System and methods for predicting epileptic seizures using electrocardiographic (ECG) signals

Abstract

Systems and methods for predicting an epileptic seizure with a sufficient window time before its onset to allow the patient to locate a safe place or take appropriate precautionary actions. The method is presented through adroit integration of maximal overlap wavelet packet transform, homogeneity index, and a K-Nearest Neighbors classifier to predict an epileptic event twenty minutes before its onset using electrocardiogram (ECG) signals. The effectively predicts an epileptic seizure 20 minutes prior to its onset with an accuracy of 93.25%.

Description

Atty. Dkt. No.103361-540WO1 SYSTEM AND METHODS FOR PREDICTING EPILEPTIC SEIZURES USING ELECTROCARDIOGRAPHIC (ECG) SIGNALS CROSS REFERENCE TO RELATED APPLICATION

[0001] This application claims priority to U.S. Provisional Patent Application No. 63 / 590,807, filed October 17, 2023, entitled “System and Methods for Predicting Epileptic Seizures,” which is expressly incorporated herein by reference in its entirety. BACKGROUND

[0002] Epilepsy, a complex pathology with various etiological origins, is characterized by producing an electrical activity in the brain called neuronal hyperexcitability, which can have multiple disruptive symptoms in a patient, e.g., uncontrollable muscular contractions, sensory hallucinations, lapses of attention, among others. This neurological disease is one of the most common, impacting more than 40 million people worldwide. From 20 to 30% of epileptic patients (EPs) have chronic and intractable seizures, which increase by 9% the probability of presenting an emergency that can require hospitalization because of a seizure-related injury (e.g., fractures, burns, submersion accidents, and soft-tissue injuries). In addition, EPs can suffer from other comorbid conditions, e.g., learning and psychological problems produced by depression, anxiety, and social isolation, among other issues, which increase their susceptibility to injury as well as fear of death during an ictal state because the risk of sudden death increases to three times that of the general population. This illness significantly decreases the quality of life in all senses.

[0003] Research has been published on electroencephalogram (EEG)-based seizure detection and epilepsy diagnosis, such as epileptic seizure detection using pre-trained deep convolutional neural network (CNN) and transfer learning. Others have investigated epileptic seizure prediction using deep neural networks via transfer learning and multi-feature Fusion; using a dual-modal information bottleneck network for seizure detection; using a six-center assessment of CNN-transformer with belief matching loss for patient-independent seizure detection in EEG; and using CNN with a multi-headed attention mechanism for seizure prediction.

[0004] Seizure Detection Using Electrocardiographic (ECG)

[0005] Shamim et al. investigated the temporal characteristics (i.e., activity, mobility, and complexity) and spectral characteristics (i.e., mean absolute deviation of the discreteAtty. Dkt. No.103361-540WO1 Fourier transform and spectral entropy) extracted from the instantaneous heart rate (IHR) of ECG signals. They utilized a threshold-based classifier for seizure detection, achieving an accuracy of 94.5%. Fractality analysis has been used successfully in the automated computer- aided diagnosis of various neurological and psychiatric disorders such as Autistic Spectrum Disorder (ASD), Alzheimer's disease, and Major Depressive Disorder and determining severity in disorders of consciousness. Chakraborty et al. investigated monofractal analysis, multiscale analysis, and multifractal detrended fluctuation analysis (MFDFA) for characterizing ECG signals to detect seizures. They discovered that MFDFA identifies a lower value of post-ictal heart rate compared to the control group. Furthermore, they concluded that the connection between autonomic control and cardiac functions is deeper than previously believed. Similar to previous work, Ghosh et al. investigated MFDFA to quantitatively determine the degree of multifractality in cardiac dynamics. Their study revealed significant cardiac changes experienced by the patients during a seizure.

[0006] Seizure Prediction Using ECG

[0007] Bai et al. investigated various time domain features (i.e., root mean square value, standard deviation, mean, and the number of pairs of adjacent RR intervals) and frequency domain attributes (i.e., the power of low and high-frequency bands; and the ratio between these bands). They employed a support vector machine (SVM) classifier to predict an epileptic crisis 12 minutes before its onset with an accuracy of 86%. Perez-Sanchez et al. integrated the wavelet packet transform and 17 statical time features (e.g., kurtosis, standard deviation, variance, among others) with a classifier based on if-else rules to predict an epileptic seizure 15 minutes before its starting point. The authors claimed an accuracy of 100% for epileptic prediction 15 minutes before seizure onset. They noted, however, that further investigation is needed to increase the prediction time and evaluate other indicators or indices to reduce the computational complexity. Hadipour et al. examined the ECG signals considering various statistical parameters in the time domain (i.e., mean, standard deviation, root mean square, and total power), frequency domain (power spectrum density), RR interval features, and wavelet coefficients, combined with a short-term long memory machine learning to predict 3 minutes before the seizure onset. They reported an accuracy of 88% in distinguishing between seizure and pre-seizure states. Mahmoudi et al. used the Pan Tompkins algorithm to calculate the RR intervals of the QRS complex in the heart rate value (HRV) of the examined ECG signals and used them as inputs to train a multilayer perceptron to distinguish between pre-ictal state and seizure-free conditions. They reported an accuracyAtty. Dkt. No.103361-540WO1 of 90% in predicting an epileptic crisis before its onset using the pre-ictal form (3 minutes before the crisis onset).

[0008] Thus, while previous research presented results for epileptic prediction employing the ECG signals, they have two main limitations: (1) inadequate prediction time, which limits the transfer of a patient to a hospital or searching for a safe place, and (2) intensive computations, which hinder an epileptic prediction in real time. Therefore, it is vital to develop new methodologies or techniques with the adeptness of forecasting a seizure some minutes prior to its beginning, which will reduce anxiety in EPs due to the fear of having a sudden unexpected crisis and the possible consequences, allowing the EPs to identify and go to a safe place without objects, avoid dangerous situations, or receive proper medical care.

[0009] The investigation of new methodologies or techniques with the capability of increasing the prediction time with good accuracy and low computational complexity is needed because this will allow patients to take the required precautions to safeguard their well-being and increase their quality of life. SUMMARY

[0010] Described herein are systems and methods of predicting a seizure with sufficient window time before its onset is highly desirable because it will allow the patient to locate a safe place or take appropriate precautionary actions. The method is presented through adroit integration of maximal overlap wavelet packet transform, homogeneity index, and a K- Nearest Neighbors classifier to predict an epileptic event twenty minutes before its onset using electrocardiogram (ECG) signals.

[0011] In an aspect, a method for training a machine learning model to predict the onset of epileptic seizure events for a subject is disclosed, The method may include recording an electrocardiogram (ECG) signal from the subject, wherein the ECG signal includes a first signal portion recorded at least a first time period prior to an onset of an epileptic seizure event and a second signal portion recorded within a second time period of the onset of the epileptic seizure event; parsing the ECG signal into a series of signal segments of a predefined length; decomposing each of the series of signal segments into a plurality of frequency bands using a maximal overlap discrete wavelet packet transform (MODWPT) function; evaluating each of the plurality of frequency bands using a homogeneity index (HOI) to determine an HOI value for each of the plurality of frequency bands; evaluating the HOI value of each of the plurality of frequency bands using Kruskal-Wallis analysis (KWA) to identify a subset of the plurality of frequency bands that are most useful in predicting theAtty. Dkt. No.103361-540WO1 onset of an epileptic seizure event; and training the machine learning model to predict the onset of an epileptic seizure event using the HOI value of each of the subset of the plurality of frequency bands. A corresponding system is also disclosed.

[0012] This summary is provided to introduce a selection of concepts in a simplified form that is further described below in the detailed description. This summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The foregoing summary, as well as the following detailed description of illustrative implementations, is better understood when read in conjunction with the appended drawings. To illustrate the implementations, there are shown in the drawings example constructions; however, the implementations are not limited to the specific methods and instrumentalities disclosed. In the drawings:

[0014] FIG.1 illustrates segmentation of a pre-ictal state in 20 intervals of 1 minute;

[0015] FIGS.2A and 2B respectively illustrate a three-minute interval of the ECG signal centered on the ictal state, and an ECG signal extracted from the monitoring group;

[0016] FIG.3 illustrates the steps of the new multi-step methodology for forecasting an epileptic convulsion based on the adroit integration of diverse techniques;

[0017] FIGS.4A and 4B illustrate decomposition at L=3, using Discrete Wavelet Packet Transform and maximal overlap discrete wavelet packet transform (MODWPT);

[0018] FIGS.5A and 5B illustrate obtained frequency bands, from 1 to 16, for the monitoring group and 5 minutes before an ictal state;

[0019] FIGS.6A and 6B illustrate obtained frequency bands, from 17 to 32, for the monitoring group and 5 minutes before an ictal state;

[0020] FIGS.7A, 7B and 7C illustrate obtained distributions of the HOI values of MG and SG in 1-min intervals, twenty minutes prior to the epileptic episode for diverse j values and for FBs no.1, 16, and 32, respectively;

[0021] FIGS.8A and 8B illustrate a distribution of the HOI values obtained by two selected patterns for the ECG signals of MG and 1-minute intervals before a seizure starting from 1 to 20 minutes, FB-3, HOI-27 and FB-6, HOI-13, respectively;

[0022] FIG.9 illustrates an average accuracy for KKN with the different characteristics;Atty. Dkt. No.103361-540WO1

[0023] FIG.10 illustrates the average accuracy of the KNN classifier for each 1- minute interval graphically using a weighted distance and k=7;

[0024] FIG.11 illustrates an accuracy comparison among KNN, SVM, and NB classifiers;

[0025] FIG.12 illustrates a comparison between the methods of the present disclosure and prior methodologies; and

[0026] FIG.13 illustrates an example computing device. DETAILED DESCRIPTION

[0027] Introduction

[0028] Electrocardiographic (ECG) signals have morphologic variations caused by brain and respiratory activity. These variations inspired investigations that related the ECG to respiratory movements and brain diseases such as epilepsy. For example, these signals have been used recently for seizure detection as well as seizure prediction because epileptic discharges in the brain alter the cardiac autonomic nervous system and impact its regular operation, resulting in a promising signal source that can be employed for epileptic prediction. In this regard, the database known as Post-Ictal Heart Rate Oscillations in Partial Epilepsy (PIHROPE) has been used by a number of investigators to detect and forecast epileptic seizures.

[0029] Hereinbelow, a novel method is described through adroit integration of maximal overlap discrete wavelet packet transform (MODWPT), homogeneity index, and a K-Nearest Neighbors (KNN) classifier to predict an epileptic event twenty minutes before its onset using ECG signals.

[0030] Feature selection / detection is an element in the development of a reliable machine-learning model. The Homogeneity Index (HOI), which is a nonlinear index, can be a reliable tool for discovering relevant features in the ECG signals features that linear methods might not easily capture due to the signal nature. MODWPT is first applied to decompose the ECG signals into various modes or frequency bands; then, each mode is analyzed by employing HOI to obtain appropriate features for forecasting an ictal state. To identify the most discriminative mode(s) in combination with the values of HOI, the Kruskal-Wallis analysis (KWA), an analysis of variance, is performed. Once the most discriminant mode(s) with the values of HOI's are identified, they are used as the inputs of a K-Nearest Neighbors (KNN) classifier for automated epileptic prediction. In addition, two other machine learning- based classifiers are also examined and compared, i.e., Naïve Bayes (NB) and Support VectorAtty. Dkt. No.103361-540WO1 Machine (SVM) with different conditions (e.g., kernels, distances, etc.). The effectiveness of the method of the present disclosure is examined below using the ECG signals monitored experimentally from 7 patients that experienced seizures from the PIHROPE database.

[0031] Dataset

[0032] The ECG signals of two men and five women between 31 to 48 years are employed to verify the method's potential for predicting an ictal state. It should be noted that seven participants were analyzed; however, three participants, denoted by no.2, 3, and 6 in Table 1, presented two epileptic episodes, resulting in a total of 10 seizure episodes. In addition, the studied patients present partial seizures with or not a secondary generalization from frontal or temporal foci and without clinical evidence of heart illness. The original survey was made for Beth Israel Hospital and published in the PIHROPE open-access database. Table 1 summarizes the exact time starting each epileptic seizure.Atty. Dkt. No.103361-540WO1 Table 1. Time recording of epileptic seizures was presented during the ECG study. Participant Seizure Onset Time (hours: minutes: seconds)

[0033] Preparatig

[0034] The ECG signals evaluated in this research were approximately acquired within 3 hours utilizing a sampling frequency of 200 Hz. However, signals were resampled to the sampling frequency of 128 Hz to reduce the data to be analyzed. Further, this frequency range has also been demonstrated to be suitable for finding heart disorders. The ECG signals are convolved to perform this task with a finite impulse response (FIR) low-pass filter. From each resampled ECG signal, two-time windows are snipped as follows: (1) a time window of 20 minutes of preictal state (time window before seizure onset, which is more significant than previous studies reported in the literature) is segmented into 1-minute lapses or intervals (employed as seizure group, or SG), resulting in twenty 1-minute intervals (see FIG.1), and (2) 1-minute lapses or intervals 1 hour after or before the epileptic crisis, a time where cardiac rhythm is considered as standard or normal, which is employed as control or monitoring group (MG). FIG.1 illustrates segmentation of the pre-ictal state in 20 intervals of 1 minute. It should be noted that the selection or segmentation of time windows in 1- minute intervals is based on the prior knowledge that this time interval has allowed detecting or identifying reliable characteristics or patterns in the ECG signals for identifying and predicting illness in previous research.Atty. Dkt. No.103361-540WO1

[0035] FIG.2A illustrates a three-minute interval of the ECG signal centered on the ictal state, and FIG.2B illustrates the ECG signal extracted from the monitoring group (data acquired one hour prior to the ictal state). For the sake of illustration, FIGS.2A and 2B show the first 1-minute of the preictal state (represented by yellow zone), the ictal state (indicated by red zone), the post-ictal lapse (represented by blue area), and three 1-minute intervals of the MG, respectively. Based on this illustration, it is not possible to see visually significant discrepancies or inconsistencies between both groups. This leads us to develop a new methodology and algorithm that can identify the hidden patterns in the ECG signals of both groups with the goal of forecasting an epileptic event with reasonable accuracy and prediction time.

[0036] FIG.3 illustrates the steps of the new multi-step methodology 300 for forecasting an epileptic convulsion based on the adroit integration of diverse techniques or methods such as MODWPT, HOI, KWA, and KNN. In particular, FIG.3 is a diagram illustrating the multi-step methodology of the present disclosure. As an overview, MODWPT is part of the signal pre-processing, and HOI is the technique used for feature extraction. KWA is employed for feature selection and machine learning algorithm (MLA) utilized to forecast a seizure.

[0037] At 302A and 302B, the first 20 minutes of ECG signals before the seizure onset, a time window longer than previously reported in the literature, are isolated and divided into 1-minute lapses / intervals. Additionally, the MG is obtained by isolating 1- minute intervals of the ECG signals 1 hour prior to or after the seizure onset. It should be noted that this time prior to or after seizure onset guarantees that the patients present a normal heart rate.

[0038] At 304A and 304B, each 1-minute lapse is decomposed by utilizing MODWPT in diverse frequency bands (FB), obtained according to the chosen level, L (i.e., if the fifth level is selected, 32 FBs are obtained since each level provides 2L FB). The MODWPT is a scheme characterized by estimating a more detailed analysis of signals without a shift in the time-frequency domain, unlike traditional wavelet packet transform (WPT), where the obtained signal is decimated or down-sampled by factor 2 according to the selected level; hence, MODWPT enables the preservation of relevant patterns in the decomposed / analyzed signals, unlike traditional WPT and digital filters, where low amplitude patterns can be diminished or suppressed.

[0039] At 306A and 306B, each FB calculated by MODWPT is analyzed using HOI, where a variant index (a time-invariant integer representing the separation between samples)Atty. Dkt. No.103361-540WO1 from 1 to 100 is investigated. The capability of the nonlinear indicator HOI is being used for the estimation or extraction of reliable features from the ECG signals (a signal with nonlinear properties produced by the complex interactions within the cardiovascular system), can vary when a person will experience or undergo an epileptic convulsion.

[0040] At 308, the HOI values for all FB are evaluated using KWA, a nonparametric method, to identify and rate the most discriminant HOIs. According to the KWA results, a pattern reduction is carried out for choosing one or more patterns to anticipate an epileptic convulsion. Finally, at 310, the HOI values determined by KWA are employed as inputs for designing a supervised classifier known as KNN to predict an epileptic seizure automatically. In addition, two different classification algorithms, i.e., SVM and NB, are also investigated to compare their results with the proposed classifier. Details of the employed algorithm are explained in the following sub-sections.

[0041] Maximal Overlap Wavelet Packet Transform (WPT)

[0042] WPT is a method capable of processing time signals with non-stationary properties, e.g., EEG and ECG signals. In particular, it allows decomposing the time signal in diverse frequency bands (FB) or modes known as detail (high-pass filter) and approximation (low-pass filter) coefficients, where these coefficients are again decomposed, i.e., a multi- resolution process, in a new level of detail and approximation coefficients to form a complete binary tree as shown in FIG.4A. In particular, FIG.4A illustrates decomposition at L=3 using Discrete Wavelet Packet Transform. Though WPT has been applied in various applications, the obtained modes are down-sampled, producing distortions in (1) the edge of the estimated FB as well as (2) in their obtained signals in the details and approximations. To lessen these limitations, an improved version of WPT named Maximal Overlap Discrete Wavelet Packet Transform (MODWPT) was introduced by Percival. This variation is characterized by decomposing the signal similar to the WPT, but MODWPT allows obtaining uniform frequency bandwidths with a property known as time-invariant (this property indicates that the algorithm does not down-sample the obtained FB, see FIG.4B), which allows avoiding the edge effects in the estimated FB as well as maintaining the relevant patterns encountered in the decomposed signals. FIG.4B illustrates decomposition at L=3 using MODWPT.

[0043] According to the multi-resolution process, the MODWPT decomposes the input signalx ^ sin several coefficients, s, at level j, by convolving the analyzed signal withAtty. Dkt. No.103361-540WO1a set of low-,l ( n )and high-pass,h ( n )filters. These coefficients are estimated by thefollowing:sj ^k^^ 1^ 2z z2n^^ ^ ^l ( n ) sj^1( k ^ n )following equations are^ a2z k ^ 1 l ( 2 zj^ ^^n ) sj( n ^ k )^ ^ ^ correct selection of themother into the estimated frequency bands. In this regard, the mother wavelet known as Daubechies wavelet with order 44 is employed since it has been demonstrated to be adequate for processing time signals obtained from the human body, particularly the ECG signals.

[0046] Homogeneity Index

[0047] The Homogeneity index is known as a texture indicator, which allows for estimating the gray level variability in the pixels of an image. The mathematical formulation of HOI is a nonlinear indicator to analyze time signals. In particular, the HOI value ranges from 0 to 1, representing the degree of variability in the analyzed signal, where higher values indicate a time-varying signal and lower values indicate a quasi-stationary signal. Based on this relevant characteristic of Homogeneity, it has been employed for predicting a sudden cardiac death, detecting multiple faults in induction motors, and monitoring the health condition of mechanical and electrical systems. Hence, the variability provided by HOI in this disclosure in the context of the prediction of epileptic crisis since the ECG presents an established pattern when the cardiac rhythm is regular. Still, it can be altered due to epileptic discharges.Atty. Dkt. No.103361-540WO1

[0048] HOI of a time series signal is determined by means of the following procedure. Firstly, a new time signal or vector of differences from the input signal,xwithNsamples, is computed by: Dj^ s ^ ^ x( s ) ^ x ^ s ^ j ^(5)a separation betweensamples in terms of uniformity / homogeneity,svaries from 0 toN ^ 1according to thequantizedqlevels. It is essential to mention that this quantization process is carried out bytransforming the original time signal into a discrete-time signal, which is specified according to the resolution (number of bits) of an analog-digital converter (ADC) utilized for obtaining the original time signal. For example, if the employed ADC contains a resolution of 8 bits, it will allow quantifying / transforming the monitored time signal in 28=256 possible levels.

[0049] Once the new time signal is estimated, a histogram or a dimensionlessincidence rateh ( s )fromD ( s )is calculated, varying from^q ^ 1toq ^ 1, where q is thequantization levels of x. Subsequently, a probability functionPD ( s )of the normalizedh ( s )is estimated as follows:hj( s ) PD( s ) ^ Dj ^ s ^. Finally, the HOIisHOI^^1^ 2 PD(s)varying signals can vary according to the selected value of j. A higher value of j indicates a dispersed histogram and, therefore, a low HOI value. Conversely, a lower value of j corresponds to a uniform or compact histogram, resulting in a high HOI value. Consequently, this paper investigatesdiverse values forjto determine which one(s) is the most appropriate for epileptic crisisprediction.

[0051] Feature selection

[0052] To determine if the datasets are statistically independent, the KWA, a nonparametric algorithm, is performed. It evaluates the capability of a feature to distinguish between groups by considering both variations within each group and variations amongAtty. Dkt. No.103361-540WO1 groups. To assess this capability, the KWA is performed in four steps described in the following paragraphs.

[0053] In the first step, the rank in the overall dataset for each value is estimated; for example, the minor data obtains a status of 1, the second receives a rank of 2, and so on. Thesum of the ranksR iis calculated for each groupi ( i ^ 1, 2..., z )of sizeg i, with z samples.

[0054] In the second step,Hrepresenting the variance of the groups iscalculated by : zR2 H^ 12i ^ ^^ 3(G^ 1)chi-square distribution, where alevel significance of 0.01 is generally employed. In this regard, ifHis higher than the chi- square distribution value, the null hypothesis (group A is equal to group B) is refused, indicating that the analyzed datasets can be associated with the investigated phenomenon. On the contrary, ifHis lower than the chi-square distribution value, then the null hypothesis is accepted, indicating that the evaluated datasets have similar values; hence, they cannot be suitable for differentiating among the conditions assessed.

[0056] Finally, if the null hypothesis is refused, a probability value (pvalue ) iscomputed:^^^ ^^^^n d^^ 2^^^ 1 ^^^^Ha gamma function whichdepends onn d; i.e., for two groups, MG and SG, the gamma function value is equal to 1.

[0057] It should be noted that thepvalue indicates the capacity of the features ordatasets evaluated to be associated with the phenomenon studied. Thepvalue varies from 0 to1; hence, a value near 0 indicates a high probability ofgroups becauseAtty. Dkt. No.103361-540WO1the data are encountered in different regions ; on the contrary, apvalue of near 1 indicates thatthe evaluated datasets have similar values; so, the feature examined cannot be used fordifferentiating diverse conditions. In this research, thepvalue is employed to select and reducethe most reliable features or patterns estimated by to train and validate asupervised classifier for predicting an epileptic an automatic manner.

[0058] K-Nearest Neighbors classifier

[0059] KNN, a nonparametric technique, is known as a classifier of low computational complexity. It has presented reliable results in biomedical applications, e.g., identification of heart and brain diseases, among others.

[0060] KNN uses an input vector with k closest training samples in the pattern space. For performing the classification, it categorizes the most common class among the k nearest neighbors, defined according to the distance between the test sample and the number of neighbors to consider, determining which category the test sample belongs to. Constructing an effective KNN classifier depends on two parameters: (1) the distance measure and (2) the number of neighbors, k, employed to determine the analyzed sample class. In this research, four different distance measures have been investigated: (a) Euclidean distance, represented by Eq. (11), a straight line range between the samples in the vector space, (d) cubic distance, Eq. (12), is the distance between samples in the vector space that accounts for the covariance between samples, (c) cosine distance, Eq. (13), is calculated by subtracting the cosine similarity (measurement of the angle (cosine based) between two vectors projected in a multi- dimensional) from one, and (d) weight distance, represented by Eq. (14), is distinguished by the closest neighbors being weighted more heavily than the farther ones; in this distance measure, the nearest neighbor gets a weight of 1, the most distant neighbor a value of 0, and the other neighbors' weights are scaled linearly to the intermediate intervals. n ^ ^^2Atty. Dkt. No.103361-540WO1 n d^^ w i ( x 1 nv^x 2 nv ) 2 ^ wiis the weight for i-thnearest1 2are

[0061] On the other hand, diverse values of neighbors, from 1 to 11 (only odd numbers are investigated), are evaluated because the selection of odd numbers allows avoiding ties (i.e., the situation that occurs when the two classes get the same number of neighbors); in addition, it should be mentioned that a more extensive range can produce a misclassification of the new data because the list of neighbors may include data points far away from its neighborhood. The present disclosure evaluates these points to investigate their efficacy in predicting an epileptic seizure accurately.

[0062] Results

[0063] Following the first phase of the proposed methodology, the segmented ECG signals for both groups MG (1-minute interval) and SG (twenty 1-minute intervals) are investigated by MODWPT to decompose the analyzed EEG signals in diverse FBs. It should be pointed out that the MODWPT requires a decomposition level, which varies from 1 to 8 in this investigation. It was discovered that the fifth decomposition level is the most adequate for evaluating the ECG signals since a lower level provided wide FB, which generated a low accuracy because of the frequencies mixture; on the contrary, a higher level did not improve the epileptic prediction. FIGS.5A, 5B, 6A and 6B show the thirty-two FBs at the fifth level for the MG and the fifth 1-minute interval before seizure onset, respectively. As no notable difference between MG and the first 1-minute interval before seizure onset is observed, in the next phase of the methodology, the obtained FBs are analyzed employing the HOI to identify reliable features or patterns in the decomposed ECG signals. In particular, FIGS.5A and 5B show obtained frequency bands, from 1 to 16, for monitoring group and 5 minutes before an ictal state, respectively. FIGS.6A and 6B illustrate obtained frequency bands, from 17 to 32, for monitoring group and 5 minutes before an ictal state, respectively.

[0064] Each FB (there are 32 FBs) for all 1-minute intervals of ECG signals (211- minute intervals per each participant) estimated by MODWPT is analyzed utilizing the HOI, where the time-invariant integer, j, is varied from 1 to 100, resulting in a total of 672,000 features, which should be evaluated for discovering the most relevant features to associate them with predicting a seizure episode. For illustrative purposes, FIG.7 shows the distribution of HOI values (with a j value of 1, 50, and 100, respectively) for FBs no.1, 16,Atty. Dkt. No.103361-540WO1 and 32. Observing this figure, it is noted that diverse HOI values are obtained for each j and FB, indicating that it is essential to perform an exhaustive analysis to identify the most reliable HOI values for predicting an epileptic crisis.

[0065] Once the HOI values for all FBs are obtained, they are analyzed through KWA to discover the most appropriate patterns / features for predicting an epileptic crisis. After an exhaustive analysis of the distributions of HOI values and the p-values provided by KWA, it is identified that the HOI values of third (j=27) and sixth (j=13) FBs are the mostdiscriminant features because they present the lowestpvalues in all the 1-minute intervals,indicating that they can be employed to distinguish the ECG signal of a patient with a normal cardiac rhythm and a patient with the risk of suffering an epileptic crisis 20 minutes beforethe event. Table 2 presents thepvalue of each 1-minute interval of both selected features.Table 2. Obtainedfor the most discriminant patterns using KWA.Minutes before a pvaluesS i 3Atty. Dkt. No.103361-540WO1 17 16.8x10-31.9x10-318 14.92x10-213.7x10-3

[0066] Onlhe computation cost without altering the prediction capability. FIG.8 shows the distribution of HOI values for the third and sixth FB employing MODWPT, respectively, for the ECG signals of the control group represented by the abbreviation MG and seizure group in 1-minute intervals, 20 minutes before the seizure onset. From this figure and KWA, it is observed that the selected features present slight overlaps between both groups in various 1-minute intervals, generating uncertainty for a proper prediction of a seizure. Hence, the capabilities of the KNN classifier with overlapping features in this disclosure. In particular, FIGS.8A and 8B illustrate a distribution of the HOI values obtained by two selected patterns for the ECG signals of MG and 1-minute intervals before a seizure starting from 1 to 20 minutes for FB-3, HOI-27 and FB-6, HOI-13, respectively.

[0067] Lastly, a KNN classifier is trained and evaluated according to the selected features FB3-HOI (j=27) and FB6-HOI (j=13), chosen by KWA results. To test the efficacy and accuracy of the proposed methodology, 70% of the data is employed for the training stage, and the remaining 30% is employed for the validation purpose. A cross-validation of 10 k-fold is used to guarantee that each sample is trained and tested, which reduces the generalization error. FIG.9 illustrates the average accuracy of the twenty 1-minute intervals obtained by four types of distance and the different numbers of neighbors for the KNN classifier.

[0068] According to this figure, it can be seen that the KNN classifier with a weighted distance presents the highest accuracy, 93.25%, for all 1-minute intervals, and the averaged accuracy is superior to other distance measures. In addition, it is also observed that the value of k = 7 is suitable for predicting an epileptic seizure. Table 3 presents the accuracy obtained using the weighted distance for each evaluated k (from 1 to 11) and the 1-minute interval before the seizure onset. Table 3. Parameters used for Machine Learning Algorithms. Minutes Weight distanceAtty. Dkt. No.103361-540WO1 before a Number of Neighbors seizure 1 3 5 7 9 11 5

[0069] According to this table, it can be observed that k values of 7, 9, and 11 present the best accuracy for predicting an epileptic seizure; however, a value of 7 is used since it allows for maintaining low computational complexity and good accuracy along all 1-minute intervals. FIG.10 illustrates the average accuracy of the KNN classifier for each 1-minute interval graphically using a weighted distance and k=7. This figure shows that the identified configuration of KNN presents a reliable global accuracy; in particular, it is observed that the proposed methodology is capable of predicting an epileptic seizure with high accuracy in the first 1-minute intervals, but its accuracy is variable in the subsequent intervals. This can beAtty. Dkt. No.103361-540WO1 explained by the fact that there exists a relation between epileptic seizures and sympathetic hyperactivity because an epileptic seizure is characterized by generating slight transients or alterations in the cardiac rhythm in special EPs with partial epilepsy present variations in the lapse between ventricular depolarization and repolarization.

[0070] Therefore, the results presented in Table 3 and FIG.10 indicate that the KNN classifier with seven neighbors and weighted distance exhibits a reliable capability for distinguishing between patients with and without risk of suffering an epileptic crisis 20 minutes before its onset with an accuracy of 93.25%. The time window of 20 minutes is longer than any research reported previously, which will allow patients to take appropriate actions, such as getting away from dangerous places or situations.

[0071] Two other machine learning-based classifiers, NB and SVM, are also examined and evaluated with different conditions (e.g., kernels, distribution, etc.) for comparing the results with the proposed methodology using KNN. These machine-learning classifiers with their different parameters are investigated because they are the most employed in medical applications reported in the literature. The investigated kernels of the SVM classifier are: linear, quadratic, cubic, fine gaussian, medium gaussian, and coarse. On the other hand, in the NB classifier, two distributions, Gaussian distribution, and kernel distribution, are investigated. Both classifiers use the same data distribution used for KNN (70% for training and 30% for validation). FIG.11 shows the average accuracy obtained by the KKN classifier (blue bar) and NB and SVM classifiers with different conditions for predicting an epileptic seizure twenty minutes before its onset. According to this figure, the KNN classifier presents the highest accuracy, 93.25%, followed by SVM (91.25%) and NB (87%) for predicting an epileptic crisis 20 minutes before its onset. Hence, the proposed KNN classifier provides the best results for differentiating an MG and SG.

[0072] Comparison with Previous Works and Improvements

[0073] As described above, other researchers have presented seizure prediction models using the same ECG signal data used in the present disclosure. A comparison of the disclosed method with other known methods, including the type of the methodology employed, prediction period prior to the onset of a seizure, and accuracy and are summarized in Table 4. FIG.12 illustrates a comparison between the proposed approach and recent methodologies. Table 4. Comparative analysis of the proposed approach and alternative methodologies.Atty. Dkt. No.103361-540WO1 Work Proposed methodForecastingAccuracy time% ^

[0074] According to Table 4 and FIG.12, three notable advances of the proposed methodology can be highlighted: (1) it exhibits a reliable tool for predicting an epileptic crisis 20 minutes before its onset, where a global accuracy of 93.25% is reached, (2) increasing theAtty. Dkt. No.103361-540WO1 prediction time window by at least 25% compared with the 12 and 15 minutes of the advance prediction reported by Billeci et al., and Perez-Sanchez et al., and (3) it utilizes a nonlinear index, Homogeneity, which has demonstrated to be reliable to measure the nonlinear variations present in ECG signals, thereby extending the prediction time window. In addition, it exhibits a low computational complexity and computational burden, which can contribute to the opportunity of implementing it in a low-end processor for predicting a seizure in real time. In contrast, the other methods introduced in the literature employ an ECG signal transformation to RR intervals or heart rate value and numerous linear indices (e.g., mean, standard deviation, and root mean square, among others) to predict an epileptic seizure, requiring a high computational burden, which limits their implementation and real-time predictions.

[0075] Hence, the systems and methods disclosed herein are an improvement to predict an epileptic crisis suitably and accurately.

[0076] Computing Device

[0077] FIG.13 is an example computing device 1300 upon which the methods described herein may be implemented is illustrated. It should be understood that the example computing device 1300 is only one example of a suitable computing environment upon which the methods described herein may be implemented. Optionally, the computing device 1300 can be a well-known computing system including, but not limited to, personal computers, servers, handheld or laptop devices, multiprocessor systems, microprocessor-based systems, network personal computers (PCs), minicomputers, mainframe computers, embedded systems, and / or distributed computing environments including a plurality of any of the above systems or devices. Distributed computing environments enable remote computing devices, which are connected to a communication network or other data transmission medium, to perform various tasks. In the distributed computing environment, the program modules, applications, and other data may be stored on local and / or remote computer storage media.

[0078] In its most basic configuration, computing device 1300 typically includes at least one processing unit 1306 and system memory 1304. Depending on the exact configuration and type of computing device, system memory 1304 may be volatile (such as random access memory (RAM)), non-volatile (such as read-only memory (ROM), flash memory, etc.), or some combination of the two. This most basic configuration is illustrated in FIG.13 by dashed line 1302. The processing unit 1306 may be a standard programmable processor that performs arithmetic and logic operations necessary for operation of the computing device 1300. The computing device 1300 may also include a bus or otherAtty. Dkt. No.103361-540WO1 communication mechanism for communicating information among various components of the computing device 1300.

[0079] Computing device 1300 may have additional features / functionality. For example, computing device 1300 may include additional storage such as removable storage 1308 and non-removable storage 1310 including, but not limited to, flash media, magnetic or optical disks, or tapes. Computing device 1300 may also contain network connection(s) 1316 that allow the device to communicate with other devices. Computing device 1300 may also have input device(s) 1314 such as a keyboard, mouse, touch screen, etc. Output device(s) 1312 such as a display, speakers, printer, etc. may also be included. The additional devices may be connected to the bus in order to facilitate communication of data among the components of the components of the computing device 1300. All these devices are well known in the art and need not be discussed at length here.

[0080] The processing unit 1306 may be configured to execute program code encoded in tangible, computer-readable media. Tangible, computer-readable media refers to any media that is capable of providing data that causes the computing device 1300 (i.e., a machine) to operate in a particular fashion. Various computer-readable media may be utilized to provide instructions to the processing unit 1306 for execution. Example tangible, computer-readable media may include, but is not limited to, volatile media, non-volatile media, removable media and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. System memory 1304, removable storage 1308, and non- removable storage 1310 are all examples of tangible, computer storage media. Example tangible, computer-readable recording media include, but are not limited to, an integrated circuit (e.g., field-programmable gate array or application-specific IC), a hard disk, flash media, an optical disk, a magneto-optical disk, a floppy disk, a magnetic tape, a holographic storage medium, a solid-state device, RAM, ROM, electrically erasable program read-only memory (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices.

[0081] In an example implementation, the processing unit 1306 may execute program code stored in the system memory 1304. For example, the bus may carry data to the system memory 1304, from which the processing unit 1306 receives and executes instructions. The data received by the system memory 1304 may optionally be stored on the removable storageAtty. Dkt. No.103361-540WO1 1308 or the non-removable storage 1310 before or after execution by the processing unit 1306.

[0082] The present disclosure further contemplates a system that includes the computing device of FIG.13 as well as an ECG sensor array – such as an arrangement of electrodes configured to record ECG signals. In some implementations, the system is realized as a portable device that can be carried and / or worn by a user. For example, the system – which, again, can include the computing device shown in FIG.13 and, in some implementations, an electrode array for recording ECG signals – can be implemented as a portable heart monitor. In some implementations, the computing device shown in FIG.13 may be separate from a portable heart monitor or other device. For example, the computing device described above may be a server or other computer with sufficient computational capacity to analyze ECG signals and / or train a machine learning model. In some such implementations, a first computing device (e.g., a portable controller or computer) may record ECG signals, which are then transmitted to the computing device described above for analysis and / or training of a machine learning model. In some implementations, the computing device described above then transmits the trained machine learning model back to the first computing device, e.g., for use in monitoring for subsequent events. In other implementations, all processing of ECG signals is handled remotely from the first computing device, e.g., to reduce computational requirements of the first computing device.

[0083] The construction and arrangement of the systems and methods as shown in the various exemplary implementations are illustrative only. Although only a few implementations have been described in detail in this disclosure, many modifications are possible (e.g., variations in sizes, dimensions, structures, shapes and proportions of the various elements, values of parameters, mounting arrangements, use of materials, colors, orientations, etc.). For example, the position of elements may be reversed or otherwise varied, and the nature or number of discrete elements or positions may be altered or varied. Accordingly, all such modifications are intended to be included within the scope of the present disclosure. The order or sequence of any process or method steps may be varied or re- sequenced according to alternative implementations. Other substitutions, modifications, changes, and omissions may be made in the design, operating conditions, and arrangement of the exemplary implementations without departing from the scope of the present disclosure.

[0084] The present disclosure contemplates methods, systems, and programming products on any machine-readable media for accomplishing various operations. The implementations of the present disclosure may be implemented using existing computerAtty. Dkt. No.103361-540WO1 processors, or by a special purpose computer processor for an appropriate system, incorporated for this or another purpose, or by a hardwired system. Implementations within the scope of the present disclosure include programming products including machine- readable media for carrying or having machine-executable instructions or data structures stored thereon. Such machine-readable media can be any available media that can be accessed by a general purpose or special purpose computer or other machine with a processor. By way of example, such machine-readable media can comprise RAM, ROM, EPROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to carry or store desired program code in the form of machine-executable instructions or data structures, and which can be accessed by a general purpose or special purpose computer or other machine with a processor.

[0085] When information is transferred or provided over a network or another communications connection (either hardwired, wireless, or a combination of hardwired or wireless) to a machine, the machine properly views the connection as a machine-readable medium. Thus, any such connection is properly termed a machine-readable medium. Combinations of the above are also included within the scope of machine-readable media. Machine-executable instructions include, for example, instructions and data which cause a general-purpose computer, special purpose computer, or special purpose processing machines to perform a certain function or group of functions.

[0086] Although the figures show a specific order of method steps, the order of the steps may differ from what is depicted. Also, two or more steps may be performed concurrently or with partial concurrence. Such variation will depend on the software and hardware systems chosen and on designer choice. All such variations are within the scope of the disclosure. Likewise, software implementations could be accomplished with standard programming techniques with rule-based logic and other logic to accomplish the various connection steps, processing steps, comparison steps and decision steps.

[0087] It is to be understood that the methods and systems are not limited to specific synthetic methods, specific components, or to particular compositions. It is also to be understood that the terminology used herein is for the purpose of describing particular implementations only and is not intended to be limiting.

[0088] As used in the specification and the appended claims, the singular forms “a,” “an” and “the” include plural referents unless the context clearly dictates otherwise. Ranges may be expressed herein as from “about” one particular value, and / or to “about” anotherAtty. Dkt. No.103361-540WO1 particular value. When such a range is expressed, another implementation includes one particular value and / or to the other particular value. Similarly, when values are expressed as approximations, by use of the antecedent “about,” it will be understood that the particular value forms another implementation. It will be further understood that the endpoints of each of the ranges are significant both in relation to the other endpoint, and independently of the other endpoint.

[0089] “Optional” or “optionally” means that the subsequently described event or circumstance may or may not occur, and that the description includes instances where said event or circumstance occurs and instances where it does not.

[0090] Throughout the description and claims of this specification, the word “comprise” and variations of the word, such as “comprising” and “comprises,” means “including but not limited to,” and is not intended to exclude, for example, other additives, components, integers or steps. “Exemplary” means “an example of” and is not intended to convey an indication of a preferred or ideal implementation. “Such as” is not used in a restrictive sense, but for explanatory purposes.

[0091] Disclosed are components that can be used to perform the disclosed methods and systems. These and other components are disclosed herein, and it is understood that when combinations, subsets, interactions, groups, etc. of these components are disclosed that while specific reference of each various individual and collective combinations and permutation of these may not be explicitly disclosed, each is specifically contemplated and described herein, for all methods and systems. This applies to all aspects of this application including, but not limited to, steps in disclosed methods. Thus, if there are a variety of additional steps that can be performed it is understood that each of these additional steps can be performed with any specific implementation or combination of implementations of the disclosed methods.

Claims

Atty. Dkt. No.103361-540WO1 WHAT IS CLAIMED IS:

1. A method for training a machine learning model to predict the onset of epileptic seizure events for a subject, the method comprising: recording an electrocardiogram (ECG) signal from the subject, wherein the ECG signal includes a first signal portion recorded at least a first time period prior to an onset of an epileptic seizure event and a second signal portion recorded within a second time period of the onset of the epileptic seizure event; parsing the ECG signal into a series of signal segments of a predefined length; decomposing each of the series of signal segments into a plurality of frequency bands using a maximal overlap discrete wavelet packet transform (MODWPT) function; evaluating each of the plurality of frequency bands using a homogeneity index (HOI) to determine an HOI value for each of the plurality of frequency bands; evaluating the HOI value of each of the plurality of frequency bands using Kruskal- Wallis analysis (KWA) to identify a subset of the plurality of frequency bands that are most useful in predicting the onset of an epileptic seizure event; and training the machine learning model to predict the onset of an epileptic seizure event using the HOI value of each of the subset of the plurality of frequency bands.

2. The method of claim 1, wherein the plurality of frequency bands are evaluated using HOI according to a variant index from 1 to 100.

3. The method of claim 1, wherein evaluating the HOI value of each of the plurality of frequency bands using KWA comprises rating each HOI value, wherein the subset of the plurality of frequency bands that are most useful in predicting the onset of an epileptic seizure event are identified based on a rating their respective HOI value.

4. The method of claim 1, wherein a number of the plurality of frequency bands is predefined prior to decomposing each of the series of signal segments.

5. The method of claim 1, wherein the predefined length is one (1) minute.Atty. Dkt. No.103361-540WO1 6. The method of claim 1, wherein the machine learning model is a supervised classifier.

7. The method of claim 6, wherein the supervised classifier is a k-nearest neighbors (KNN) model.

8. The method of claim 6, wherein the supervised classifier is one of a support vector machine (SVM) or a Naïve Bayes (NB) model.

9. The method of claim 1, wherein the ECG signal is a first ECG signal, the method further comprising: recording a second ECG signal from the subject; and using the trained machine learning model to evaluate the second ECG signal to predict the onset of subsequent epileptic seizure events for the subject.

10. A system comprising: one or more processors; and memory having instructions stored thereon that, when executed by the one or more processors, causes the system to: record an electrocardiogram (ECG) signal from a subject, wherein the ECG signal includes a first signal portion recorded at least a first time period prior to an onset of an epileptic seizure event and a second signal portion recorded within a second time period of the onset of the epileptic seizure event; parse the ECG signal into a series of signal segments of a predefined length; decompose each of the series of signal segments into a plurality of frequency bands using a maximal overlap discrete wavelet packet transform (MODWPT) function; evaluate each of the plurality of frequency bands using a homogeneity index (HOI) to determine an HOI value for each of the plurality of frequency bands; evaluate the HOI value of each of the plurality of frequency bands using Kruskal- Wallis analysis (KWA) to identify a subset of the plurality of frequency bands that are most useful in predicting the onset of an epileptic seizure event; and train a machine learning model to predict the onset of an epileptic seizure event for the subject using the HOI value of each of the subset of the plurality of frequency bands.Atty. Dkt. No.103361-540WO1 11. The system of claim 10, wherein the plurality of frequency bands are evaluated using HOI according to a variant index from 1 to 100.

12. The system of claim 10, wherein evaluating the HOI value of each of the plurality of frequency bands using KWA comprises rating each HOI value, wherein the subset of the plurality of frequency bands that are most useful in predicting the onset of an epileptic seizure event are identified based on a rating their respective HOI value.

13. The system of claim 10, wherein a number of the plurality of frequency bands is predefined prior to decomposing each of the series of signal segments.

14. The system of claim 10, wherein the predefined length is one (1) minute.

15. The system of claim 10, wherein the machine learning model is a supervised classifier.

16. The system of claim 15, wherein the supervised classifier is a k-nearest neighbors (KNN) model.

17. The system of claim 15, wherein the supervised classifier is one of a support vector machine (SVM) or a Naïve Bayes (NB) model.

18. The system of claim 10, wherein the ECG signal is a first ECG signal, the instructions further causing the system to: record a second ECG signal from the subject; and predict the onset of subsequent epileptic seizure events for the subject using the trained machine learning model to evaluate the second ECG signal.

Images