WO2015079264A1

WO2015079264A1 - A system and method for detecting or predicting the onset of a neurological episode

Info

Publication number: WO2015079264A1
Application number: PCT/GB2014/053558
Authority: WO
Inventors: Walid Juffali; Jamil El-Imad
Original assignee: Neuropro Limited; Hoarton, Lloyd
Priority date: 2013-11-29
Filing date: 2014-12-01
Publication date: 2015-06-04
Also published as: GB2522752B; GB201421292D0; GB201321124D0; GB2522752A

Abstract

A system and method of layered pattern matching comprises: detecting or predicting the onset of a neurological episode: receiving a neurological electrical input comprising a digital signal representation of a neurologically derived signal over a recording length L; converting the digital signal into a digital data string of characters; deriving a layer of digital data from the digital signal, the layer comprising: multiple intervals n having an interval length il from the recording length L, where n = Uil; for a layer: identifying recurring patterns and/or anomalies in the digital data string of characters within each interval; and measuring a pattern-derived parameter for intervals in the layer.

Description

Title: A system and method of detecting or predicting the onset of a neurological episode

Field of the invention

This invention relates to a monitoring or predicting system and to a method of monitoring or predicting neurological electrical signals. Background

Major attempts are constantly being made to monitor and predict epileptic seizures. Most predictive methods analyse electrical signals representing neuronal activity in the brain using electrode pickups located at the level of the scalp, externally, under the scalp or deep within the brain.

An electroencephalogram (EEG) is a system for recording electrical activity in the brain produced by the firing of neurons within the brain. Multiple electrodes are placed around the scalp but electrodes can also be placed in direct contact with the brain or within the brain. The EEG signal is composed of different wave patterns operating in a spectrum going from below 4Hz to over 100Hz. There are other mechanisms for detecting and recording neuronal activity such as an electrocorticogram (ECoG) where the signal is derived directly from the cerebral cortex or functional magnetic resonance imaging (FMRI).

Epilepsy is just one example of a potential neurological episode.

WO2012/025765 discloses the following technology which is described with reference to accompanying Figures 1 to 18 and Tables I to VII:

Figure 1 shows a first set of potential electro positions for use with a device or system embodying the present invention; Figure 2 shows another potential set of electro positions for use with a device or system embodying the present invention; Figure 3 is a graph showing a relationship between seizure risk and neuronal activity signal anomalies;

Figure 4 is a block diagram of a monitoring or predicting device not embodying the present invention;

Figure 5 is a schematic block diagram showing a system not embodying the present invention, for gathering and analysing neuronal activity signals;

Figure 6 is a table of anomaly results achieved using the system of figure 5;

Figure 7 shows an example of the pattern count during pre-itcal and ictal periods, the ictal period being shaded;

Figure 8 is a graph showing an example of the pattern count varying over time;

Figure 9 shows a graphical user interface designed to analyse neuronal activity data signals; Figure 10 is another block diagram representation of a monitoring or predicting device not embodying the present invention.

Table I shows a 10 nibble pattern matrix; Table II shows an example of a first part of a 9 nibble pattern matrix; Table III shows an example of a first part of an 8 nibble pattern matrix;

Table IV shows an example of a first part of a 7 nibble pattern matrix; Table V shows an example of a first part of a 6 nibble pattern matrix;

Table VI shows some initial results distinguishing pre-ictal from ictal periods;

Table VII shows historical results related to pattern count changes when analysing full data sets;

Figure 1 1 is a snap-shot of over 12000 electronic readings taken from a single patient, the bold trace reflecting normal state and the fainter trace

representing readings taken from the same patient during seizure;

Figure 12 is a detail of the electronic readings from figure 1 1 running from electronic readings 4000 to 5200;

Figure 13 is a detail of the electronic readings from figure 1 1 running from electronic readings 4800 to 5040;

Figure 14 shows a case list of available data for different patients;

Figure 15 shows raw data and process data for case 7;

Figure 16 gives a pattern count for 10 nibble patterns identified during seizure conditions - run 13 and a summary of the anomaly percentage; and

Figure 17 gives the same information as the pattern count shown in figure 16 but for run 14 taken from the same patient during a normal state; and Figure 18 shows a second graphical user interface designed to analyse neuronal activity data signals.

The neurological signal monitoring or predicting device disclosed in WO2012/025765 does not embody the present invention. The device receives input from one or more sensors which are suitable for receiving signals indicative of neuronal activity from the brain.

Electrodes or electrical contacts are the preferable form of sensors to detect neuronal activity from the brain, i.e. neuronal activity sensors. For the sake of convenience, this specification refers to neuronal activity sensors as electrodes or electrical contacts but non-electrical sensors to detect or derive neuronal activity are possible alternatives or equivalents to electrical sensors. As well as being used as inputs, the electrical contacts may also be configured as outputs to provide neuronal stimulation to a part or parts of the brain.

There are conventions for positioning and fixing of EEG electrodes (see figures 1 and 2) so that aspect will not be discussed further here.

The electrical signals from the brain comprise rhythmic patterns and anomalies. By anomalies, we are referring to electrical signals which are random in nature and do not conform to rhythmic signal patterns. As the proportion of anomalies to rhythmic patterns in the electrical signal increases, then the likelihood of a neurological episode such as an epileptic fit also increases. This relationship is shown graphically in figure 3.

Specific identification of individual anomalies, such as signatures, is not necessary to provide useful information to predict or monitor the likelihood of a neurological event such as an epileptic seizure. Some specific anomalies are, however, indicators of the onset of a neurological event. As well as detecting patterns using signal processing techniques and creating pattern ratios to identify a potential threshold between patterns indicating a normal state and using those patterns to distinguish from a seizure state, one can also use more observational techniques to distinguish between the different classes of signal pattern.

The electrical signals received from EEGs are received as floating point data. The floating point data is then digitised and weighted in accordance with predetermined characteristics which can be pre-set or controlled by a user. Figure 1 1 shows such a weighted graph derived from the floating point data. In figure 1 1 the electronic readings are taken at a rate of 256 per second. The bolder line in figure 1 1 represents a floating point data which has been digitised and weighted, taken from the patient when in a normal state. The fainter trace represents electronic readings taken from the same patient pre- seizure and during seizure. Exactly the same scaling and weighting has been applied to the processed floating point data. It is clear from figure 1 1 that there is an almost rhythmic nature to the electronic reading when in the normal state. When in the seizure state, the electronic reading is clearly more erratic. An observation can be made looking at this data that the rhythmic electronic readings are characteristic of a normal state and the almost pseudorandom electronic readings are characteristic of a seizure state. These characterisations can be used through electronic processing/signal processing to determine a likelihood of the patient being in the normal state or in the seizure state. Usefully, when the electronic reading characteristics decay from the almost rhythmic pattern, observance of this decay can be used as a trigger to provide an alert that the patient is moving from a normal state towards a seizure state. WO2012/025765 discloses a number of different measures to make threshold decisions and some of those measures are discussed below. This disclosure bases decisions on pattern-derived parameters which may involve thresholding or reacting to a profile of a particular pattern-derived parameter. Thus, if a pattern derived parameter exceeds or falls below a predetermined or learned threshold, then a decision can be taken in response to that and an indicator given. Similarly, pattern-derived parameters can be profiled so when a parameter follows a particular trend such as decaying, then a decision can be taken in response to that and an indicator given.

A pattern-derived parameter is a parameter derived from an observation of or operation on a digital data string which gives information about one or more patterns that recur in the digital data string. Examples of pattern-derived parameters are: the number of patterns identified in a data run; the proportion of patterns of a certain length compared to the total data payload; and combinations of these and including profiles or signatures of pattern-derived parameters such as monitoring the rate of change of a particular pattern- derived parameter.

The thresholds or profiles of pattern-derived parameters can be learnt by the monitoring or predicting system and varied according to individual characteristics of the user being monitored. Monitor learning uses known heuristics, neural network and artificial intelligence techniques.

A basic signal gathering and analysing system takes a neuronal activity signal either in digital form or converts it from analog to digital and then presents the signal as a character string. The character string may be in binary, hexadecimal or other base. The character string is preferably of the characters 0...9; A...F making up the hexadecimal character set. What is important is that the characters can provide a pattern of characters. A sliding window of predetermined bit length or nibble length is placed over the data string and the data characters sitting within the window are considered to be a pattern. The pattern and the number of further occurrences of that pattern are logged as the window is slid over the entire data string. The window may be stepped incrementally through the data string bit by bit, in steps of multiple bits or potentially even pseudo-randomly. In basic terms, the system counts the number of occurrences of each pattern and creates various parameters or characterisations of the data based on pattern count. Variations in pattern count have been shown to provide an indication of whether or not the brain is in a pre-ictal or ictal period. The system also includes an output giving an indication of the onset of an ictal state based on the parameters derived from or characterisation of the pattern count.

The most basic monitoring or predicting system makes use of this relationship between pattern count and changes in neuronal activity to provide a monitoring or predicting system to provide a warning to a user based on an analysis which determines whether there has been a change in pattern count indicative of a change in neuronal activity indicative of onset of a seizure or the like. The analysis is based on internally stored historical ratios of pattern counts or can be processed by the monitoring or predicting device on the fly and compared with predetermined thresholds given the different parameters for the incoming data and the user.

The output of the monitoring or predicting system can be a wired output, a wireless output, a Bluetooth™ output, an optical output, an audio output or any other mechanism of alerting a user or reporting to a user. A particularly preferred method is the use of a traffic light indicator giving an alert status continually. The status of the indicator goes from green where there is no indication of onset of an ictal period, through amber where there is a potential risk of onset of an ictal period; to red where an ictal period is indicated as being imminent or ongoing. The monitoring or predicting device is configured as a piece of electronic hardware with input connections to one or more neuronal activity sensors such as EEG electrodes which form part of a skull cap or an array of electrodes positioned on and attached to the skull. The device is preferably located on headgear or attached to the skull so that the path or distance from the or each sensor to the monitoring or predicting device is as short as possible. The device preferably has an internal power source but can be connected to an external power source. The monitoring or predicting device 1 as shown in figure 4 comprises a number of modules defined by their functionality. In various examples, the modules are: either all held in a common housing of the monitoring or predicting device; or some modules are remote from the skull or body-located monitoring or predicting device and connected thereto by a wired or wireless connection.

There are four basic modules making up the monitoring or predicting device 1 : a signal sourcing module 2 which receives input signals representing neuronal activity from sensors; a pre-processing module 3 which takes a sampled signal and creates a data string; a pattern search module 4 which analyses the data string and shows repeated patterns; and a pattern monitor module 5 which analyses the patterns and generates a monitor and/or predictor output in dependence on the analysed patterns. Figure 14 shows a case list of available historical EEG data taken from patients in various conditions, usually either normal or abnormal, abnormal indicating pre-ictal or ictal state.

In the device of figure 4, a neural stimulator is provided to furnish electrical or other stimuli to a part or parts of the brain. The stimuli are preferably furnished in response to the monitor and/or predictor output of the device. Figure 15 shows the raw data and the process data for case 7. The raw data comprises the original floating point data from the EEG before it has been digitised and weighted. The process data shows the hexadecimal characters representing the digitised and weighted data from which patterns can be derived.

Figure 16 shows the patterns identified in case 7 in run 13 for which the data was captured during seizure. The size of the file is 40732 bits and for a 10 nibble pattern 4156 patterns were identified leaving 36576 anomalies giving an anomaly density or ratio of 89.8%.

Figure 17 shows the results for run 14 of case 7 which is data captured when the same patient in case 7 was in a normal state. Again, the file size is 40732 bits but the number of patterns identified is 39090 leaving only 1642 anomalies, giving an anomaly density or ratio of 4.03%. This conveys an immediate distinction between the pattern/anomaly density or ratio allowing immediate characterisation of the data signals as being either captured during a normal state or during a seizure state. The percentage of anomalies present during a seizure state is vastly greater than the percentage of anomalies present during a normal state. A threshold can be determined or even learned by the monitoring or predicting device which can constantly monitor, for example 10 second readings in real time and make a judgement on whether the pattern ratio or pattern threshold has been decayed or passed and provide an alert or prediction in response to monitoring of this pattern- derived parameter. Conventional pattern analysis and pattern derivation mechanisms can be used to derive, identify, count and monitor patterns.

Figure 5 shows the modules 2,3,4 and 5 of a monitoring or predicting device 1 as part of a larger and more detailed network which includes the facility to stream live data or run stored data through the modules. Referring to figure 4, the signal sourcing module 2 has an amplifier 100 or pre-amplifier to receive neuronal activity input signals (an analog signal) preferably from EEG electrodes. Downstream of the amplifier there are one or more analog to digital converters 105 (or a multiplexed analog to digital converter) operating at a sampling frequency fs and having as their input the respective amplified EEG signals from the electrodes 10.

The sampled output of the analog to digital converters 1 10 is a binary string which is preferably converted to hexadecimal by HEX converter 1 15. The use of hexadecimal is particularly helpful to gain a visible and direct appreciation of the presence of patterns in the signal being monitored.

An analog-to-digital converter is used with typical sampling frequencies (fs) of 128-512Hz for EEG and ECoG to 10-30KHz for single neuron and local field potential (LFP) signals. The conversion, depending upon the application can result in 8- 6bit data. When stored for software (and microcontroller hardware) this information is represented at its lowest level in binary, but in a higher level of abstraction in hexadecimal (HEX). Hence, the data is already available in an alphanumeric format.

The hexagonal output is fed to the pattern search module 4 which is configured in this example as an n-gram model. Additionally, we can adjust the level of lossiness of the data representation by dividing the data (an N bit number) by 2^D where D is an integer, to result in a reduced data format - i.e. reducing a 16bit number to an 8bit one by dividing down with D = 8. The n-gram process in the pattern search module 4 extracts any patterns in the signals. Once patterns are extracted the number of significant patterns are counted. A significant pattern is a pattern that has occurred more than 2 times but other threshold limits can be selected and may be usefully varied for different pattern sizes. The greater the pattern size, i.e. string length, the less repeating patterns there will be.

The pattern count is monitored and when the pattern count drops below a historically derived threshold stored in the pattern monitor, the pattern monitor outputs a change of status. A significant pattern count is quantified in two ways: (1) to count out of the number of significant patterns the total number of occurrences of all these patterns and (2) out of the patterns found what percentage were significant. The former is shown in the below results, the latter method quantified similar results so is not shown here. These pattern counts can then be quantified as a ratio between a current window of analysis and a previous window during an inter-ictal state (ictal refers to the state during a seizure).

The hexadecimal output is sampled and patterns identified and counted.

In figure 6, there are four sets of results 6A, 6B, 6C and 6D. The "NC" columns are data taken in the time prior to a neurological event (pre-ictal). The "ANC2" columns are data taken during a seizure onset and during the event (ictal) - see also the timing diagram at the foot of the table in figure 6.

6A gives the raw results. 6B recognises that certain patterns occur very frequently particularly those patterns representing a saturated signal for a null signal which in hexadecimal terms would equate to "00" or "FF". These patterns are therefore excluded from the list of patterns. 6C removes all repeat patterns from the list of patterns. A repeat pattern is a sub-set of a pattern which occurred in a larger pattern size pattern list. The other figures give similar pattern matrices for 9, 8, 7 and 6 nibbles taken for the same data string. Tables I to V show the first page of patterns and frequency of occurrence for the five pattern sizes of 10 to 6 nibbles. Preferably the data is sampled as 6, 7, 8, 9 or 10 nibbles from a sliding window applied to the hexadecimal data output string and the occurrence of each individual distinct nibble pattern is logged. In the 10 nibble pattern matrix shown in Table I, the two most popular occurring 10 nibble patterns in the "NC" data acquisition period are 020100FFFE and 20100FFFEF which patterns both occur 5 times in the "NC" data acquisition period. Many other 10 nibble patterns occur during the "NC" period.

The signal sourcing module receives input signals S1 -S7 representing neuronal activity from one or more EEG sensors 10 (see figures 1 and 2) attached to the skull in a conventional manner (of both attachment and/or array). The input signals in this example are electrical signals S1 -S7 direct from EEG sensors 10. In other examples, the input signals may be remotely streamed from a live feed or a recorded data set. In the preferred example, the number of repeated patterns in NC is compared to the number of repeated patterns in ANC2. There are usually more repeated patterns in NC rather than ANC2 during the actual seizure. As a consequence, there are less patterns identifiable during a seizure, meaning that there are also more anomalies occurring during seizure hence an increase in the proportion of anomalies to repeated patterns is an indicator or predictor of the onset of a neurological event such as an epileptic seizure.

A relative increase in the number of repeated patterns is a direct indicator of the onset of a neurological episode and is useful information to allow the device to perform an episode prediction function. The likelihood of the onset of a neurological episode increases as the number of repeated patterns increases.

Aspects of the disclosure deal with one of the bottlenecks of analysis of epileptic seizure activity. An aspect of the disclosure allows the ability to work with consistent data which is well annotated and databased to establish a framework for future work and storage of results into the same framework. The system shown at figure 5 provides this framework. Figure 5 shows data acquisition, user interface and processing blocks. In theory each of these components could be placed in a different technological implementation, such as the acquisition being an implantable neural monitoring or predicting device, the user interface being on a mobile phone or PC and the processing units being a web-accessed cloud (such as the Amazon Elastic Compute Cloud). The distribution of these elements will vary depending upon the signal processing requirements (computational complexity) and application space.

Pattern analysis of historical data yields sets of parameters concerning the patterns. Predictions or decisions on whether a neural event is upcoming can be taken by comparing in either relative or absolute terms real-time patterns with stored parameters, pre-determined patterns and thresholds. The monitoring or predicting device provides an output indicative of whether a neural event is unlikely, likely or imminent, much like a traffic light output: red, amber and green.

The electrodes or electrical contacts that are used to detect neuronal activity from the brain are the inputs to the monitoring or predicting device. These inputs may be reversed to provide a stimulus output. The disclosure also includes the provision of neuronal stimulation to a part or parts of the brain. The stimulus may be provided in response to any of the parameters measured or monitored by the monitoring or predicting device, such as a change in pattern count indicative of a change in neuronal activity, an onset, a seizure or the like. Thus, the traffic light output from the device can be used to trigger a neural stimulation, perhaps targeted stimulation, in an attempt to ameliorate, offset, delay or avoid entirely a neurological episode such as an epileptic seizure. The neural stimulation is provided by a neural stimuli generator 21 which can be a part of the device or connectable to a device output, potentially wirelessly, or wired.

Referring to figure 6, this data was obtained from the University Hospital of Freiburg Epilepsy Centre, Germany. The data used were pre-sampled at 256Hz and quantised using a 128 channel 16-bit data acquisition. All 21 patients' first seizure was used from this data set. There is a pre-ictal period of up to 1 hour in most cases with seizure durations varying from 15-170 seconds. There were multiple types of seizures present including: simple partial, complex and general tonic-clonic.

Conventional pattern analysis techniques were used. The disclosure does not relate to the analysis technique but in the identification that patterns and pattern ratios of the digitised and sampled data are characteristic of a patient being in normal, pre-ictal and seizures states.

Two sets of tests were conducted with this data. The first set aimed to quantify whether there were pattern differences between seizure/ictal areas when compared against inter and pre-ictal periods. To do this, sample pre- ictal periods equal in size to the seizure period were extracted. The average pattern count between each of 10 sections of pre-ictal data is computed in order to compare against the seizure pattern count. For this analysis we used D = 8 and analysed the data for n-gram sizes of 12 and 14, where one token was one electrical reading in 1 byte or 2HEX characters. These results (shown in Table VI) show that in most cases the pattern count (P) compared to the seizure count (S) was considerably different; 18 out of 21 cases showed a ratio that indicated a greater than 25% change in pattern count. These results aim to distinguish pre-ictal from ictal periods.

The second set of tests analysis used the full pre-ictal period, segmented into 5 and 10 second windows. We analysed the data using D = 8 and with n- gram sizes 10 and 14. A typical pattern count found in these data sets is shown in figure 7. The results for all 21 patients are shown in Table VII. Table VII shows heuristic results (related to pattern count changes) when analysing the full data sets. The descriptions refer to the changes that occur that are visible changes compared to the pre-ictal period.

It is clear patterns exist and interestingly some patterns change or exist for certain n-gram sizes but not for others. An interesting pattern was an increase followed by a slow decrease over several minutes - see sharp increase at 60 minutes in figure 8. These findings conclude that out of the 21 patients, 18 can be detected using the features outlined in figure 6. Table VII shows patterns at different n-gram sizes. Patterns are identical using different n-gram sizes while making sure that patterns in an n-gram of 12 are not replicated in an n-gram size of 10, i.e. making sure that patterns are unique and not a subset of a larger pattern in the data. Also, one needs to identify data parameters (e.g. types of seizure) to identify if certain patterns correlate with patient-specific information.

To gather further historical data and develop pattern parameters and thresholds there is disclosed a modular analysis framework as shown in figure 5. The system of figure 5 is an open online and/or real-time analysis tool (www.winam.net) with an SQL database that is used to examine multiple cases and runs of data sets and presents in a webpage form as shown I figure 9. As in figure 5 the structure is such that the data sourcing can be through an RSS feed, or offline data source. The processing (n-gram) is implemented through a separate processing cluster allowing multiple parallel processing efficiently. This is an open system that allows users to freely access and carry out the algorithmic techniques described in this paper. The database structure itself is designed to allow users to input multiple patient cases, and for each case run particular data sets (EEG, ECoG, ECG etc..) or parts of these data sets.

Figure 18 shows a second graphical user interface designed to analyse neuronal activity data signals. This interface is used to set the parameters, in real-time or offline, of the logics module 4 and/or the downstream analysis module.

In Figure 18, a number of parameters may be monitored and/or adjusted and applied to the data in real time (and not just post-analysis). The listed parameters are not exclusive and other parameters or sub-parameters can also be modified.

The "weighting" parameter refers to the rounding of the EEG signal. For example, the EEG signal may be sampled at 5Hz, and then the number of samples may be divided by 128 to remove noise - effectively "zooming in" on the results. The signal then needs to be rounded as a 5Hz sample frequency divided by 128 will not give an integer number of samples. In a preferred example, the signal is rounded up.

The "interval" parameter is the window length the user wishes to process cycles in. In one example, the interval length may be 1 minute. The "frequency" parameter is a frame frequency and relates to the number of frames to be processed. If, for example, the user has 1 hour of data and the 1 1 th to 20th minutes are of interest, the user can skip to the 1 1 th minute and select how many frames will be read - e.g. 20 frames at a 30s interval length.

The "optimiser" parameter determines the optimum pattern length to determine a suitable anomaly ratio. For example, with a pattern length of 2 bits it is likely there will be very few or no anomalies, but with a pattern length of 20 bits, it is likely there will be several anomalies. The optimiser effectively sets a benchmark for the anomaly ratio. In a preferred example, the optimiser parameter determines the pattern length of each pattern type A, B, C, D, (see, for example, Figure 17, where the length of pattern A is 10 nibbles, and patterns B,C and D are 0) such that anomaly ratio is less than 10%. The optimiser can automatically determine the ideal settings for each of the pattern groups (shown as GNBP(A-D) on the user interface). The pattern group settings may be overridden manually by adjusting, respectively, GNBP(A-D) individually or jointly.

The "SD" parameter control relates to the threshold standard deviation of the results.

Other importing functionality is readily implemented as is reporting documentation to further visualise, analyse the results and further develop pattern-based parameters on which to base neuronal event monitoring or predicting decisions.

Neural Binary Pattern Storage Table Each pattern group from(A, B, C or D) has a length defined by Patter n_n_length (where n = a, b, cor d) will have a storage table where'unique' patterns are stored. Each pattern stored has an associated count and the last offset (from the start of frame).

Pattern A - (validation A>B>C>D)

Pattern_a_name: xxxxxxxx (Key)

Pattern_a_count: 32,768 (i6bit/2byte)

Pattern a last offset: 32,768 (i6bit/2byte)

Sample:

Pattern B

Pattern_b_name: xxxxxxxx (Key)

Pattern_b_count: 32,768 (i6bit/2byte)

Pattern b last offset: 32,768 (i6bit/2byte)

Pattern C

Pattern_c_name: XXXXXX(Key)

Pattern_C_count: 32,768 (i6bit/2byte)

Pattern c last offset: 32,768 (i6bit/2byte)

Pattern D

Pattern_d_name: XXXXXX(Key)

Pattern_d_count: 32,768 (i6bit/2byte)

Pattern d last offset: 32,768 (i6bit/2byte)

NBP Working Storage

CMBPWS!

Pattern_a_length: .■H. f.X 2-24 byte patterns

Pattern_b_length: 2-24 byte patterns

Pattern_c_length: 2-24 byte patterns Pattern_D_length: 2-12 byte patterns

Each data value is 1 byte in length (8bits). Pattern_N_length determines the number of 1 byte patterns that constitute a pattern - example: Pattern_A_length = 3 refers to 'XX XX XX', such as '10 7F 33'.

Offset: 32,768 (i6bit / 2byte), initial value -1

LLA: 32,768 (i6bit / 2byte), initial value 0 BP Process

Open Input File -> Apply weight factor (1

to 1 or 2 byte (2 Character Hex)-

2. Read Input File -> Using user inputs: Skip to. Frame Length Frame Fr

Gap

3. Write unformatted data to input_storage_table

4. Open Exclude List File (Using user input Exclude List):

Write to Exclude List Storage Table (ELST) 1. Start logic

Offest = Offset+1

If LLA = 0 skip to Update Pattern Table Routine Compare LLA with Offset range A-D

Overlap? Y: Start Logic

N: Continue 2. Update Pattern Table Routine

IF Pattern_d_Length > 0 then

Compare offset + Pattern_d_Length of input_storage_table to Pattern_d_name table in BPST. Duplicate key? Yes: Set duplicate_flag to d, continue

No: Continue

IF Pattern_c_Length > 0 then

Compare offset + Pattern_c_Length of input_storage_table to Pattern c name table in N BPST. Duplicate key? Yes: Set duplicate_flag to c, continue

No: Continue IF Pattern_b_Length > o then

Compare offset + Pattern_b_Length of input_storage_table to Pattern b name table in N BPST. Duplicate key? Yes: Set duplicate_flag to b, continue

No: Continue

IF Pattern_a_Length > o then

Compare offset + Pattern_a_Length of input_storage_table to Pattern_a_name table in N BPST.

Duplicate key? Yes: Set duplicate_flag to a, continue

No: Continue Duplicate_flag = ON? (will equal a, b, c or d) a. NO

Move offset to pattern_a_last_offest

Move 1 to pattern_a_count

Move offset + Pattern_a_length of input_storgae_table to Pattern_a_name

If Pattern_name in ELST Return to Startjogic

Else Write table entry into N B PST

Move offset to pattern_b_last_offest

Move 1 to pattern_b_count

Move offset + Pattern_b_length of input_storgae_table to Pattern_b_name

If Pattern_name in ELST Return to Startjogic

Else Write table entry into N BPST Move offset to pattern_c_last_offest

Move 1 to pattern_c_count

Move offset + Pattern_c_length of input_storgae_table to Pattern c name If Pattern_name in ELST Return to Startjogic

Else Write table entry into N BPST

Move offset to pattern_d_last_offest

Move 1 to pattern_d_count

Move offset + Pattern_d_length of input_storgae_table to Pattern_d_name

If Pattern_name in ELST Return to Startjogic

Else Write table entry into N BPST

Return to Start Logic Loop Till EOF

YES

Move offset to pattern_n_last_offest (n = a, b, c or d - found from

Duplicate_flag)

Move offset to LLA

Add 1 to pattern_count

Re-Write table entry

If Pattern_count not > ELST count Return to Startjogic

Else Update correlation bit map table

Return to Start Logic Loop Till EOF

Additional background

WALID JUFFALI et.al: "The WiNAM project: Neural data analysis with applications to epilepsy" BIOMEDICAL CIRCUITS AND SYSTEMS

CONFERENCE (BIOCAS), 2010 IEEE, IEEE 3 November 2010, pages 45-48, (XP031899695, D01 : 10.1 109/BIOCAS, 2010, 5709567, ISBN: 978-1 -4244- 7269-7) discloses a novel algorithmic method based on an ngram approach and applies it to ECoG and deep brain neural data for analysis of epileptic seizures.

US 6,658,287 discloses a method, and system for predicting the onset of a seizure prior to electrograph onset in an individual.

During an "off-line" mode, signals representing brain activity of an individual (either stored or real time) are collected, and features are extracted from those signals. A subset of features, which comprise a feature vector, are selected by a predetermined process to most efficiently predict (and detect) a seizure in that individual. An intelligent prediction subsystem is also trained "off-line" based on the feature vector derived from those signals.

During "on-line" operation, features are continuously extracted from real time brain activity signals to form a feature vector, and the feature vector is continuously analyzed with the intelligent prediction subsystem to predict seizure onset in a patient.

US 2006/111644 discloses methods and systems for patient-specific seizure onset detection.

In one embodiment, at least one EEG waveform of the patient is recorded, and at least one epoch (sample) of the waveform is extracted. The waveform sample is decomposed into one or more subband signals via a wavelet decomposition of the waveform sample, and one or more feature vectors are computed based on the subband signals. A seizure onset can then be identified based on classification of the feature vectors to a seizure or a non- seizure class by comparing the feature vectors with a decision measure previously computed for that patient. The decision measure can be derived based on reference seizure and non-seizure EEG waveforms of the patient.

US 2005/197590A discloses a system (10) which analyzes signals

representative of a subject's brain activity in a signal processor (12) for information indicating the subject's current activity state and for predicting a change in the activity state.

One preferred embodiment uses a combination of nonlinear filtering methods to perform real-time analysis of the electro-encephalogram (EEG) or electro- corticogram (ECoG) signals from a subject patient for information indicative of or predictive of a seizure, and to complete the needed analysis at least before clinical seizure onset. The preferred system then performs an output task for prevention or abatement of the seizure, or for recording pertinent data.

WO 2010/115939 discloses a method for the real-time identification of seizures in an Electroencephalogram (EEG) signal.

The method provides for patient-independent seizure identification by use of a multi-patient trained generic Support Vector Machine (SVM) classifier. The SVM classifier is operates on a large feature vector combining features from a wide variety of signal processing and analysis techniques. The method operates sufficiently accurately to be suitable for use in a clinical environment. The method may also be combined with additional classifiers, such a

Gaussian Mixture Model (GMM) classifier, for improved robustness, and one or more dynamic classifiers such as an SVM using sequential kernels for improved temporal analysis of the EEG signal. PCT/GB2013/051096 published as WO/2013/160706 discloses applying optimising parameters to the system and method of WO2012/025765. PCT/GB2013/051096 discloses the following technology which is described with reference to accompanying Figures 19 to 29:

FIGURES 19-24 show 6 case studies of WO2012/025765; and FIGURES 25-29 show five hours of a surface EEG of a patient.

This disclosure relates to the optimising parameters of WO2012/025765, outlined above. In WO2012/025765, the optimising parameter "determines the optimum pattern length to determine a suitable anomaly ratio".

In PCT/GB2013/051096, the optimising parameters incorporate at least one of several settings such as: the pattern length for each pattern type, weightings (involving filtering and scaling as discussed below), frequency, intervals, SD and any thresholds. These settings are shown in figure 18 and described above. The optimising parameters may further incorporate any other variable disclosed herein.

In WO2012/025765, the optimising parameter was determined from the pre- ictal and ictal data only. The optimising parameter was not calculated from "normal" i.e. inter-ictal data, data that is taken at a time distant from any ictal event. In other words, the optimising parameter was previously calibrated by the pre-ictal and/or ictal data only.

The applicant has noted a surprising technical effect in that if the inter-ictal data is optimised with at least one optimising parameter, and the optimising parameter(s) for the inter-ictal data are applied to the pre-ictal and/or ictal data, the onset of a neurological episode is more readily identified. The optimiser parameter may be applied to the raw digital data directly, or to the digital data string, which comprises patterns which recur with a distribution, and anomalies. The optimised inter-ictal data may be patient specific and applied to the pre-ictal and/or ictal data for the same patient, or alternatively be applied to pre-ictal and/or ictal data for another patient.

PCT/GB2013/051096 also relates to calculating a statistical spread such as a standard deviation from the digital data string. The applicant has also identified that calculating the statistical spread or volatility (standard deviation, variance, SEM and confidence intervals, including combinations of the same, with which the person skilled in the art would be familiar) of the distribution or recurrence of patterns or anomalies in the digital data string (or of any other collected data) provides a further method for identifying the onset of a neurological episode.

The standard deviation may be determined for an individual ("Z-Score") and thus be individual-specific, or be calculated for a group (or sub-group with preset criteria) as a whole. Other statistical measures such as skewness and kurtosis may be considered and applied to the data during analysis. The calculation of a statistical measure of spread such as standard deviation has not previously been used as a method for identifying the onset of a neurological episode.

Preferably, the statistical spread is a standard deviation or variance is calculated from:

the distribution of the recurring patterns in the digital data string;

a number or percentage of anomalies in the digital data string; or an average of the number or percentage of anomalies from two or more time intervals; or

a rolling average of the number or percentage of anomalies from two or more time intervals. The prediction logic outlined below details a method of calculating a rolling standard deviation of the percentage of anomalies using a 7-interval average. In other embodiments, any other number of intervals may be used. The interval average is used in the calculation to provide a sample window.

In some embodiments of PCT/GB2013/051096, the standard deviation is calculated from an interval of a number or percentage of anomalies in the digital data string. In other embodiments, the standard deviation is calculated from the distribution of the recurrence of patterns in the digital data string.

Furthermore, the applicant has recognised that if standard deviation data for an inter-ictal period is optimised and the optimising parameters are applied to standard deviation data from the pre-ictal and/or ictal periods, the onset of a neurological episode is further more readily identified.

The standard deviation data may be optimised by applying at least one optimising parameter to the digital data to provide optimised digital data and calculating a standard deviation of the optimised digital data, or applying at least one optimising parameter to the standard deviation data directly.

It is logical to apply optimising parameters from a pre-ictal state to an ictal state since both data windows relate to the ictal event. It is not obvious to apply optimising parameters from inter-ictal/"normal" data to ictal data, since one would not expect that optimising parameters for a normal state would also aid in identifying an abnormal, ictal state. Nevertheless, the applicant has undertaken this and has identified a surprising resultant technical effect.

In a preferred embodiment of PCT/GB2013/051096, the optimising parameters are determined from the percentage of anomalies of the inter-ictal data. In one embodiment, the optimising parameters are selected to provide an anomaly ratio of less than 10%. In another preferred embodiment, the optimising parameters are derived from a standard deviation of the inter-ictal data. In other embodiments, the optimising parameters are determined from other functions or aspects of the inter-ictal or "normal" data. Figures 19-24 show 6 case studies of PCT/GB2013/051096. Each figure represents an individual case study and shows:

(a) INTER-ICTAL PERIOD: the anomalies (as a percentage) against time;

(b) INTER-ICTAL PERIOD: the standard deviation of the percentage of anomalies against time;

(c) ICTAL PERIOD: the anomalies (as a percentage) against time, with the optimiser parameters for the inter-ictal period applied to the ictal period; and

(d) ICTAL PERIOD: the standard deviation of the percentage of anomalies against time, with the optimiser parameters for the inter-ictal period applied to the ictal period.

Figures 19-24 clearly show the surprising technical effect. Part (c) of each figure clearly shows the dramatic change in percentage anomalies at the onset of a neurological event when the optimising parameters for the inter- ictal period are applied to the ictal period.

Similarly, part (d) of each figure clearly shows the dramatic change in standard deviation at the onset of a neurological event when the optimising parameters for the inter-ictal period are applied to the ictal period.

In particular, figures 23(c) and (d) show that PCT/GB2013/051096 can predict the onset of a neurological episode. These figures show substantial peaks of the percentage of anomalies and the standard deviation of the percentage of anomalies, respectively, well ahead of the actual onset of the neurological episode and therefore either/both measurements can be used as predictors. As shown in figures 19-24, PCT/GB2013/051096 has the clear technical effect of making the identification and prediction of the onset of a neurological episode significantly easier.

Figures 25 to 29 show five hours of surface EEG of a patient that had a secondarily generalized tonic-clonic seizure after an unusually long focal phase of 5 min. The data consists of five 1 -hour windows of EEG of channel T4-T6. The sampling frequency is 256 Hz. The time at the beginning of each window respectively is 20:32:00, 21 :32:00, 22:32:00, 23:32:00, and 00:32:00 the next day. The seizure onset is at 23:23:30, the change to a generalized seizure at 23:28:30, and at the end at 23:31 :00.

Accordingly, the first and second windows (figs 25 and 26) are pre-ictal, the third window (fig 27) is ictal, showing the seizure at the end, and the fourth and fifth windows (figs 28 and 29) are post-ictal. These figures clearly show the surprising technical effect.

Each figure represents an hour window and shows:

(a) the percentage of anomalies against time;

(b) the standard deviation of the percentage of anomalies against time; and

(c) the EEG amplitude.

The EEG amplitude is between +-50 μν most of the time, except at seizures. The filter used is 0.3 Hz (there is no change in amplitude if the filter is off). The results of figures 25-29 clearly show the algorithm first predicting the onset of a neurological episode at around 22:00, again at 22:30 and 22:42, and reconciling with the onset at 23:28. Both the percentage of anomalies and the standard deviation of the percentage of anomalies are clear predictors of the onset of a neurological episode. Figure 25 shows the percentage of anomalies and standard deviation of the percentage of anomalies as normal. Figure 26 shows a dramatic fluctuation (approximately by a factor of 4) in the percentage and standard deviation of anomalies at around 22:00, indicating the onset of a neurological episode. Figure 27 shows further dramatic fluctuation (approximately by a factor of 5) in the number and standard deviation of the percentage of anomalies in the first 10 minutes of the window, i.e. around 22:30, and a seizure occurs in the last 10 minutes of the window at 22:28, as indicated by the very large fluctuation in both the percentage and standard deviation of anomalies.

Figures 28 and 29 show the percentage and standard deviation of the percentage of anomalies returning to normal.

For the avoidance of any doubt, any calculations using the count or proportion of recurring patterns in the digital data string may of course instead use the count or proportion of anomalies in the digital data string since the digital data string comprises recurring patterns and anomalies.

The principles outlined above are not only applicable to predicting the onset of neurological episodes from EEG data, but also to predicting other events such as earthquakes from seismic data, heart attacks from ECG data or events from EMG data.

The present invention relates to the following aspects, individually and in any combination: One aspect of the invention provides a method of detecting or predicting the onset of a neurological episode comprising:

i) receiving a neurological electrical input comprising a digital signal representation of a neurologically derived signal over a recording length L;

ii) converting the digital signal into a digital data string of characters;

iii) deriving a layer of digital data from the digital signal, the layer comprising: multiple intervals n having an interval length il from the recording length L, where n = L/il;

iv) for a layer: identifying recurring patterns and/or anomalies in the digital data string of characters within each interval; and

v) measuring a pattern-derived parameter for intervals in the layer. Preferably, further comprising analysing patterns and/or anomalies from at least two different layers.

Advantageously, further comprising deriving multiple layers and selecting for analysis one of the layers.

Conveniently, further comprising applying different weights to respective layers.

Preferably, further comprising providing a weighted output from analysis of multiple layers.

Advantageously, wherein the interval start points are offset from one another.

Conveniently, the method further provides an output giving an indication of the onset or occasion of a neuronal activity. Preferably, the converting step ii) is performed after layering step iii).

Conveniently, the converting step ii) is performed before layering step iii). Advantageously, the pattern-derived parameter relates to one or more of: a count of said recurring patterns or anomalies in the digital data string; a proportion of said recurring patterns in the digital data string; a rate of change of a count of said recurring patterns or anomalies in the digital data string; a rate of change of a proportion of said recurring patterns in the digital data; and a statistical spread of the distribution of the recurring patterns in the digital data string.

Another aspect of the present invention provides a monitoring or predicting system to detect or predict the onset of a neurological episode, the system comprising: a neurological electrical input, the input being a digital signal representation of a neurologically derived signal over a recording length L; a converter to convert the digital signal into a digital data string of characters; a layer generator to derive a layer of digital data from the digital signal, the layer comprising: multiple intervals n having an interval length il from the recording length L, where n = L/il; a pattern analyser to identify, for a layer, recurring patterns and/or anomalies in the digital data string of characters within each interval; and a monitor to measure a pattern-derived parameter for intervals in the layer. Preferably, an output from the monitor gives an indication of the onset or occasion of a neuronal activity.

Conveniently, the pattern analyser is operable to identify patterns and/or anomalies from at least two different layers. Advantageously, there are multiple layers and the pattern analyser is operable to select for analysis one of the layers.

Conveniently, a weighting mechanism is operable to apply different weights to respective layers.

Preferably, the weighting mechanism is operable to provide a weighted output from analysis of multiple layers. Conveniently, the system is operable to offset interval start points from one another.

In order that the present invention may be readily understood, embodiments thereof will now be described, by way of example, with reference to the accompanying drawings, in which:

Figure 30 is a block diagram of an embodiment of the present invention;

Figure 31 is an example of different predetermined interval lengths being applied to a recording of EEG data embodying the present invention;

Figure 32 is an example of the anomaly ratio outputs for each of the intervals based on the same recording length; Figure 33 is an example of the use of staggered intervals embodying the present invention; and

Figure 34 is an example of the use of staggered intervals of different interval lengths embodying the present invention. The monitoring system 1 shown in figure 30 contains similar modules to the monitoring or predicting device of figure 4. A signal sourcing module 2 receives input signals representing neuronal activity from sensors attached to a patient. The input to the signal sourcing module 2 can also be recorded EEG data, for example from a data library or remotely recorded data.

A pre-processing module 3 takes a sampled signal and creates a data string. A pattern search module 4' analyses the data string over a predetermined interval length and shows repeated patterns and anomalies in the patterns. A pattern monitor module 5' analyses the patterns and anomalies for each given interval and generates a monitor and/or predictor output independence on the analysed patterns. In a preferred embodiment, the monitor and/or predictor output is a percentage ratio of the number of anomalies divided by the total number of patterns identified.

The preceding work has based the monitoring and predicting analysis on the use of this ratio as a threshold level. When the ratio exceeds a threshold level, then this is an indicator of neuronal activity. The analysis is conducted of the patterns during a fixed interval length. The previous work has involved varying the interval length and optimising the results based on different interval lengths depending on the particular patient being monitored. The interval length can be tuned to a particular patients patterns. Anomalous patterns show themselves differently depending on the interval length. The prior work methodology can lead to data being misinterpreted. For example, if an epileptic seizure begins just before the end of an interval and finishes just after the start of the next interval, then the seizure will have bridged the two intervals and will not necessarily show through as significantly anomalous patterns in either of the adjacent intervals. Another example of data being misinterpreted is an electrode being jogged perhaps by a sudden movement or the like resulting in an increase in the number of anomalous patterns which register above the threshold and indicate an epileptic seizure. This data gives a false impression that an epileptic seizure has occurred.

Anomalous patterns arising from a seizure (abnormal neuronal activity) should be evident to the monitoring and/or predicting system irrespective of the interval length.

The pattern search module 4¹ in embodiments of the present invention set one or more predetermined interval lengths and interval start points. The interval start points may be offset from one another.

The pattern search module 4' calculates the anomaly ratio for each predetermined interval lengths and interval start points. The pattern search module 4' is operable to provide the anomaly ratio output for each predetermined interval either serially or in parallel. Turning to figure 31 which shows a selection of potential intervals in the form of "layers", the recording length of the sample data L is divided into sixteen equally sized intervals in the first layer. The same data is processed in the second layer into eight equal sized intervals, the third layer processes the same data into four equally sized intervals, the fourth layer processes the data into two equally sized intervals; and the final layer has the entire recording length L as the interval length - i.e. just one interval. Referring now to figure 32, this is an example of the values for the anomaly ratio output for each intervals. Along the timeline there are a series of incidents A, B, C and D. A is a verified seizure; B is a verified seizure; C is an electrode being dislodged momentarily due to a sudden impact; and D is a short burst or anomalies not related to a seizure. In A, the seizure occurs in the transition between the first and the second interval in the first "layer". As such, the anomaly ratio in these adjacent intervals does not cross the threshold level previously indicated as critical. Thus, the real seizure is not picked up in either of these two adjacent intervals since its contribution to the anomaly ratio is dilated across the two intervals. In contrast, the seizure shows strongly in the first interval of the second layer where the seizure occurs substantially mid-interval. The increase in the anomaly ratio caused by the seizure shows through in the second layer interval.

In example B, another real seizure, this occurs again at the transition between fourth and fifth intervals, between second and third intervals in the second layer and between first and second intervals in the third layer. However, at the fourth layer, the seizure B occurs substantially mid-interval and shows through.

Event C is a very short term disruption of the active electrode which causes spurious signals to be generated. There is a spike of spurious signals at event C but nothing previous to event C is abnormal and nothing after event C is abnormal. There is a high proportion of anomalies in this smallest seventh interval of the first layer but as soon as the same time period is analysed I n the second layer, the effects of the spike are smoothed out over the longer interval lengths and the spurious data spike does not show through in the second or subsequent layers.

In figures 31 and 32 the intervals all start at the beginning of the recordal length of data but the interval start times can be staggered or offset from the beginning of the recording length as shown in figures 33 and 34. In figure 33, the interval length for all the intervals are the same but the intervals are offset from one another in each of the respective layers. In figure 34, the intervals are of two different lengths and some of them are offset. Within each layer, the lengths of the intervals are preferably the same but it is possible to have intervals of different lengths within the same layer. Any combination of interval lengths and/or offsets can be used.

As shown in figure 32, the relevant layers can be assigned different weights by the weighting module 100.

Different weights can be applied to respective layers, heuristicly or using particular algorithms to assign respective weights to respective layers.

The weighted output is a weighted anomaly ratio output which a pattern monitor can use for comparison with a threshold level to indicate a warning of the onset of a neurological offset.

Layered pattern matching techniques include: i) Direct layering o same baseline (same recording)

o different intervals (1 , 2, 4, 8 min sizes) ii) Derived layering o anomaly-based

Amplitude logic

o a seizure is an increase in frequency & amplitude

o anomaly ratio: expression of f & a

o can identify seizures using frequency, amplitude and patterning

methods and combinations thereof

When used in this specification and claims, the terms "comprises" and "comprising" and variations thereof mean that the specified features, steps or integers are included. The terms are not to be interpreted to exclude the presence of other features, steps or components.

The features disclosed in the foregoing description, or the following claims, or the accompanying drawings, expressed in their specific forms or in terms of a means for performing the disclosed function, or a method or process for attaining the disclosed result, as appropriate, may, separately, or in any combination of such features, be utilised for realising the invention in diverse forms thereof. Protection may be sought for any features disclosed in the earlier application WO2012/025765 in combination with the above disclosure. Seizure Prediction Logic

Background;

Variables; Value ilv Interval I annomaiies va! ue base 100 7.00 i2v interval 2 anriorrsaiies va! ue base 100 9.00 i3v interval 3 annomaiies val ue base 100 4.00 s4v interval 4 annomaiies va! ue base 100 14.00 i5v interval 5 annomaiies va! ue base 100 12.00 i6v interval 6 annomaiies va! ue base 100 6.00 s7v interval 7 annomaiies va! ue base 100 15.00 ira The roiiing a erage of the last 7 interval anomaiies 9.57143

Logic; Sample Vaiue

Step 1 (,'. ; · - ira)² = (la (7 - 9.57143)² = ila

{flv — ira)⁷ = i2 (9 - 9.57143)² = i2a

(13 v - ira)² — i3a (4 - 9.57143)² = i3a

4v - ira)² - iAa (14 - 9.57143)² = iAa

5 v — ira)² - i.5a (12 - 9.57143)² = iSa

(i6v — ira)² — i6a (6 - 9.57143)² = i6«

7v — ira)² = i7a (15 - 9.57143)² = 17

Step 2

i i a ÷ i 2 + i 3 a ÷ ϊ4α 4- i 5 ÷ i ha + 17 a

Step 3

x = Roiiing Standard Deviation jiow=good, high=bad)

Step 4 Output

Step 5 Function = Predict Preferred Features

1 . A method of gathering data to detect or predict the onset of a neurological episode comprising:

receiving a neurological electrical input comprising a digital representation of a neurologically derived signal; and

selecting at least one optimising parameter derived from analysis of data collected during an inter-ictal period.

2. A method of detecting or predicting the onset of a neurological episode comprising:

receiving a neurological electrical input comprising a digital representation of a neurologically derived signal;

converting the digital signal into a digital data string;

providing at least one optimising parameter derived from digital data collected during an inter-ictal period;

applying the at least one optimising parameter to the digital data to provide optimised digital data;

identifying recurring patterns and/or anomalies in the optimised digital data;

measuring a pattern-derived parameter; and

providing an output giving an indication of the onset or occasion of a neuronal activity in dependence on the pattern-derived parameter. 3. The method of any preceding clause, wherein the pattern-derived parameter relates to one or more of:

a count of said recurring patterns or anomalies in the digital data string;

a proportion of said recurring patterns in the digital data string;

a rate of change of a count of said recurring patterns or anomalies in the digital data string; a rate of change of a proportion of said recurring patterns in the digital data; and

a statistical spread of the distribution of the recurring patterns in the digital data string.

4. The method of any preceding clause, further comprising:

monitoring the rate of change of the pattern-derived parameter; and/or counting significant recurring patterns; and/or

excluding patterns in the data string that are identified as null signals from the significant recurring pattern count; and/or

detecting or identifying the type of neurological episode.

5. A method of detecting or predicting the onset of a neurological episode comprising:

converting the digital signal into a digital data string, the digital data string comprising patterns which recur with a distribution, and anomalies; calculating a statistical spread from the digital data string; and providing an output giving an indication of the onset or occasion of a neuronal activity in dependence on the statistical spread of the digital data string.

6. The method of clause 5, wherein the statistical spread is a standard deviation or variance calculated from:

the distribution of the recurring patterns in the digital data string;

a number or percentage of anomalies in the digital data string;

an average of the number or percentage of anomalies from two or more time intervals; or

a rolling average of the number or percentage of anomalies from two or more time intervals. 7. The method of any of clauses 5 to 6, further comprising:

providing at least one optimising parameter derived from the digital data collected during an inter-ictal period; and

applying the at least one optimising parameter to the digital data to provide optimised digital data,

wherein calculating a statistical spread of the digital data comprises calculating an optimised statistical spread of the optimised digital data; and wherein the output depends on the optimised statistical spread.

8. The method of any of clauses 5 to 7, further comprising:

providing at least one optimising parameter derived from the digital data collected during an inter-ictal period;

applying the at least one optimising parameter to the statistical spread to provide optimised statistical spread,

wherein the output depends on the optimised statistical spread.

9. The method of any preceding clause, wherein the output giving an indication of the onset or occasion of a neuronal activity is determined based on either:

analysing internally stored historical data or analysing data in real-time; and/or

comparing with a predetermined threshold. 10. The method of clause 9, wherein the predetermined threshold is learned from the user profile using heuristics, neural networks and/or artificial intelligence, or determined by the total number of significant patterns, a statistical spread of the distribution of significant patterns and/or a percentage of significant patterns found.

1 1 . The method of any of clauses 2 to 10, further comprising: stimulating a part of a brain using a neural stimuli generator.

12. The method of any of clauses 1 to 4, 7 to 1 1 , wherein at least one of the optimising parameters is user-determined, heuristically determined or predetermined.

13. The method of any of preceding clause, wherein the digital data string is a character data string, a binary data string or a hexadecimal data string. 14. The method of any of clauses 1 to 4, 7 to 13, wherein the optimising parameter comprises one or more of: a pattern length for a pattern type, a weighting, a frequency, an interval and a threshold.

15. A monitoring or predicting system to detect or predict the onset of a neurological episode, the system comprising:

a neurological electrical input, the input being a digital representation of a neurologically derived signal;

a converter to convert the digital signal into a digital data string;

an optimiser to apply at least one optimising parameter derived from analysis of data collected during an inter-ictal period to the digital data to provide optimised digital data;

a pattern analyser to identify recurring patterns in the optimised digital data; and

a monitor to measure a pattern-derived parameter, wherein an output from the monitor gives an indication of the onset or occasion of a neuronal activity in dependence on the pattern-derived parameter.

16. The system of clause 15, wherein the pattern-derived parameter relates to one or more of:

a count of said recurring patterns or anomalies in the digital data string; a proportion of said recurring patterns in the digital data string;

a rate of change of a count of said recurring patterns or anomalies in the digital data string;

a rate of change of a proportion of said recurring patterns in the digital data; and

17. A monitoring or predicting system to detect or predict the onset of a neurological episode, the system comprising:

a converter to convert the digital signal into a digital data string, the digital data string comprising patterns which recur with a distribution, and anomalies;

a pattern analyser to identify the recurring patterns in the digital data string;

a calculator to calculate a statistical spread from the digital data string; and

an output giving an indication of the onset or occasion of a neuronal activity in dependence on the statistical spread of the digital data string.

18. The system of clause 17, wherein the statistical spread is a standard deviation or variance calculated from:

the distribution of the recurring patterns in the digital data string;

a rolling average of the number or percentage of anomalies from two or more time intervals. 19. The system of any of clauses 17 to 18, further comprising:

an optimiser to apply at least one optimising parameter derived from analysis of data collected during an inter-ictal period to the digital data to provide optimised digital data, wherein the calculator calculates the optimised statistical spread of the optimised digital data and the output depends on the optimised statistical spread; or

an optimiser to apply at least one optimising parameter derived from analysis of data collected during an inter-ictal period to the calculated statistical spread to provide an optimised statistical spread, wherein the output depends on the optimised statistical spread.

20. The system of any of clauses 15 to 19, further comprising:

a neural stimuli generator for stimulating a part of a brain.

Claims

CLAIMS:

1 . A method of detecting or predicting the onset of a neurological episode comprising:

ii) converting the digital signal into a digital data string of characters;

iv) for a layer: identifying recurring patterns and/or anomalies in the digital data string of characters within each interval;

v) measuring a pattern-derived parameter for intervals in the layer.

2. The method of claim 1 further comprising analysing patterns and/or anomalies from at least two different layers.

3. The method of any preceding claim, further comprising deriving multiple layers and selecting for analysis one of the layers.

4. The method of any preceding claim, further comprising applying different weights to respective layers.

5. The method of claim 4, further comprising providing a weighted output from analysis of multiple layers.

6. The method of any preceding claim, wherein the interval start points are offset from one another.

7. The method of any preceding claim, wherein the method further provides an output giving an indication of the onset or occasion of a neuronal activity.

8. The method of any preceding claim, wherein converting step ii) is performed after layering step iii).

9. The method of any one of claims 1 to 7, wherein converting step ii) is performed before layering step iii).

10. The method of any preceding claim, wherein the pattern-derived parameter relates to one or more of:

a count of said recurring patterns or anomalies in the digital data string;

a proportion of said recurring patterns in the digital data string;

1 1 . A monitoring or predicting system to detect or predict the onset of a neurological episode, the system comprising:

a neurological electrical input, the input being a digital signal representation of a neurologically derived signal over a recording length L; a converter to convert the digital signal into a digital data string of characters;

a layer generator to derive a layer of digital data from the digital signal, the layer comprising: multiple intervals n having an interval length il from the recording length L, where n = L/il; a pattern analyser to identify, for a layer, recurring patterns and/or anomalies in the digital data string of characters within each interval; and

a monitor to measure a pattern-derived parameter for intervals in the layer.

12. The system of claim 1 1 , wherein an output from the monitor gives an indication of the onset or occasion of a neuronal activity.

13. The system of claim 1 1 or 12, wherein the pattern analyser is operable to identify patterns and/or anomalies from at least two different layers.

14. The system of any one of claims 1 1 to 12, wherein there are multiple layers and the pattern analyser is operable to select for analysis one of the layers.

15. The system of any one of claims 1 1 to 14, wherein a weighting mechanism is operable to apply different weights to respective layers.

16. The system of claim 15, wherein the weighting mechanism is operable to provide a weighted output from analysis of multiple layers.

17. The system of any one of claims 1 1 to 16, wherein the system is operable to offset interval start points from one another.