Analysis of Electroencephalograms: Application of Artificial Neural Networks for Detection of Epileptic Discharges Rokas Mykolas Deveikis Tadas Meškauskas Faculty of Mathematics and Informatics, Vilnius University, Faculty of Mathematics and Informatics, Vilnius University, Lithuania Lithuania rokas.deveikis@mif.stud.vu.lt tadas.meskauskas@mif.vu.lt Abstract — Algorithms based on multilayer feedforward and machine learning and in particular artificial neural networks recurrent neural networks were employed to automatically detect can be employed for the automatic EEG analysis. The main epileptic discharges (spikes) in unprocessed motivation behind selection of this method is an opportunity to electroencephalograms (EEG). Results were compared to two develop an algorithm that can potentially solve problems separate benchmarks: analysis provided by an expert and output mentioned earlier and improve identification of various EEG of algorithm based on mathematical morphological filters [3]. elements. Nonetheless, it would be much easier to adjust such Feedforward neural network was able to on average detect ~95% algorithm for diagnostic of other medical conditions. Case of spikes detected by the expert and ~60% of spikes detected by study is done with spikes common to benign childhood an aforementioned algorithm. Recurrent neural network was epilepsy with centrotemporal spikes (referred as Rolandic able to on average detect ~93% and 80% spikes respectively. While in some cases artificial neural network was able to epilepsy from now on), yet described methods can potentially outperform algorithm based on morphological filter in terms of be applied for diagnostic purposes of other types of epilepsy detected spikes, the main issue remains the high number of false and brain injuries alike. positives and in particular low-amplitude spike-like waveforms The paper is organized into three main parts: first of all we that cannot be utilised for diagnostic purposes. will discuss the current state of research in an automatic EEG analysis. Secondly we will describe data collection and Keywords— Electroencephalograms, Epileptiform discharge, preparation procedures as well as main research methods. Artificial neural network, Machine Learning, Epilepsy Finally we will present how multilayer feedforward neural network and recurrent neural network was able to deal with I. INTRODUCTION aforementioned research goals. Early and accurate diagnostics of epilepsy is one of the key factors in order to prescribe proper treatment and improve the II. RELATED WORK quality of life for patients and caretakers alike. Idea to develop algorithm for automatic EEG analysis dates Electroencephalograms (EEG) remain one of the main tools back to 1970s. In 1972 Carrie JR. [5] published paper used for diagnostic purposes and can be accurately analysed by describing one of the first automatic algorithms aimed to detect professionals; however, such process is time consuming and epileptic discharges (also called spikes) visible in EEG. inefficient. While many algorithms are employed to support Algorithm was based on calculation of certain spike features, manual EEG analysis, issues like classification quality or namely specifications of spike peak and sharpness. Later inflexibility still limit practical application. Guedes de Oliveira, et al. [6] described method to identify EEG Despite that currently there are many algorithms designed spikes based on spike curves and a certain threshold was for automatic EEG analysis, in many cases it’s still done applied to separate spikes from non-spikes. One of more recent manually by visually inspecting recording [2]. Given that EEG works was published by A.V. M. Misiūnas, et al. where recordings can last from few minutes to several hours, mathematical morphological filters were employed to identify nonetheless are composed of 21 or more channels and can be EEG spikes associated with Rolandic epilepsy [3]. Based on affected by various noises [2], such analysis remains difficult expert evaluations – this algorithm is able to identify ~90% and time consuming. While aforementioned algorithms EEG spikes. This algorithm was also used in this paper for certainly help in EEG analysis, they are not always reliable to comparison purposes. identify elements of interest and in most cases are hardcoded to While aforementioned algorithms present good results in detect specific elements thus making them non-reusable for certain cases they all face several issues: other diagnostic purposes (see [2], [3], [4]). The main purpose of this research is to analyse how 1) Sensitivity to noise. In most cases results are highly Copyright held by the author(s). 43 affected by noise, which highly depends not only on how distinct sharp waveforms also called spikes. A typical spike is recording was performed but on device itself characterized by high amplitude (more than twice higher than 2) Being hardcoded for one specific problem. Therefore average amplitude of normal brain activity prior spike) and application for other purposes remains limited and in many change of phase (see [13], [14]). There are also non-typical EEG elements associated to epileptic discharges yet given that cases – impossible. those only constitute from 1% to 7% of all cases (see [15], Machine learning algorithms (and especially artificial [16]); therefore, those are not further investigated. While neural networks) are in many cases employed to address aforementioned spikes might slightly differ, common spike similar problems. Main advantage of such method is that it’s characteristics remains similar among different patients. not necessary to specify spike duration and morphology [8] Spikes can be classified into two main categories based on therefore neural network can be quickly adjusted (trained) to duration [22]: analyse signals recorded with different electroencephalographs and employed to detect spikes with different durations and - Spikes – duration of 20 – 70 ms. morphologies. Research related to application of neural networks for EEG analysis can mostly fall within two distinct - Sharp waves – duration of 70 – 200 ms. categories: pre-processed EEG data (see. Gabor and Seyal, Both can be described as short term elements, clearly 1992 [11]; Webber et al., 1996 [9]) and raw EEG data (see distinguishable from normal brain activity with the main Ozdamar and Kalayci, 1998 [10]). Ozdamar and Kalayci was component having negative phase in most cases. Both (spike able to design a neural network that can correctly classify EEG and sharp waves) have similar initial (or elevation) stage; elements (with main purpose of spike-detection) with however, sharp waves have longer demotion stage. In this sensitivity of ~94%, however Webber et al. [9] was not able to paper spikes are not distinguished from sharp waves and are replicate such results (and found true positives to be ~76% and analysed together. sensitivity ~40%). Second distinct characteristic can be defined as a Cheng-We and Hsiao-Wen [12] used feedforward neural combination of three key measurements: amplitude, duration network composed of 3 layers (with 30, 6 and 1 neuron and sharpness. Frost J.D. [17] analysed aforementioned accordingly). They also tried to replicate research of Ozdamar characteristics among spikes with a duration up to 70 ms and and Kalayci [10] and concluded that initial results was possibly developed CPS (composite spike parameter) index that can biased due to errorous data preparation. One of the main issues help to detect Rolandic EEG spikes. The author defined authors encountered is overemphasis on 10th neuron. Authors sharpness as second derivative of voltage at spike peak and argue that this happened due to fixed location of spike centres, normalized by amplitude. Among all spikes analysed by author where 10th element of supplied time-series always aligned to mean values of amplitude, duration and sharpness were spike centre, thus resulting I high numbers of false positives. 160,9μV, 74ms. and 0.022 respectively. Authors tried to train neural network with varying coordinates of spike centre relative to analysed section, yet were B. Data description and preparation unsuccessful to achieve satisfactory results and concluded that at that point in time computing technologies were not viable to For this research EEG readings of 20 patients was used. apply neural networks for EEG analysis. Nonetheless this Among those data of 3 patients was analysed by expert. In each emphasizes importance of data preparation and training of the files aforementioned expert (a paediatric neurologist) strategy. marked spikes and elements that visually resemble spikes but are definitely not. In total coordinates of 499 spikes were While it‘s rather difficult to analyse raw EEG data, much collected. better results were achieved while working with pre-processed datasets. Nigam, and Graupe [21] employed feedforward Considering spike characteristics discussed previously, also neural networks for spike detection, however signal was pre- given that sharp waves and spikes can be both associated with processed with non-linear filters. Authors were able to achieve Rolandic epilepsy – a fixed spike length of 200 ms (maximum ~96% classification accuracy thus implying that data pre- length of sharp wave) will be used in this research. Nonetheless processing is a good alternative. However, given that such sub-elements will not be separated in any way and will be filters must be individually prepared for different EEG this treated as one typical epileptic discharge associated with method has the same issues as discussed in previous chapters Rolandic epilepsy. EEG readings used in this paper was therefore it’s not further analysed in this paper. presented in .edf (European data format) type files recorded in 256Hz sample rate (1 element in time series is equal to 3.90625 ms). Further data analysis was done in fixed-duration sections. III. CHARACTERISTICS OF ROLANDIC EPILEPSY EEG In order to prepare such sections, 200 ms. (maximum spike SPIKES AND DATA COLLECTED FOR THIS REASERCH duration used in this research) was approximated to 50 elements in time series or roughly 196 ms. Each section with A. EEG spikes specific for Rolandic epilepsy spike was prepared by choosing 25 elements backwards from Rolandic epilepsy is one of the most common types of spike centre and 25 elements forward, thus ensuring that spike childhood epilepsy with ~23% of early school age (mean age centre (peak) and centre of section will align. ~7 years) children being affected [13]. In most cases, the While expert tried to mark spikes at the exact centre it was disease affects boys more than girls (with ration 5 to 1). EEG proven to be difficult to point exact coordinates, therefore data readings of patients diagnosed with this type of epilepsy have 44 was further post-processed by shifting positions of spike minimal adjustments as well as to potentially work with centres to lowest voltage value in the section. While extraction of other EEG features. Nonetheless, working with theoretically spike peaks can have positive phase it was not the raw EEG allows to evaluate neural network’s ability to deal case for this data therefore no further adjustments was done. with variability in data. Considering aforementioned arguments Example is presented in figure 1 only 50Hz notch filter was applied in order to reduce noise created by power supply. FIG 1. EXAMPLE OF INITIAL (SOLID VERTICAL LINE) AND ADJUSTED (DASSHED VERTICAL LINE) SPIKE CENTERS IV. NEURAL NETWORK DESIGN In this paper we employ and compare two neural network architectures: multi-layered feedforward neural network and recurrent neural network based on LSTM model. A. Training data As discussed in paragraph III data from 3 patients and 4 EEG channels were used for initial training of neural networks. Given that main purpose of this analysis is to be able to identify spikes in full EEG channel data was prepared in following manner. Each channel was divided into sections each consisting of 50 elements based on rolling window i.e. each section starts with element n and ends at n+50 while next section starts at n+1 and ends at n+1+50. Under such rules each channel is represented by number of sections equal to number of elements in channel. In the next step segments was classified into two main categories with corresponding numeric values: 0. Non-spikes Waveforms that morphologically resemble spikes, however have lower amplitude, were not included to initial sample. 1. Spikes Those can be considered a non-typical spikes. While those Data of full EEG channel was used for training. Each sometimes can be associated with Rolandic epilepsy, however, section was classified as spike, if centre of that section falls as discussed with expert, final diagnosis is never based on such within interval: “non-typical“ elements. Summary of amplitude characteristics of all spikes is presented in table I where it is evident that [v – Section size * 0,425; v + section size* 0, 425] average amplitude is approximately 12% lower when Where v – centre of actual spike, section size – section size compared to research of Frost J.D. [17] with ~35% lower or in this case 200 ms. standard deviation. Both can be explained by usage of different measuring device, however deviations remains insignificant. Coefficient 0,425 represents 42,5 % and is calculated under main assumption that at least ¾ of spike must be visible in TABLE I. AMPLITUDE CHARACTERISCITS OF EEG SPIKES order to correctly classify segment it as spike. Therefore USED IN THIS REASERCH (assuming that all elements of specific segment falls within interval [0 ms.; 200 ms.] with spike centre being at 100 ms.) Mean 142,8 Standard deviation 44,2 the earliest point when ¾ of shortest possible spike (20ms) can Minimum 51 be visible in this segment is when x coordinate of spikes’ Maximum 325,7 centre is equal to 15 ms. (3/4 * 20 ms. = 15 ms.) and latest Aforementioned data was used for all further researches point is 185 ms. (200 ms. – 15 ms.). This gives interval of [15 presented in this paper. ms.; 185 ms.] or ~21 element. Summary of data used for training is presented in TABLE II. As discussed in previous chapter data pre-processing can potentially improve detection quality, however we decided to keep it to a minimum, mostly because EEG data pre-processing TABLE II. SUMMARY OF TRAINNG DATA methods used for feature extraction or denoising are in most Patient Chanel Spikes Segments Total cases designed to work with specific data (eg. classified as spikes segments Electrooculogram (EOG) induced noise can be successfully 1 T4 177 4248 47310 removed in pre-procesing stage by applying regression based 2 T3 144 3456 185550 methods [23] or filters based on independent component 2* T5 28 672 185550 analysis [24] however such approach requires to have data with 3 P4 150 3600 66510 separate EOG measurements which is not always available). * Due to low number of spikes, data from patient 2, cahnnel T5 was used only for training purpose and not utilized for testing. However, main purpose of this research is to develop algorithm that can be employed to analyse data from various sources with 45 B. Measuring Classification quality was employed (with k = 3). Due to non-standard measurement, At this point it‘s important to establish criteria to evaluate data was not shuffled but instead neural network was trained algorithm performance. Typical binary classification accuracy with data from two patients, while tested on data from third is not suitable here for several reasons: one. As in typical cross validation such technique was repeated 3 times. In order to have a starting point several researches First of all, given that in typical 60 second EEG channel were taken into consideration: (that is consisted of 15360 segments), only less than 1% of those segments can be typically classified as spikes, simple Based on research by Cheng-Wen Ko and Hsiao-Wen Chu binary accuracy will provide relatively good results despite that [12] a network architecture consisting of 3 layers with 30 spikes was not necessarily found. Nonetheless main purpose of (segment size), 6 and 1 neurons accordingly were chosen; such algorithms is spike detection. Also when using binary Cheng-wen et. al. [18] used four layered neural network classification accuracy, a case where multiple sections was with 4,5 and 1 neurons (input layer is excluded since it always detected next a single spike, is treated in a same manner as holds number of neurons equals to segment size) when multiple sections were detected next to different spikes. Yet again – given that main purpose is to detect all spikes this Weng and Khorasani [19] used four layered neural network would not represent desired performance. with 9, 90 and 1 neurons. Therefore 3 main criteria was defined: Mirchandani and Cao [20] presented an idea that number of neurons in hidden layer can be calculated using following 1) Ratio of all detected spikes with spikes detected by formula: expert (true spikes) H = log2 M 2) Elements (group of sections) falsely classified as spikes 3) Spikes falsely classified as non-spikes Where H – number of neurons in hidden layer, M – largest number of linear separable regions in input data (in this M = 50 – 1 = 49). Given that exact coordinates of spikes were well-known following methodology was employed: Using aforementioned researches as a starting point and employing k-fold cross validation optimal feedforward network 1) All algorithms were design in a way that only sections architecture was found to be 4 layers with following that were classified as spikes will be outputted. If distribution of neurons in each corresponding layer: 50, 25, 90, section was classified as spike it‘s coordinate is 1. Averaged results of cross validation is presented in table III adjusted in following manner: (TP and FP refers to True Positives and False Positives xn = xo + section size / 2 respectively). where xn – new coordinate of segment; TABLE III. CLASSIFICATION RESULTS OF FEEDFORWARD xo – old coordinate of segment (number of first NEURAL NETOWRK WHEN COMPARING TO ANALYSIS PROVIDED element); BY EXPERT Average Δ found Average FP section size – 200 ms detected spikes spikes sections Δ FP sections 2) Each xn is then compared to known spike coordinates. 95% 4% 30% 48% Whenever it falls within aforementioned interval of: [v – section size * 0,425; v + section size* 0, 425] While algorithm was able to detect ~95% spikes detected (where v – centre of actual spike) by expert, there were ~30% of segments (not spikes) falsely classified as spikes. Nonetheless ~48% difference of FP and TP that section is considered correctly classified and spike segments in channels indicated that results are highly is considered detected. dependent upon channel. Nonetheless at this point analysis was According to our evaluation methodology - each spike can only performed on 3 EEG files therefore results are not good be only detected once, therefore more sections (classified as representation of algorithm’s performance. spikes) around single spike will not inflate result. All sections In order to expand analysis results from single channel outside interval are classified as non-spikes. Those sections were further processed. In order to ensure that each false were further grouped in order to assess number of false positive segment will represent a new spike results were positives where each group was composed of up to 42 sections grouped. First of all true positives were properly identified and starting from first false positive section, with last one being not matched to corresponding spikes. Then each false positive is further than 42 elements while ignoring true positives. grouped so that each group was filled with segments that are no further from the false positive (that is not part of any other group) than 50 elements. E.g. whenever segment is classified as false positive a new group is formed where that segment is a C. Application of Multilayer Feedforward Neural Network first group member and all other segments that are distanced no In order to design optimal structure of neural network more than 50 elements from the first one are assigned to the (number of neurons in hidden layers) a k-fold cross validation same group. One distant segment can have its’ own group. 46 T3 channel of patient 2 was further processed using max 99% 56% 100% aforementioned technique thus resulting in 53 false positive groups that will be referred as 53 false positives from now on. Further analysis of false positives revealed that such elements On average feedforward neural network was able to detect can be classified into 3 main categories: 60% of spikes, detected by algorithm based on mathematical morphological filters, however a lot of elements were falsely 1) Segments that are close to spike, however falls outside identified as spikes (average of ~78%). Nonetheless in 11 out pre-defined range. of 17 (with 3 channels removed from initial sample) cases 2) Spike-like waveforms with low amplitude neural network was able to detect more than 50% spikes with 3) Elements that are not related to spikes 86% detected spikes on average. Technically 1st group is classified correctly, however given At this point false several examples of positives were sent that spike peak is on the edge of the segment (at least 3/4 of to expert for detailed analysis. Due to limited time and spike is not visible) it would be impossible to tell without any resources expert was able to analyse only 12 false positives further references that this is actually a spike therefore such however this potentially reveals certain trend. Among 12 elements won‘t be included or reclassified. See fig.2 for results were distributed in following order: example.  2 were identified as spikes FIG 2. BARELY VISIBLE SPIKE CLASSIFIED AS FALSE POSITIVE  2 were identified as non-spikes  8 were identified as either spike-like waveforms with low amplitude or true spikes, however expert was not able to correctly classify those with certainty. FIG 2. EXAMPLE OF SPIKE (CICRLED ON THE LEFT SIDE) AND LOW AMPLITUDE SPIKE-LIKE WAVEFORM (ON THE RIGHT) Therefore we conclude that the main problem remains low amplitude or low sharpness spike-like waveforms that are constantly identified as spikes although are unwanted for Despite all efforts – due to low amount of high quality data diagnostic purposes. it is difficult to came up to conclusions therefore data sample To sum up we conclude that multilayer feedforward neural (of 3 patients) was increased by 17 more EEG files without network can potentially be used for primary analysis of EEG spike coordinates. To prepare more results for validation an with a purpose to identify spikes associated with Rolandic algorithm based on morphological filters were employed [3]. epilepsy and can detect ~60% of spikes on average when All 17 files were analysed and results were compared with ones comparing with algorithm based on morphological operations produced by algorithm based on neural network. Comparison or ~95% when comparing to data prepared by expert. However was done in the same manner as previously discussed. main problem remains low amplitude, spike-like waveforms Summary of results are presented in table IV. All numbers identified as spikes that can potentially be removed by certain (True positives or false positives) are in relation to results of filters. Finally, algorithm discussed in this paragraph did not algorithm based on morphological filters. Also 3 channels, detected spikes in channels where there potentially were no where aforementioned algorithm identified less than 10 spikes, spikes (or low number of spikes, eg. 1). were removed from final results. D. Recurrent Neural Network TABLE IV. CLASSIFICATION RESULTS OF FEEDFORWARD Given that too many low-amplitude spike-like waveforms NEURAL NETOWRK COMPARED TO ALGORITHM BASED ON were falsely classified as spikes, a recurrent neural network MORPHOLOGIC FILTERS was employed to address this issue. Main rationale for this type Detected TP FP of neural network is that it captures information from previous spikes elements therefore it‘s actually possible to estimate spike St. dev 37% 19% 19% amplitude relative to the background brain activity. mean 60% 22% 78% Jezusefovich et. al [7] analysed more than 10000 recurrent min neural network architectures and proven that there is no 0% 0% 44% alternative that can consistently outperform LSTM model and 47 GRU (gated recurrent unit). While there are not many V. RESULTS AND DISCUSSION researches where LSTM was applied for EEG analysis, considering previous statement this model was selected as potentially offering the best performance. Two algorithms aimed to detect spikes in unprocessed EEG data were analysed in this research. While working with the same data, first step was to identify optimal network architecture. Same three-fold cross validation The algorithm based on recurrent neural network was able was applied as previously discussed with feedforward neural to on average detect 78% of spikes (see table VI) when network. Optimal network depth (layers) was proven to be 1 comparing results to an algorithm based on mathematical input layer, 2 LSTM layers and 1 output layer (feedforward morphological filters. Also, it outperformed algorithm based on neural network). Each LSTM layers had 90 units (also referred feedforward neural network by ~10% (see table V) on average as cells). Results of cross validation (when algorithm output in terms of detected spikes, but increased number of false was compared to analysis performed by expert) are presented positives by 3% on average. While neither algorithm developed in Table V. Only results of network with aforementioned in this paper is sufficient on its’ own for diagnostic purposes, optimal structure is included. It is evident that while recurrent neural networks can be used for data preparation or primary neural network performed slightly worse in comparison to analysis. feedforward neural network (eg.93% detected spikes versus While neural networks can effectively classify EEG spikes 95%), deviation is minor and given low number of test and in some cases, the main issue is false positives and in particular training samples in each case we conclude that there are no - low amplitude spike-like waveforms (see figure 2). While in major differences in performance when testing on similar type some cases such waveforms can be classified as spikes, those of data. are not used for diagnostic purposes thus are undesirable and should be removed from the result set. It can possibly be TABLE V. CLASSIFICATION RESULTS OF RECURRENT NEURAL achieved by applying amplitude based filters thus making such NETOWRK WHEN COMPARING TO ANALYSIS PROVIDED BY combined algorithm an effective tool for EEG analysis, EXPERT however, this is beyond scope of this paper and is Average Δ found Average FP recommended for future research. detected spikes spikes sections Δ FP sections 93% 10% 30% 10% ACKNOWLEDGMENT We deeply acknowledge paediatric neurologist MD Rūta Furthermore, research was repeated in a same manner as Samaitienė for her input and expertise related to epilepsy and previously (compared to results of algorithm based on data collection efforts. Also, I would like to thank MSc morphological filters [3]). Data of same 20 patients were used. Andrius Vytautas Misiukas Misiūnas for support and Results are presented in TABLE VI. Same as previously - all methodological guidance. numbers (True positives or false positives) are in relation to results of algorithm based on morphological filters. Also 3 REFERENCES channels where aforementioned algorithm identified less than 10 spikes was excluded. [1] EEG SIGNAL PROCESSING. Saeid Sanei and J.A. Chambers Centre of Digital Signal Processing Cardiff University, UK TABLE VI. CLASSIFICATION RESULTS OF RECURRENT NEURAL [2] Samaitienė R. Rolando epilepsija sergančiu vaiku EEG pakitimu, miego NETWORK COMPARED TO ALGORITHM BASED ON MORPHOLOGIC bei elgesio sutrikimu ir klinikiniu charakteristikų sąsajos. Vilniaus FILTERS universitetas, 2013. Detected TP FP [3] Misiūnas, A. V. M., Meškauskas, T., & Juozapavičius, A. On the spikes implementation and improvement of automatic EEG spike detection St. dev algorithm.In: Proceedings of the Lithuanian Mathematical Society, Ser. 26% 16% 16% A, Lietuvos matematikos rinkinys, 56, pp. 60–65, 2015. ISSN: 0132- mean 78% 19% 81% 2818. min [4] Wilson, S. B., & Emerson, R. (2002). Spike detection: a review and 8% 2% 49% comparison of algorithms. Clinical Neurophysiology, 113(12), 1873- max 100% 51% 98% 1881. [5] Carrie JR. A hybrid computer technique for detecting sharp EEG transients.Electroencephalography and clinical Neurophysiology 1972;33(3):336–8. At this point recurrent neural network was able to [6] De Oliveira, P. G., Queiroz, C., & Da Silva, F. L. (1983). Spike detect mores spikes on average, however at the expense of detection based on a pattern recognition approach using a false positives with a small increase from 78% to 81%. After microcomputer. Electroencephalography and clinical neurophysiology, visual analysis of false positives it was clear that same low 56(1), 97-103. amplitude waveforms are dominant in this group as well [7] Jozefowicz, R., Zaremba, W., & Sutskever, I. (2015). An empirical therefore it is evident that shift from feedforward to recurrent exploration of recurrent network architectures. In Proceedings of the neural network alone can’t significantly improve results. 32nd International Conference on Machine Learning (ICML-15) (pp. 2342-2350). 48 [8] Valenti, P., Cazamajou, E., Scarpettini, M., Aizemberg, A., Silva, W., & Kochen, S. (2006). Automatic detection of interictal spikes using data mining models. Journal of neuroscience methods, 150(1), 105-110. [9] Webber, W. R. S., Lesser, R. P., Richardson, R. T., & Wilson, K. (1996). An approach to seizure detection using an artificial neural network (ANN). Electroencephalography and clinical Neurophysiology, 98(4), 250-272. [10] Ozdamar O, Kalayci T. Detection of spikes with artificial neural networks using raw EEG. Comput Biomed Res 1998;31:122–142. [11] Gabor, A. J., & Seyal, M. (1992). Automated interictal EEG spike detection using artificial neural networks. Electroencephalography and clinical Neurophysiology, 83(5), 271-280. [12] Beritelli, F., Capizzi, G., Sciuto, G.L., Napoli, C. and Scaglione, F. (2018) Automatic heart activity diagnosis based on gram polynomials and probabilistic neural networks. Biomedical Engineering Letters, 8(1), pp.77-85. [13] Ko, C. W., & Chung, H. W. (2000). Automatic spike detection via an artificial neural network using raw EEG data: effects of data preparation and implications in the limitations of online recognition. Clinical neurophysiology, 111(3), 477-481. [14] Kramer U (July 2008). "Atypical presentations of benign childhood epilepsy with centrotemporal spikes: a review". J. Child Neurol. 23 (7): 785–90. doi:10.1177/0883073808316363. PMID 18658078 [15] Gloor P. Contributions of electroencephalography and electrocorticography inthe neurosurgical treatment of the epilepsies. Adv Neurol 1975;8:59–105. [16] Fejerman, N., Caraballo, R., & Tenembaum, S. N. (2000). Atypical Evolutions of Benign Localization‐Related Epilepsies in Children: Are They Predictable?. Epilepsia, 41(4), 380-390. [17] Panayiotopoulos, C. P. (1999). Benign childhood partial seizures and related epileptic syndromes (Vol. 15). John Libbey Eurotext. [18] Frost, J. D., Hrachovy, R. A., & Glaze, D. G. (1992). Spike morphology in childhood focal epilepsy: relationship to syndromic classification. Epilepsia, 33(3), 531-536. [19] Ko, C. W., Lin, Y. D., Chung, H. W., & Jan, G. J. (1998). An EEG spike detection algorithm using artificial neural network with multi-channel correlation. In Engineering in Medicine and Biology Society, 1998. Proceedings of the 20th Annual International Conference of the IEEE (Vol. 4, pp. 2070-2073). IEEE. [20] Damaševičius, R., Napoli, C., Sidekerskienė, T. and Woźniak, M. (2017). IMF mode demixing in EMD for jitter analysis. Journal of Computational Science, 22, pp.240-252. [21] Weng, W. D., & Khorasani, K. (1996). An adaptive structure neural networks with application to EEG automatic seizure detection. Neural Networks, 9(7), 1223-1240. [22] Mirchandani, G., & Cao, W. (1989). On hidden nodes for neural nets. IEEE Transactions on circuits and systems, 36(5), 661-664. [23] Nigam, V. P., & Graupe, D. (2004). A neural-network-based detection of epilepsy. Neurological Research, 26(1), 55-60. [24] IFSECN, International Federation of Societies for Electroencephalography and Clinical Neurophysiology. A glossary of terms most commonly used by clinical electroencephalographers. Electroencephalogr Clin Neurophysiol 1974; 37:538–53. [25] Princy, R., Thamarai, P., & Karthik, B. (2015). Denoising EEG signal using wavelet transform. International Journal of Advanced Research in Computer Engineering & Technology, 3. [26] Vorobyov, S., & Cichocki, A. (2002). Blind noise reduction for multisensory signals using ICA and subspace filtering, with application to EEG analysis. Biological Cybernetics, 86(4), 293-303. 49