Seizure identification system and method

Abstract

Seizure identification system 5 is worn or carried by a subject 10 and comprises biosensor 15, e.g. electroencephalography (EEG) electrodes 25. It may have a processing system 30 having data storage 35, a communications module 40 and a battery 45, as could be embedded in a smartphone. The method preprocesses EEG signals and splits them into non-overlapping time slots e.g. epochs of around 3 seconds. These are decomposed, possibly by Discrete Wavelet Transform (DWT) or Fast Fourier Transform (FFT) into sub-bands, using a series of high-pass and low-pass filters. Statistical features can then be extracted, producing a feature vector which is passed to e.g. a convolutional neural network (CNN). This classifier (1500, fig. 15) may comprise one-dimensional convolution layers. The system may identify epileptic seizures.

Description

Field The present disclosure relates generally to seizure identification, such as a system and method for identifying epileptic seizures. Background Epilepsy is a neurological disorder characterized by recurrent episodes of brain activity associated with abnormal brain synchronization and fast neuronal activities. Excessive electrical activity in brain cells manifests itself as epileptic seizures. In many some cases, epileptic seizures can lead to serious cognitive, neurological, and physiological implications, including loss of consciousness or even death if rapid and thorough diagnosis and monitoring are not conducted. EEGs (Electroencephalograms) can be used to detect epileptic seizures. EEG signals are captured by placing multiple electrodes on different areas of the patient’s scalp. However, the signals recorded using EEG are difficult to analyze. Continuous monitoring and analysis of EEG signals is a significant challenge since it requires the presence of a trained neurophysiologist. Furthermore, accurate identification of seizures is time-consuming and requires expert knowledge. At least one example of the present disclosure seeks to provide improved seizure identification. Summary Various aspects of the present invention are defined in the independent claims. Some preferred features are defined in the dependent claims. According to a first example of the present disclosure is a seizure identification system for identifying seizures in a subject, the seizure identification system comprising: at least one input for receiving at least one input signal indicative of brain activity of the subject; a decomposition engine configured to decompose at least part of the at least one input signal into one or more frequency sub-bands; a classifier configured to identify if the at least one input signal is indicative of a seizure based at least on the one or more decomposed signal components, the classifier being configured to identify if the at least one input signal is indicative of a seizure. The classifier may comprise at least one model. The model may comprise an artificial intelligence model. The model may comprise or be comprised in a neural network, which may be a random neural network (RNN) or a convolutional neural network (CNN). The input signal may be, comprise or be representative of an electrical signal. The at least one input may comprise at least one EEG (Electroencephalogram) signal. The system may comprise at least one EEG sensor system for collecting the at least one EEG signal from the subject. The system may be configured to communicate, e.g. wirelessly communicate, with at least one EEG sensor system for collecting the at least one EEG signal from the subject. The at least one EEG sensor system may be comprised in or integral with the seizure identification system, e.g. in the same device and / or with a shared housing and / or configured for mutual movement together, or the at least one EEG sensor system may be separate, e.g. provided in a separate device to the seizure identification system, or otherwise spaced apart from the seizure identification system in use. The at least one input signal may be received and / or processed by the seizure identification system in use, e.g. on or immediately after collection. The at least one input may be or comprise data retrieved from computer readable storage. The at least one input signal may comprise a plurality of channels, e.g. each channel may comprise an EEG reading obtained from a different combination of electrodes. The decomposition engine may be configured to decompose at least part of the at least one input signal into the one or more frequency sub-bands so as to generate a decomposed signal component for each frequency sub-band. The decomposition engine may be implemented in software, in hardware or any combination thereof. The decomposition engine may comprise or be configured to implement a frequency analysis technique. The decomposition engine may comprise or be configured to implement a wavelet transform, such as a discrete wavelet transform (DWT). The wavelet transform may comprise using Daubechies-4 (DB-4) wavelet function. The decomposition engine, e.g. the wavelet transform or other suitable operation, may extract any, some or all of the following EEG frequency bands: gamma (30-60 Hz), beta (15-30 Hz), alpha (8-15 Hz), theta (4-8 Hz) and delta (0-4 Hz). The decomposition engine may comprise or be configured to implement the wavelet transform on the at least part of the at least one input signal before applying the classifier, e.g. the neural network. The decomposition engine may comprise or be configured to implement a Fourier transform, such as a fast Fourier transform (FFT). The decomposition engine may be configured to decompose at least part of the at least one input signal into a plurality of the frequency sub-bands. The frequency sub-bands may be non-overlapping, i.e. each frequency sub-band may not overlap any other frequency sub-band. The frequency sub-bands may be contiguous with each other. The decomposition engine may be configured to decompose at least part of the at least one input signal into between three and ten frequency sub-bands, e.g. between three and seven frequency sub-bands, such as five frequency sub-bands. The frequency sub bands may be sub-bands of a frequency range from 0Hz to 500Hz, e.g. between 0Hz and 200Hz, such as between 0Hz and 100Hz, and in some examples 60Hz or less, or 15Hz or less. The isolation of certain specific frequency sub-bands from the EEG data and feeding the frequency sub-bands into the neural network gives significant improvements in accuracy of seizure detection and reduces false results. The seizure identification system may comprise or be configured to implement a pre-processor. The pre-processor may be implemented as software or hardware or a combination of both. The pre-processor may be configured to receive and / or process the at least one input signal. The pre-processor may be provided in a data path between the sensors, e.g. the EEG sensors, that collect the at least one input signal and the decomposition engine. The output of the pre-processor may be provided directly or indirectly to the decomposition engine. The pre-processor may be configured to identify time slots in the at least one input signal and / or segment the at least one input signal into the time slots, each time slot being associated with a different time range at which the at least one input signal was collected. Each time slot may not overlap any other time slot, e.g. each time slot may cover a non-overlapping time range. The pre-processor may be configured to segment each channel of the at least one input signal into the time slots. The preprocessor may be configured to segment each channel of the at least one input signal into time slots using a sliding window. The time slots may be partially overlapping with the preceding and / or following time slot. The overlap between adjacent time slots may be less than half of the time slot, e.g. a third of the time slot or less. The time slots may have a duration that is in a range from 3 to 30 seconds, e.g. from 5 to 20 seconds, such as around 10 seconds. The overlap may be 10 seconds or less, e.g. 5 seconds or less, e.g. around 3 seconds. The time slots for each channel may correspond. The time slots may be contiguous, e.g. each time slot for a given channel of the at least one input signal may be contiguous with one or two other time slots. By segmenting the input signal as identified above, it can be ensured that the various phases of seizure activities including pre-ictal, ictal, and inter-ictal stages are captured. The decomposition engine may configured to decompose at least part of the at least one input signal into one or more frequency sub-bands in order to produce a decomposed input signal. The decomposition engine may be configured to decompose at least part of at least one or each channel of the at least one input signal in at least one or each time slot into a plurality of frequency sub-bands in order to produce the decomposed electrical signal. The decomposition engine may be configured to decompose at least part of at least one or each channel of the at least one input signal in at least one or each time slot into between three and ten frequency sub-bands, e.g. between three and seven frequency bands. The frequency bands may be contiguous and / or non-overlapping. The frequency bands for each time slot may have the same ranges of frequencies, e.g. the decomposition engine may be configured to decompose at least part of at least one or each channel of the at least one input signal in each time slot into set or pre-set frequency bands, having a set or pre-set range of frequencies. The pre-processor may comprise at least one filter, such as a high pass filter, a low pass filter and / or a band pass filter. The pre-processor may comprise a high pass filter, which may be configured to remove frequencies above a threshold value from at least one or each channel of the at least one input signal or a band pass filter configured to remove frequencies above a first threshold value from at least one or each channel of the at least one input signal and to remove frequencies below a second threshold value from at least one or each channel of the at least one input signal. For example, the second threshold may be in a range from 30 and 100 Hz, e.g. from 30 and 50 Hz, such as around 40Hz. The first threshold may be 20Hz or less, e.g. 10Hz or less, such as around 0Hz. This may help to remove certain noise that manifests itself as high frequency noise and optionally also noise that manifests itself as very low frequency noise and leave a signal that is focused on seizure activity. This can significantly improve the efficacy of the resulting method. The seizure identification system may comprise or be configured to implement a feature extractor. The feature extractor may be configured to extract at least one feature of the at least one input signal, e.g. from each decomposed input signal. Each decomposed input signal may comprise one of the frequency sub-bands of one of the channels of the at least one of the input signals produced by the decomposition engine. The feature extractor may be configured to extract at least one, and preferably a plurality of features, e.g. statistical features, from the at least one input signal, e.g. from each of the frequency sub-bands of the at least one input signal produced by the decomposition engine. The feature extractor may be configured to form a feature map from the extracted at least one feature. The features extracted by the feature extractor may comprise at least four, e.g. six or more features. The features extracted by the feature extractor may comprise at least one or each or any of: standard deviation, mean, skewness, kurtosis, minimum and / or maximum of the at least one decomposed input signal. Extracting any, some or all of these particular features has been found to improve the accuracy, speed and / or efficiency of seizure detection by reduce the dimensionality of the EEG data while capturing significant characteristics that can indicate seizure activities. The seizure identification system may be configured to normalise the values of the extracted features. This may ensure uniformity and comparability across the extracted features. The normalisation may comprise applying min-max scaling to the feature vectors. The normalization may transform the data into a range between 0 and 1. The features extracted by the feature extractor may be provided to the classifier. The classifier, e.g. the model of the classifier, may receive at least one or each of the features extracted from at least one or each of the one or more frequency sub-bands as inputs. The classifier may be configured to identify if the at least one input signal is indicative of a seizure based at least on the features extracted by the feature extractor, e.g. the features extracted from each of the frequency sub-bands of the at least one input signal produced by the decomposition engine. The model may be configured to perform further feature extraction based on the inputs to the model. The model may be configured to learn and / or by learning patterns and relationships, e.g. higher level patterns and relationships, between the input features over time. The neural network may be trained on training data comprising labelled input signals, e.g. labelled EEG data. The labelling may be done manually or using another classification technique. The labels may indicate portions of the input signal collected whilst a subject was undergoing a seizure, and optionally what type of seizure and / or part of the brain giving rise to the seizure. The training data may have been collected from a range of patients. The training data may be labelled with patient demographic information, such as age, gender, and / or the like, but may be devoid of personal identifying information that could be used to identify the individual subject. The neural network may be trained using a portion of the data to train the neural network and a portion of the data to test the training of the neural network. The neural network may be trained on a training to testing ratio (i.e. the ratio of the training data used to train the neural network to the training data used to test the neural network) of at least 50:50, such as at least 60:40 or at least 80:20. The neural network may be trained with a learning rate of at least 1x10-5, such as at least 1x10-4 or 1x10-3 or more. The neural network may comprise an input layer, an output layer and one or more hidden layers between the input and output layers. The input layer may have a dimensionality or number of neurons corresponding to a number of channels or dimensionality of an input vector derived from the at least one input signal, e.g. corresponding to a factor of the number of channels of EEG data, the number of frequency sub bands and the number of features extracted. The output layer may have a dimensionality or number of neurons of one (in the case of the neural network being configured to provide a binary output such as having a seizure or not having a seizure) or may comprise a dimensionality or number of neurons of greater than one (in the case that the neural network is configured to classify the seizure into one of a plurality of different seizure types). The hidden layer may have a dimensionality or number of neurons that is less than that of the input layer but more than that of the output layer. The hidden layer may have a dimensionality or number of neurons that is 5 or more, e.g. 10 or more and / or 50 or less, e.g. 30 or less. In examples, a hidden layer having a dimensionality or number of neurons of between 15 to 25, e.g. 20, gives an output with a beneficial compromise of accuracy and low error rate against computational resource (e.g. processor, memory and / or network) requirements. The neural network may be or comprise a 1-D CNN. The neural network may comprise a plurality of layers, such as convolutional layers, at least one or two or more of which may be 1D convolutional layers (e.g. convID layers). The 1D convolutional layer may be a temporal convolution. At least one or each of the convolutional layers, e.g. the 1D convolutional layers, may comprise a plurality of filters, e.g. 64 filters, and a kernel size greater than 1, e.g. a kernel size of 3. The 1D convolutional layers may be configured for initial feature learning from the at least one input signal, e.g. from the pre-processed EEG data. The convolutional layers, e.g. the 1D convolutional layers, may be configured to perform the further feature extraction based on the inputs to the model. The convolutional layers, e.g. the 1D convolutional layers, may be configured to learn and / or by learning the patterns and relationships, e.g. higher level patterns and relationships, between the input features over time. The convolutional layers, e.g. the 1D convolutional layers, through their filters, can identify complex, non-linear relationships and temporal patterns in the data that are indicative of seizures. This includes learning from the changes and interactions among the statistical features over time, which has been found to be highly beneficial for predicting seizures. The neural network may comprise one or more layers configured to reduce the dimensionality of the input, e.g. of the pre-processed EEG data. The neural network may comprise one or more pooling layers, such as MaxPoolinglD layers. The MaxPooling 1D layers may down-sample the input, e.g. by taking a maximum value over a spatial window of a size equal to the pool size. In examples, the pool size may be 2 but other pool sizes could be used. The neural network may comprise one or more dropout layers. Each dropout layer may be configured to disregard some nodes in a layer at random during training, wherein the number of nodes disregarded may be according to a dropout rate. The dropout layer may be between 0.1 and 0.5, e.g. a dropout rate of 0.2. The one or more pooling layers and one or more dropout layers may be configured for dimensionality reduction and / or to mitigate overfitting. The neural network may comprise one or more flattening layers. The one or more flattening layers may be configured to reduce the dimensionality of the feature map. The one or more flattening layers may be configured to transform the feature map from a 2D feature map to a 1D feature map. The neural network may comprise one or more dense layers, which may follow the one or more flattening layers. The one or more dense layers may be configured such that each neuron of the dense layer receives input from all neurons of a preceding layer. The dense layer may comprise between 5 and 500 neurons, e.g. between 10 and 100 neurons, such as between 30 and 70 neurons, and may have around 50 neurons. The output layer may receive the output of the one or more dense layers. The output layer may be configured to generate a probability distribution of possible outputs. The output layer may comprise a softmax or softargmax function to generate the probability distribution of possible outputs. The softmax function may match a number of classes in a training set. The seizure identification system may be configured to optimize the model using a model optimizer such as, but not limited to, RMSprop. The seizure identification system may be configured to analyse each channel of the input electrical signal in each time slot to determine if it is indicative of a seizure occurring or likely to occur. This may comprise decomposing at least the at least one or each channel of at least one or each of the input signals for each time slot into a plurality of different frequency bands of the at least one input signal in order to derive the decomposed input signal. This may further comprise extracting the features from the at least one decomposed input signal for each frequency band. This may further comprise analysing at least the features using the neural network in order to classify the features, e.g. as indicative of a likely seizure or not or indicative of a specific type of seizure or a seizure associated with a particular location in the brain. This may be beneficial in that it can identify onset of a seizure down to an accuracy of the duration of the time slot and can also identify the duration of the seizure at least to the accuracy of the duration of a time slot. The seizure identification system may be configured to take an action upon detecting that a seizure is likely to occurring. The action may be to raise an alarm, alert or flag. The action may comprise recording the seizure in a data log and / or in a physician’s database. The action may comprise communicating the seizure to a stored contact, e.g. via message, email, phone call, or the like. The stored contact may comprise an emergency service and / or at least one designated contact, e.g. family or friends. The seizure identification system may be implemented in software, hardware or any combination of the two. The seizure identification system may comprise at least one processor, and at least the decomposition engine and classifier may be implemented at least in part using the at least one processor. The at least one processor may be configured to at least partially implement at least one or more or each of: the preprocessor and the feature extractor. The seizure identification system may be implemented, at least in part or in full, on a user device, which may be or comprise a wearable device and / or a hand held, carried or otherwise mobile computing device such as a mobile phone or tablet computing device or the like. At least one or more or each of: the decomposition engine, the classifier, the pre-processor and / or the feature extractor may be implemented at least in part by a remote server or other computing system or distributed between the remote server or other computing system and a local device worn or carried by the subject or other user, such as the user device. The EEG sensor may be comprised in or wirelessly or physically connectable to the user device or the other local device. According to a second example of the present disclosure is a method for identifying seizures in a subject, the method comprising: receiving at least one input signal indicative of brain activity of the subject; decomposing at least part of the at least one input signal into one or more frequency sub-bands; and identifying if the at least one input signal is indicative of a seizure based at least on the one or more decomposed signal components using a classifier, which may comprise a model, which may comprise or be comprised in a neural network, which may be a random neural network (RNN) or a convolutional neural network (CNN). The model (e.g. the neural network, may be configured to identify if the at least one input signal is indicative of a seizure. The method may comprise using the seizure identification system of the first aspect and may comprise use of any feature thereof described above. According to a third example of the present disclosure is a computer program configured such that when implemented by a processing system causes the processing system to identify seizures in a subject by: receiving at least one input signal indicative of brain activity of the subject; decomposing at least part of the at least one input signal into one or more frequency sub-bands; and identifying if the at least one input signal is indicative of a seizure based at least on the one or more decomposed signal components using a model configured to identify if the at least one input signal is indicative of a seizure. The model may comprise or be comprised in a neural network, which may be a random neural network (RNN) or a convolutional neural network (CNN). The computer program product may be configured to implement the seizure identification system of the first aspect and may implement any feature thereof described above. According to a fourth example of the present disclosure is a processing device configured to implement the computer program of the third example. According to a fifth example is user device comprising or being connectable to at least one EEG sensor and comprising the processing device of the fourth example or the seizure identification system of the first example. The individual features and / or combinations of features defined above in accordance with any aspect of the present invention or below in relation to any specific embodiment of the invention may be utilised, either separately and individually, alone or in combination with any other defined feature, in any other aspect or embodiment of the invention. Furthermore, the present invention is intended to cover apparatus configured to perform any feature described herein in relation to a method and / or a method of using or producing, using or manufacturing any apparatus feature described herein. Brief Description of the Drawings Various examples of the present disclosure will now be described by way of example, and with reference to the accompanying drawings, of which: Figure 1 is a schematic of a seizure identification system; Figure 2 is an example of EEG input data that can be obtained using the seizure identification system of Figure 1; Figure 3 is a flowchart showing a method of identifying a seizure using a seizure identification system, such as that of Figure 1; Figure 4 is a more detailed overview of the method of Figure 3; Figure 5 is a schematic of an application of a discrete wavelet transform (DWT) that can be used in the methods of Figures 3 and 4; Figure 6 shows EEG signals decomposed into different frequency bands using a DWT; Figure 7 shows the accuracy of the method of Figure 2 for different training I testing ratios and frequency bands; Figure 8 shows the accuracy of the method of Figure 2 for different training I testing ratios and learning rates with a FFT used to decompose the signals; Figure 9 shows the accuracy of the method of Figure 2 for different training / testing ratios and learning rates with a DWT used to decompose the signals; Figure 10 shows an example of a confusion matrix for the method of Figure 2 based on an example data set; Figure 11 shows the accuracy of the method of Figure 2 for different training I testing ratios and learning rates with a FFT used to decompose the signals for a different data set; Figure 12 shows the accuracy of the method of Figure 2 for different training I testing ratios and learning rates with a DWT used to decompose the signals for the different data set; Figure 13 shows an example of a confusion matrix for the method of Figure 2 based on the different data set; Figure 14 shows a classifier for classifying between different types of seizure; and Figure 15 is an illustration of structure of a CNN classifier model. Detailed Description of the Drawings Examples described herein relate generally to seizure identifier systems and methods. In examples, EEG data is analysed to identify seizures. This is conventionally done manually by a highly trained neurophysiologist. However, this is generally too time consuming and expensive to perform monitoring over extended periods for many subjects and it is challenging to continuously provide quick, real time results. To address these challenges, machine learning methods can be used to automate the analysis of EEG data. However, it can be challenging to implement a machine learning model that provides sufficiently accurate results quickly, so that it can be used in real time or near real time, and on-the fly and without over use of computing resources such as processor capacity, memory and bandwidth utilization, and without an unacceptably high level of false detections. That is, in some cases false detections and other errors, can be undesirably high. Some models can take too long to arrive at a result or require too much computing resource (e.g. processor capacity, memory or network usage, battery or power, etc.). This is particularly an issue when fast, real time alerts are required or hen the system is embodied in a portable device or the like. Overcoming these problems is not trivial, particularly given the importance of providing results accurately, quickly enough for on the fly identification, and doing so without excessive use of computing resources (e.g. processor capacity, memory, power / battery, and / or network bandwidth). Furthermore, imbalanced data is a common issue with epilepsy, where the normal class can be many hours long and epileptic events are relatively very short, e.g. a few seconds / minutes. Raw EEG data usually contains artefacts and noise, making it very difficult to identify the necessary patterns or signatures in the signal that are required to effectively classify seizures. Therefore, selection of the most suitable statistical features can contribute to efficient detection and low computational complexity. As such, improvements in systems for seizure identification would be beneficial. Figure 1 shows an example of a seizure identifier system 5 for identifying seizures. The seizure identifier system 5 is carried or worn by a subject 10. The seizure identifier system comprises a biosensor 15, which in examples is in the form of an Electroencephalogram (EEG) sensor for collecting EEG data from the subject 10 but could be another suitable form of brain activity sensor. The seizure identifier system 5 further comprises a processing system 20 and the biosensor 15 is configured to communicate the EEG data to the processing system 20 that analyses the EEG data to determine if the subject is having, or is likely to imminently have, a seizure. The biosensor 15 in this example comprises an EEG sensor having a plurality of electrodes 25. The EEG data obtained from each different combination of electrodes 25 (e.g. combinations of electrodes used as a signal source and a receiver pair) forms a different channel of the EEG signal. However, in other examples, the biosensor 15 may be configured to collect EEG data from a single channel, e.g. a two electrode system. EEG sensors per se are known and any suitable EEG sensor could be used. In the example shown in Figure 1, the EEG data is wirelessly transmitted from the biosensor 15 to the processing system 20, but in other examples, the biosensor 15 may be connected by a wire or data cable to the processing system 20. The processing system 20 comprises at least one processor 30, data storage 35, a communications module 40 and a battery 45 or other energy storage or creation device. In examples, the processing system 20 could be embodied in a suitably configured user device such as a smartphone or tablet computer. The processing system 20 is configured to process the EEG data from the biosensor 15 in order to identify a seizure in the subject, which could be a likely imminent seizure or a seizure in process. The at least one processor 30 performs the processing outlined below in relation to Figures 3 and 4. The data storage 35 is configured to store the EEG data, to store any programs or applications required to perform the processing and to store any models, parameters and anything else required to perform the processing. The communications module 40 is configured to communicate with the biosensor 15 and with a remote computing resource, such as a cloud server (not shown). The battery 45 or other energy storage or generation is configured to power the processing system 20. In examples, the processing system 20 is provided locally to the subject and is in short range communication with the biosensor 15, e.g. via Bluetooth, Bluetooth low energy (BLE), Zigbee or WiFi communication. In examples, the processing system 20 may be a remote computing resource, e.g. a server or cloud computing resource and communications are via a WAN or over the internet. In other examples, the processing involved in identifying the seizure or likely seizure may be distributed over the user device and the remote computing resource. Figure 2 shows the EEG data sent from the biosensor 15. The EEG data includes a plurality of EEG traces, each EEG trace corresponding to a different channel, i.e. to EEG data collected from different combinations of electrodes 25. However, in other examples, the EEG data may comprise a single channel or more or less channels that those shown in Figure 2. Figures 3 and 4 show an overview of the process of identifying seizures or likely seizures from the EEG data. The process could be performed by the at least one processor 30, or by a remote computing resource, or distributed between the two. Beneficially, the processing comprises decomposing the EEG data into frequency bands and applying an artificial intelligence algorithm, in this case preferably a convolutional neural network (CNN) or random neural network (RNN) to process the decomposed EEG to identify the seizures or likely seizures. The EEG data is received at step 305. At step 310, the EEG data is subject to pre-processing. An initial step may comprise standardizing the EEG data if it is not already in a standardized form, for example by limiting the EEG data to defined channels. The pre-processing 310 optionally further comprises applying a filter such as a band-pass filter to the EEG data to remove the high-frequency artefacts from the signal. In an example, the band pass filter passes frequencies in the range of 0.5 Hz to 197.50 Hz but is not limited to this, although in some applications passing frequencies in the range of Hz up to hundreds of Hz may be preferable. For example, the filter could pass frequencies in a range from 0-40Hz in each channel of the EEG data. This filtering passes those frequencies most associated with seizure activity, attenuating irrelevant frequencies, and thereby greatly improving the accuracy of the seizure likelihood determination. The result of the band pass filtering on the EEG data is shown as 405 in Figure 4. The pre-processing 310 comprises splitting the EEG data into contiguous, nonoverlapping time slots. In examples, the EEG data is divided into epochs of 3 seconds each with no overlap, but the disclosure is not limited to this and other durations of time slots could be used depending on the required accuracy and processing resource available. A single time slot for each of 23 EEG channels (i.e. for EEG data collected via 23 different electrode combinations) is shown as 410 in Figure 4 (n.b. only 6 EEG channels 1 to 5 and 23 are shown for clarity in Figure 4, with channels 6 to 22 hidden). In other examples, the time slots could be overlapping. For example, the EEG data can be segmented into time slots, such as partially overlapping time slots. This could be achieved by using a sliding window or other suitable approach. The time slots in some examples have an overlap of one third of the duration of the time slot or less, e.g. 30% of the duration of the time slot, with each preceding and / or following time slot. In some examples, the sliding window can be used to create 10 second time slots with a 3 second overlap with each neighbouring time slot. This segmentation I splitting processes can help ensure capturing that the relevant phases of seizure activities, such as pre-ictal, ictal, and inter-ictal stages, are captured. After the EEG data has been pre-processed in step 310, the EEG data for each channel in each time slot is decomposed using a frequency decomposition to generate decomposed EEG data in step 315. In some examples, this could be carried out by applying a fast Fourier transform (FFT) to each channel in each time slot. However, in certain beneficial examples, this is carried out by applying a discrete wavelet transform (DWT), which can have benefits in greater accuracy, specificity and selectivity over a wider range of measurement parameters. In a specific example of an application of the DWT, each time slot / epoch is decomposed using four levels of decomposition of DWT utilizing Daubechie’s mother wavelet of the 4f" order (DB4) on all channels of the EEG signal. The DWT-based decomposition partitions all channels in each epoch into five frequency sub-bands, namely: Delta Ac-4 (0.0-4.0 Hz), Theta Dc-4 (4.0-8.0 Hz), Alpha Dc-3 (8.0-15.0Hz), Beta Dc-2 (15.0-30.0 Hz), and Gamma Dc-1 (30.0-60.0 Hz). However, the present disclosure is not limited to this and other applications of DWT and / or different frequency bands could be used. Particularly, DWT is used to decompose the EEG signal into multiple frequency bands and provides multi-resolution analysis and time-frequency localization of the signal. An overview of an exemplary DWT process is shown in Figure 5. In this example, a series of high-pass filters 505a-505d and low-pass filters 510a-510d are applied to the time-series EEG signal. The DWT is applied to the respective EEG signal channels in each time slot to extract approximation coefficients (ACs) and detail coefficients (DCs). The ACs can be further divided iteratively using the same process to obtain finer frequency bands if required. The high-frequency and low-frequency components of the signal are obtained by stretching or compressing the mother wavelet. In the first iteration of DWT, the EEG signal is simultaneously filtered using the low pass (LP) filter 510a and High Pass (HP) filter 505a to extract approximation and detail coefficients. The outputs of the low pass (LP) filters 510a-510d and High Pass (HP) filters 505a-505d are denoted in Figure 5 as Ac-1 and Dc-1, respectively. The same process is repeated at each step, the approximation coefficient is further divided into sub-bands. In this specific example, DWT has been used to decompose the EEG signal using Daubechies-4 (DB-4) wavelet function to extract five EEG frequency bands: gamma (30-60 Hz), beta (15-30 Hz), alpha (8-15 Hz), theta (4-8 Hz) and delta (0-4 Hz). The mother wavelet DB-4 is suitable for analyzing the EEG signal in epilepsy diagnosis due to the orthogonal shape of the wavelet that resembles the spike waves in EEG signal. Although this is provided as a working example of how a DWT could be applied, other applications of a DWT could be used. This allows key frequency components critical for seizure prediction to be isolated. The five frequency bands for one particular EEG channel and time slot are shown as 415 in Figure 4 (along with the original raw EEG signal for comparison only). For the avoidance of doubt, the feature extraction is only from the decomposed frequency bands and not from the raw EEG signal, which is only shown for comparison. Example decompositions of the EEG signal into frequency bands are also shown in Figure 6. The left hand side of Figure 6 shows the raw EEG signal indicative of “no seizure” at the top with the five decomposed EEG signals that have been decomposed from this into five different frequency bands below. Similarly, the right hand side of Figure 6 shows the raw EEG signal indicative of “seizure” at the top with the five decomposed EEG signals that have been decomposed from this into five different frequency bands below. Step 320 comprises a feature extraction phase. This comprises analyzing the decomposed EEG data generated by the decomposition in step 315 to extract statistical features from the decomposed EEG data in each frequency sub-band, for each channel in each time slot I epoch. In this example, four statistical features are extracted from each frequency sub-band in each channel in each epoch. The statistical features obtained are standard deviation, mean, kurtosis, and skewness. The statistical features extracts for one frequency band of one EEG channel in one time slot are shown as 420 in Figure 4. Although the four statistical features identified above have been identified as being particularly indicative of seizures, potentially other or different features or combinations of features cold be used. In an alternative example, six statistical features are extracted, namely standard deviation, mean, kurtosis, skewness, minimum, and maximum. This reduces the dimensionality of the data while encapsulating the significant relevant characteristics of the EEG signals. This in turn may reduce the time taken to identify a likely seizure and increase the time available to take action. Furthermore, use of some or all of the above features may help the subsequent neural network arrive at the correct determination by reducing less relevant or irrelevant data that may otherwise guide the neural network to an incorrect result. In step 325, the values for the extracted statistical features is normalized. To ensure uniformity and comparability across the extracted features, min-max scaling is applied to the feature vectors. This normalization process transforms the data into a range between 0 and 1. This may maintain data consistency. In the example shown in Figure 4 in which the EEG signal consists of 23 channels, the processes in steps 305 to 325 result in a feature vector of size 23 (EEG channels) x 5 (frequency sub-bands) x 4 (statistical features). Therefore, the corresponding feature vector (size 460 x 1) for each EEG time slot I epoch is extracted and used as an input to the neural network, e.g. to a convolutional neural network (CNN) or random neural network (RNN) based classification model. However, it will be appreciated that this number would be different for different numbers of channels of EEG data, for different numbers of frequency sub-bands into which each channel is decomposed, or for different numbers of types of features extracted from each frequency sub-band. In step 330, the classification model is applied to the input feature vector obtained after steps 305 to 325. As indicated above, the classification model could comprise a convolutional neural network (CNN) or random neural network (RNN) based classification model. Examples of both CNNs and RNNs will be described. However, variations on the specific implementation described herein could be used and features described in relation to one approach could be used in relation to the other. In the case of the described RNN, the RNN described herein is fundamentally different from memory-based algorithms such as Recurrent Neural Network architecture. Therefore, the RNN does not possess memory or temporal context, meaning that it does not consider or retain any information from previous inputs. In RNNs, each neuron transitions between excitatory or inhibitory states based on the potential of the received signal, which can be either positive or negative. In RNN layers, the neurons probabilistically send and receive excitation and inhibition signals. After receiving a +1 signal, the neuron will enter an excitatory state. Similarly, a neuron will enter an inhibitory state after receiving a -1 signal. A vector containing non-negative integers km (0 represents the potential state of a neuron m at time t. A neuron enters an excited state when km(t) >0, and it is considered idle when km(t) <0. In the excited state, a neuron sends a potential spike signal at a rate of r(m) >0, which reduces its excitation potential by 1. The mathematical expression of the activation function of a RNN fm is as follows. fm — ~ i 'm ' where A+ and are excitatory and inhibitory inputs, respectively: N ^m = XfnrnPtm n=l and N — fnrnPn,m n=l The firing rates of neurons m and n are denoted as rm and rn, respectively. The probabilities of excitatory or inhibitory spikes sent from neuron n to neuron m are represented as p*m or p~m respectively. The activation function fm for neuron m can be expressed as follows: and also In the above equation, the sum of probabilities of all the signals leaving the network must be equal to 1 where the probability of each signal is denoted by d(m), and N refers to the sum of all neurons in the network. The rate at which an excited neuron sends positive or negative signals is as follows: w'rimfn) = >0, n) - rmp'^n >0, The firing rate of neuron m can be derived from the above equations as follows: N Tf; :::: (1 — + The weights used in RNNs, i.e. w+(m,n) and w(m,n) follow the same conventions as those used in traditional artificial neural networks (ANNs) and may be trained using conventional optimization techniques such as gradient descent or others. The RNN is configured to label each time slot / epoch as either 1 or 0 for “seizure” or “no seizure” data, respectively. The RNN comprises an input layer, a hidden layer and an output layer. The first (input) layer is set at the size of the input feature vector and will depend on the number of channels in the EEG signal, the number of frequency bands and the number of features extracted. So, for example, the first (input) layer in the example illustrated in Figure 4 is set at 460 neurons due to the size of the input vector generated at the feature extraction phase. The output layer in this example has only one neuron as the model is performing binary classification between ictal (seizure) and inter-ictal (non-seizure) data, but in other examples that classify the seizure between different types of seizure, the output layer would have a higher number of nodes (e.g. four nodes for classifying between no seizure, generalized seizure, aware focal seizure and impaired awareness focal seizure). The optimal number of neurons for the hidden layer has been examined and it has been found that between 5 and 50 neurons, particularly between 10 and 30 neurons, e.g. 20 neurons is optimal, but could conceivably be different for different applications. The RNN-based model can be trained and tested multiple times using different sets of training and testing ratios and learning rates and the like. The learning rate is a hyper-parameter that determines the size of the steps taken during the optimization process. It plays a crucial role in controlling how quickly or slowly a neural network learns. The concept and use of learning rates is, in itself, known in the art but in some examples described herein can be used by training and testing the RNN-based model using different learning rates. Optionally, k-fold cross-validation can be used, which can improve performance and accuracy. This comprises dividing the dataset into k subsets, with each subset serving as validation data for one model and the remaining k - 1 subsets serving as training data. This is repeated k times, with a different validation subset for each model. So, for instance, a 10-fold cross-validation could be used, resulting in the dataset being divided into 10 subsets. This means that 10 unique RNN models are trained using the following procedure: a first model uses the first subset for validation and is trained on the subsequent 9 subsets. Similarly, a second model uses the 2nd subset for validation and is trained on subsets 1, and 3-9, and so on for all 10 models. Finally, the model with the highest validation accuracy, indicating stronger generalization capabilities, was chosen. Once the best model has been trained and chosen, this is used to process the input feature vector for each time slot I epoch in order to identify seizures (e.g. by labelling each time slot I epoch as being indicative of a seizure or no-seizure, or by classifying the seizure into different types of seizure such as classifying between no seizure, generalized seizure, aware focal seizure and impaired awareness focal seizure). In some examples that use a CNN, the classification model uses a reduced 1-D CNN design that is configured to strike a balance between computational efficiency and predictive performance. However, variations on this can be used, e.g. depending on processing resource available. An example of a suitable CNN classifier model 1500 is shown in Figure 15. The CNN classifier model 1500 can be used in 330 in Figure 3. The classifier model 1500 comprises a plurality of convolutional layers 1510, such as 1-dimensional convolution layers (e.g. convl D layers), configured for initial feature learning from the pre-processed EEG data 1505. In examples, each of the convolutional layers may comprise 64 filters and have a kernel size of 3, but other arrangements, e.g. other numbers of filters, dimensionalities and / or kernels, could be used. The convolutional layers, e.g. the 1D convolutional layers, perform further feature extraction based on the inputs to the model, i.e. based on at least one or each of the features extracted from at least one or each of the one or more frequency sub-bands. The convolutional layers, e.g. the 1D convolutional layers, learn the patterns and relationships, e.g. higher level patterns and relationships, between the input features over time. The filters of the convolutional layers can identify complex, non-linear relationships and temporal patterns in the data that are indicative of seizures. This includes learning from the changes and interactions among the statistical features over time, which has been found to be highly beneficial for predicting seizures. The convolutional layers (e.g. the ConvID layers) in the CNN model perform feature extraction, but they work on a different level compared to the initial decomposition and statistical feature extraction. They build on the pre-processed data to learn deeper, more abstract features that are not directly computable from the raw EEG signals or the initial feature set. This layered approach to feature extraction contributes to the model's high performance. The output of the convolutional players can be optionally and beneficially down sampled. In some examples, the convolutional layers 1510 output to one or more pooling layers 1515 (such as maxpoolingID layers). Each pooling layer 1515 is configured to down sample the input (i.e. the output of the convolutional layers), e.g. by taking a maximum value over a spatial window of a size equal to a pool size. In examples, the pool size is 2, but could be higher or lower. Each dropout layer 1520 is configured to disregard some nodes in a layer at random during training according to a dropout rate. In some examples, the dropout layer has a dropout rate of between 0.1 and 0.5, e.g. a rate of 0.2. The one or more pooling layers and dropout layers allow for dimensionality reduction and mitigate overfitting. The model is optionally configured to flatten or otherwise reduce the dimensionality of the data. In examples, a flatten layer 1525 is provided after the dropout layers 1520. The flatten layer 1525 in this example is configured to transform 2D feature maps into a 1D vector. The flatten layer 1525 outputs to a dense layer 1530, which could have, by way of non-limiting example, 50 neurons. The dense layer 1530 may output to the output layer 1535. In examples, the output layer 1535 is configured to generate a probability distribution of possible outputs. Example implementations of this include the output layer 1535 comprising a softmax or softargmax function to generate the probability distribution of possible outputs. The softmax function may match a number of classes in a training set. The model can be optimized using a model optimizer such as, but not limited to, RMSprop. Beneficially, categorical cross-entropy can be used as the loss function in the optimization. An optimization using accuracy of the model in identifying seizures or likely seizures has been found to be beneficial. Although a CNN comprising a variety of specific layers is described above, not all of these layers are essential, and the CNN cold comprise fewer or alternate layers. For example, at least the flatten layer, the pooling layer, dense layer and / or the drop-out layer are optional. Figures 7 to 13 illustrate the efficacy of the method described above and various implementation options thereof. Figure 7 illustrates the variation in RNN classification accuracy based on training / testing ratios and frequency bands for a standard EEG dataset when the RNN is used with DWT decomposition. The lower frequency sub-bands, such as Delta (0-4 Hz), Theta (4-8 Hz), and Alpha (8-15 Hz), contain more discriminating features for epileptic seizure classification compared to the higher frequency sub-bands. The RNN model yielded more than 70% accuracy on all of the frequency sub-bands. However, the delta band (0-4 Hz) showed the highest accuracy of 83.63%. Table 1 below shows the performance of the RNN model above for each frequency sub-band. Frequency sub-band Accuracy Sensitivity Specificity Gamma - Dc-1 (30-60 Hz) 72.86% 55.73% 90.22% Beta - Dc-2 (15-30 Hz) 73.18% 57.09% 89.46% Alpha - Dc-3 (8-15 Hz) 79.10% 67.21% 91.11% Theta - Dc-4 (4-8 Hz) 81.09% 70.51% 91.82% Delta - Ac-4 (0-4 Hz) 83.63% 79.38% 87.86% Table 1: Performance of RNN-based classification model on frequency sub-bands Figure 8 shows the comparison of RNN classification accuracy when the RNN model is used with FFT decomposition for different learning rates on the X-axis and for different testing to training ratios as different bars on the chart. This data was obtained by using a standard single channel EEG data set (the Bonn dataset, BONN EEG dataset (https: / / repositori.upf.edu / handle / 10230 / 42894)). It can be seen that accuracy increases with increasing learning rate, with a learning rate of 0.001 giving the best accuracy regardless of training to testing ratio. Figure 9 shows the comparison of RNN classification accuracy when the RNN model is used with DWT decomposition for different learning rates on the X-axis and for different testing to training ratios as different bars on the chart. All parameters are otherwise the same as for the data shown in Figure 8. It can be seen that accuracy again increases with increasing learning rate, with a learning rate of 0.0001 and above giving the best accuracy. Figure 10 shows a confusion matrix for the RNN model used with DWT decomposition for the standard single channel EEG dataset (the Bonn dataset). The results were obtained with a learning rate of 0.001 and a training / testing ratio of 70% / 30%. The RNN-based model converges relatively quickly on the simple Bonn dataset, even with a lower learning rate (0.000001-0.001). It can be seen that the RNN+DWT approach achieved an accuracy of 99.30% in classifying seizures on the simpler Bonn dataset while the false positive rate remained below 1%. This represents a good performance. Table 2 below indicates a comparison of the performance of the RNN approaches described herein (both RNN + DWT and RNN + FFT) with other methods or models that could be used. All results are obtained using the same single channel standard EEG dataset (the Bonn dataset). It can be seen that the RNN based methods described herein outperform the other options listed across all metrics (accuracy, sensitivity and specificity). Method Accuracy Sensitivity Specificity ANN+DWT using 5 statistical 98.67% 98.00% 99.00% features SSTFT + GNMF-based LoqE 98.99% 98.53% 99.27% + FKNN LSTM 99.46% 99.62% 99.31% 2D CNN STFT + LSTM 98.20% 98.20% 98.99% Gaussian mixture models 98.43% 98.46% 97.48% fractional Fourier transform 98.76% 98.34% 98.40% CNN+DWT 97.36% 96.17% 96.36% MVAR Sample entropy STFT 99.15% 98.80% 98.45% SVM+DWT 99.30% 99.50% 99.0% RNN+DWT 99.84% 100% 99.80% RNN+FFT 99.77% 100% 99.71% Table 2: Comparison of the RNN model on the Bonn dataset with other methods The same general trends also apply when other, more detailed datasets are used. The CHB-MIT standard EEG dataset comprises multi-channel EEG data and is publically available as the CHB-MIT Scalp EEG Dataset (https: / / phvsionet.Org / contenVchbrnit / 1.0.0 / ). The RNN classification accuracy with different training / testing ratios and learning rates when applied to the CHB-MIT standard dataset is shown in Figures 11 and 12 for a combination of RNN + FFT and RNN + DWT respectively. The RNN-based model coupled with DWT in the feature extraction preprocessing phase demonstrated a gradual increase in accuracy with each increment in the learning ratio, as depicted in Figure 12, while the accuracy for RNN + FFT shown in Figure 11 remained steady for the initial increments in the learning rate. Overall, the use of RNN + DWT shows better accuracy with the more detailed CHB-MIT dataset than RNN + FFT. The confusion matrix for the RNN + DWT approach applied to the CHB-MIT standard dataset is shown in Figure 13. It can be observed that the CHB-MIT dataset, with a training and testing ratio of 30%-70% and a learning rate of 0.05, identifies 91% of seizures accurately while maintaining a low false positive rate of only 9.1%. This represents a very good performance on this highly detailed dataset. Overall, the DWT-based RNN model achieved better performance on both datasets. For comparison, Table 3 below indicates a comparison of the performance of the RNN approaches described herein (both RNN + DWT and RNN + FFT) with other methods or models that could be used. All results are obtained using the CHB-MIT dataset The comparative models are the same as those used to generate Table 2. Method Accuracy SensitivitySpecificity ANN+DWT using 5 statistical features 89.01% 88.39% 89.62% SSTFT + GNMF-based LoqE 88.92% 88.43% 89.22% + FKNN LSTM 87.46% 87.42% 86.11% 2DCNNSTFT + LSTM 89.21% 89.30% 88.69% Gaussian mixture models 86.93% 86.26% 87.58% fractional Fourier transform 89.67% 89.5% 89.75% CNN+DWT 83.27% 74.08% 92.46% MVAR Sample entropy STFT 90.10% 88.60% 91.23% SVM+DWT 90.68% 83.37% 88.03% RNN+DWT 93.27% 90.10% 96.53% RNN+FFT 84.60% 79.17% 89.60% Table 3: Comparison of the RNN model on the CHB-MIT dataset with other methods Table 3 provides a comparison of the results obtained from the above RNN-based classification model on the CHB-MIT dataset. The results indicate that the RNN-based model described above performed with an accuracy of 93.27% as compared to ANN and SVM, which achieved an accuracy of 86.10% and 90.68% respectively. In addition, the presently disclosed RNN-based model converged faster as it was trained for 50 epochs, while the ANN-based model was trained for 200 epochs on the same data. As such, the RNN model described above can provide significantly better accuracy, sensitivity and specificity than the comparative approaches, and was able to be trained and deployed more efficiently. Experiments corresponding to those above but using a CNN instead of the RNN, as described above, have also been performed. In testing using the CHB-MIT dataset, the approach above using the CNN model described above achieved an accuracy of 97.18%, showcasing its high predictive capability. Sensitivity, or the true positive rate, was recorded at 97.18%, indicating the above CNN model's proficiency in correctly identifying seizure events. Additionally, the above CNN model achieved a specificity of 98.59%, illustrating its ability to accurately identify non-seizure events and minimize false positives. The combination of decomposition (preferably wavelet decomposition, such as DWT decomposition) and statistical feature extraction, with the use of a model comprising convolutional layers, such as ConvID layers, allows the model to benefit from both manually engineered features and automatically learned features. This dual approach ensures a comprehensive representation of the EEG signals, capturing both predefined characteristics and allowing the model to uncover additional patterns that improve prediction accuracy. Furthermore, while decomposition techniques, particularly DWT and other wavelet transformations, and / or (preferably in combination with) statistical feature extraction provide a focused and computationally efficient input to the seizure identification model, particularly where that model comprises a CNN or RNN. The convolutional layers, e.g. the ConvID layers, elevate the model's predictive capacity by exploiting the complex temporal dynamics within the EEG signals, which are crucial for accurate seizure prediction and identification. This approach greatly improves the efficiency and performance of the model in seizure prediction and identification. This synergistic approach described above leverages the strengths of the identified signal processing techniques (such as DWT and statistical feature extraction) and deep learning (such as ConvID layers) to create a robust and efficient model for seizure prediction. It showcases an innovative methodology that could enhance the model's suitability for real-time applications and its overall effectiveness in a clinical setting. Any of the identifications or predictions of seizures described above can be used to take set or pre-set actions responsive to seizure identification or prediction. For example, responsive to identifying or predicting a seizure, the processing system 20 may be configured to raise an alarm, alert or flag, e.g. by calling or messaging a set or pre set contact or contacts, such as friends, family or medical assistance. Another example of an action is to log the seizure and any required properties of the seizure such as day, time, duration, type, and / or the like. This can then be reviewed by the subject medical professional to assist in treatment options, for example. Although various examples of implementation of the methods above are descried, variations to these are possible. For example, different numbers of channels of EEG data, different numbers or ranges of frequency bands, different durations of time slots other than the 3 second slots used in the example, and / or less or more statistical features or different statistical features could be used. Furthermore, although specific implementations of the DWT and RNN are described above, alternative implementations may be apparent to a skilled person. In addition, although examples above describe applying (i.e. training or otherwise configuring) the RNN to make a binary decision, i.e. seizure or no-seizure, in other examples, the RNN could be trained or otherwise configured or combined with other classifier models in order to classify the seizures between no-seizure, and different types of seizure such as one or more of: generalized seizures, aware focal seizures or impaired awareness focal seizures, as indicated in Figure 14. As such, the specific examples described above in relation to the Figures are to aid understanding and the scope of protection is defined by the claims.

Claims

1. A seizure identification system for identifying seizures in a subject, the seizure identification system comprising:at least one input for receiving at least one input signal indicative of brain activity of the subject;a decomposition engine configured to decompose at least part of the at least one input signal into one or more frequency sub-bands in order to generate a decomposed signal component for each frequency sub-band; anda classifier, wherein the classifier comprises a neural network configured to identify if the one or more decomposed signal components is indicative of a seizure.

2. The seizure identification system of claim 1, wherein the at least one input signal comprises at least one EEG (Electroencephalogram) signal; andthe seizure identification system comprises at least one EEG sensor system for collecting the at least one EEG signal from the subject; orthe seizure identification system is configured to communicate with at least one EEG sensor system for collecting the at least one EEG signal from the subject and / or communicate with data storage for storing the at least one EEG signal from the subject.

3. The seizure identification system according to claim 1 or claim 2, wherein the neural network is a convolutional neural network.

4. The seizure identification system of any preceding claim, wherein the decomposition engine comprises or is configured to implement a discrete wavelet transform (DWT).

5. The seizure identification system of any of claims 1 to 3, wherein the decomposition engine comprises or is configured to implement a fast Fourier transform (FFT).

6. The seizure identification system of any preceding claim, wherein the decomposition engine is configured to decompose at least part of the at least oneinput signal into a plurality of the non-overlapping, contiguous frequency subbands.

7. The seizure identification system of any preceding claim, wherein the frequency sub bands are sub-bands of a frequency range from 0Hz and 200Hz.

8. The seizure identification system of any preceding claim, comprising or configured to implement a pre-processor, the pre-processor being configured to segment the at least one input signal into time slots, each time slot being associated with a different time range at which the at least one input signal was collected.

9. The seizure identification system of claim 8, wherein the decomposition engine is configured to decompose at least part of at least one or each channel of the at least one input signal in at least one or each time slot into a plurality of frequency sub-bands in order to produce the decomposed electrical signal.

10. The seizure identification system of claim 8 or claim 9, wherein the seizureidentification system is configured to apply a sliding window to segment the at least one input signal into the time slots, wherein the time slots partially overlap at least one other time slot.

11. The seizure identification system of any preceding claim, comprising at least one band pass filter or high pass filter configured to remove frequency components above a threshold or above a first threshold and below a second threshold from the at least one input signal.

12. The seizure identification system of any preceding claim, comprising or configured to implement a feature extractor, the feature extractor being configured to extract at least one feature from each decomposed input signal produced by the decomposition engine.

13. The seizure identification system of claim 12, wherein the at least one feature extracted by the feature extractor comprises at least one of: standarddeviation, mean, skewness, kurtosis minimum and / or maximum of the at least one decomposed input signal.

14. The seizure identification system of claim 12 or claim 13, wherein the at least one feature extracted by the feature extractor is provided to the classifier as inputs and the classifier is configured to identify if the at least one input signal is indicative of a seizure based at least on the features extracted by the feature extractor.

15. The seizure identification system of any preceding claim, wherein the neural network is trained on training data comprising labelled EEG data, the labels indicating at least which portions of the EEG data was collected whilst a subject was undergoing a seizure and / or what type of seizure and / or part of the brain giving rise to the seizure.

16. The seizure identification system of claim 15, wherein the neural network is trained using a portion of the labelled EEG data to train the neural network and a portion of the labelled EEG data to test the training of the neural network, and the neural network is trained on a training to testing ratio that is the ratio of the training data used to train the neural network model to the training data used to test the neural network model of at least 60:40.

17. The seizure identification system of any preceding claim, wherein the neural network model comprises an input layer, an output layer and one or more hidden layers between the input and output layers, wherein the input layer has a dimensionality or number of neurons corresponding to a number of channels or dimensionality of an input vector derived from the at least one input signal; the output layer has a dimensionality or number of neurons of one or more; and the hidden layer has a dimensionality or number of neurons that is less than or the same that of the input layer but more than or the same as that of the output layer.

18. The seizure identification system of claim 17, wherein the hidden layer has a dimensionality or number of neurons that is in a range from 5 to 30.

19. The seizure identification system of any preceding claim, wherein the neural network comprises one or more of:one or more convolutional layers configured to learn features from the at least one input signal;one or more pooling layers configured to down sample the at least one input signal;one or more dropout layers configured to disregard some nodes in a layer at random during training;one or more flattening layers, which may be configured to transform into 1D data; and / orone or more dense layers.

20. The seizure identification system of any preceding claim, wherein the neural network model is configured to output a probability distribution of possible outputs.

21. The seizure identification system of claim 8 or any claim dependent thereon, wherein the seizure identification system is configured to analyse each channel of the input electrical signal in each time slot to determine if it is indicative of a seizure occurring or likely to occur and / or indicative of a specific type of seizure or a seizure associated with a particular location in the brain.

22. The seizure identification system of any preceding claim, configured to take an action responsive to detecting that a seizure is likely to occurring, wherein the action comprises at least one of:raising an alarm, alert or flag;recording at least one of: the seizure, the time of onset of the seizure, the duration of the seizure and / or the type of seizure in a data log and / or in a physician’s database; and / orcommunicating the seizure to a stored contact via message, email, or phone call.

23. A method for identifying seizures in a subject, the method comprising operating a processing system to:receive at least one input signal indicative of brain activity of the subject;decompose at least part of the at least one input signal into one or more frequency sub-bands; andidentify if the at least one input signal is indicative of a seizure based at least on the one or more decomposed signal components using a neural network configured to identify if the one or more decomposed signal components is indicative of a seizure.

24. The method of claim 23, wherein the processing system is comprised in the seizure identification system of any of claims 1 to 22, and the method comprises using the seizure identification system to receive the at least one input signal indicative of brain activity of the subject; decompose the at least part of the at least one input signal into one or more frequency sub-bands; and identify if the at least one input signal is indicative of a seizure based at least on the one or more decomposed signal components using the neural network configured to identify if the one or more decomposed signal components is indicative of a seizure.

25. A computer program configured such that when implemented by a processing system causes the processing system to identify seizures in a subject by:receiving at least one input signal indicative of brain activity of the subject; decomposing at least part of the at least one input signal into one or more frequency sub-bands; andidentifying if the at least one input signal is indicative of a seizure based at least on the one or more decomposed signal components using a neural network configured to identify if the one or more decomposed signal components is indicative of a seizure.

26. A processing device configured to implement the computer program of claim 25.

27. A user device comprising or being connectable to at least one EEG sensor and comprising the processing device of claim 26 or the seizure identification system of any of claims 1 to 22.30

Images