typically, the movement of an infant is video-recorded and then the movement patterns are observed by a doctor, physiotherapist, etc. Observation and classification of such movement patterns may predict later neurological outcomes such as CP as early as 3-5 months post term. In research settings, this method has resulted in a diagnosis of CP with a sensitivity and specificity of around 93%.

However, the qualitative nature of the GMA technique is a problem, particularly where the clinician is working alone, and the accuracy is very much dependent on the clinician's ability and experience. Furthermore it is a time-consuming process, requiring the clinician to spend a lot of time studying the movement of the infant.

What is needed is a simple-to-use and economical way of estimating whether an infant has CP, which is accurate even at an early age. An objective method is particularly needed, which is not too dependent on a clinician's ability, and not prohibitively time consuming.

Some work has been conducted in this area by Meinecke et al, as disclosed in "Movement analysis in early diagnosis of a developing spasticity in newborns with infantile cerebral palsy", Gait & Posture Vo. 18, Suppl. 2, 96-98, 2003; and

"Procedures for the classification of movement analysis data in early diagnosis of a developing spasticity in newborns with ICP", Gait & Posture Vo. 20, Suppl. 1, 61- 112, 2004. As far as can be understood from these documents, a method for estimating whether a baby is likely to have CP or not has been developed based on 'real-world' indications from 3D movement data of the baby. Experienced physicians identified movement features that they look for when visually assessing a baby, and from this, 125 parameters based on movements or combinations thereof were determined which are relevant to CP, such as movement speed, trajectory smoothness, periodicity, range of motion and acceleration. In this approach, in order to estimate whether or not the baby is healthy or at risk, parameter data for babies which have already been classified by a physician are used to select optimum parameter combinations. For example, five optimum parameters are selected using cluster analysis based on Euclidian distances. In order to estimate whether a baby is likely to have CP, these five parameters are measured for the baby, and it is then determined whether they are within the range of the standard deviation for the norm collective in respect of each parameter. Depending on the number of parameters within or outside the standard deviation, classification is effected.

A second method using quadratic discriminant analysis with eight parameters and a cost function is also disclosed. Here, the parameters are measured for a baby, and the values are compared with parameter set values for babies that have already been classified. The baby being tested is classified according to the classification of the 'known' parameter set showing the greatest similarity with the parameter data of the test baby. In other words, this method looks for the 'nearest neighbour' parameter set.

However, there are a number of drawbacks to these methods. Firstly, they rely on 'parameters' being pre-determined using the expertise of many physicians, which may be suggestive of CP. Thus only 'real-world' parameters which can actually be recognized by a human as clinically meaningful are considered, and any other movements are ignored. This is a serious disadvantage because the differences in movement between babies with and without CP can be extremely subtle and complex and not easily quantified by the human eye. As such, the inventors have recognised that much important data is not used in these prior art methods and this can affect accuracy. Further, particular movement parameters may only be valid for certain ages of a child, and thus either investigations must be limited to certain ages of children, or the age dependency of the parameters has to be determined and only those suitable for the test child used. This adds further complexity and room for error in the methods.

As such, there still remains a need for an improved method of objectively, simply and accurately estimating whether an infant has (or is likely to develop) CP. In fact, there is a need for a simple, objective and effective technique for analyzing movement in order to evaluate motor-related neurological conditions in general.

Indeed, such a need also extends to the general analysis of movement data for other medical and non-medical purposes.

According to a first aspect, the present invention provides a method of categorising data derived from the movements of a living subject, comprising: processing the data to extract information; and classifying the extracted information into one of a plurality of categories using a classification model; wherein the classification model is trained using data derived from the movements of other subjects whose category is known.

The extracted information relates (albeit indirectly) to patterns of movement which may, or may not, be readily recognisable to a human observer. However, the invention does not involve or require these patterns to be defined or recognised as such. There is no definition or determination of the "real world" observable parameters used in the prior method discussed above. Thus, in contrast to the prior art, the method is not dependent on particular human-defined parameters. Instead, information is simply extracted from the movement data in order to categorise data. In this way, the present invention can take into account movement phenomena that are otherwise incomprehensible to humans (or at least not readily recognisable or describable), for example because they involve complex inter-relationships of the movements of a plurality of limbs.

Such a method can be used to categorise movement data of infants according to whether or not they have, or are susceptible to, CP. For example, movement data may be categorised as 'normal', 'CP' or 'at risk¹. It is a particularly robust technique for estimating whether an infant has CP, as noted above. However, the invention is not restricted to categorising movement data to determine whether an infant has CP, and is in fact suitable for categorising any sort of movement data. For example, the method could be used to examine whiplash injuries, Parkinson's disease and ADHD. In the case of detecting CP in an infant, it is data relating to the spontaneous movements that should be used in the categorisation process. Preferably, the categorisation process takes place during 6-20 weeks post-term (46-60 weeks postmenstrual age), or more preferably 5-10 weeks post-term. This is the time during which spontaneous movement can best be used to detect CP. The information extracted from such data and used in the categorisation relates to patterns in the signals from the spontaneous movements.

According to the invention, movement data is used both in the categorisation process and for training the model. Movement data for use in both these stages can be collected and processed in a similar way. Normally, the training data will be processed in the same manner as the movement data that is to be classified.

Preferably, information is extracted from the 'training data' in order to train the model. Thus, the methods described below for collecting, pre-processing and extracting information from movement data are applicable to both the categorisation and training. Preferably, movement data includes data of how different parts of the subject's body are moving over time. Movement data can be collected in any known way. For example, electromagnetic sensors can be connected to different parts of the body, and the data fed to a tracking system. Alternatively, the subject may be videoed with a video camera, and the data analysed using image-processing software to determine the movement of different parts of the body. Two or three-dimensional movement data may be collected. The data may correspond to more dimensions, for example three linear dimensions (x, y, z) and, say, three axes of rotation. If a simple system is desired, two-dimensional data can be collected using only one camera. Alternatively, a more complex system using more than one video camera may be used to provide movement data in three dimensions. Data reduction to two dimensions may form part of the data pre-processing. In one embodiment, reflective elements are attached to different parts of the body, to assist in tracking movement. Preferably, when training a model/categorising data to determine CP status, the movements of the ends of the limbs are monitored in particular, since this gives a good representation of the spontaneous movement. Movement data may be used in real-time to perform categorisation or to create the model, or alternatively it may be stored in a database and used at a later date.

Generally, the movement signal will be sampled at regular time intervals and it is this sampled signal that forms raw movement data to be processed. For example, if the signal is sampled at 25Hz, every 40ms a sample will be recorded.

Raw movement data may be pre-processed before the information is extracted. This can comprise a number of stages. Firstly, a 'region of interest' in the data may be extracted. For example, in the case of assessing CP, data collected when the infant is crying or playing is not useful, and consequently should be excluded. Regions of interest may be selected automatically, for example by a computer, or manually by a user.

The movement data may be subject to principle component analysis (PCA), in order to reduce the data set to its most important components, thereby selectively retaining the most important data. In one preferred embodiment, two principal components are chosen and calculated over time, although a greater number, such as three or four or more can be used. It will be appreciated that using a greater number of components increases complexity and the amount of processing power used.

Time frequency decomposition may be carried out on the movement data, whether PCA has been performed or not, although it is clearly preferable for it to be applied to the output from the PCA. Thus, preferably, time frequency representation is performed for each principal component, in order to determine the amount of each frequency present in the signal over time. Any or all of the above pre-processing stages are preferably carried out before the data is processed to extract information.

In one embodiment, the information extracted from the movement data is the entropy. This can be found using many known methods, for example Renyi entropy (preferably with alpha = 2 or 3), or Tsallis entropy (preferably with alpha = 3).

Entropy may be calculated for the time frequency representation of each principal component. Alternatively, the Holder exponent may be evaluated. These methods essentially extract patterns, or complexity in the signal. This information is then used to train the model/classify data. The present inventors have discovered another type of information that can be extracted and used to train the model/classify data. This is a vector of 'period length'. In order to find this, the signal is divided into windows of equal size, so that each window contains a number of signal samples. For each window, the number of samples Λ: between the time the signal changes its sign are counted. This number of samples is called 'period length¹. Typically, in each window the signal might cross the zero line several times, thus several values of x will be obtained. A vector is then created from these values. For example, if the signal crosses the zero line n times, vector X = [x; , %₂ ... X_n] where *,^• is the number of samples between each zero crossing. X may then be processed to remove unnecessary noise. In a preferred approach, the vector X of period length is found by detrending the signal by subtracting from each sample the average signal value of a region centred around the current sample, creating a vector of the number of samples between consecutive zero-crossings of the detrended signal, using the vector of the number of samples to compute a vector of local periodicity using a sliding window, thresholding the vector of local periodicity, and calculating the sum of all local periodicity.

This vector measures contain information about the movement that depends on the CP status of the subject, and as such can be used to train a model/classify data according to CP status. However, 'period length¹ information is not limited to use in the context of CP, it can be used when training a model/classifying data for other purposes as well. In one embodiment, more than one method is used to extract information from the data, and all the information is used to train the model/classify data.

The classification model may be any suitable model capable of being trained. Preferably, the model comprises an algorithm that can be trained using machine- learning techniques. Suitable algorithms include linear/nonlinear discriminant analysis algorithms, decision trees, clustering algorithms and neural networks such as Fuzzy ARTMap. Statistical models can be used to find a probability distribution that can discriminate between children with CP and those without CP.

The trainable algorithms can 'learn' how to classify data using training data (input signals), i.e. data for which the classification (category) is already known.

Such training data may be 'marked' with the known category. In one embodiment, a database is provided containing the known category of each training input signal, and this is interrogated as necessary during the training process to find the correct category for a certain piece of data. Thus, when the invention is applied to the diagnosis of CP, training data is required for which it is known which of the children went on to be firmly diagnosed as having CP. The 'known' category of the movement data may be found in any way, but is preferably found by obtaining movement data from a population and then waiting for a period of time, for example until the infants are around two years old, when it is normally clear which have CP. This is the most effective way of determining the 'correct' CP status of an infant. Alternatively, an experienced clinician may use the GMA technique to estimate whether the infant has CP or not, thus averting the need to wait two years to determine the CP status.

In order to train the model, training data may used to estimate the model parameters. In one embodiment, the chosen algorithm is used in an initial state to classify data into a particular category according to the extracted information. This category is then compared with the known correct category for the data, and if the chosen category does not match the correct category, the parameters of the model are modified until the chosen category matches the correct category. The accuracy of the classification model may be tested using a test dataset of movement data, for which the correct category is already known. (A sub-set of the training data is preferably not used for the training process and instead forms the test dataset.) The sensitivity and specificity may be evaluated, and the confidence interval may be determined.

In one embodiment of the invention, a number of different sets of training movement data are used to create different classification models, and the model which provides the best test results is chosen. In another embodiment, a combination of a number of models is used.

Generally, using more training data to train the model will result in a more accurate model. Therefore in the case of CP, the classification model will become more refined as datasets from more children whose CP status is known is used to train the model.

The trained algorithm can be used to classify further data for which the category is not known. For example, if a discriminant analysis algorithm is used, the result will be a trained discriminant analysis algorithm that can then be used to classify data. The method of creating a classification model for categorising data relating to the movement of a living subject is considered as an invention in its own right.

Thus, viewed from another aspect, the present invention provides a method of creating a classification model for categorising data derived from the movements of a living subject into one of a plurality of categories, comprising: providing a set of data derived from a population of subjects whose category is known; processing the data to extract information; and using this information to train a classification model.

The methods of categorising movement data and creating a classification model involve a significant amount of data processing and so they will normally be implemented using a computer. For example, a computer may be programmed to receive movement data, process the movement data and categorise it using the classification model. The computer may also be programmed to train the classification model. As such, a program may be provided with two modes: training mode and production mode. In the training mode, data is received and processed to train the algorithm, and in the production mode new data is received and is categorised according to the so-trained algorithm. Alternatively, the classification algorithm may be trained separately. The invention also provides a method of training a classification algorithm for categorising data relating to the movement of a living subject, comprising: using the classification algorithm to classify movement data into a particular category according to information extracted from the movement data; comparing the chosen category with a known correct category for the data; and if the chosen category does not match the correct category, modifying the classification algorithm until the chosen category matches the correct category.

In another aspect, the invention provides a method of creating a classification model for categorising data relating to the movement of a living subject, comprising: obtaining data corresponding to the movement of at least one part of the subject over time; performing time frequency decomposition of this data; processing movement data to extract patterns; and training a classification algorithm using this movement data; wherein the trained classification algorithm forms the classification model.

The term "patterns" is used herein to encompass measures or features such as periodicity or entropy as discussed above.

The invention also extends to a classification model created according to any of the methods described above.

The present invention also provides a method of categorising data relating to the movement of a living subject, comprising: creating a classification model according to one of the methods described above; and categorising new data using the classification algorithm. Also provided is a method of categorising data relating to the movement of a living subject, comprising: processing movement data to extract information/patterns from the signal; and classifying the data into a particular category according to the extracted information, using a classification model created as described above.

The invention also extends to apparatus configured to carry out any of the methods described above.

Another aspect of the invention provides an apparatus for categorising data derived from the movements of a living subject, comprising: a processor for processing the data to extract information; and a classifier for classifying the extracted information into one of a plurality of categories using a classification model; wherein the classification model comprises a classification algorithm trained using data derived from the movements of other subjects whose category is known.

The processor can be any suitable means for processing data, and the classifier is any suitable means for classifying data. Preferably both of these are implemented as a computer program.

In yet another aspect, the invention provides an apparatus for creating a classification model for categorising data relating to the movement of a living subject, comprising: a means for processing movement data to extract patterns in the signal; and a means for training the classification model using the processed data. Preferably, the classification algorithm comprises a classification model.

According to another aspect, the invention provides an apparatus for creating a classification model for categorising data relating to the movement of a living subject, comprising: a means for processing movement data to extract patterns in the signal; and a means for training the classification algorithm using the processed data; wherein the trained classification algorithm forms the classification model.

Preferably, said means for processing and said means for training are computer programs.

Preferably the apparatus is configured to operate in accordance with any of the methods described above. In another aspect, the invention provides a software product comprising instructions which when executed by a computer cause the computer to process data relating to the movements of a living subject in order to extract information, and classify the extracted information into one of a plurality of categories using a classification model; wherein the classification model has been trained using data derived from the movements of other subjects whose category is known.

In yet another aspect, the invention provides a software product comprising instructions which when executed by a computer cause the computer to process a set of movement data derived from a population of subjects whose category is known; extract information from this data; and use this information to train a classification model for categorising data relating to the movement of a living subject.

Preferably, a software product is provided comprising instructions which when executed by a computer cause the computer to train a classification model and use a classification model to classify data according to a user's requirements, as described previously.

The software product may be a physical data carrier, or it may comprise signals transmitted from a remote location. In a further aspect, the invention provides a method of manufacturing a software product which is in the form of a physical carrier, comprising storing on the data carrier instructions which when executed by a computer cause the computer to process data relating to the movements of a living subject in order to extract information, and classify the extracted information into one of a plurality of categories using a classification model; wherein the classification model has been trained using data derived from the movements of other subjects whose category is known.

In yet another aspect, a method of manufacturing a software product which is in the form of a physical carrier, comprising storing on the data carrier instructions which when executed by a computer cause the computer to process a set of movement data derived from a population of subjects whose category is known; extract information from this data; and use this information to train a classification model for categorising data relating to the movement of a living subject.

In another aspect, the invention provides a method of providing a software product to a remote location by means of transmitting data to a computer at that remote location, the data comprising instructions which when executed by the computer cause the computer to process data relating to the movements of a living subject in order to extract information, and classify the extracted information into one of a plurality of categories using a classification model; wherein the classification model has been trained using data derived from the movements of other subjects whose category is known.

In yet another aspect, the invention provides a method of providing a software product to a remote location by means of transmitting data to a computer at that remote location, the data comprising instructions which when executed by the computer cause the computer to process a set of movement data derived from a population of subjects whose category is known; extract information from this data; and use this information to train a classification model for categorising data relating to the movement of a living subject.

The invention further extends to a method of diagnosing cerebral palsy (or another neurological condition) using the methods and/or apparatus discussed above. It will be appreciated that the preferred features described in relation to particular aspects above may also be applicable to other aspects. Further, it will be appreciated that the various features described may be used in isolation or in combination with other preferred features.

Preferred embodiments of the present invention will now be described by way of example only and with reference to the accompanying drawings, in which:

Figure 1 is a schematic view of the elements of an apparatus for categorising movement data according to one embodiment of the invention;

Figure 2 shows in more detail the elements of an apparatus for categorising movement data according to one embodiment of the invention; Figure 3 illustrates sensors attached to a subject's body;

Figure 4 is a schematic view of the process of training a classification algorithm according to an embodiment of the invention;

Figure 5 illustrates the steps of processing movement data and extracting information, according to an embodiment of the invention; Figure 5 a illustrates a matrix of movement data;

Figure 6a is a graph of first and second principal components v. time of the data signal from a sensor on a 'normal' patient.

Figure 6b is a graph of first and second principal components v. time of the data signal from a sensor on a patient having CP. Figure 7 schematically shows the step of training a classification algorithm according to an embodiment of the invention;

Figure 7a schematically illustrates a principle of training a classification algorithm;

Figure 8 schematically illustrates the Fuzzy ARTMAP algorithm; Figure 9 schematically illustrates the steps of training classification models according to an embodiment of the invention; Figure 10 shows a welcome screen of a computer program for categorising movement data, according to an embodiment of the invention;

Figure 11 shows a system status check screen displayed by computer program for categorising movement data, according to an embodiment of the invention;

Figure 12 shows a main menu displayed by a computer program for categorising movement data, according to an embodiment of the invention;

Figure 13 shows an ID number input box displayed by a computer program for categorising movement data, according to an embodiment of the invention; Figure 14 shows a patient record displayed by a computer program for categorising movement data, according to an embodiment of the invention;

Figure 15 shows recording information displayed by a computer program for categorising movement data, according to an embodiment of the invention;

- ■ Figure 16 shows video capture controls displayed by a computer program for categorising movement data, according to an embodiment of the invention;

Figure 17 shows a region of interest editing screen displayed by a computer program for categorising movement data, according to an embodiment of the invention;

Figure 18 shows the results of the categorisation displayed by a computer program for categorising movement data, according to an embodiment of the invention; and

Figure 19 shows the detailed results of the categorisation displayed by a computer program for categorising movement data, according to an embodiment of the invention. Figure 1 schematically illustrates different elements of an apparatus 1 for categorizing movement data according to one embodiment of the invention. The apparatus 1 comprises an input device 5 connected to a CP detection unit 8 which is connected to display 11. CP detection unit 8 may also be connected to printer 12 (this is indicated by a dotted connection line). Input device 5 collects movement data and provides this to CP detection unit

8. Unit 8 estimates whether the movement is suggestive of CP, and outputs the result to computer display 11 and/or printer 12. Alternatively, instead of input device 5 that collects movement data, a database 7 may be provided in which movement data of a subject is already stored.

In a first embodiment, as illustrated in Figure 2, the input device is a video camera combined with a motion tracking system. Video camera 4 is pointed at mat 2 upon which an infant 3 is lying. The camera is typically raised about 105 cm above the level of the mat and is located near one end thereof, pointing down at the mat and inclined about 25 degrees from the vertical. Video camera 4 is linked to a motion tracking system 6 connected to CP detection unit 8, which outputs results 10. The video camera 4 collects two- dimensional video footage of the infant, and sends it to tracking system 6. This analyses the video footage using motion detection software in order to determine how the different parts of the infant are moving.

Figure 3 shows an alternative, second, embodiment where a camera is not used. Instead, electromagnetic sensors 8 are attached to key sites of movement and the output data is fed to the tracking system 6. Such sensors might monitor position, muscle activity or another kind of signal that provides information on relative movement.

In both embodiments the movement data output of the tracking system is input to the CP detection unit 8, which pre-processes the data, extracts information and classifies the movement data as represented by the extracted information into a particular category, for example, 'healthy', 'CP', or 'at risk'. The result is then output. Classification of the movement data is performed by a classification model which categorises the data using a trained classification algorithm.

The CP detection unit and results display are implemented as elements of a computer program running on a conventional computer, with which a user can interact to categorise movement data. The GUI of the computer program is described later with reference to Figures 10-19.

The process of creating a classification model for use by CP detection unit 8 to categorise movement data will now be described with reference to Figures 4-9.

Figure 4 is a schematic diagram illustrating the process of training a classification algorithm for categorizing data that is used in CP detection unit 8. Movement data 20 is data relating to the movement of different parts of an infant. This is collected using the apparatus and methods described above in relation to categorizing movement data. It is collected in advance and then stored in a database, before being used to train a classification algorithm. The 'true' CP status of the infant is known, having been determined by waiting until the infant is about two years old, when it is normally clear whether an infant has CP or not. The raw movement data 20 (sometimes called 'training data') is input to a processing module 21 shown in more detail in Figure 5 and 5a. The processed data is then used to train a classification model, as described in more detail below in relation to Figure 7.

Figure 5 illustrates the processing steps performed by processing module 21. Raw movement data of an infant has been collected from six different sensors placed at different parts of the body. This data comprises the position of the sensor in three dimensions (x, y and z coordinates), over time. A matrix 26 of such movement data is illustrated in Figure 5a. This raw data has been selected according to a Region of Interest (ROI) in time, i.e. corresponding to when spontaneous movement is occurring, which is crucial in terms of CP assessment. For example, when the infant is crying or playing, the data at those times is not useful, and so falls outside the ROI and is not selected. In this embodiment, the ROI has been selected manually by a human operator who has viewed the video of the infant.

As shown in Figure 5, principal component analysis (PCA) 27 is carried out on the movement data, in order to reduce the data matrix 26 to a matrix 28 of the principle components of the data, over time. Temporal spectral decomposition is then carried out on the principal components, using time frequency representation 29, to find the amount of each frequency present in each principal component at a given time, as shown in matrix 30. Fourier transforms may be used for this step.

Figure 6a is a graph showing how the first and second principal components of a raw data signal from a sensor placed on a 'normal' patient, i.e. one which does not have CP, vary over time. Figure 6b on the other hand is a graph showing the first and second principal components of a raw data signal from a sensor placed on a ¹CP' patient.

As shown in Figure 5, at information extraction stage 31, the entropy of the TFR is found for each principle component, using Renyi entropy. In other embodiments, other types of information are extracted, the common feature being that the information is related in some way to patterns of movement in the data. The processing steps illustrated in Figure 5 are also performed by CP detection unit 8 on incoming test movement data from tracking system 6, during the categorisation process.

The output of the information extraction stage 31 together with the correct category of the data is then input to the classifier 32, that comprises a trainable classification algorithm, as shown in Figure 7. The algorithm is trained using the output from the information extraction together with the desired output, i.e. the known CP status. The result is a trained classification model.

A general principle of training a model by supervised learning is shown in Figure 7a. In this embodiment, both the extracted information and the correct class (category) for each subject are stored in a database, and are retrieved when the classifier is to be trained. Initial parameters of the 'learning algorithm' 22 (e.g. the classification algorithm) are pre-set, and a chosen class is computed for the extracted information. Supervisor 27 processes the desired output so that it can be compared to the class chosen by learning algorithm 22 at step 25. If the class is not correct, then feedback 26 is provided to the algorithm and the internal parameters are changed. The process is repeated until the learning algorithm gives the correct result.

This is then repeated for a number of different training datasets, so that the algorithm is refined.

In the case of a neural network algorithm, for example, the output networks may contain 2 neurons. If the input pattern is indicative of normal CP status, one neuron should be active, if abnormal then the other should be active. Each time a normal pattern is presented, the output of the "normal" neuron is checked to see if it is correct, and the output of the "abnormal" neuron is also checked. If there are errors, the system learns to put those neurons in the right state.

In one embodiment of the invention, the classification algorithm is a Fuzzy ARTMAP (FAM) artificial neural network, illustrated in Figure 8. This is a known type of neural network, and is described in more detail, for example, in "Use of reliability measures to improve the performance of fuzzy ARTMAP network",

International Joint Conference on Neural Networks (IJCNN '99), 1999, Ramuhalli et al.; "Fuzzy ARTMAP: A neural network architecture for incremental supervised learning of analog multidimensional maps", IEEE Trans. Neural Networks, Vol. 3, No. 5, pp. 698-713, September 1992, Carpenter et al.; and 'A comparison of Fuzzy ARTMAP and Multilayer Perception for Handwritten Digit Recognition, Busque et al, citeseer.csail.mit.edu/busque97 comparison.html A Fuzzy ART neural network is an unsupervised neural network capable of incremental learning. Fuzzy ARTMAP on the other hand is a supervised learning network comprising two Fuzzy ART modules, ARTa and ARTb. The two modules are interconnected through a single layer of weighted connections (called the 'Map' layer) between the 'F2' layers of the modules. The layer Fl is the input layer ('hypothesis'), and the layer F2 is the output/category layer.

In order to train the FAM network, firstly, all weights are initialised. Then, the input (i.e. the extracted information — the entropy etc.) is presented to ARTa. The desired output (i.e. the known class) is presented to ARTb. Category formation takes place in both modules. This means that in each module, the input activates a search mechanism for a matching neuron in the F2 layer. In ARTa, a neuron is selected according to the extracted information. In ARTb, a neuron is selected according to the desired output. Once neurons are selected, a vigilance criterion is evaluated for each module. The vigilance criterion is a parameter representing how similar the input patterns need to be in order to be classified as belonging to the same category.

The vigilance criterion will generally be set to 1 in ARTb, so as to perfectly distinguish between the desired output vectors. This process in ARTb essentially 'encodes' the desired output information into a format that can be compared with the output of ARTa. The vigilance criterion of ARTa will vary during the learning process. If the vigilance criterions are not satisfied, then category formation is repeated until they are (shown as 'Reset' in Figure 8).

The selected neurons are then compared using the Map interconnections to determine whether the neuron selected by ARTa corresponds to the desired output presented to ARTb. If this is not satisfied, then the vigilance criterion of ARTa is increased and category formation is repeated, such that a different neuron is selected. When a correct match is found, the Map layer 'learns' the association between the input and the desired output by updating its weights, so it can correctly classify a similar input in the future.

This process is repeated for each new input/desired output presented to the FAM network. When the trained classifier is used to classify new data, the extracted information is input to ARTa. ARTb is not used and learning is deactivated. ARTa selects an output neuron, the Map field associates a class with this selected neuron, and this chosen class is output from the classifier.

Movement data of a population of many different infants are used in order to train the classification algorithm in an optimum manner. The classifier is then tested using additional movement data for which the correct CP status is known, and the results compared with the true status. The sensitivity and specificity is commonly used to evaluate the results, wherein:

Sensitivity = TP/(TP+FN)

Specificity = TN/(TN+FP)

TP = No. of 'true positives' TN = No. of 'true negatives' FP = No. of 'false positives'

FN = No. of 'false negatives'

Typically, the ratio of training data to testing data may be 8:2, i.e. given a set of raw movement data with known outcomes, 80% may be used for training and 20% may be used for testing.

Once the classification algorithm has been trained to the required degree and produces acceptable results, it is maintained in that format. It can then be used to categorise new movement data as previously described in relation to Figures 1 and 3. As illustrated in Figure 9, outputs from different sensors may be used to provide input raw movement data for training the algorithm. In one embodiment, a number of different sensors are grouped into different sensor sets 41. The data from each sensor set is separately used to train a classification algorithm (termed in this figure a 'learning algorithm') in the manner described above. The result is a differently trained classification algorithm for each sensor set, which can be considered as different classifiers. The model creation module 43 evaluates the different classifiers using test data, and chooses the one that performs best, hi fact, the chosen classifier may be a compound of different classifiers. In that case, when categorizing data, the chosen category will be that given the most 'votes' by the different classifiers.

The CP detection unit for categorising movement data described previously is implemented using a computer program, with which a user can interact to categorise the movement of infants and display the results. Figures 10 to 19 illustrate an example of a GUI of a suitable computer program which takes movement data as its input, categorises this and then displays the output. Figure 10 illustrates a welcome screen. Before the system starts, a system status check is performed to check various parameters, such as disk capacity, video camera (if used), position of subject in field of view etc. If the subject is outside the screen or one part of the body is outside the screen, a warning message may be given. This is shown in Figure 11. Figure 12 illustrates a main menu for the system with various options, and Figure 13 shows a box in which a patient's ID number can be input. Figure 14 shows what a patient record may look like. Figure 15 visualises the recording date as a particular time during the period of 5-10 weeks post term, the time during which spontaneous movement can be used to detect CP. The cursor should always be in the central zone, otherwise the interpretation of the result might be invalid. In this example, the computer program is used to control the video capture, and this is shown in Figure 16. The video is on the left, and the activity of different sensors in the centre. Any number of sensors can be displayed. The elapsed time and recording length is given in the 'status' box. The user can control the recording using the buttons in the 'control panel'. Figure 17 is a view of the editing screen. The user can choose to remove a part of the input data, for example if the infant is crying. The desired data is termed the region of interest, and can be selected using the buttons in the 'ROI list' box. Figure 18 is a screen showing the output of the CP detection process. The classification result shows the outcome both graphically and as text.

Figure 19 illustrates the 'detailed result' screen displayed by clicking the 'detailed result' tab of Figure 18. A classification graph is given which represents two-dimensionally the position of the tested subject in the feature space. For the graph, two main features have been selected, for example entropy and periodicity.

Claims

1. A method of categorising data derived from the movements of a living subject, comprising: processing the data to extract information and classifying the extracted information into one of a plurality of categories using a classification model, wherein the classification model is trained using training data derived from the movements of other subjects whose category is known.

2. A method as claimed in claim 1, wherein the extracted information relates to patterns in the data.

3. A method as claimed in claim 1 or 2, wherein the extracted information is a measure of entropy.

4. A method as claimed in claim 3, wherein the entropy is Renyi entropy.

5. A method as claimed in claim 3, wherein the entropy is Tsallis entropy.

6. A method as claimed in claim 1 or 2, wherein the extracted information is the Holder exponent.

7. A method as claimed in claim 1 or 2, wherein the data to be categorised comprises samples of a signal relating to the movements of the subject, and wherein the signal has been sampled at regular time intervals.

8. A method as claimed in claim 7, wherein the extracted information is a vector X of period length.

9. A method as claimed in claim 8 wherein the vector X of period length is found by: dividing the data into time windows of equal size so that each window contains a number of signal samples; for each window, calculating the period length as the number of samples x between the time the signal changes its sign; and creating a vector X of period length from this information.

10. A method as claimed in claim 8 or 9, wherein the vector X of period length is found by: detrending the signal by subtracting from each sample the average signal value of a region centred around the current sample; creating a vector of the number of samples between consecutive zero-crossings of the detrended signal; using said vector of the number of samples to compute a vector of local periodicity using a sliding window; thresholding the vector of local periodicity; and calculating the sum of all local periodicity.

11. A method as claimed in any preceding claim, wherein more than one type of information is extracted and used to categorise the data.

12. A method as claimed in any preceding claim, wherein the classification model comprises an algorithm that can be trained using machine-learning techniques.

13. A method as claimed in any preceding claim, wherein the classification model comprises a linear and/or nonlinear discriminant analysis algorithm.

14. A method as claimed in any preceding claim, wherein the classification model comprises a decision tree.

15. A method as claimed in any preceding claim, wherein the classification model comprises a clustering algorithm.

16. A method as claimed in any preceding claim, wherein the classification model comprises a neural network.

17. A method as claimed in claim 16, wherein the neural network is Fuzzy ARTMap

18. A method as claimed in any preceding claim, further comprising the step of collecting the data to be categorised.

19. A method as claimed in claim 18, wherein the data is collected using electromagnetic sensors connected to different parts of the body which feed data to a tracking system.

20. A method as claimed in claim 18, wherein the data is collected by videoing the subject and using image-processing software to determine the movement of different parts of the body.

21. A method as claimed in any preceding claim, wherein the data to be categorised and/or the training data is two-dimensional.

22. A method as claimed in any preceding claim wherein the data to be categorised and/or the training data is three-dimensional.

23. A method as claimed in any preceding claim, wherein the data represents three linear (x, y, z) dimensions and also rotational parameters.

24. A method as claimed in any of claims 19 to 23, wherein categorisation takes place in real-time as the data is collected.

25. A method as claimed in any of claims 19 to 23, wherein the collected data is stored and categorisation is performed at a later time.

26. A method as claimed in any preceding claim, further comprising the step of pre-processing data before the information is extracted.

27. A method as claimed in claim 26, wherein the pre-processing comprises selecting a region of interest in the data comprising only useful data.

28. A method as claimed in claim 26, wherein the pre-processing comprises performing principal components analysis (PCA) on the data.

29. A method as claimed in claim 26, wherein time frequency decomposition is performed on the data.

30. A method as claimed in claim 28, wherein time frequency decomposition is performed on each principal component.

31. A method as claimed in any preceding claim, wherein the classification model is trained by: extracting information from the training data; using the classification model in an initial state to classify data into a particular category according to the extracted information; comparing the chosen category with the known category; and if the chosen category does not match the known category then adjusting the parameters of the model until a correct match is found.

32. A method as claimed in any preceding claim, wherein the training data includes data corresponding to all of said plurality of categories.

33. A method as claimed in any preceding claim, wherein a database is provided containing the known category of each piece of training data and wherein this database is interrogated during training process to find the correct category for certain piece of data.

34. A method as claimed in any preceding claim, wherein a number of different classification models are created from different sets of training data and a combination of these models is used to create a signal optimum classification model.

35. A method as claimed in any preceding claim, wherein the training data is collected using electromagnetic sensors connected to different parts of the body which feed data to a tracking system.

36. A method as claimed in any of claims 1 to 34, wherein the training data is collected by videoing the subject and using image-processing software to determine the movement of different parts of the body.

37. A method as claimed in any preceding claim, further comprising the step of pre-processing the training data before the model is created.

38. A method as claimed in claim 37, wherein the pre-processing comprises selecting a region of interest in the training data comprising only useful data.

39. A method as claimed in claim 38, wherein the pre-processing comprises performing principal components analysis (PCA) on the training data.

40. A method as claimed in any of claims 37 to 39, wherein time frequency decomposition is performed on the training data.

41. A method as claimed in any preceding claim, wherein the classification model is adapted to classify extracted information according to whether the subject has, or is susceptible to, Cerebral Palsy (CP).

42. A method as claimed in claim 41 , wherein the subject is a human infant of the age 6-20 weeks post term.

43. A method as claimed in claim 41 or 42, wherein the movements of the subject are spontaneous movements.

44. A method as claimed in claim 41, 42 or 43, wherein data relates to movements of one or more of the following: the end of the subject's limbs; head; chest; and/or other part(s) of the subject's body.

45. A method of creating a classification model for categorising data derived from the movements of a living subject into one of a plurality of categories, comprising: providing a set of data derived from a population of subjects whose category is known; processing the data to extract information; and using this information to train a classification model.

46. A method of creating a classification model for categorising data relating to the movement of a living subject, comprising: obtaining data corresponding to the movement of at least one part of the subject over time; performing time frequency decomposition of this data; processing the data to extract patterns; and training a classification algorithm using this movement data; wherein the trained classification algorithm forms the classification model.

47. A method of training a classification model for categorising data relating to the movement of a living subject, comprising: using the classification model to classify movement data into a particular category according to information extracted from the movement data; comparing the chosen category with a known correct category for the data; and if the chosen category does not match the correct category, modifying the classification model until the chosen category matches the correct category.

48. A method of categorising data relating to the movement of a living subject, comprising: creating a classification model according to claim 45, 46 or 47; and categorising new data using the classification model.

49. A classification model created according to the method of claim 45, 46 or 47.

50. An apparatus configured to carry out the method of claims 1 to 48.

51. An apparatus for categorising data derived from the movements of a living subject, comprising: a processor for processing the data to extract information; and a classifier for classifying the extracted information into one of a plurality of categories using a classification model; wherein the classification model comprises a classification algorithm trained using data derived from the movements of other subjects whose category is known.

52. An apparatus for creating a classification model for categorising data relating to the movement of a living subject, comprising a means for processing movement data to extract patterns in the signal and a means for training the classification model using the processed data.

53. An apparatus for creating a classification model for categorising data relating to the movement of a living subject, comprising a means for processing movement data to extract patterns in the signal and a means for training the classification algorithm using the processed data, wherein the trained classification algorithm forms the classification model.

54. A software product comprising instructions which when executed by a computer cause the computer to process data relating to the movements of a living subject in order to extract information and classify the extracted information into one of a plurality of categories using a classification model, wherein the classification model has been trained using data derived from the movements of other subjects whose category is known.

55. A software product comprising instructions which when executed by a computer cause the computer to process a set of movement data derived from a population of subjects whose category is known, extract information from this data, and use this information to train a classification model for categorising data relating to the movement of a living subject.

56. A software product comprising instructions which when executed by a computer cause the computer to create a classification model according to the method of claims 45, 46 or 47 and use the classification model according to the method of any of claims 1 to 44 in order to categorise movement data.

57. A software product as claimed in claims 54, 55 or 56, wherein the software product is a physical data carrier.

58. A software product as claimed in claims 54, 55 or 56, wherein the software product comprises signals transmitted from a remote location.

59. A method of manufacturing a software product which is in the form of a physical carrier, the method comprising: storing on the data carrier instructions which when executed by a computer cause the computer to: process data relating to the movements of a living subject in order to extract information; and classify the extracted information into one of a plurality of categories using a classification model; wherein the classification model has been trained using data derived from the movements of other subjects whose category is known.

60. A method of manufacturing a software product which is in the form of a physical carrier, comprising storing on the data carrier instructions which when executed by a computer cause the computer to: process a set of movement data derived from a population of subjects whose category is known; extract information from this data; and use this information to train a classification model for categorising data relating to the movement of a living subject.

61. A method of providing a software product to a remote location by means of transmitting data to a computer at that remote location, the data comprising instructions which when executed by the computer cause the computer to: process data relating to the movements of a living subject in order to extract information; and classify the extracted information into one of a plurality of categories using a classification model; wherein the classification model has been trained using data derived from the movements of other subjects whose category is known.

62. A method of providing a software product to a remote location by means of transmitting data to a computer at that remote location, the data comprising instructions which when executed by the computer cause the computer to: process a set of movement data derived from a population of subjects whose category is known; extract information from this data; and use this information to train a classification model for categorising data relating to the movement of a living subject.