DESC:FIELD OF THE INVENTION:
This invention relates to the field of biomedical engineering.

Particularly, this invention relates to the field of signal processing in the realm of biomedical engineering and, additionally, relates to epileptic seizure detection.

Specifically, this invention relates to a system for complex partial epileptic seizure detection.

BACKGROUND OF THE INVENTION:
Epilepsy is the most common neurological illness impacting 50 million people globally. Epilepsy is exhibited by recurrent seizures in the brain due to its association to central nervous system. Electroencephalogram (EEG) signals are used to explore the functioning of the brain during an epileptic seizure.

Epileptic persons are three times more prone to die early as compared to a non-epileptic human being. In addition, complex partial epileptic seizures generally start and end instinctively without any exterior intrusion. Also, they may remain unobserved. Hence, analysis and detection of complex partial epileptic seizure EEG signals has always been of great interest to researchers.

Epileptic EEG signals obtained have multiple components such as time, frequency information, and different ictal data in addition to the epileptic spike behaviour. These characteristics are difficult to analyse and it makes hard to detect the epileptic zone present in the signal. In order to efficiently interpret seizure information, segment-wise analysis is proposed in prior art literature.

Clinically, visual interpretation of the complex partial epileptic seizure EEG information is performed by brain experts. But the task of examining these long-term seizure signals to find the proper temporal information of the brain is very exhausting and time consuming. Thus, there is a need for advancement of automated epileptic seizure detection systems assists neuro-experts to correctly identify the EEG epileptic region in the brain.

There is a need to provide a seizure identification system in order to attain high performance in terms of a high true positive rate and a false positive rate. In order to obtain this, it is important to further increase the overall seizure diagnosis rate. This would help neurologists in coming up with an effective diagnosis and further treatment of epileptic persons.

OBJECTS OF THE INVENTION:
An object of the invention is to provide an epileptic seizure detection system; specifically, a complex partial seizure in a patient.

Another object of the invention is to provide automatic detection and localization of origination of complex partial epileptic seizure in Electroencephalogram (EEG) signals.

Another object of the invention is to provide an automated epileptic seizure detection system which assists neuro-experts to correctly identify an EEG epileptic region in a brain.

Yet another object of the invention is to provide a seizure identification system in order to attain high performance in terms of a high true positive rate and a false positive rate.

An additional object of the invention is to provide a system for automatically detecting and localizing the origination of the complex partial epileptic seizures in a brain region.

SUMMARY OF THE INVENTION:
According to this invention, there is provided a system for complex partial epileptic seizure detection.

In at least an embodiment, the system of the present invention performs the following steps:
Acquisition of EEG signal using 21- channel electrodes from a patient.
The signal acquired, channel wise, is supplied to a computing system.
Detection and localization of the origination of high frequency complex partial epileptic seizure zone by eliminating the undesired low frequency physiological disturbances.
Computation of Critical Spectral Verge (CSV) as an optimized feature derived using bio-inspired Flower Pollination Algorithm (FPA) method in combination with spike statistical features for each EEG frequency band, segment wise.
Classifying the extracted feature into complex partial epileptic seizure and non- seizure segments, channel-wise.
Mapping of the classified signals in the brain region using topographic maps.

According to this invention, there is provided a system for complex partial epileptic seizure detection, said system comprises:
EEG signal input module consisting of multi-channel electrodes configured to acquire an EEG signal from a patient;
segmenting module configured to segment said EEG signal into smaller segments to identify a seizure region in said acquired EEG signal;
a signal pre-processor configured to pre-process said EEG signal input module’s (101) recorded multi-channel EEG signal, to provide a pre-processed signal, using Multiresolution based Adaptive Filtering (MRAF) for removing low frequency physiological artefacts, said a pre-processed signal preserving seizure information;
a feature extractor configured such that, from said pre-processed signal, Critical Spectral Verge (CSV), is computed, along with spike and statistical related features being extracted from seizure preserved information in order to form extracted feature set;
a combinative based feature selector configured to pass said extracted feature set, using a feature extractor, in order to select a best feature set from said extracted feature set;
a Support Vector Machine (SVM) classifier configured to receive said best feature set for classifying multichannel-wise segments in order to obtain a determined output of said signal being a ‘complex partial seizure’ signal or a ‘non-seizure’ signal;
a segment-wise processor configured to detect complex partial epileptic seizure signals in a segment-wise manner.

In at least an embodiment, said segment-wise processor is configured to detect complex partial epileptic seizure signals, said signals being characterised with High Frequency Oscillations (HFOs), based on extracted Critical Spectral Verge (CSV) feature in a channel-wise manner and in a segment-wise manner, said segment-wise processor, being configured with instructions, configured to perform the steps of:
processing one or more segments, of said acquired EEG signal;
processing one or more bands, of said acquired EEG signal;
checking, if all segments, of said EEG signal, are processed;
checking, if all bands, of said EEG signal, are processed;
calculating Power Spectral Density (PSD)of all processed segments and all processed bands;
computing average of calculated Power Spectral Density (PSD);
applying Flower Pollination Algorithm (FPA)to optimise and compute Critical Spectral Verge (CSV) from said calculated Power Spectral Density (PSD); and
adding values of each calculate Critical Spectral Verge (CSV) to a feature map.

In at least an embodiment, said feature extractor is configured with instructions, configured to perform the steps of:
computing Power Spectral Density (PSD) values for each frequency band (b), for each segment (k), of each of said EEG signals;
computing average of EEG Power Spectral Density (PSD) values for each frequency band (b), for each segment (k), of each of said EEG signals;
computing, segment-wise, a Spectral Verge (SV) value, which is a maximal frequency point at which the Power Spectral Density (PSD) value of that segment is greater than the average Power Spectral Density (PSD) values of the corresponding frequency band;
seeding the computed Spectral Verge (SV) value, for a segment, as a seed solution value;
optimizing, using a Flower Pollination Algorithm, by comparing each of said seed solution values, with a probability switch (P), to decide between local pollination optimisation and global pollination optimisation basis random path length, thereby, obtaining a Critical Spectral Verge (CSV) value, for each of said segments; and
providing each of said Critical Spectral Verge (CSV) values, as a best feature set per segment, to said Support Vector Machine (SVM) classifier.

In at least an embodiment, said Support Vector Machine (SVM) classifier configured to receive said best feature set for classifying multichannel-wise segments in order to obtain a determined output of said signal being a ‘complex partial seizure’ signal or a ‘non-seizure’ signal, is configured with instructions, configured to perform the steps of:
forming a cluster of Power Spectral Density (PSD) values for each segment of each channel of said EEG signal;
computing, segment-wise, a Spectral Verge (SV) value, which is a maximal frequency point at which the Power Spectral Density (PSD) value of that segment is greater than the average Power Spectral Density (PSD) values of the corresponding frequency band;
determining said computed Spectral Verge (SV) value as a reference value;
searching for High Frequency Oscillations (HFOs) component in each of said formed clusters;
checking, for each of said searched High Frequency Oscillation (HFO) component with said reference value;
determining a seizure signal if said searched High Frequency Oscillation (HFO) component matches with said reference value;
jumping to a next cluster, using a Flower Pollination Algorithm, if said searched High Frequency Oscillation (HFO) component does not match with said reference value; and
said jump being defined by Levy’s path, said Levy’s path being determined by a probability switch in order to determine quantum of said jump.

In at least an embodiment, each signal consists of 21-channel EEG readings.

In at least an embodiment, said Multiresolution based Adaptive Filtering (MRAF) filtered EEG segment is characterised by high frequencies.

In at least an embodiment, said Critical Spectral Verge (CSV) is a point in a higher frequency range at which Power Spectral Density (PSD) is greater than average Power Spectral Density (PSD).

BRIEF DESCRIPTION OF THE ACCOMPANYING DRAWINGS:
The invention will now be described in relation to the accompanying drawings, in which:
Figure 1 depicts a block diagram of the proposed system for the complex partial epileptic seizure detection and identification of the origination of the seizure according to one of the embodiments of the present invention;

Figure 2 depicts a flowchart illustrating a method for optimized feature map extraction for the accurate automatic detection and identification of the origination of the complex partial epileptic seizure according to one of the embodiments of this present invention;

Figure 3 depicts a topographic map illustrating the detection of the origination of the complex partial epileptic seizure activity in the brain; and

Figure 4 depicts a heat map is a representation of correlation values exiting between different features

DETAILED DESCRIPTION OF THE ACCOMPANYING DRAWINGS:
The following description is provided to facilitate an understanding to the embodiments and the accompanying drawings which form a part hereof, in which the invention can be practiced. These embodiments are described in sufficient detail to enable the practice of the invention. The following detailed description is therefore not to be taken in a limiting sense, and the scope of the present invention is defined by the included claims.

According to this invention, there is provided a system for complex partial epileptic seizure detection.

Clinically, visual interpretation of a complex partial epileptic seizure from an EEG is performed by neurologists. But, the task of examining these long-term EEGs to find the proper temporal information of the brain is very exhausting and time consuming. The present disclosure is therefore directed to the automatic complex partial epileptic seizure detection and localization of the origination of seizures which assists neurologists to correctly identify the epileptic region.

Furthermore, it is difficult to locate the exact epileptic seizure region in long duration EEG recordings. Therefore, segmenting a signal into small sections helps to identify a seizure region properly in an EEG signal. The segment-wise variation in the spectral power helps to track the normal EEG segments from the abnormal seizure segments particularly the complex partial epileptic seizure. Hence, this specification utilizes spectral changes to trace this abnormal seizure information. In order to identify this behaviour, segment-wise multi- channel EEG analysis help in localizing the exact seizure origin of the brain, which is missing in the surrogate channel related analysis. In this present invention, segment wise multi-channel classification is performed using optimized feature extracted using bio-inspired Flower Pollination Algorithm (FPA); this is proposed for the detection of the origination of complex partial epileptic seizure. Uniform distribution of spectral power is observed for high frequency range for seizure segment whereas for non- seizure segment high spectral power is observed for low frequency range. So, in order to represent this information properly, the average of Power Spectral Density (PSD)is computed to differentiate between seizure and non-seizure segments. This PSD for each segment is used as a threshold to identify seizure and non-seizure areas. However, it has been observed that some segments were misclassified, since to calculate average PSD, frequency information is not considered. To capture frequency information, we have to consider a point in the higher frequency range at which PSD is greater than the average PSD. This point is refereed as critical spectral verge (CSV). It is observed that the average PSD of some non-seizure segments is greater than PSD, which led to the misclassification of those EEG segments. In order to overcome this misclassification result, local and global abnormalities are defined. As local abnormalities are considered as abnormalities introduced due to seizure information and global are undesired region. FPA is a powerful technique which can iterate through local and global abnormalities effectively and optimize the solution to correctly identify the seizure segments using pollination algorithm. Hence, CSV point is further optimized using FPA in order to extract the actual seizure activity for every segment, and every channel. This point is given to Radial Basis Function (RBF) based SVM classifier to classify the segment as seizure or non-seizure. The proposed system and methodology performs automated complex partial epileptic seizure detection and segment-wise multi-channel classification to find the exact location of the occurrence of the epileptic seizure in the brain region.

Figure 1 depicts a block diagram of the proposed system for the complex partial epileptic seizure detection and identification of the origination of the seizure according to one of the embodiments of the present invention.

In at least an embodiment, the invention provides a system (100) for complex partial epileptic seizure detection and identification of the origination of the seizure activity:

In at least an embodiment, an EEG signal input module (101) is configured to record an EEG signal from a patient. Each signal consists of 21-channel EEG readings.

In at least an embodiment, a signal pre-processor (102) is configured such that the input module (101) recorded multichannel EEG is pre-processed. Typically, the signal pre-processor (102) is a Multiresolution based Adaptive Filtering (MRAF). MRAF is one of the more suitable processes that can remove low frequency physiological artefacts and correctly preserve the high frequency complex partial epileptic seizure information. Frequently seizure EEG details are affected by the presence of physiological artifacts. These physiological artifacts change true seizure information available in the epileptic EEG signal. Hence, it is very difficult to analyze seizure information. Also, seizure data may overlap with these EEG artifacts which raises a complication in the exact localization of the seizure region. So, it is very crucial to reduce such EEG interferences which hamper identification of the epileptogenic zone accurately. MRAF helps to remove the physiological artifacts by preserving the information of the epileptic seizure
MRAF is implemented in three main steps:
i) To begin with, the EEG signal that is input is decomposed into multiple frequency levels making use of DWT.
ii) Next, soft thresholding is brought into use with the DWT coefficients to generate a signal with minimized abrupt changes.
iii) Finally, MRAF is undertaken to produce an EEG signal that contains minimal physiological EEG artifacts.
Here, soft thresholding is executed on the obtained high frequency wavelet coefficients. Soft thresholding is rendered by (1) as follows:

T_l=s_l v2logR (1)
where T_l is the threshold number, s_l is the standard deviation of the input signal, l is the limit of the DWT decomposition, and R is the number of samples of the EEG.
MRAF is then performed to produce the pre-processed information of the seizure zone. The data thus obtained is expressed by the MRAF expression [39]:

X_o (n)_per(b) =?_(i=0)^(s-1)¦?w_i (n) x_i (n-j) ? (2)
whereX_o (n)_per(b) is the output of MRAF which can be obtained for each band b. Here, bands such as delta, theta, alpha, beta, and gamma are considered at the output stage, where all bands represent different resolutions of frequency and hence multiresolution filtering is achieved. w_i (n)represents the weight of the MRAF approach and represents the newly adapted components of the adaptive filter.

In at least an embodiment, a feature extractor (103) is configured such that, from the pre-processed signal, obtained at the output of the signal pre-processor (102), the calculate Critical Spectral Verge (CSV)along with spike and statistical related features is extracted from seizure preserved information.

After pre-processing the EEG signal using MRAF, features are extracted for each EEG segment. The first step of feature extraction is the computation of the average of spectral values (average PSD) for each frequency bands such as delta, theta, alpha, beta, and gamma. For each frequency band, for each segment PSD is calculated. Then the maximal frequency point at which the PSD of the segment is greater than the average PSD of the corresponding frequency band is computed which is termed as Spectral Verge (SV). However, due to the involvement of undesired glitches in the low frequency region most of the seizure related information is missed. In order to ensure that no seizure details are missed, the SV is optimized further using the FPA. Then, by considering the SV point as the seed point, optimization is performed by the FPA utilizing Levy’s path. The optimized point thus obtained is called CSV. The CSV is obtained for each segment of the multichannel EEG signal. Detailed process to obtain CSV is depicted in Fig. 2

Feature Extraction: The presence of seizure is distinguished by a lowering in the frequency, a rise in the amplitude, and most importantly, a sharpness of the EEG wave. In this stage, seventeen features derived from Spike Statistical (SS) -CSV such as spike height, width, prominence, count, spike width of heights, mean, variance, standard deviation, skewness, entropy, kurtosis along with frequency related details such as alpha (??), beta (??), gamma (??), theta (??), delta (??), and full bands are extracted from the MRAF processed signal

Spike derived features: Epileptic spike related information enables the detection, categorization, and analysis of spikes depending on the intrinsic time and frequency domain characteristics of the signal. Following spike related features are considered:
Spike height: It is computed as the distinction of the maximum spike voltage and the baseline value.
Spike width: It is the difference between two subsequent peaks of the spikes.
Spike prominence: It describes the strength of the spike present in the seizure signal.
Spike count: The total number of spikes which occur within the seizure signal window is termed as the spike count.
Spike width of heights: The height of the contour lines at which the widths are evaluated is called the spike width of heights.

Statistical based features: EEG signals associated with seizures show variability to a greater extent in the time domain, and the spectral information during the duration of the seizure as compared to the non-seizure signals. For multichannel seizure classification, signal statistics like mean, variance, standard deviation, entropy, skewness, and kurtosis are considered.
The skewness, kurtosis, and entropy are expressed by (3), (4), and (5) respectively for
C1, C2, …, Cm
sample points of seizure signal, A.

s= -?_(m=1)^A¦?c^2-log??c^2 ? ? (5)
where ?1and ??2 are the measures of skewness and kurtosis respectively. ?? represents the standard deviation, and ??y is the mean of the signal.

3) Critical Spectral Verge: The prominent frequencies available in all bands of spectrum are utilized as features. The derivation of these spectrum associated details is acquired by the estimation of Power Spectral Density (PSD) from the processed MRAF wave. In order to obtain the critical frequency from all the frequency bands, the bio-inspired Flower Pollination Algorithm (FPA) technique is used. In order to obtain the critical frequency using FPA, global and local optimization is used. Global optimization is attained by the initialization of population containing possible solution and problem variables.

In at least an embodiment, a combinative based feature selector (104) is configured to pass the extracted feature set, from the feature extractor (103), in order to select a best feature set.
Combinative based best feature selection: Combinations are used to compute the total conclusions of an event where the order of the resultants will not be considered. Combinations of an event is computed by (6) as,
(6)
where m represents the total number of features and s expresses the number of elements in the combination.

A heat map is a representation of correlation values existing between different features. The features which are highly correlated indicates the correlation values are nearly one or minus one and are redundant. In Figure 4, spike height, width, prominence, and standard deviation provides such redundant information. So, any one feature out of these can be used for further analysis. From the analysis it is found that spike height extracts most significant seizure information when compared to other redundant features. Hence, spike height is selected and utilized for classification along with CSV. Lesser correlation values which are near to zero indicates that there is no relation existing between a pair of features. It is also necessary to consider all such features for classification.

In at least an embodiment, the extracted best feature set is classified multichannel-wise using Support Vector Machine (SVM) classifier (105) which is given 80% training and 20% testing dataset. The output of this classifier is either determination of a ‘complex partial seizure’ (107a) or determination of a ‘non-seizure’ (107b).
Classification: Finally, robustness of the suggested method is tested by using classification techniques. This method helps differentiate the undisclosed testing dataset into the proper classes depending on the training set details. As a result of subsequent analysis from the literature, to check the accuracy of the proposed approach, this analysis utilizes a well- known classifier, SVM.

SVM (Support Vector Machine) is a machine learning methodology utilized for non-linear signal analysis. SVM applies kernel functions for it to be able to project samples that are non-linear and separable onto some other higher-dimensional space. Next it locates the optimal separating hyperplane in the projection space. The hyperplane maximizes the distance between the hyperplane and the nearest points from each class which is known as support vectors. Radial basis functions (RBF) as kernel functions are reported to have a better performance in the EEG analysis. In this work, RBF is used to distinguish between seizure and non-seizure signals, segment-wise for multichannel EEGs.

In at least an embodiment, the system (100), of this invention, detects complex partial epileptic seizure in a segment-wise manner using a segment-wise processor (106). The segment-wise multichannel analysis helps medical experts to detect the exact complex partial epileptic seizure area of the brain in the real time complex partial epileptic seizure analysis.

Figure 2 depicts a flowchart illustrating a method for the automatic detection and identification of the origination of the complex partial epileptic seizure according to one of the embodiments of this present invention.

MRAF is able to preserve the complex partial epileptic seizure onset zone.

In at least at embodiment, the system is configured to facilitate a method wherein a complex partial epileptic seizure activity is characterised with High Frequency Oscillations (HFOs) based on extracted CSV feature in a channel-wise manner and in a segment-wise manner.

Step 1: Process segment k, of an EEG signal
Step 2: Process band b, of an EEG signal
Step 3: Calculate Power Spectral Density (PSD) values of all segments and all bands
Step 4: Take the average of the calculated Power Spectral Density (PSD) values
Step 5: Apply the Flower Pollination Algorithm (FPA) to optimise and calculate Critical Spectral Verge (CSV) from said calculated Power Spectral Density (PSD)
Step 6: Add values of each calculated Critical Spectral Verge (CSV) to the feature map.

The MRAF (equation 2) filtered EEG segment is characterised by high frequencies.
Now signal is divided into small segments in order to perform analysis on small parts of the signal.
Let x_o (n) be x(n) and x(n)={x_o1 (n),x_o2 (n),…,x_on (n)}
x_o1 (n),x_o2 (n),…,x_on (n)are the segments of x(n).
Uniform distribution of spectral power is observed for high frequency range for complex partial epileptic seizure segment. To represent this information the average spectral power is computed.
Let ?PS?_1 be the power spectrum of x_1 (n).
?PS?_avg=?_(k=1)^P¦(PS (k))/P(7)
P is the signal frequency.
This average spectral power is then used as a threshold for each segment to identify the complex partial epileptic seizure zone and non-complex partial epileptic seizure zones.
At ?PS?_1 (k)>?PS?_avg,
max?(k_s )fork_s={k??PS?_1 (k)>?PS?_avg}
In order to observe frequency information at which EEG Power Spectral Density (PSD) a point is defined in the higher frequency range at which PSD is greater than the average. In order to differentiate between the local and global seizure abnormalities FPA is used which will iterate through local and global abnormalities effectively and optimize the solution to correctly identify the complex partial epileptic seizure segments using pollination algorithm.
Let us define optimal solution by,
S_i=bestfit {f_si }(8)
wherei=1,2,…,d
The resultant solution will be in upper bound (256Hz) to lower bound (0Hz).
Criteria for best fit is the solution for which the stability is achieved over possibilities of all solutions or maximum iterations are reached for solution generations.
In order to reach best solution, define population function map give as,
map (g^* )={S_i}(9)
FPA can now be applied to population of solutions above to get the optimized solution.
Biotic or abiotic process of search for the optimized solution among the map is performed as follows:
Biotic process is given as,
x_i (t^+ )=x_i^((i))+L(g^*-x_1^((i) ))(10)
Levy’s distribution is expressed as,
L~(? ?(?)sin?(p?/2))/(p (S^1+?))(11)
where?(?)is the gamma function with ??=0.5.
Abiotic process is given as,
x_i (t^(+1) )=x_i^((t))+ ?(x_i^((t) )-x_k (t)(12)
The probability switch defines which type of process to perform from the above computed paths. The probability switch is a randomly chosen solution at each iteration.
Finally, the system and method gets the best solution as CSV which is the optimized of all the solutions in the map.
Now CSV can be obtained similarly for all the EEG segments as ?CSV?_1,?CSV?_2,…,?CSV?_n for each segment k, where k= 1,2,…,n.
For each band, CSV can be obtained similarly. The feature map thus formed can be represented as,
[¦(?CSV?_a1,?CSV?_a2,…,?CSV?_an@?CSV?_ß1,?CSV?_ß2,…,?CSV?_ßn@?CSV?_?1,?CSV?_?2,…,?CSV?_?n@?CSV?_?1,?CSV?_?2,…,?CSV?_?n@?CSV?_d1,?CSV?_d2,…,?CSV?_dn@?CSV?_fb1,?CSV?_fb2,…,?CSV?_fbn )](13)

This research proposes the computation of this CSV based optimal feature map.
Combinative based feature selection method is used to compute the best feature set. Classifier SVM is trained using the feature set extracted from SS-CSV based combinative method. The classifier detects and localizes the origination of the complex partial epileptic seizure in the brain.

The proposed complex partial epileptic seizure detection system detects 68 complex partial epileptic seizure EEG correctly out of 70 complex partial epileptic seizure EEG signals. Therefore, the detection rate of this complex partial epileptic seizure dataset is 97.142%. This invention achieves 2.08% of the rejection rate for the complex partial epileptic seizure EEG dataset. The average detection rate (ADR) for the proposed invention is 95%. The proposed invention has achieved 93.24% of sensitivity, 98.15% of specificity, and 97.074% of accuracy. The performance efficiency will vary with respect to the number of EEG dataset.

Steps for Critical Spectral Verge (CSV) and Optimization using Flower Pollination Algorithm (FPA):
After Multiresolution based Adaptive Filtering (MRAF), the system and method, of this invention, was tried to analyse the behaviour of EEG Power Spectral Density (PSD).
The system and method, of this invention, computed EEG Power Spectral Density (PSD) of Multiresolution-based Adaptive Filtering (MRAF) output.
The system and method, of this invention, then, tested EEG Power Spectral Density (PSD) for various EEG signals obtained from different patients.
Here, the system and method, of this invention, was used to observe that non-seizure is characterised with f region, and seizure in the region.
Then, the system and method, of this invention, was used to differentiate between Normal and Seizure activity with the help of EEG Power Spectral Density (PSD).
So, system and method, of this invention, was used to determine whether EEG Power Spectral Density (PSD) average will help.
Therefore, in order to decide a point which corresponds to High Frequency Oscillations (HFOs), we took EEG Power Spectral Density (PSD) average.
Then, the system and method, of this invention, was used to analyse the points above and below the EEG Power Spectral Density (PSD) average.
From here, the system and method, of this invention, computed the maximal ‘f’ point at which the EEG Power Spectral Density (PSD) average is greater than the EEG Power Spectral Density (PSD) average.
This point, is referred to, and called, as Spectral Verge (SV).
Spectral Verge (SV) is computed, segment-wise by the system and method, of this invention,
Thus, the system and method, of this invention, provides ‘N’ no. of Spectral Verge (SV) points for each channel.
Then, the system and method, of this invention, tried to classify these Spectral Verge (SV) points using k-means clustering
No. of clusters = 2
k = 3 & 5; for 5, we got good results
It was observed that classification accuracy = 80%
But, here, some segments were misclassified.
Reason: Due to the presence of unwanted spikes and glitches in f region (range).
So, the system and method, of this invention, was used to optimize this Spectral Verge (SV) point, in order to extract the actual seizure activity.
Therefore, Flower Pollination Algorithm (FPA) algorithm was used by the system and method of this invention, in which:
Spectral Verge (SV) point is taken as a seed solution.
This was considered as a current best solution.
In order to optimize the solution, Flower Pollination Algorithm (FPA) algorithm uses Levy’s path, which is a stochastic (i.e. random) distribution.
In each iteration, it is compared with a probability switch (P) to decide between local and global (pollination) optimization.
stochastic process is well / better suited to find optimized solutions in time variant (s/ms)
Initially, the system and method, of this invention, sets p = 0.5.
If, random path length > P:
Then, do global pollination;
Else, do local pollination.
The algorithm searches whether it is able or possible to compute / find the optimized SV point (as a fitness criteria).
If, yes, and not, the system and method, of this invention, reaches maximum iterations:
Then, continue to next iteration;
Else, If, yes, and max iterations reached;
Then, Stop;
If, no, repeat Step (viii).
Extract best solution for every segment, band, and channel.
These best solutions are given to Support Vector Machine (SVM) classifier (105).
Classification accuracy obtained is 90%.
SVM – RBF, gamma = 1.0

Figure 3 depicts a topographic map illustrating the detection of the origination of the complex partial epileptic seizure activity in the brain.

Based on dataset annotations, the complex partial epileptic seizure emerges from the right and left lobes. From the topographic map obtained, using the proposed invention, it is noted that strong signals emerging from the right and left ear lobes are captured by the gamma and full frequency bands by the CSV. Full band contain all the frequency information and gamma band contains high frequency related epileptic seizure information. The details corresponding to alpha and delta frequency represents the indication of involuntary related details exhibited in the brain region. These visualizations evidences the effect of CSV as an effective feature for complex partial epileptic seizure origination detection.

According to a non-limiting exemplary embodiment, the system, of this invention, achieved 93.239% of sensitivity, 98.151% of specificity, and 97.074% of accuracy. The proposed complex partial epileptic seizure detection system has an average detection rate of 95%. The system, of this invention, achieves 2.8% of rejection rate for complex partial epileptic seizure signal dataset.

The TECHNICAL ADVANCEMENT of this invention lies in providing a system and method for automatic detection of complex partial epileptic seizure from Electroencephalogram (EEG). This system has a benefit of illustrating origination of complex partial epileptic seizure to diagnose epileptic behaviour from segment-wise multichannel EEG. This system is used in the analysis of seizure details from different regions of the brain for the neuro experts. Using this system, the automatic complex partial epileptic seizure detection of the EEG has given an advantage of localization of the origination of seizure.

While this detailed description has disclosed certain specific embodiments for illustrative purposes, various modifications will be apparent to those skilled in the art which do not constitute departures from the spirit and scope of the invention as defined in the following claims, and it is to be distinctly understood that the foregoing descriptive matter is to be interpreted merely as illustrative of the invention and not as a limitation.
,CLAIMS:WE CLAIM,

1. A system for complex partial epileptic seizure detection, said system comprising:
- EEG signal input module (101) consisting of multi-channel electrodes configured to acquire an EEG signal from a patient;
- segmenting module configured to segment said EEG signal into smaller segments to identify a seizure region in said acquired EEG signal;
- a signal pre-processor (102) configured to pre-process said EEG signal input module’s (101) recorded multi-channel EEG signal, to provide a pre-processed signal, using Multiresolution based Adaptive Filtering (MRAF) for removing low frequency physiological artefacts, said pre-processed signal preserving seizure information;
- a feature extractor (103) configured such that, from said pre-processed signal, Critical Spectral Verge (CSV), is computed, along with spike and statistical related features being extracted from seizure preserved information in order to form extracted feature set;
- a combinative based feature selector (104) configured to pass said extracted feature set, using a feature extractor (103), in order to select a best feature set from said extracted feature set;
- a Support Vector Machine (SVM) classifier (105) configured to receive said best feature set for classifying multichannel-wise segments in order to obtain a determined output of said signal being a ‘complex partial seizure’ (107a) signal or a ‘non-seizure’ (107b) signal;
o a segment-wise processor (106) configured to detect complex partial epileptic seizure (107a) signals in a segment-wise manner.

2. The system as claimed in claim 1 wherein, said segment-wise processor (106) being configured to detect complex partial epileptic seizure (107a) signals, said signals being characterised with High Frequency Oscillations (HFOs), based on extracted Critical Spectral Verge (CSV) feature in a channel-wise manner and in a segment-wise manner, said segment-wise processor (106), being configured with instructions, configured to perform the steps of:
- processing one or more segments, of said acquired EEG signal;
- processing one or more bands, of said acquired EEG signal;
- checking, if all segments, of said EEG signal, are processed;
- checking, if all bands, of said EEG signal, are processed;
- calculating Power Spectral Density (PSD)of all processed segments and all processed bands;
- computing average of calculated Power Spectral Density (PSD);
- applying Flower Pollination Algorithm (FPA)to optimise and compute Critical Spectral Verge (CSV) from said calculated Power Spectral Density (PSD); and
- adding values of each calculate Critical Spectral Verge (CSV) to a feature map.

3. The system as claimed in claim 1 wherein, said feature extractor (103) being configured with instructions, configured to perform the steps of:
- computing Power Spectral Density (PSD) values for each frequency band (b), for each segment (k), of each of said EEG signals;
- computing average of EEG Power Spectral Density (PSD) values for each frequency band (b), for each segment (k), of each of said EEG signals;
- computing, segment-wise, a Spectral Verge (SV) value, which is a maximal frequency point at which the Power Spectral Density (PSD) value of that segment is greater than the average Power Spectral Density (PSD) values of the corresponding frequency band;
- seeding the computed Spectral Verge (SV) value, for a segment, as a seed solution value;
- optimizing, using a Flower Pollination Algorithm, by comparing each of said seed solution values, with a probability switch (P), to decide between local pollination optimisation and global pollination optimisation basis random path length, thereby, obtaining a Critical Spectral Verge (CSV) value, for each of said segments; and
- providing each of said Critical Spectral Verge (CSV) values, as a best feature set per segment, to said Support Vector Machine (SVM) classifier (105).

4. The system as claimed in claim 1 wherein, said Support Vector Machine (SVM) classifier (105) configured to receive said best feature set for classifying multichannel-wise segments in order to obtain a determined output of said signal being a ‘complex partial seizure’ (107a) signal or a ‘non-seizure’ (107b) signal, being configured with instructions, configured to perform the steps of:
- forming a cluster of Power Spectral Density (PSD) values for each segment of each channel of said EEG signal;
- computing, segment-wise, a Spectral Verge (SV) value, which is a maximal frequency point at which the Power Spectral Density (PSD) value of that segment is greater than the average Power Spectral Density (PSD) values of the corresponding frequency band;
- determining said computed Spectral Verge (SV) value as a reference value;
- searching for High Frequency Oscillations (HFOs) component in each of said formed clusters;
- checking, for each of said searched High Frequency Oscillation (HFO)component with said reference value;
- determining a seizure signal (107a) if said searched High Frequency Oscillation (HFO)component matches with said reference value;
- jumping to a next cluster, using a Flower Pollination Algorithm, if said searched High Frequency Oscillation (HFO)component does not match with said reference value; and
- said jump being defined by Levy’s path, said Levy’s path being determined by a probability switch in order to determine quantum of said jump.

5. The system as claimed in claim 1 wherein, each signal consists of 21-channel EEG readings.

6. The system as claimed in claim 1 wherein, said Multiresolution based Adaptive Filtering (MRAF) filtered EEG segment being characterised by high frequencies.

7. The system as claimed in claim 1 wherein, said Critical Spectral Verge (CSV) being a point in a higher frequency range at which Power Spectral Density (PSD) is greater than average Power Spectral Density (PSD).

Dated this 15th day of March, 2022

CHIRAG TANNA
of INK IDÉE
APPLICANT’S PATENT AGENT
REGN. NO. IN/PA – 1785