System and method for seizure prediction and detection

Abstract

This disclosure provides a system and method for detecting epileptic seizures. The method may include receiving multiple electroencephalogram (EEG) signals for a subject via multiple channels, preprocessing the multiple EEG signals by segmenting the multiple EEG signals of each channel into multiple time data segments, extracting multiple features from each time data segment of each channel, and applying a machine learning algorithm to the multiple features to perform a binary classification of seizures for each time data segment of each channel. A control strategy may be employed to determine the seizure burden of the aggregated seizure binary classification. A notification may be generated when the seizure burden is equal to or exceeds a threshold. Healthcare practitioners may use the notification to assess whether the subject is likely at risk of experiencing a seizure.

Description

[0001] Cross-reference

[0002] This PCT application claims priority to U.S. Patent Application No. 16 / 578,032, filed September 20, 2019, the contents of which are incorporated herein by reference in their entirety. Background Technology

[0003] Monitoring electroencephalogram (EEG) signals is an important task in the early diagnosis of epileptic seizures. While analyzing EEG signals plays a crucial role in monitoring a patient's brain activity, expert analysis of all EEG recordings is required to detect seizure activity. This can be lengthy and time-consuming, and timely and accurate diagnosis of seizure activity is essential for initiating treatment and reducing the risk of future seizures and seizure-related complications.

[0004] Currently, machine learning algorithms offer a way to classify EEG signals while minimizing expert intervention. However, these algorithms require high-quality EEG signals to provide effective classification results. Typically, the EEG signals provided to machine learning algorithms need to be optimized to enable them to predict epileptic seizures more effectively. Summary of the Invention

[0005] The quality of the EEG signals provided to machine learning algorithms needs to be optimized to make them more effective in preventing seizures. To optimize EEG signal quality, EEG signals can be used to construct features. Each derived feature can focus on the unique characteristics of the EEG signal, allowing machine learning algorithms to more easily identify the measured EEG signal. By combining these features and then classifying them using machine learning algorithms, more effective means of predicting seizures are possible. Furthermore, by implementing control strategies and seizure burden calculations after classification, even more powerful means are possible to accurately identify seizures by removing false positives or inaccurate readings from the provided EEG signals.

[0006] One aspect of this disclosure provides a method for detecting epileptic seizures. The method may include receiving multiple electroencephalogram (EEG) signals for a subject via multiple channels. The method may further include preprocessing the multiple EEG signals by segmenting the multiple EEG signals of each channel into multiple time data segments. In some cases, the method may extract multiple features from each time data segment of each channel. In some cases, the method may apply a machine learning algorithm to the multiple features to perform a binary classification of epileptic seizures for each time data segment of each channel.

[0007] In some embodiments, preprocessing of the multiple EEG signals may further include applying filters to the multiple EEG signals on multiple channels before segmenting the multiple EEG signals. In some cases, the filters may include bandpass filters configured to filter the multiple EEG signals between 1 Hz and 35 Hz.

[0008] In some embodiments, the binary classification of seizures may include classifying each time data segment of each channel as (1) seizure-positive or (2) seizure-negative. In some cases, each time data segment may be associated with a timeepoch. In some cases, a seizure-positive classification group may indicate a potential electrocardiographic seizure at the corresponding timeepoch. In some cases, the method may further include sequentially comparing classifications across multiple times on each channel and discarding a subset of classifications if it includes fewer than three seizure-positive classifications in a single row. In some cases, the method may further include classifying a particular timeepoch as associated with a potential electrocardiographic seizure if time data segments of a subset of multiple channels are classified as seizure-positive. In some cases, the subset may include at least half of the multiple channels.

[0009] In some embodiments, the method may further include binary classification of seizures by aggregating multiple time segments from multiple channels within a moving time window. In some cases, the moving time window may range from about one minute to ten minutes. In some cases, the moving time window may be about five minutes. In some cases, the multiple channels may include at least three channels. In some cases, the multiple channels may include eight channels. In some embodiments, each time segment may have a duration ranging from about one second to twenty seconds. In some cases, the duration of each time segment may be about ten seconds.

[0010] In some embodiments, the multiple features may include time-domain and / or frequency-domain features inherent in multiple EEG signals. In some cases, the multiple features may include at least twenty different time-domain and / or frequency-domain features. In some cases, the multiple features may include multiple discrete values ​​associated with time-domain and / or frequency-domain features.

[0011] In some embodiments, the machine learning algorithm may include random forest, boosting decision tree, classification tree, regression tree, bagging tree, neural network, or rotation forest. In some cases, the machine learning algorithm may be applied individually to multiple features extracted for each channel, such that each channel has its own iteration of the machine learning algorithm.

[0012] In some embodiments, the method may further include determining the seizure burden within a moving time window based on an aggregated binary seizure classification. In some cases, the seizure burden may include the percentage of periods classified as seizure-positive. In some cases, determining the seizure burden may include averaging the seizure-positive classifications within the moving time window.

[0013] In some embodiments, the method may further include generating one or more notifications when the seizure burden is equal to or exceeds one or more thresholds. In some cases, the one or more notifications may be used by a healthcare practitioner to assess whether a subject is at risk of having a seizure. In some cases, the one or more notifications may be generated in the form of visual, audio, and / or text alerts. In some cases, a first notification indicating frequent seizure activity may be generated when the seizure burden is equal to or exceeds a first threshold of 10%. In some cases, a second notification indicating a large number of seizures may be generated when the seizure burden is equal to or exceeds a second threshold of 50%. In some cases, a third notification indicating continuous seizure activity may be generated when the seizure burden is equal to or exceeds a third threshold of 90%.

[0014] On the other hand, this disclosure provides a seizure detection system. The seizure detection system may include a preprocessing module configured to receive multiple electroencephalogram (EEG) signals for a subject via multiple channels. The preprocessing module may also be configured to preprocess the multiple EEG signals by segmenting the multiple EEG signals of each channel into multiple time data segments. The seizure detection system may further include a processing module in communication with the preprocessing module. The processing module may be configured to receive the multiple time data segments corresponding to the multiple channels. The processing module may also be configured to extract multiple features from each time data segment of each channel. The processing module may also be configured to apply a machine learning algorithm to the multiple features to perform a seizure binary classification for each time data segment of each channel.

[0015] In some embodiments, the seizure detection system may further include an output module in communication with a preprocessing module. The output module may be configured to aggregate binary classifications of seizures from multiple time segments across multiple channels within a moving time window. The output module may also be configured to determine the seizure burden within the moving time window based on the aggregated seizure binary classifications. The output module may further be configured to generate one or more notifications when the seizure burden equals or exceeds one or more thresholds.

[0016] On the other hand, this disclosure provides a system including one or more computer processors. The system may further include a memory containing machine-executable instructions that, when executed by the one or more computer processors, implement a method for detecting epileptic seizures. The method may include receiving multiple electroencephalogram (EEG) signals for a subject via multiple channels. The method may further include preprocessing the multiple EEG signals by segmenting the multiple EEG signals of each channel into multiple time data segments. The method may further include extracting multiple features from each time data segment of each channel. The method may further include applying a machine learning algorithm to the multiple features to perform a binary classification of epileptic seizures for each time data segment of each channel.

[0017] On the other hand, this disclosure provides a non-transitory computer-readable medium including machine-executable instructions that, when executed by one or more computer processors, implement a method for detecting epileptic seizures. The method may include receiving multiple electroencephalogram (EEG) signals for a subject via multiple channels. The method may include preprocessing the multiple EEG signals by segmenting the multiple EEG signals of each channel into multiple time data segments. The method may include extracting multiple features from each time data segment of each channel. The method may further include applying a machine learning algorithm to the multiple features to perform a binary classification of epileptic seizures for each time data segment of each channel.

[0018] On the other hand, this disclosure provides a method for detecting epileptic seizures. The method may include receiving multiple electroencephalogram (EEG) signals for a subject via multiple channels. The method may also include preprocessing the multiple EEG signals by segmenting the multiple EEG signals of each channel into multiple time data segments. The method may further include extracting multiple features from each time data segment of each channel. The method may further include applying a machine learning algorithm to the multiple features to perform a seizure binary classification for each time data segment of each channel. The method may further include determining the seizure burden of a moving time window based on the aggregated seizure binary classification. The method may further include presenting the seizure burden to a user.

[0019] On the other hand, this disclosure provides a method for detecting epileptic seizures. The method may include receiving multiple electroencephalogram (EEG) signals for a subject via multiple channels. The method may further include preprocessing the multiple EEG signals by segmenting the multiple EEG signals of each channel into multiple time data segments. The method may also include extracting multiple features from each time data segment of each channel, wherein each time data segment is associated with a period. The method may further include applying a machine learning algorithm to the multiple features to perform a seizure binary classification for each time data segment of each channel, thereby generating multiple classifications for the multiple time data segments, wherein the seizure binary classification for each time data segment includes classifying each time data segment of each channel as (1) seizure positive or (2) seizure negative, wherein the multiple classifications are sequentially compared across multiple periods on each channel, and if a subset of the classifications includes fewer than three seizure positive classifications in a row, then the subset is discarded. The method may further include aggregating the remaining undiscarded classifications of the multiple time data segments of the multiple channels within a moving time window. The method may further include determining the seizure burden of the moving time window based on aggregated classification, wherein the seizure burden comprises the percentage of the time data segment classified as seizure-positive, and wherein the seizure burden is a metric providing a measure of the severity or likelihood of a seizure. The method may further include generating one or more notifications when the seizure burden equals or exceeds one or more thresholds, wherein the one or more notifications indicate different seizure activities and can be used to assess whether the subject is at risk of having a seizure.

[0020] In some embodiments, a positive seizure taxonomy indicates a potential electrocardiographic seizure at the corresponding time period. In some cases, each time data segment has a duration ranging from about one second to twenty seconds. In some cases, the duration of each time data segment is about ten seconds. In some cases, the moving time window ranges from about one minute to ten minutes.

[0021] In some embodiments, one or more notifications are generated in the form of visual, audio, and / or text alerts. In some cases, a first notification indicating frequent seizure activity is generated when the seizure burden is equal to or exceeds a first threshold of 10%. In some cases, a second notification indicating a large number of seizures is generated when the seizure burden is equal to or exceeds a second threshold of 50%. In some cases, a third notification indicating continuous seizure activity is generated when the seizure burden is equal to or exceeds a third threshold of 90%.

[0022] In some embodiments, the multiple channels include at least three channels. In some cases, the multiple channels include eight channels. In some cases, the multiple features include time-domain and / or frequency-domain features inherent in the multiple EEG signals. In some cases, the multiple features include at least twenty different time and / or frequency features.

[0023] In some embodiments, the plurality of features include a plurality of discrete values ​​associated with time-domain and / or frequency-domain features.

[0024] In some embodiments, the machine learning algorithm includes random forest, boosting decision tree, classification tree, regression tree, bagging tree, neural network, or rotation forest. In some cases, the machine learning algorithm is applied individually to multiple features extracted for each channel, such that each channel has its own iteration of the machine learning algorithm.

[0025] In some embodiments, the preprocessing of the multiple EEG signals further includes applying filters to the multiple EEG signals on multiple channels prior to segmentation of the multiple EEG signals. In some cases, the filters include bandpass filters configured to filter the multiple EEG signals between 1 Hz and 35 Hz.

[0026] In some embodiments, the method may further include classifying a specific period as associated with a potential electrocardiographic seizure if a subset of time data segments from multiple channels is classified as seizure-positive. In some cases, the subset includes at least half of the multiple channels. In some cases, the sequential time periods formed by the time windows do not overlap.

[0027] In some embodiments, the method may further include outputting seizure burden as a graphical visual element to a display. In some cases, the method may include displaying one or more thresholds in the graphical visual element. In some cases, the graphical visual element includes a time-series plot, bar chart, or graph. In some cases, the time-series plot is configured to change color as the seizure burden exceeds a threshold among one or more thresholds.

[0028] In some embodiments, the method may further include using graphic visual elements to (i) assess the condition of the subject, (ii) determine the course of treatment, (iii) monitor the effectiveness of the course of treatment when it is provided to the subject, or (iv) monitor the progress of the subject's condition over time.

[0029] Additional aspects and advantages of this disclosure will readily become apparent to those skilled in the art from the following detailed description, which shows and describes only illustrative embodiments of the disclosure. It will be appreciated that other and different embodiments of the disclosure are possible, and that several details thereof can be modified in various obvious respects, all without departing from the disclosure. Therefore, the drawings and descriptions should be considered illustrative in nature and not restrictive.

[0030] Incorporate by reference

[0031] All publications, patents, and patent applications mentioned in this specification are incorporated herein by reference to the extent that each individual publication, patent, or patent application is specifically and individually indicated as being incorporated by reference. Where a publication, patent, or patent application incorporated by reference contradicts the disclosure contained in this specification, the specification is intended to supersede and / or give precedence to any such contradictory material. Attached Figure Description

[0032] The novel features of the invention are particularly set forth in the appended claims. A better understanding of the features and advantages of the invention will be obtained by referring to the following detailed description and accompanying drawings (also referred to herein as “drawings” and “figures”) illustrating illustrative embodiments in which the principles of the invention are utilized, wherein:

[0033] Figure 1 An EEG device configured to provide EEG signals to a seizure detection module according to an embodiment of the present disclosure is shown.

[0034] Figure 2 A description illustrating the workflow of an EEG signal for epileptic seizure detection according to embodiments of the present disclosure.

[0035] Figure 3 A seizure burden curve is shown according to an embodiment of the present disclosure.

[0036] Figure 4 A description of an EEG device with a display visualization of epileptic seizure detection output according to an embodiment of the present disclosure.

[0037] Figure 5 Description of EEG device software with epileptic seizure detection output according to embodiments of the present disclosure.

[0038] Figure 6 Computer systems that are programmed or otherwise configured to implement the methods provided herein, according to embodiments of the present disclosure. Detailed Implementation

[0039] While various embodiments of the invention have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many variations, modifications, and substitutions will occur to those skilled in the art without departing from the invention. It should be understood that various alternatives may be made to the embodiments of the invention described herein.

[0040] Whenever the term "at most about" or "at least about" precedes the first value in a series of two or more values, the term "at most about" or "at least about" applies to each value in the series. For example, at most about 3, 2, or 1 is equivalent to at most about 3, at most about 2, or at most about 1.

[0041] Overview

[0042] Manually examining electroencephalogram (EEG) signals can be a time-consuming and laborious process. When a patient may be experiencing a seizure, manually analyzing such EEG signals can waste valuable time. Automated technologies that can analyze EEG signals can help provide timely and accurate diagnoses of seizure activity, assisting clinicians in initiating treatment and reducing the risk of future seizures and seizure-related complications.

[0043] Machine learning algorithms offer a way to automatically classify EEG signals with minimal expert intervention. However, machine learning algorithms require high-quality EEG signals to provide effective classification results. Furthermore, the EEG signals provided to machine learning algorithms may need to be given as features describing the characteristics of EEG signals related to seizure activity. Moreover, post-classification of features through machine learning algorithms, control strategies including a set of rules, and seizure burden calculation allows methods and / or systems to more accurately depict whether a subject is experiencing or may experience seizures.

[0044] The seizure detection system and method described herein provide a user with an EEG detection device coupled to a seizure detection module capable of automatically and accurately detecting seizures. Furthermore, the seizure detection module can notify the user of an impending / active seizure. The seizure detection module acquires and preprocesses EEG signals from the EEG device. Then, the EEG signals are analyzed and valuable EEG features are extracted. The EEG features are classified using a machine learning algorithm module. Next, the classification of features for a given period is controlled by a control strategy to calculate a seizure burden value. If the seizure burden value is equal to or exceeds one or more thresholds, the method and / or system generates one or more notifications that can be used by healthcare practitioners to assess whether a subject is likely at risk of having a seizure. Healthcare practitioners can use the increased seizure burden value as an indication of increased seizure severity. Healthcare practitioners can use a seizure burden value equal to or exceeding one or more thresholds as an indication of a medical condition (e.g., status epilepticus). Healthcare practitioners can use the changes in seizure burden over time or the characteristic shape of a seizure burden curve to determine the course of treatment for a subject or to assess the effectiveness of the course of treatment.

[0045] Epilepsy Seizure Detection Module

[0046] I. Signal Acquisition and Preprocessing

[0047] II. Signal Analysis

[0048] III. Seizure Burden Calculation and Output

[0049] IV. Post-seizure testing

[0050] V. Computer Systems

[0051] I. Signal Acquisition and Preprocessing

[0052] (a) EEG signal / acquisition

[0053] For ease of explanation, the following figures and corresponding descriptions are described with reference to the analysis of signals representing brain activity (e.g., electroencephalogram (EEG) signals) and / or cardiac activity (e.g., electrocardiogram (ECG) signals) in living organisms. However, those skilled in the art will recognize that signals representing other bodily functions (e.g., electromyography (EMG) signals, or electronystagmography (ENG) signals, pulse oximetry signals, carbon dioxide maps, and / or photoplethysmography signals) may be used instead of (e.g., in combination with) one or more signals representing brain activity and / or cardiac activity.

[0054] Systems for measuring bioelectrical signals typically include one or more electrodes electrically coupled to a controller and / or output device via corresponding conductive wires. In other variations, the electrodes may be wirelessly coupled to the controller and / or output device. The electrodes may be contained within an electrode carrier system fixed around the patient's head. The electrode carrier system may be configured as a headband or incorporated into any number of other platforms or positioning mechanisms for holding the electrodes on the patient's body. Individual electrode assemblies may be spaced apart from each other so that, when the headband is positioned on the patient's head, the electrode assemblies are optimally aligned for receiving EEG signals.

[0055] The controller and / or output device may typically include any number of devices for receiving electrical signals (e.g., electrophysiological monitoring devices) and may also be used in combination with any number of brain imaging devices (e.g., fMRI, PET, NIRS, etc.). In a particular variant, the electrode embodiments described herein may be used in combination with, for example, devices configured to receive and process electrical signals from the electrodes.

[0056] In some embodiments, signals corresponding to brain electrical activity are obtained from the human brain, and these signals correspond to electrical signals obtained from a single neuron or from multiple neurons. In some embodiments, the sensor comprises one or more sensors (e.g., extracranial sensors) externally fixed (e.g., glued, attached, adhesive) to the human scalp. For example, an extracranial sensor may comprise electrodes (e.g., electroencephalogram (EEG) electrodes) or multiple electrodes (e.g., EEG electrodes) externally fixed to the scalp (e.g., adhesive to the skin via conductive gel), or more generally positioned at a corresponding location outside the scalp. Alternatively, in some embodiments, dry electrodes (e.g., conductive sensors mechanically placed on the body of a living person rather than implanted within the body of a living person, or held in place by conductive gel) may be used. An example of a dry electrode is a headband with one or more metallic sensors (e.g., electrodes) worn by a living person during use. Signals obtained from extracranial sensors are sometimes referred to as EEG signals or time-domain EEG signals. In some cases, the sensor may be an accelerometer or inertial measurement unit (IMU) capable of measuring the mechanical movement of the subject and / or device (e.g., generating one or more electrical signals corresponding to the mechanical movement of the subject and / or device). The system may be configured to utilize one or more sensors to assist in the detection of seizures as described elsewhere herein.

[0057] On one hand, this disclosure provides a method for detecting epileptic seizures. In some cases, the method may include receiving multiple signals (e.g., EEG signals, EKG signals, EMG signals, etc.) for a subject via multiple channels. The method may include receiving multiple electroencephalogram (EEG) signals for a subject via multiple channels. The multiple EEG signals may be provided to an epileptic seizure detection module. Figure 1This document describes the workflow of collecting EEG signals from the EEG device module 110 to the seizure detection module 120. Figure 2 This section provides a detailed explanation of the workflow for using the EEG device module and the seizure detection module for seizure prediction. (For example...) Figure 2 As shown, the seizure detection module may include a preprocessing module, a signal analysis module, and a seizure burden calculation and output module. The EEG device module 110 may have multiple channels 205 for acquiring EEG signals from the subject. In some cases, the multiple channels may have between 1 and 256 channels. In some cases, the multiple channels may have between 8 and 256 channels. In some cases, the multiple channels may have more than 256 channels. In some cases, the multiple channels may have 8, 10, 16, 20, 32, 64, 128, or 256 channels.

[0058] (b) Preprocessing of EEG signals

[0059] In some embodiments, the EEG device module may have one or more analog front ends configured to receive sensor EEG signals from sensors. The EEG signals may be preprocessed as described elsewhere herein. In some embodiments, a separate (e.g., independent) analog front end may be provided for interfacing with each of a set of sensors. In some embodiments, one or more analog front ends may be provided for interfacing with a set of EEG sensors.

[0060] In some embodiments, the method may include preprocessing multiple signals by dividing multiple signals of each channel into multiple time data segments. In some embodiments, the method may include preprocessing multiple EEG signals by dividing multiple EEG signals of each channel into multiple time data segments. Figure 2 Description of the seizure detection module 120. The seizure detection module acquires EEG signals from multiple channels of the EEG device module. The seizure detection module can preprocess the EEG signals from the multiple channels using a preprocessing module 210 configured to preprocess the EEG signals. For example... Figure 2 As shown in the figure, the preprocessing module may include a signal filtering module 215, a signal segmentation module 220, and a signal adjustment module 225.

[0061] In some embodiments, the filtering module 215 may be configured to filter EEG signals from the input channel group of the EEG device module, as described elsewhere herein. In some cases, for example, preprocessing may include segmenting the EEG signal, filtering the EEG signal based on frequency, adjusting the EEG signal, or preprocessing as described elsewhere herein.

[0062] exist Figure 2In this configuration, the signal segmentation module 220 can be configured to segment EEG signals. In some embodiments, multiple EEG signals can be segmented into data segments between 1 and 100,000. In some cases, the number of EEG data segments may depend on the duration of the EEG recording. In other cases, the number of EEG data segments may be fixed and independent of the duration of the EEG recording.

[0063] In some embodiments, each time segment may have a duration between approximately 1 second and 1 hour. In some cases, each time segment may have a duration between approximately 1 second and 30 seconds. In some cases, each time segment may have a duration between approximately 1 second and 10 seconds. In some cases, the duration of each time segment may be fixed for the entire EEG recording. In some cases, the duration of each time segment may be variable or adaptive during the EEG recording.

[0064] In some embodiments, preprocessing of multiple EEG signals may include applying one or more filtering steps to multiple EEG signals on multiple channels. Preprocessing of multiple EEG signals may include using at least one filter, two filters, three filters, four filters, five filters, six filters, seven filters, eight filters, nine filters, ten filters, fifteen filters, or more filters. Preprocessing of multiple EEG signals may include using up to about 15 filters, ten filters, nine filters, eight filters, seven filters, six filters, five filters, four filters, three filters, two filters, or fewer filters. Preprocessing of multiple EEG signals may include using between one and 15 filters, between one and 10 filters, between one and five filters, or between one and three filters as needed.

[0065] In some embodiments, one or more filtering steps may be applied before, during, and / or after the segmentation of multiple EEG signals. One or more of the filtering steps may include, for example, digital filters, analog filters, or combinations thereof. One or more of the filtering steps may include, for example, bandpass filters, low-pass filters, high-pass filters, band-stop filters, all-pass filters, Kalman filters, adaptive filters, or notch filters, etc. In some cases, the low-frequency cutoff of the filter may be between 0.1 Hz and 5 Hz. In some cases, the high-frequency cutoff of the filter may be between 5 Hz and 200 Hz. In some cases, the notch filter frequency may be matched to the local power line frequency. In some cases, the notch filter frequency may be 50 Hz or 60 Hz to match the local power line frequency.

[0066] In some embodiments, each time data segment may be associated with a period. For each corresponding period, a positive seizure taxonomy may indicate a potential electrocardiographic seizure. In some cases, a positive seizure taxonomy may include between approximately 1 and 50 positive seizure categories. In some cases, a positive seizure taxonomy may include between 1 and 10 positive seizure counts.

[0067] In some embodiments, the method may further include sequentially comparing classifications across multiple periods in each channel. In some cases, sequential classifications across multiple periods in each channel may be discarded before / after / during the sequential comparison of classifications across multiple periods in each channel. In some cases, a subset of classifications may be discarded. In some cases, a subset of fewer than about 1 to 20 classifications may be discarded. In some cases, a subset of fewer than 3 classifications may be discarded.

[0068] In some embodiments, a subset of positive seizure classifications may be discarded because (for example) it may be random readings, have low reliability, inaccurate classification, incorrect classification, calibration, systematic errors, disconnected electrodes, human signals, system interference, or other signals, etc.

[0069] In some embodiments, a subset of positive seizure classifications may be discarded to (for example) save memory space, improve processing speed, reduce energy consumption, reduce system heat, reduce computing costs, save processing power, save processing time, increase reliability, or reduce random access memory usage, etc.

[0070] In some embodiments, a larger number of positive seizure classifications in a row can indicate high reliability. The greater the reliability of the positive seizure classifications, the more accurate the determination of a patient's seizure. In some cases, greater reliability of positive seizure classifications can indicate the accuracy of the machine learning algorithm, the quality of the data (EEG signal), or the health of the EEG detection system. In some embodiments, if a subset of time data segments from multiple channels is classified as seizure-positive, then a specific period can be classified as associated with a potential EEG seizure. In some cases, the subset may be at least 5%, 10%, 20%, 30%, 40%, 50%, or more from multiple channels. In some cases, the subset may be at most about 50%, 40%, 30%, 20%, 10%, 5%, or less from multiple channels.

[0071] (c) EEG signal adjustment

[0072] Figure 2A signal adjustment module 225 configured to adjust EEG signals is shown. In some embodiments, the method can adjust any EEG signal. Adjusting an EEG signal may include, for example, increasing and / or decreasing the amplitude of the EEG signal, adding or decreasing the noise level of the EEG signal, increasing and / or decreasing the duration of the EEG signal, increasing and / or decreasing the intensity of the EEG signal, increasing and / or decreasing the signal frequency of the EEG signal, increasing and / or decreasing the voltage of the EEG signal, changing the morphology of the EEG signal (e.g., the shape of the EEG signal), increasing and / or decreasing the periodicity of the EEG signal, increasing or decreasing the synchronicity of the EEG wave, spectral attenuation, normalization, etc.

[0073] In some cases, the EEG signal can be reduced. In other cases, the EEG signal can be downsampled to a lower sampling frequency. For example, EEG data recorded at a sampling frequency of 500 Hz can be downsampled by a factor of 2 to 250 Hz.

[0074] In some cases, the EEG signal can be reduced in bit width. In some cases, the method may not require a high resolution level for recording the EEG signal to achieve accurate seizure detection. In some cases, bit width reduction can be achieved by standardizing the EEG signal to reduce the signal to a lower number of bits per sample, for example, from 32 bits per sample to 12 bits per sample. In some cases, bit width reduction can be advantageous if the method is to be implemented in a portable system, as it may be useful for reducing power consumption due to the reduced processing load.

[0075] In some cases, spectral subtraction can be used to reduce the amount of additive noise in an EEG signal. In some cases, the noise may be caused by the external environment. In some cases, the noise may be caused by the measuring equipment. In some cases, the noise may be caused by the user. In some cases, the average spectrum of the non-epileptic EEG signal can be calculated over a period of time to provide a basic estimate of the noise spectrum level. In some cases, the EEG signal can be converted to the frequency domain when it is recorded. In some cases, the average noise spectrum can then be subtracted from the EEG spectrum. In some cases, the resulting spectrum can be combined with phase information from the original noise signal. In some cases, the resulting spectrum can be converted back to the time domain to produce a denoised signal.

[0076] In some embodiments, EEG signals can be normalized by eliminating the montage effect used to collect the EEG signal. In some cases, independent component analysis (ICA) or principal component analysis (PCA) methods can be used to provide montage elimination. In some cases, ICA or PCA methods can separate the EEG signal into a set of sources independent of the montage used to record it. In some cases, using normalized EEG data can remove errors introduced by different clinicians' practices.

[0077] In some cases, nonnegative matrix factorization (NMF) can be applied to each channel as a form of artifact removal. In other cases, the signal's spectrum can be decomposed into extracted bases to obtain weights. In still other cases, the spectrum can be reconstructed using the artifact bases and the corresponding weights removed from the initial EEG signal.

[0078] II. Signal Analysis

[0079] (a) Feature extraction

[0080] exist Figure 2 In this system, the seizure detection module may include a signal analysis module. The signal analysis module may include a feature extraction module 245 and a machine learning classification module 250. The feature extraction module 245 may be configured to acquire preprocessed measurement data (e.g., EEG signals) from the preprocessing module 210 to construct derived values ​​(e.g., features). In some embodiments, feature extraction may begin with a set of initial measurement data (e.g., EEG signals, EEG signals at a given time period, etc.) and may construct non-redundant derived values ​​(e.g., features) intended to provide information. In some cases, the feature extraction module may include extracting multiple features individually from each time segment of each channel. In some cases, the feature extraction module may include extracting multiple features together from each time segment of all channels. In some cases, the feature extraction module may include extracting multiple features from each time segment of one or more groups, where each group consists of one or more channels. Figure 2 As shown, the extracted features can be forwarded to the machine learning classification module 250, which can be configured to analyze and classify the extracted features, as described elsewhere in this document. In some cases, feature extraction can facilitate subsequent learning and generalization steps of the machine learning algorithm. In some cases, feature extraction can lead to better human interpretation. In some cases, feature extraction may be related to dimensionality reduction.

[0081] In some cases, when the input data for a machine learning algorithm (e.g., EEG signals) is too large to process and appears redundant (e.g., identical measurements in Hz and seconds, or repetitive characteristics), the data can be converted into a set of simplified features.

[0082] In some cases, determining a subset of initial features may be referred to as feature selection. In some cases, it may be expected that the selected features contain relevant information from the input data (e.g., EEG signals). In some cases, it may be expected that the selected features contain relevant information from the input data, making it possible to perform the desired task by using this simplified representation instead of the complete initial data.

[0083] In some embodiments, feature extraction may involve reducing the amount of resources required to describe large datasets (e.g., EEG signals). In some cases, analyses with a large number of variables may require significant memory and computational power. In some cases, this can lead to machine learning algorithms overfitting the training samples and failing to generalize to new samples. In some cases, feature extraction can construct combinations of variables to accurately describe the data with sufficient precision. In some cases, feature extraction can construct combinations of variables to accurately describe the data with sufficient precision while preventing overfitting.

[0084] In some embodiments, a constructed application-relevant feature set can be used to improve the results. In some cases, the constructed set can be built by experts. In some cases, general dimensionality reduction techniques can be used. In some cases, general dimensionality reduction techniques may be (for example) independent component analysis, isometric mapping, core PCA, latent semantic analysis, partial least squares, principal component analysis, multifactor dimensionality reduction, nonlinear dimensionality reduction, multilinear principal component analysis, multilinear subspace learning, semidefinite embedding, autoencoders, etc.

[0085] In some cases, a set of numerical features can be described by feature vectors. In some cases, a feature vector can be an n-dimensional vector representing the numerical features of an object.

[0086] In some embodiments, the data analysis software package can provide feature extraction. In some cases, the data analysis software package can provide dimensionality reduction. In some cases, the data analysis software package may include a programming environment, such as MATLAB, SciLab, NumPy, or R. In some cases, programming language scripts can be used to extract features from EEG signals. In some cases, the programming language script may be (for example) MATLAB, Python, Java, JavaScript, Ruby, C, C++, or Perl.

[0087] In some cases, multiple features may be inherent in multiple EEG signals. Inherent features can be measurable characteristics of the EEG signal, such as the amplitude, duration, variation, power, local maxima / minimum, pattern, regularity, spectral power distribution, or frequency. In some cases, multiple features may be a measurement of the power of a signal with a specific frequency. The frequency may be (for example) from approximately 0 Hz to 100 Hz. In some cases, the signal power may be normalized to the total power. In some cases, the signal power may be the power ratio between one or more signals. In some cases, features may be functions applied to the signal to obtain a value. For example, the function may measure the root mean square (RMS) of the signal (e.g., the EEG signal) to obtain the RMS value of the signal. In some cases, features may compare a signal (e.g., the EEG signal) with one or more signals. In some cases, features can compare one or more signals (e.g., an EEG signal) with one or more other signals. In some cases, features can measure properties of a signal (e.g., an EEG signal). In some cases, features can compare one or more properties of a signal (e.g., an EEG signal) with one or more other properties of the signal. Properties can be, for example, inherent properties of the EEG signal. In some cases, the features of the EEG signal can be continuous and / or discrete in time.

[0088] In some cases, multiple features may include at least twenty different time and / or frequency features. In some cases, multiple features may include up to one thousand time and / or frequency features. In some cases, multiple features may include between approximately 10 and 200 features. In some cases, multiple features may include between approximately 10 and 100 features. In some cases, multiple features may include between approximately 10 and 50 features.

[0089] In some cases, multiple features may contain multiple discrete values ​​associated with time-domain, frequency-domain, time-frequency-domain, information theory, and nonlinear dynamics system theory features. In some cases, multiple features may contain multiple discrete values ​​associated with time-domain and / or frequency-domain features. Multiple features may contain multiple continuous values ​​associated with time-domain and / or frequency-domain features.

[0090] In some cases, multiple signals can be converted into digital signals. In other cases, multiple signals can be converted into digital signals and then into analog signals.

[0091] In some cases, features can be sampled from a portion of the EEG signal. Sampling from a portion of the EEG signal can reduce the required processing time and power.

[0092] In some embodiments, features may be associated with a weight value. A weight value may assign a higher score to a feature for detecting seizures. A higher score may indicate that the feature is more relevant to predicting seizure activity. The method may adjust the weight value of any feature at any given time. The method may be adjusted by increasing and / or decreasing the weight value of any feature at any given time.

[0093] (b) Classification using machine learning

[0094] In some embodiments, the method may include applying a machine learning algorithm to multiple features to perform seizure classification individually for each time segment of data for each channel. In some cases, the machine learning classification module may include performing seizure classification together for each time segment of data for all channels. In some cases, the machine learning classification module may include performing seizure classification for each time segment of data for one or more groups, wherein each group consists of one or more channels. Figure 2 This demonstrates a machine learning classification module 250, which can acquire / extract features from a preprocessing step and classify those features. In some cases, features can be extracted without a preprocessing step.

[0095] In some cases, machine learning algorithms may need to extract and visualize relationships between features because conventional statistical techniques may be insufficient. In other cases, machine learning algorithms can be used in conjunction with conventional statistical techniques. In still others, conventional statistical techniques can provide preprocessed features to machine learning algorithms.

[0096] In some embodiments, multiple features can be classified into any number of categories. Time periods can be classified (for example) as seizure-positive, seizure-negative, seizure-like, indeterminate seizure activity, etc. In some cases, multiple features can be classified into between 1 and 20 categories. Individual categories can also be divided into subcategories. For example, time periods classified as seizure-positive can be further subdivided into focal and generalized seizure events.

[0097] In some embodiments, the method may include applying a machine learning algorithm to multiple features to perform a binary classification of seizures for each time segment of data in each channel.

[0098] In some embodiments, one or more features collected may be discarded before or during machine learning classification.

[0099] In some embodiments, a human can select and discard features before / during machine learning classification. In some cases, a computer can select and discard features. In some cases, features can be discarded based on a threshold.

[0100] In some embodiments, any number of features can be classified using a machine learning algorithm. The machine learning algorithm can classify at least 10 features. In some cases, the number of features may range from approximately 10 to 200. In some cases, the number of features may range from approximately 10 to 100. In some cases, the number of features may range from approximately 10 to 50. In some embodiments, the machine learning algorithm may be, for example, an unsupervised learning algorithm, a supervised learning algorithm, or a combination thereof. Unsupervised learning algorithms may be, for example, clustering, hierarchical clustering, k-means, mixture models, DBSCAN, OPTICS algorithm, anomaly detection, local outlier factorization, neural networks, autoencoders, deep belief networks, Hebbian learning, generative adversarial networks, self-organizing maps, expectation-maximization (EM) algorithm, moment estimation, blind signal separation techniques, principal component analysis, independent component analysis, nonnegative matrix factorization, singular value decomposition, or a combination thereof. Supervised learning algorithms can be, for example, support vector machines, linear regression, logistic regression, linear discriminant analysis, decision trees, k-nearest neighbors, neural networks, similarity learning, or combinations thereof. In some embodiments, machine learning algorithms can include deep neural networks (DNNs). Deep neural networks can include convolutional neural networks (CNNs). CNNs can be, for example, U-Net, ImageNet, LeNet-5, AlexNet, ZFNet, GoogleNet, VGGNet, ResNet18, or ResNet. Other neural networks can be, for example, deep feedforward neural networks, recurrent neural networks, LSTM (Long Short-Term Memory), GRU (Gated Recurrent Unit), autoencoders, variational autoencoders, adversarial autoencoders, denoising autoencoders, sparse autoencoders, Boltzmann machines, RBMs (Restricted BMs), deep belief networks, generative adversarial networks (GANs), deep residual networks, capsule networks, or attention / transformer networks, etc.

[0101] In some embodiments, the machine learning algorithm may be, for example, a random forest, a boosting decision tree, a classification tree, a regression tree, a bagging tree, a neural network, or a rotating forest. The machine learning algorithm may be applied individually to multiple features extracted for each channel, such that each channel can have a separate iteration of the machine learning algorithm.

[0102] In some embodiments, the method may apply one or more machine learning algorithms. In some embodiments, the method may apply one or more machine learning algorithms to each channel.

[0103] exist Figure 2In this embodiment, the machine learning classification module 250 may include any number of machine learning algorithms. In some embodiments, the random forest machine learning algorithm may be a set of bagged decision trees. In some cases, the set of bagged decision trees may classify each time segment of data in each channel as (1) seizure positive or (2) seizure negative. The set may be at least about 1, 2, 3, 4, 5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 120, 140, 160, 180, 200, 250, 500, 1000 or more bagged decision trees. The set may be at most about 1000, 500, 250, 200, 180, 160, 140, 120, 100, 90, 80, 70, 60, 50, 40, 30, 20, 10, 5, 4, 3, 2 or fewer bagged decision trees. The set can be approximately 1 to 1000, 1 to 500, 1 to 200, 1 to 100, or 1 to 10 bagged decision trees.

[0104] In some embodiments, the method may include applying a machine learning classifier to any number of channels. The method may include applying a machine learning classifier to at least about 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, 25, 50, 100, 500, 1000 or more channels. The method may include applying a machine learning classifier to at most about 1000, 500, 100, 50, 25, 20, 15, 10, 9, 8, 7, 6, 5, 4, 3, 2 or fewer channels. The method may include applying a machine learning classifier to about 1 to 1000, 1 to 100, 1 to 25, or 1 to 5 channels.

[0105] In some embodiments, the method may include applying a machine learning classifier to a subset of channels. The subset of channels may be at least about 1%, 5%, 10%, 20%, 30%, 40%, 50%, or more of the total number of channels. The subset of channels may be at most about 50%, 40%, 30%, 20%, 10%, 5%, 1%, or less of the total number of channels. The subset of channels may be from about 1% to 50%, 1% to 40%, 1% to 30%, 1% to 20%, 1% to 10%, or 1% to 5% of the total number of channels.

[0106] In some embodiments, the machine learning algorithm may have various parameters. These parameters may include, for example, the learning rate, mini-batch size, number of training epochs, momentum, learning weight decay, or neural network layers.

[0107] In some embodiments, the learning rate may be between about 0.00001 and 0.1.

[0108] In some embodiments, the small batch size can be between approximately 16 and 128.

[0109] In some embodiments, the neural network may include neural network layers. The neural network may have at least about 2 to 1000 or more neural network layers.

[0110] In some embodiments, the number of training epochs may be at least about 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 25, 30, 35, 40, 45, 50, 55, 60, 65, 70, 75, 80, 85, 90, 95, 100, 150, 200, 250, 500, 1000, 10000 or greater.

[0111] In some embodiments, the momentum may be at least about 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9 or greater. In some embodiments, the momentum may be at most about 0.9, 0.8, 0.7, 0.6, 0.5, 0.4, 0.3, 0.2, 0.1 or less.

[0112] In some embodiments, the learning weight decay may be at least about 0.00001, 0.0001, 0.001, 0.002, 0.003, 0.004, 0.005, 0.006, 0.007, 0.008, 0.009, 0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, 0.1 or greater. In some embodiments, the learning weight decay may be up to about 0.1, 0.09, 0.08, 0.07, 0.06, 0.05, 0.04, 0.03, 0.02, 0.01, 0.009, 0.008, 0.007, 0.006, 0.005, 0.004, 0.003, 0.002, 0.001, 0.0001, 0.00001 or less.

[0113] In some embodiments, the machine learning algorithm may use a loss function. The loss function may be, for example, regression loss, mean absolute error, mean bias error, hinge loss, Adam optimizer, and / or cross-entropy.

[0114] In some embodiments, the parameters of the machine learning algorithm can be adjusted with the help of a human and / or computer system.

[0115] In some embodiments, machine learning algorithms may prioritize certain features. For example, a machine learning algorithm may prioritize features that are more relevant to detecting seizures. If a feature has a higher classification frequency than another feature, then that feature is likely more relevant to detecting seizures. In some cases, a weighted system may be used to prioritize features. In some cases, features may be prioritized probabilistically based on their frequency and / or quantity. Machine learning algorithms may prioritize features with the assistance of humans and / or computer systems.

[0116] In some embodiments, one or more of the features may be used in conjunction with machine learning or conventional statistical techniques to determine whether a segment may contain artifacts. Figure 2 The artifact rejection module 255 is demonstrated to identify segments containing artifacts. Identified artifacts can be the result of electrical interference, electrode instability or motion, subject movement, subject eye movement or blinking, subject chewing, subject muscle tension, subject electrocardiogram artifacts, etc. In some cases, motion sensors or other sensors can be used as additional inputs to the artifact rejection module. In some cases, identified artifacts can be rejected for use in seizure classification. In some cases, identified artifacts can be reduced, canceled, or eliminated while still being able to process the remaining signal for seizure classification.

[0117] In some cases, machine learning algorithms can prioritize certain features to reduce computational costs, save processing power, save processing time, increase reliability, or reduce random access memory usage.

[0118] III. Seizure Burden Calculation and Output

[0119] (a) Control strategies and seizure burden

[0120] In some embodiments, seizure binary classification may include classifying each time segment of data for each channel as (1) seizure positive or (2) seizure negative. Seizure binary classification may use machine learning algorithms as described elsewhere herein. The method may include aggregating seizure binary classifications of multiple time segments from multiple channels within a moving time window. The aggregated seizure classifications may be subject to a control strategy module 275 of the seizure burden calculation and output module 270, such as… Figure 2 It is displayed in the middle. Figure 2 Demonstrates the epileptic seizure burden calculation and output module 270. (For example...) Figure 2As shown, the seizure burden calculation and output module 270 may include a control strategy module 275, a seizure burden calculation module 280, a seizure burden plotting module 285, and a seizure burden notification module 290. Specifically, the control strategy module 275 can be configured to implement control strategies, the seizure burden calculation module 280 can be configured to calculate seizure burden values, the seizure burden plotting module 285 can be configured to plot seizure burden values, and the seizure burden notification module 290 can be configured to provide notifications as described elsewhere herein.

[0121] In some cases, the moving window may have a time period between 1 minute and 1 hour. In some cases, the time period of the moving window may be dynamic or adjustable rather than fixed. In some cases, the time period of the moving window may depend on the subject.

[0122] In some embodiments, a positive seizure classification on one or more channels may be subject to a control strategy module 275 to result in an overall determination of the patient’s seizures at the corresponding time.

[0123] A control strategy may be a set of rules that lead to an overall determination of a patient's seizure frequency at a corresponding period. The control strategy may take a set of parameters as input and act on those parameters according to the set of rules to lead to an overall determination of a patient's seizure frequency at a corresponding period. The set of rules may be as described elsewhere herein. The set of rules may be adjusted at any point in time to act on more or fewer parameters. The set of rules may be adjusted at any point in time to include more rules or remove rules. The set of rules may consist of at least about 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, 25, 50, 100, 500, 1000 or more rules. The set of rules may consist of at most about 1000, 500, 100, 50, 25, 20, 15, 10, 9, 8, 7, 6, 5, 4, 3, 2 or fewer rules. The group rules can be from approximately 1 to 1000, 1 to 500, 1 to 100, 1 to 25, 1 to 10, 1 to 5, or 1 to 3 rules.

[0124] In some embodiments, the parameter inputs of the control strategy may include the number of channels classified as seizure-positive, the number of channels classified as seizure-negative, the classification of channels as seizure-positive, the classification of channels as seizure-negative, the corresponding period, the number of channels, the machine learning algorithm used for classification, the length of the moving window, the connection quality of each channel, information derived from the EKG signal, information derived from the EMG signal, information about the patient's demographics, information about the patient's current or previous condition, information about the treatment or medication applied to the patient, information derived from motion sensors (e.g., accelerometers or inertial measurement units), etc.

[0125] In some embodiments, the control strategy may have any number of parameter inputs. The control strategy may have at least about 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 20, 25, 50, 100, 500, 1000 or more parameter inputs. The control strategy may have at most about 1000, 500, 100, 50, 25, 20, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2 or fewer parameter inputs. The control strategy may have from about 1 to 1000, 1 to 500, 1 to 100, 1 to 50, 1 to 25, 1 to 15, 1 to 10 or 1 to 5 parameter inputs.

[0126] In some embodiments, the group rule may specify the classification of the channel to be discarded by the control strategy. For example, if the control strategy receives input of a single positive seizure classification for a corresponding period, the group rule may discard the positive seizure classification for that period. In some cases, the control strategy may receive at least about 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 50, 100, 500 or more positive seizure classifications, and the group rule may discard each positive seizure classification for the corresponding period. In some cases, the control strategy may receive up to about 500, 100, 50, 10, 9, 8, 7, 6, 5, 4, 3, 2 or fewer positive seizure classifications, and the group rule may discard each positive seizure classification for the corresponding period. In some cases, the control strategy may receive from about 1 to 500, 1 to 100, 1 to 50, 1 to 10 or 1 to 5 positive seizure classifications, and the group rule may discard each positive seizure classification for the corresponding period.

[0127] In some embodiments, the group rule may specify that the control strategy outputs a set of positive seizure classifications for a corresponding period. For example, if the control strategy receives a set of four or more channels, where each channel is registered with a positive seizure classification for a corresponding period, then the group rule may output the positive seizure classification for that period. In some cases, the control strategy may receive a set of positive seizure classifications for at least about 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, 25, 50, 100, 1000 or more channels, and the group rule may output the positive seizure classification for that set of positive seizure classifications for the corresponding period. In some cases, the control strategy may receive a set of positive seizure classifications for at most about 1000, 100, 50, 25, 20, 15, 10, 9, 8, 7, 6, 5, 4, 3, 2 or fewer channels, and the group rule may output the positive seizure classification for that set of positive seizure classifications for the corresponding period. In some cases, the control strategy may receive a set of positive seizure classifications ranging from approximately 1 to 1000, 1 to 500, 1 to 100, 1 to 50, 1 to 25, 1 to 10, or 1 to 5, and the set of rules may output a positive seizure classification for the set of positive seizure classifications for the corresponding period.

[0128] In some embodiments, the method may include calculating a patient's seizure burden as a percentage of positive seizure classifications over a specified time period. For example... Figure 2 As shown, the seizure burden calculation module 280 can be configured to calculate a patient's seizure burden. In some cases, the time period used for seizure burden calculation can be between 1 minute and 1 hour. In some cases, the time period used for seizure burden calculation can be the entire recording session. In some cases, the time period used for seizure burden calculation can be dynamic or adjustable rather than fixed.

[0129] In some embodiments, seizure burden can be calculated by moving the seizure burden within a time window to obtain seizure burden values ​​for individual sequential time periods, forming a continuous output measurement. In some cases, the time period of the moving window can be between 1 minute and 1 hour. In some cases, the time period of the moving window can be dynamic or adjustable rather than fixed. In some cases, the sequential time periods formed by the moving window can overlap. In some cases, the sequential time periods formed by the moving window can be non-overlapping. In some cases, the moving window can move in time increments between 1 second and 1 hour. In some cases, the moving window can pause or skip time periods, resulting in seizure burden values ​​that are discontinuous or non-sequential in time.

[0130] Figure 2 The seizure burden mapping module 285 is displayed, configured to map the seizure burden of a subject. (Example) Figure 3 The diagram illustrates that the seizure burden output can be displayed to a user as a time series plot 310, where each point represents the seizure burden over a period of time. In some embodiments, the seizure burden output can display one or more thresholds (e.g., 10%, 50%, 90%, etc.) to the user on the time series plot. In some embodiments, the seizure burden output can be displayed to the user as a time series plot, bar chart, or graph. In some embodiments, a certain color can be used to depict the time series plot to indicate the thresholds that have been exceeded. Figure 3 As shown, when the time-series plot of seizure burden exceeds a 50% threshold over a period of time, the plot changes from gray to orange. In some cases, when the time-series plot exceeds 90%, it changes from orange to red. The time-series plot can have any color and any color can indicate exceeding the threshold. In some cases, the seizure burden plotting module can display a wide variety of information, such as the time period of measurement, date, or initial time of collection. In some cases, healthcare practitioners can use seizure burden plotting to assess a subject's condition and determine treatment duration. Healthcare practitioners can also use seizure burden plotting to monitor the progression of a subject's condition over time or to monitor the effectiveness of treatment.

[0131] Figure 2 A seizure burden notification module 290 configured to generate notifications is shown. In some embodiments, the method may include generating one or more notifications when a positive seizure classification has been performed or when a seizure burden value is equal to or exceeds one or more thresholds. Figure 4 The system demonstrates that when the seizure burden value equals or exceeds a threshold (e.g., a 90% threshold), it can display a notification 410 indicating that continuous seizure activity has been detected to the subject (e.g., a patient) or user (e.g., a healthcare professional, doctor, nurse, etc.). The notification can also contain any color. For example, the background of the screen displaying the notification could be red. The text of the notification can be any color, for example, white. The color of the screen background can be related to the calculated seizure burden value. For example, if the seizure burden equals or exceeds a certain threshold, then the selected color of the screen background could indicate that the seizure burden equals or exceeds the threshold. The color of the notification text can also be related to the calculated seizure burden value. For example, if the seizure burden equals or exceeds a certain threshold, then the selected color of the notification text could indicate that the seizure burden equals or exceeds the threshold.

[0132] In addition to notifying the detected continuous seizure activity, the system can display a variety of information to the subject or user. The system may display a seizure burden graph 415, a calculated percentage of seizure burden 420, the time period during which continuous seizure activity was detected (e.g., 7:40 pm to 7:50 pm), etc. Healthcare practitioners can use one or more notifications to assess the subject's condition and determine treatment duration. In some cases, one or more notifications may be generated when the seizure burden value is equal to or exceeds one or more thresholds as described elsewhere herein. In some cases, one or more notifications may be generated in the form of visual, audio, and / or text alerts. The device may include a speaker 425 for providing audio notifications. In some cases, one or more notifications may be delivered via networked communication technologies such as the Internet, telephone, fax, pager, SMS service, etc. In some cases, the form, content, or delivery mechanism of the generated one or more notifications may depend on the seizure burden value. In some cases, the user may be able to select the form, content, or delivery mechanism of the generated one or more notifications.

[0133] Figure 5 This document demonstrates the epileptic seizure detection output 510 provided by the method and system described herein. The interface provides indications of EEG signal activity from multiple channels of the EEG device module 110. Figure 5 Examples of user-adjustable parameters are displayed, such as time display, scale, high-pass frequency, low-pass frequency, or notch value. The interface may also provide seizure burden plotting, as described elsewhere in this document. The interface may also provide seizure burden results for different time periods. The interface may also depict seizure determinations for each time period. The interface may also provide a mechanism for users to accept or reject seizure determinations or seizure burden calculations derived by the algorithm. The interface may also provide a mechanism for users to input their own determinations of segments containing seizures or seizure burden. In some cases, the seizure burden may be adjusted based on user-input information about seizures or seizure burden. The displayed seizure burden and seizure burden notifications may be based solely on seizure determinations derived by the algorithm, solely on seizure determinations input by the user, or a combination of algorithmic and user-input seizure determinations.

[0134] In some embodiments, the seizure burden calculation module can calculate a seizure burden value. If the seizure burden value exceeds a threshold, the seizure burden notification module can output a notification. For example, if the seizure burden value exceeds a threshold of 10%, a notification can be generated. In another instance, if the seizure burden value exceeds a threshold of 50%, a notification can be generated. In yet another instance, if the seizure burden value exceeds a threshold of 90%, a notification can be generated. In some cases, when the seizure burden is equal to or exceeds a first threshold of 10%, a first notification indicating frequent seizure activity can be generated. In some cases, when the seizure burden is equal to or exceeds a second threshold of 50%, a second notification indicating a large number of seizures can be generated. In some cases, when the seizure burden is equal to or exceeds a third threshold of 90%, a third notification indicating continuous seizure activity can be generated. In some cases, a notification can be generated for a specific person to whom the method is programmed to make the notification.

[0135] In some embodiments, the notification threshold can be any percentage. The notification threshold can be at least about 1%, 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50%, 55%, 60%, 65%, 70%, 80%, 90%, 95%, 99%, or greater. The notification threshold can be at most about 99%, 95%, 90%, 85%, 80%, 75%, 70%, 65%, 60%, 55%, 50%, 45%, 40%, 35%, 30%, 25%, 20%, 15%, 10%, 5%, 1%, or less. The notification threshold can be from approximately 0% to 100%, 1% to 99%, 5% to 95%, 10% to 90%, 20% to 80%, 30% to 70%, or 40% to 60%. The notification threshold can also be user-adjustable.

[0136] In some embodiments, the seizure burden notification module may have any number of thresholds. The seizure burden notification module may have at least about 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, 25, 50, 100, 500, 1000, 5000 or more thresholds. The seizure burden notification module may have at most about 5000, 1000, 500, 100, 50, 25, 20, 15, 10, 9, 8, 7, 6, 5, 4, 3, 2 or fewer thresholds. The seizure burden notification module may have from about 1 to 5000, 1 to 1000, 1 to 500, 1 to 100, 1 to 50, 1 to 25, 1 to 10 or 1 to 5 thresholds. The number of thresholds may be increased or decreased at any point in time.

[0137] In some embodiments, the seizure burden notification module can provide any number of notifications. The seizure burden notification module can provide at least about 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, 25, 30, 35, 40, 50, 100, 500, 1000 or more notifications. The seizure burden notification module can provide up to about 1000, 500, 100, 50, 40, 35, 30, 25, 20, 15, 10, 9, 8, 7, 6, 5, 4, 3, 2 or fewer notifications. The seizure burden notification module can provide from about 1 to 1000, 1 to 500, 1 to 100, 1 to 50, 1 to 25, 1 to 10 or 1 to 5 notifications.

[0138] In some embodiments, the seizure burden calculation and output module may be standardized to output a notification, in addition to a seizure burden threshold. In some cases, the duration for which the seizure burden exceeds the threshold can be used to determine whether to output a notification. In some cases, the amount of time since the last time the seizure burden threshold was exceeded can be used to determine whether to output a notification. In some cases, dynamic criteria may be applied in combination with time-based, seizure burden-based, and other strategies to determine whether to output a notification.

[0139] In some embodiments, the system may be coupled to other systems. In some cases, the system may be an eye tracker, a motion sensor (e.g., an accelerometer or inertial measurement unit), an electromyography (EMG) device, an electrocardiogram (ECG or EKG) device, etc.

[0140] IV. Post-seizure testing

[0141] In some embodiments, the method may include generating one or more notifications, as described elsewhere herein.

[0142] In some embodiments, the method may provide a user with a response that minimizes or prevents detected seizures. The method may provide a response that minimizes or reduces the risk of seizures. In some cases, a therapeutic agent may be delivered to the subject to prevent and / or mitigate predicted seizures. In some cases, a neural modulation signaling pattern may be applied to the subject to prevent and / or mitigate predicted seizures. In some cases, the method may adjust the amount of neural modulation or therapeutic agent delivered to the subject.

[0143] V. Computer Systems

[0144] This disclosure provides a computer system programmed to implement the methods of this disclosure, including control of a seizure detection system, control hardware components, receiving and processing data, and user interface. The seizure detection system and its various components may include computer hardware (and associated firmware) electrically connectable to a computer control system. A control unit may contain such a computer system.

[0145] Figure 6 A computer system 601 is shown that is programmed or otherwise configured to operate and / or control the EEG device module and the seizure detection module. The computer system 601 can regulate various aspects of the seizure detection system of this disclosure, such as (for example) processing EEG signals, segmenting EEG signals, extracting features, processing features using machine learning algorithms, implementing control strategies and seizure burden, calculating seizure burden values, plotting seizure burden, providing notifications, etc. The computer system 601 may be a user's electronic device or a computer system remotely located relative to an electronic device. The electronic device may be a mobile electronic device.

[0146] Computer system 601 includes a central processing unit (CPU, also referred to herein as a "processor" and "computer processor") 605, which may be a single-core or multi-core processor or multiple processors for parallel processing. Computer system 601 also includes memory or memory locations 610 (e.g., random access memory, read-only memory, flash memory), electronic storage units 615 (e.g., hard disks), communication interfaces 620 for communicating with one or more other systems (e.g., network adapters), and peripheral devices 625 (e.g., caches, other memories, data storage devices, and / or electronic display adapters). Memory 610, storage units 615, interface 620, and peripheral devices 625 communicate with CPU 605 via a communication bus (solid line) (e.g., motherboard). Storage unit 615 may be a data storage unit (or data repository) for storing data. Computer system 601 may be operatively coupled to computer network ("network") 630 with the aid of communication interface 620. Network 630 may be the Internet, an intranet and / or extranet, or an intranet and / or extranet communicating with the Internet. In some cases, network 630 is a telecommunications and / or data network. Network 630 may include one or more computer servers that enable distributed computing, such as cloud computing. In some cases, network 630 may implement a peer-to-peer network with the assistance of computer system 601, enabling devices coupled to computer system 601 to act as clients or servers.

[0147] CPU 605 can execute a series of machine-readable instructions that may be embodied in a program or software. The instructions can be stored in a memory location, such as memory 610. The instructions can be directed to CPU 605, which can then be programmable or otherwise configured to implement the methods of this disclosure. Examples of operations performed by CPU 605 may include instruction fetching, decoding, execution, and write-back.

[0148] CPU 605 may be part of a circuit (e.g., an integrated circuit). One or more other components of system 601 may be included in the circuit. In some cases, the circuit is an application-specific integrated circuit (ASIC).

[0149] Storage unit 615 may store files, such as drivers, libraries, and saved programs. Storage unit 615 may store user data, such as user preferences and user programs. In some cases, computer system 601 may include one or more additional data storage units outside of computer system 601 (e.g., on a remote server located and communicating with computer system 601 via an intranet or the Internet).

[0150] Computer system 601 can communicate with one or more remote computer systems via network 630. For example, computer system 601 can communicate with a user's remote computer system (e.g., a seizure detection system manager, a seizure detection system user, a seizure detection data collector, a seizure detection system recorder, etc.). Examples of remote computer systems include servers, personal computers (e.g., portable PCs), notebook or tablet PCs (e.g., ...). iPad Galaxy Tab), phone, smartphone (e.g., iPhone, Android-compatible devices This could be a personal digital assistant (PDA). The user can access computer system 601 via network 630. In some cases, the EEG device module and the seizure detection module will be located within the same computer system. In other cases, the EEG device module will be located within a computer system networked to a remote computer system containing the seizure detection module. After performing seizure detection, the remote computer system can then transmit the seizure detection results to the computer system containing the EEG device module and other remote computer systems that can be used to display the seizure detection results. In some cases, different parts of the seizure detection module may be located in different computer systems already networked together.

[0151] The methods described herein can be implemented using machine-executable code stored in an electronic storage location of computer system 601 (e.g., in memory 610 or electronic storage unit 615). The machine-executable or machine-readable code can be provided in software form. During use, the code can be executed by processor 605. In some cases, the code can be retrieved from storage unit 615 and stored in memory 610 for easy access by processor 605. In some scenarios, electronic storage unit 615 can be excluded, and machine-executable instructions can be stored in memory 610.

[0152] The code can be pre-compiled and configured for use with machines having processors suitable for executing the code, or it can be compiled during runtime. The code can optionally be supplied for a programming language that enables it to be executed pre-compiled or post-compiled.

[0153] Aspects of the systems and methods provided herein (e.g., computer system 601) can be embodied in programming. Aspects of the technology can be considered as “products” or “articles of art” generally in the form of machine (or processor) executable code and / or associated data carried or embodied on a type of machine-readable medium. Machine-executable code can be stored on electronic storage units, such as memory (e.g., read-only memory, random access memory, flash memory) or hard disks. “Storage” type media can include any or all tangible storage of a computer, processor, or the like, or associated modules thereof, such as various semiconductor memories, magnetic tape drives, hard disk drives, and the like, which can provide non-transitory storage for software programming at any time. All or part of the software can sometimes be communicated via the Internet or various other telecommunications networks. Such communication (for example) enables the loading of software from one computer or processor to another computer or processor, for example, from a management server or host computer to a computer platform for an application server. Therefore, another type of media capable of carrying software elements includes light waves, radio waves, and electromagnetic waves, such as physical interfaces between local devices, wired and optical ground networks, and various air links. Physical elements carrying such waves (e.g., wired or wireless links, optical links, or the like) can also be considered media carrying software. As used herein, unless limited to non-transitory, tangible "storage" media, the term "readable media" for example, such as computer or machine media, refers to any media involved in providing instructions to a processor for execution.

[0154] Therefore, machine-readable media, such as computer-executable code, can take many forms, including (but not limited to) tangible storage media, carrier media, or physical transmission media. Non-volatile storage media include (for example) optical discs or magnetic disks, such as any storage device in a computer or the like, for example, used to implement the database shown in the diagram. Volatile storage media include dynamic memory, such as the main memory of such computer platforms. Tangible transmission media include coaxial cables; copper wires and optical fibers, including conductors that include buses within a computer system. Carrier transmission media can take the form of, for example, electrical or electromagnetic signals generated during radio frequency (RF) and infrared (IR) data communication, or sound waves or light waves. Therefore, common forms of computer-readable media include (for example): floppy disks, floppy disks, hard disks, magnetic tapes, any other magnetic media, CD-ROMs, DVDs or DVD-ROMs, any other optical media, punched card tapes, any other physical storage media having a perforated pattern, RAM, ROM, PROM and EPROM, FLASH-EPROM, any other memory chips or cartridges, carrier waves for transmitting data or instructions, cables or links for transmitting such carrier waves, or any other media from which a computer can read programming code and / or data. Many of these forms of computer-readable media may involve loading one or more sequences of one or more instructions onto a processor for execution.

[0155] Computer system 601 may include or communicate with an electronic display 635, the electronic display 635 including a user interface (UI) 640, which is used (for example) to provide an administrator with a login screen to access software programmed to control the seizure detection system and its functionality and / or to provide information on the operational health status of the seizure detection system. Examples of UIs include (but are not limited to) graphical user interfaces (GUIs) and web-based user interfaces.

[0156] The methods and systems disclosed herein may be implemented using one or more algorithms. The algorithms may be implemented using software when executed via the central processing unit 605. The algorithms may (for example) be components of software described elsewhere herein and may modulate parameters of the seizure detection system (e.g., processing of EEG signals, machine learning algorithms, control strategies, seizure burden, notifications, etc.).

[0157] While preferred embodiments of the invention have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. The invention is not intended to be limited to the specific examples provided in the specification. Although the invention has been described with reference to the foregoing specification, the description and illustration of the embodiments herein are not intended to be limiting. Many variations, alterations, and substitutions will now occur to those skilled in the art without departing from the invention. Furthermore, it should be understood that all aspects of the invention are not limited to the specific depictions, configurations, or relative proportions set forth herein, but depend on various conditions and variables. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of the invention. Therefore, the invention should also be considered to cover any such alternatives, modifications, variations, or equivalents. The following claims are intended to define the scope of the invention and thereby cover the methods and structures within the scope of these claims and their equivalents.

Claims

1. A system comprising: One or more computer processors; and A memory containing machine-executable instructions that, when executed by the one or more computer processors, implement a method for detecting epileptic seizures, comprising: (a) Receive multiple electroencephalogram (EEG) signals for the subject via multiple channels; (b) The plurality of EEG signals are preprocessed by dividing the plurality of EEG signals of each channel into a plurality of time data segments; (c) Extract multiple features from each time segment of each channel, where each time segment is associated with a period; (d) Applying a machine learning algorithm to the plurality of features to perform a seizure binary classification for each time data segment of each channel, thereby generating a plurality of classifications for the plurality of time data segments, wherein the seizure binary classification for each time data segment includes classifying each time data segment of each channel as (1) seizure positive or (2) seizure negative, wherein the plurality of classifications are compared sequentially across a plurality of periods on each channel, and if a subset of the classifications includes fewer than three seizure positive classifications in a row, then the subset is discarded; (e) Aggregate the remaining undiscarded data segments of the multiple channels within the moving time window; (f) Determining the seizure burden of the moving time window based on the aggregated classification, wherein the seizure burden includes the percentage of the time data segments classified as seizure-positive, and wherein the seizure burden is a metric providing a measure of the severity or likelihood of seizures; and (g) When the seizure burden is equal to or exceeds one or more thresholds, one or more notifications are generated, wherein the one or more notifications indicate different seizure activities and can be used to assess whether the subject is at risk of having a seizure.

2. The system of claim 1, wherein a positive seizure cluster indicates a potential electrocardiographic seizure at the corresponding time.

3. The system of claim 1, wherein each time data segment has a duration ranging from about one second to twenty seconds.

4. The system of claim 3, wherein the duration of each time data segment is approximately ten seconds.

5. The system of claim 1, wherein the movement time window ranges from about one minute to ten minutes.

6. The system of claim 1, wherein the one or more notifications are generated in the form of visual, audio, and / or text alerts.

7. The system of claim 1, wherein a first notification indicating frequent seizure activity is generated when the seizure burden is equal to or exceeds a first threshold of 10%.

8. The system of claim 1, wherein a second notification indicating a large number of seizures is generated when the seizure burden is equal to or exceeds a second threshold of 50%.

9. The system of claim 1, wherein a third notification indicating continuous seizure activity is generated when the seizure burden is equal to or exceeds a third threshold of 90%.

10. The system of claim 1, wherein the plurality of channels comprises at least three channels.

11. The system of claim 10, wherein the plurality of channels comprises eight channels.

12. The system of claim 1, wherein the plurality of features include time-domain and / or frequency-domain features inherent in the plurality of EEG signals.

13. The system of claim 12, wherein the plurality of features comprises at least twenty different time and / or frequency features.

14. The system of claim 12, wherein the plurality of features includes a plurality of discrete values ​​associated with the time-domain and / or frequency-domain features.

15. The system according to claim 1, wherein the machine learning algorithm includes random forest, boosting decision tree, classification tree, regression tree, bagging tree, neural network or rotation forest.

16. The system of claim 1, wherein the machine learning algorithm is individually applied to the plurality of features extracted for each channel, such that each channel has a separate iteration of the machine learning algorithm.

17. The system of claim 1, wherein the preprocessing of the plurality of EEG signals further comprises: Before the segmentation of the plurality of EEG signals, filters are applied to the plurality of EEG signals on the plurality of channels.

18. The system of claim 17, wherein the filter comprises a bandpass filter configured to filter the plurality of EEG signals between 1 Hz and 35 Hz.

19. The system of claim 1, wherein the method further comprises: If a subset of the time data segments from the multiple channels is classified as epileptic seizure positive, then the specific period is classified as associated with a potential electrocardiographic epileptic seizure.

20. The system of claim 19, wherein the subset comprises at least half of the plurality of channels.

21. The system of claim 1, wherein the sequential time periods formed by the time window are non-overlapping.

22. The system of claim 1, wherein the method further comprises: The epileptic seizure burden is output as a graphic visual element to the display.

23. The system of claim 22, wherein the method further comprises: The one or more thresholds are displayed in the graphic visual elements.

24. The system of claim 22, wherein the graphical visual element comprises a time series plot, a bar chart, or a graph.

25. The system of claim 24, wherein the time-series plot is configured to change color as the seizure burden exceeds a threshold of one or more thresholds.

26. The system of claim 22, wherein the method further comprises: The graphic visual elements are used to (i) assess the condition of the subject, (ii) determine the course of treatment, (iii) monitor the effectiveness of the course of treatment when it is provided to the subject, or (iv) monitor the progress of the subject's condition over time.

27. The system of claim 5, wherein the movement time window is approximately five minutes.

Images