Adaptive systems and methods for seizure detection

Abstract

The disclosed electroencephalogram (EEG) monitoring and detection system and method can detect distinct EEG seizure features of different seizure types across diverse patient populations. The disclosed system and method can utilize EEG data collected by discrete wireless EEG sensors positioned on the patient's scalp.

Description

[Technical Field]

[0001] Incorporation by reference to any priority application This application claims priority under U.S. Provisional Patent Application No. 63 / 505678 filed June 1, 2023, and U.S. Provisional Patent Application No. 63 / 511536 filed June 30, 2023, both of which are invoked by reference in their entirety. Any foreign or domestic priority claims identified in the application data sheets filed with this application are invoked by reference pursuant to 37 CFR 1.57. [Background technology]

[0002] Electroencephalography (EEG) is a diagnostic tool that measures and records the electrical activity of a person's brain to assess brain function. Multiple electrodes are attached to a person's head and connected to a device by wires. The device amplifies the signals and records the electrical activity of the person's brain. Electrical activity is the sum of neural activity across multiple neurons. These neurons generate small voltage fields. The sum of these voltage fields forms an electrical reading, which the electrodes on the head can detect and record. By monitoring the amplitude and temporal dynamics of the electrical signals, information about the underlying neural activity and the person's medical condition can be obtained.

[0003] There are thousands of hospitals across the United States. Many of these are community or local hospitals. These community or local hospitals are usually part of a hospital system or network. One example of such a network is one that includes several community hospitals and one major tertiary care hospital. Community or local hospitals that do not belong to any large hospital network usually contract with a large tertiary care hospital for emergency and intensive care solutions in areas outside their area of ​​expertise.

[0004] EEG monitoring is typically only available in large tertiary care hospitals that support neurology departments equipped with EEG services. Many hospitals do not offer EEG monitoring. In these hospitals, if such monitoring is necessary or desirable for a patient, arrangements are made with a larger tertiary care hospital or its affiliate. This usually takes the form of referring the patient to a tertiary care hospital for specialist or physician services. Often, this involves the patient traveling or being transported to a tertiary care hospital for the service. This presents many problems, especially for patients in rural areas. Consequently, it is desirable to provide improvements to EEG monitoring systems and methods. [Overview of the project]

[0005] EEG can be performed for the diagnosis of epilepsy, confirmation of problems related to loss of consciousness or dementia, confirmation of brain activity in comatose individuals, study of sleep disorders, monitoring of brain activity during surgery, and monitoring of other physical problems. Appropriate treatment plans can be developed based on EEG.

[0006] This specification discloses a system, method, and computer-readable medium for monitoring brain activity using one or more wireless EEG sensors configured to be detachably positioned at one or more locations on a patient's scalp. One or more computing devices can communicate with the EEG sensors and facilitate the setup of the EEG sensors, the reception of EEG data collected by the EEG sensors, and the processing of that data. Advantageously, accurate EEG measurements can be obtained and processed to determine (and treat) one or more physiological conditions in a patient, such as seizures or epilepsy. In addition, the disclosed system and method enable non-professionals to set up EEG monitoring, allowing a larger patient population to benefit from such monitoring.

[0007] In the following description, various embodiments will be described. For the purpose of the description, specific configurations and details are shown to provide a thorough understanding of the embodiments. However, it will also be apparent to those skilled in the art that the embodiments can be implemented without these specific details. Furthermore, well-known features may be omitted or simplified in order not to obscure the described embodiments.

Brief Description of the Drawings

[0008] [Figure 1A] FIG. 1A is a perspective top view and a bottom view of an EEG recording wearable sensor. [Figure 1B] FIG. 1B is various views of an EEG recording wearable sensor. [Figure 1C] FIG. 1C is various views of an EEG recording wearable sensor. [Figure 1D] FIG. 1D is various views of an EEG recording wearable sensor. [Figure 2A] FIG. 2A shows an example of an attachment. [Figure 2B] FIG. 2B shows an example of an attachment arranged on top of an electrode and placed on a wearable sensor. [Figure 2C] FIG. 2C shows an example of a sensor placed on a patient's scalp [Figure 3] FIG. 3 is a diagram of an EEG monitoring system. [Figure 4A] FIG. 4A shows an example process of seizure detection. [Figure 4B] FIG. 4B shows an example process of seizure detection. [Figure 5] FIG. 5 shows an example process of detecting a seizure event using multiple distinct seizure detection paths. [Figure 6] FIG. 6 shows an example process of adaptively preprocessing EEG data collected by multiple individual wireless EEG sensors. [Figure 7] FIG. 7 is an example timeline showing how the process shown in FIG. 6 can be used to adaptively preprocess EEG data collected by multiple individual wireless EEG sensors. [Figure 8] Figure 8 shows an example of the process for generating metafeatures by expanding multiple features. [Figure 9] Figure 9 shows that multiple features can be extracted and expanded to generate one or more meta-features. [Figure 10] Figure 10 shows an example of augmenting EEG data for training. [Figure 11] Figure 11 shows an example of the process for detecting a seizure event. [Figure 12] Figure 12 shows how the probability of one EEG data segment can be modified using the probabilities of one or more adjacent EEG data segments. [Figure 13] Figure 13 shows an example of recreating a missing EEG data channel. [Figure 14] Figure 14 shows the confidence levels of various event identifiers. [Figure 15] Figure 15 shows an example process that outputs seizure events along with confidence levels. [Figure 16] Figure 16 shows the seizure event output with confidence values. [Modes for carrying out the invention]

[0009] overview Some EEG monitoring systems may involve complex, multi-component medical device systems that require technical skills for setup and adjustment, and can be uncomfortable for the wearer. When such systems are used outside of large, specialized hospitals or research hospitals, setup and adjustment are difficult, and user error is more likely. EEG monitoring systems using multiple sensor components also require time synchronization between individual devices to combine sensor data. Achieving time synchronization between multiple sensor devices can be difficult if the sensors are not wired. EEG monitoring systems are used for long-term monitoring in homes or in hospitals of any size or specialization, including, for example, small, rural general hospitals. It can also be used for other purposes. Long-term EEG recording requires a high degree of complexity in setup and adjustment, but for everyday use, it needs to be seamless and easy to use.

[0010] This specification describes improved systems, kits, and methods for EEG monitoring and seizure detection.

[0011] Wearable sensors and EEG monitoring kits Figure 1A is a top and bottom perspective view of a wearable EEG recording sensor 101 that can be used as a seizure monitoring tool. As shown in Figure 1A, the wearable sensor 101 is self-contained within a housing 102. The housing 102 may be made of a plastic, polymer, composite material, etc., that has water resistance, waterproofing, etc.

[0012] The housing 102 may contain all the electronic circuitry for recording EEG from at least two electrodes 104, 105. Electrodes 104, 105 are located on the underside (i.e., the side facing the scalp) as shown on the right side of Figure 1A. Electrodes 104, 105 may be formed from any suitable material, such as gold, silver, silver chloride, carbon, or combinations thereof. One of electrodes 104, 105 may be a reference electrode and the other a measuring (for measurement) electrode. As described above, the entire wearable sensor 101 may be self-contained within the waterproof housing 102. The wearable sensor 101 may be designed as a self-contained EEG device limited to one-time use per user and disposable. The wearable sensor 101 may contain more than two electrodes, possibly including three, and in some embodiments including four. Additional electrodes (such as a third and / or fourth electrode) may also be formed from any suitable material, such as gold, silver, silver-silver chloride, carbon, or combinations thereof. In some embodiments, the wearable sensor 101 may include additional sensors such as an accelerometer, a PPG sensor, a differential PPG sensor, a skin chemistry sensor, and a temperature sensor.

[0013] The wearable sensor 101 has two electrodes 104 and 105 and can be used alone or in combination with other wearable sensors 101 (e.g., three more wearable sensors 101) as a discrete tool for monitoring seizures (and possibly counting the number of seizures). It is desirable, but not required, for the user to have received a prior diagnosis of seizure disorder using conventional wired EEG based on 10-20 montages. This diagnosis provides clinical guidance on the optimal placement of the wearable sensor 101 to record electrical seizure activity (sometimes called “seizure activity”) in the individual user. In some cases, the spacing of electrodes 104 and 105 forms single-channel EEG data using bipolar derivation.

[0014] Figure 1B is a top perspective view of a wearable sensor 101 for EEG recording, equipped with a housing 102 having an extended, rounded shape. This shape may be called a jellybean shape and may facilitate accurate placement in the correct orientation for the subject (or patient) and promote patient comfort and prolonged wear.

[0015] In some cases, the EEG recording wearable sensor 101 has a shape that fits behind the ear. The EEG recording wearable sensor 101 may have a shape that fits along the hairline. The EEG recording wearable sensor 101 may have a shape that fits along the scalp. For example, as shown in Figure 1B, the EEG recording wearable sensor 101 has an extended rounded shape configured to fit around or complement the user's hairline, so that the extended rounded shape of the housing 102 allows for discreet wear on the user's scalp while facilitating EEG signal collection. In some embodiments, the housing 102 includes a narrow section configured to curve along the user's hairline. Figure 1C shows a cross-sectional view, and Figure 1D shows a perspective view of the EEG recording wearable sensor 101 of Figure 1B. C-1D indicates that the housing 102 includes a narrow section 110. The side of the housing 102 having the narrow section 110 can be positioned closer to the patient's ear (see Figure 2C), which may facilitate discreet fitting and EEG signal acquisition. The narrow section 110 may be thinner than the rest of the housing 102. The housing 102 may become thicker (or wider) from the end containing the narrow section 110 to the opposite end 111. Such a change in thickness may facilitate discreet fitting. The thickness of the housing 102 at its widest point may be approximately 10.0 mm, 9.5 mm, 9.0 mm, 8.5 mm, 8.0 mm, 7.5 mm, 7.0 mm, 6.5 mm, 6.0 mm, 5.5 mm, 5.0 mm, 4.5 mm, 4.0 mm, or within a range comprising these values.

[0016] In some embodiments, the EEG recording wearable sensor 101 has a shape that mimics the appearance of a hearing aid. The EEG recording wearable sensor 101 may include an antenna. The external design (jellybean shape) of the EEG recording wearable sensor 101 affects the internal shape and may require the unique design and tuning of the antenna.

[0017] In some cases, the wearable EEG recording sensor 101 includes a power supply supported by the housing and configured to power the electronic circuitry. In some cases, the wearable EEG recording sensor 101 includes a rechargeable battery. The wearable EEG recording sensor 101 may include electrodes. The wearable EEG recording sensor 101 may include at least two electrodes positioned on the outer surface of the housing and configured to detect EEG signals indicating the user's brain activity when the housing is placed on the user's scalp. The electrodes may be located inside the housing 102 of the wearable EEG recording sensor 101. Unlike conventional wired EEG systems using 10-20 montages, the wearable EEG recording sensor 101 allows for a much smaller distance between the measurement electrode and the reference electrode, which not only makes the housing 102 more compact but can also improve signal quality. The distance between electrodes can be configured to allow for the capture of a less noisy EEG signal, thereby improving signal quality. The inter-electrode distance can be reduced compared to conventional wired EEG systems using 10-10 or 10-20 montages. The inter-electrode distance may be within a range of approximately 25 mm or less, approximately 20 mm or less, approximately 18 mm or less, approximately 15 mm or less, approximately 10 mm or less, or a range comprising these values ​​between centers. The housing 102 may be configured so that the electrodes are positioned at a distance that enables better EEG signal acquisition.

[0018] The EEG recording wearable sensor 101 includes electronic circuitry that may be supported in a housing 102. The electronic circuitry can be configured to process electroencephalogram (EEG) signals detected by at least two electrodes. In some embodiments, the electronic circuitry is configured to wirelessly transmit the processed EEG signals to a remote computing device. The remote computing device may be the portable computing device described herein.

[0019] Due to the elongated and rounded shape of the wearable sensor 101 for EEG recording, one or more of the following can be provided: (a) an appropriate electrode pair spacing for acquiring EEG signals, (b) a sealed housing 102 of sufficient size to accommodate a complete set of electronic devices including an antenna and a battery to support frequent communication (e.g., Bluetooth or Bluetooth Low Energy (BLE)), (c) the curvature of the scalp, and / or the hairline, and / or the curvature around the back of the ear, and a housing 102 design that conforms to these curvatures.

[0020] In some cases, the surface area of the housing 102 is about 8.5 cm 2 , 8.0 cm 2 7.5 cm 2 7.0 cm 2 6.5 cm 2 6.0 cm 2 5.5 cm 2 5.0 cm 2 4.5 cm 2 or within a range composed of any combination of the above values. The surface area of the jelly bean-shaped housing 102 shown in Figure 1B is about 20 cm 2 19.5 cm 2 19.0 cm 2 18.5 cm 2 18.0 cm 2 17.5 cm 2 17.0 cm 2 16.5 cm 2 16.0 cm 2 15.5 cm 2 15.0 cm 2 14.5 cm 2 14.0 cm 2 13.5 cm 2 13.0 cm 2 12.5 cm 2 12.0 cm 2 I1.5 cm 2 11.0 cm 2 10.5 cm 2 10.0 cm 2 9.5 cm 2 9.0 cm 2 I8.5 cm 28.0cm 2 7.5cm 2 7.0cm 2 6.5cm 2 6.0cm 2 5.5cm 2 5.0cm 2 4.5cm 2 The following range may be comprised of any combination of the above values. The volume of the jellybean-shaped housing 102 shown in Figure 1B is approximately 8.0 cm³. 3 7.5cm 3 7.0cm 3 6.5cm 3 6.0cm 3 5.0cm 3 4.5cm 3 , 4.0cm 3 3.5cm 3 , 3.0cm 3 , 2.5cm 3 , 2.0cm 3 The range may consist of the following, or any combination of the above values. The wearable sensor 101 can be placed anywhere on the patient's scalp (e.g., behind the ear) to record EEG.

[0021] The wearable sensor 101 may be packaged such that the circuit activates when removed from the packaging. Embodiments of the wearable sensor 101 can be placed anywhere on the scalp, similar to how conventional wired EEG electrodes are positioned. The wearable sensor 101 can self-adhere to the scalp by conductive mounting parts, conductive mounting parts, and / or mechanical means such as intradermal fixation with memory-shape metal.

[0022] When attached to the scalp (for example by the attachment part described below), in some embodiments, the wearable sensor 101 can function as a seizure detection device (either alone or in combination with one or more other wearable sensors 101, for example, three more wearable sensors 101). The wearable sensor 101 can record EEG continuously and uninterrupted for up to 16 days (or more). In some embodiments, each EEG recording wearable sensor 101 is configured to detect EEG signals independently of other sensors.

[0023] The wearable sensor 101 may use capacitive coupling as a means of spot checking signal quality. By bringing a device such as a mobile terminal close to the wearable sensor 101, capacitive coupling can be established with the device, providing a means of querying the EEG or impedance signal in real time. The wearable sensor 101 may be used to warn of seizures in real time or near real time.

[0024] The wearable sensor 101 can be used to record ultra-low frequency events from the scalp, such as cortical spreading inhibition. An amplifier circuit (not shown) may be suitable for recording DC signals. Alternatively, an amplifier circuit may be suitable for recording both DC and AC signals. The wearable sensor 101 may be used as a means of monitoring for the presence or absence of cortical spreading inhibition and / or seizures or other epileptic-like activity after a suspected stroke event. The wearable sensor 101 may be placed on the patient's scalp by any type of healthcare provider, such as paramedics, physicians, or nurses.

[0025] In some embodiments, the wearable sensor 101 may use capacitive coupling to monitor cortical spreading inhibition in real time. Spreading inhibition can be analyzed over time and displayed as an EEG visualization. The wearable sensor 101 may store these EEGs (e.g., in storage) for later retrieval. These EEGs may be archived in an electronic medical record or the like.

[0026] Figure 2A shows the mounting portion 200, which has been peeled from the backing paper 201 and whose adhesive surface is exposed. The mounting portion 200 may be called a sticker or adhesive. The backing paper 201 may be made of paper, plastic, or any other suitable material. Figure 2B shows the mounting portion 200 aligned on electrodes 104 and 105. Figure 1B shows a mounting portion 200 positioned on the wearable sensor 101. Optionally, the mounting portion includes a first surface substantially corresponding to an elongated, rounded shape and is configured to be attached to the outer surface of the housing 102 of the wearable sensor 101. In some embodiments, the mounting portion includes a second surface configured to removably position the wearable sensor 101 on the user's scalp. A laminated mounting portion 200 may be used. The user removes a layer (backing 201) to expose an adhesive containing hydrogel in wells aligned to the positions of electrodes (e.g., electrodes 104, 105). The mounting portion can then be placed on the sensor (sensor 101) and then on the user's skin to adhere the sensor 101, etc., to the skin. The mounting portion 200 may be shown as rectangular, but in any embodiment disclosed herein, it may be a jellybean shape or the like, corresponding to the shape of the housing 102 shown in Figure 1B.

[0027] Figure 2C shows a sensor 101 placed on the patient's scalp. The sensor 101 is reversibly attached to the scalp by a mounting part 200. The sensor 101 is placed in an appropriate location on the user to detect and record EEG data, for example, on the scalp below the hairline. The EEG data can be analyzed onboard, for example, by applying an analysis or machine learning model stored on the sensor 101, or it may be analyzed by a local device, a remote device, or a combination thereof. For example, the sensor 101 can communicate with a local device such as a smartphone or tablet over a personal area network (PAN) using a wired or wireless protocol (e.g., secure Bluetooth Low Energy (BLE)). Similarly, the sensor 101 can communicate with a remote device using a wide area network (WAN), such as by transmitting EEG data to a remote server or cloud server over the internet (including cases where communication is via a relay device such as a local device or not).

[0028] The hydrogel is conductive and provides sufficient adhesion to the scalp to effectively record EEG even during prolonged wear. Alternatively, the wearable sensor 101 may be bonded using a combination of conductive hydrogel and an adhesive structure. After use, the attachment part 200 can be simply peeled off and discarded from the wearable sensor 101. Before the next use (e.g., after the wear period), a new attachment part 200 can be applied to the wearable sensor 101.

[0029] The use of an integrated conductive hydrogel and adhesive mounting section 200 enables consistent EEG signal data transfer from person to person. The mounting section 200 allows the wearable sensor 101 to be reversibly adhered to the scalp. The design of the mounting section 200 also reduces both water ingress into the hydrogel and water evaporation from the hydrogel during prolonged wear. In some cases, the mounting section 200 is fabricated by laminating multiple adhesive and non-adhesive layers, each having a hydrogel-filled well, and sandwiching them between release liners. In some embodiments, the mounting section 200 is further individually packaged in an airtight and watertight pouch.

[0030] EEG system setup and provisioning The systems and methods disclosed herein may include software that assists a user in setting up the system. The user may be a healthcare provider or a patient.

[0031] Figure 3 is an example of an EEG monitoring system 300. The system in Figure 3 includes a plurality of wearable sensors 301 configured to record a patient's brain activity. Each wearable sensor 301 may include at least two electrodes configured to detect signals indicating the user's brain activity when placed on the user's scalp. Each wearable sensor 301 may further include electronic circuitry configured to determine data associated with the user's brain activity based on the signals detected by the at least two electrodes and to wirelessly transmit the data associated with the user's brain activity to one or more portable computing devices 302. Any of the wearable sensors 301 are related to wearable sensor 101 It is possible.

[0032] In some cases, the system may further include, when executed by at least one processor (one or more processors of the portable computing device 302), causing the at least one processor to facilitate the activation of the multiple wearable sensors 301; causing the user to position the multiple wearable sensors 301 on the user's scalp using multiple attachments configured to detachably attach the multiple wearable sensors 301 to the user's scalp; and a non-temporary computer-readable medium that stores instructions to record data transmitted by the multiple wearable sensors 301 that are associated with the user's brain activity.

[0033] The portable computing device 302 may include communication functions such as wireless communication. The portable computing device 302 may be configured to be worn by a user. The portable computing device 302 may include a smartwatch which may have a display. The portable computing device 302 may include a smart band, smart jewelry, etc. which may not have a display. The portable computing device 302 may include other computing devices such as a tablet or a medical-grade tablet. Such a portable computing device 302 may include a display larger than that of a smartwatch. The portable computing device 302 may connect to a remote computing device 304 (which may be a cloud service) via a network such as the Internet. The remote computing device 304 may include one or more computing devices such as a server.

[0034] This specification provides a system for monitoring brain activity. In some embodiments, the system includes a plurality of wearable sensors 301 configured to detect EEG signals indicating a patient's brain activity. Each of the plurality of wearable sensors 301 may include at least two electrodes configured to monitor EEG signals when placed on the patient's scalp. Each of the plurality of wearable sensors 301 may include electronic circuitry configured to process the EEG signals monitored by the at least two electrodes. In some cases, the system described herein further includes a non-temporary computer-readable medium for storing executable instructions that can be executed by at least one processor of a portable computing device 302.

[0035] Patient EEG data collected by multiple wearable sensors 301 may include multiple EEG data channels. Each wearable sensor may include two electrodes, which may function interchangeably as a separate detection electrode and a separate reference electrode for each wireless EEG sensor. The wearable sensor may measure the differential voltage between the separate detection electrode and the separate reference electrode. This differential voltage may be measured over time and generate an EEG data channel for each of the multiple individual wireless EEG sensors. The resulting multiple EEG data channels may be provided to one or more portable computing devices, such as a portable computing device 302, and processed as patient EEG data. In some embodiments, individual wireless EEG sensors may collect patient EEG data independently without referencing each other. Each EEG sensor may independently provide an EEG data channel to one or more portable computing devices, and these multiple EEG data channels may be compiled to form patient EEG data.

[0036] Seizure detection pathway for EEG data collection and processing The approaches described herein can detect and output possible or potential electrical seizure events (sometimes referred to as seizure events). This is because 1) ongoing or This method can be applied to either or both of the following: 1) rapid detection of recent seizure events, or 2) retrospective analysis of EEG data to detect past seizure events. The advantages are that rapid detection can support patient treatment by non-specialist clinicians, while retrospective detection can support EEG data review by specialist clinicians. Rapid detection can determine the frequency of EEG data showing electrical seizures over a period of time and can be used to output ongoing or recent (e.g., with a delay of a few seconds or minutes) seizure events. For example, a rolling analysis of 1-minute interval EEG data can detect a seizure event in response to identifying that a specific electrical seizure characteristic or pattern is present for X% of the time (e.g., 1% or more, 10% or more, 50% or more, or 90% or more). X may be called the frequency threshold. Retrospective seizure detection can output individual seizure events (or entire seizure events), which may refer to a pattern of sustained seizure activity with defined start and end times. Both rapid and retrospective detection can provide notification of seizure events as alerts or alarms.

[0037] A seizure detection path can be designed and tuned to detect specific electrical seizure characteristics and output seizure events. The seizure detection path may include a set of subsystems, blocks, or steps for preprocessing, data segmentation, feature extraction, classification, and postprocessing to determine seizure events. Figure 4A shows an example of a seizure detection path 420. The path 420 may be executed by one or more computing devices, for example, a remote computing device 304 alone, or in combination with one or more portable computing devices 302 or wearable sensors 301. The seizure detection path 420 may contain more or fewer blocks. One or more blocks may be executed in a different order or simultaneously with one or more other blocks in the seizure detection path 420.

[0038] In block 422, the seizure detection pathway 420 may receive EEG data collected by multiple EEG sensors (e.g., multiple individual wireless sensors 301). In some embodiments, the EEG data may include four EEG data channels collected by four individual EEG sensors. These EEG sensors may be placed at four discrete locations on the patient's scalp (e.g., locations corresponding to the left and right frontal and temporoparietal regions). Each data channel may include multiple samples (e.g., N) associated with a voltage within a specific voltage range (e.g., ±0.2 millivolts (mV) or ±0.5 mV).

[0039] Next, the exemplary seizure detection pathway 420 moves to block 424, where the EEG data may be preprocessed. Preprocessing may include filtering out unwanted frequency ranges of the EEG data. For example, the EEG data may be filtered to remove high-frequency components (e.g., electromyography (EMG) artifacts, electromagnetic interference, etc.), filtered to remove low-frequency components (e.g., DC shift, baseline drift, etc.), filtered to remove specific frequencies (e.g., 60 Hz line noise), or filtered to allow only specific desired frequency components (e.g., delta or theta bands) to pass through using a band-pass filter. Preprocessing may include adaptive standardization to account for different EEG data characteristics over time between patients and within specific patients (e.g., impedance changes due to environmental changes (e.g., humidity and temperature changes), skin-electrode interface changes such as increased sweating or decreased adhesiveness, or sensor movement and rotation during recording, etc.), as described in the section titled "Adaptive Preprocessing". Preprocessing may include desaturating the EEG data, which may be done before or after adaptive standardization. Saturation can occur due to improper EEG sensor placement, interference, motion artifacts, etc. Saturated data near the minimum and / or maximum data amplitude can be corrected. For example, saturated data can be removed and replaced with linear interpolation between two points on either side of the minimum or maximum amplitude, or with zero or NaN values. In some cases, one of the purposes of preprocessing is to maximize seizure seizure seizure seizures while minimizing inter-subject variability.

[0040] The seizure detection pathway 420 moves to block 426, where the preprocessed EEG data is processed by the second The data can be divided into shorter segments. These segments may overlap. For example, preprocessed EEG data can be divided into multiple segments of a fixed length (e.g., 1-10 seconds, 1-2 seconds). In some cases, segments of different lengths may be used to divide the EEG data.

[0041] The seizure detection pathway 420 moves to block 428, where multiple features may be extracted from each segment. Features may be extracted for each EEG data segment. Features may include one or more time-domain (or temporal) features, frequency-domain (or spectral) features, time-frequency features, and complexity-domain features. Temporal features may include minimum, maximum, variance, standard deviation, skewness, kurtosis, autocorrelation, cross-correlation, etc. Spectral features may include power spectrum, relative bandwidth power, spectral mean, spectral variance, spectral skewness, etc., in at least one specific frequency band (e.g., about 3 Hz or about 5 Hz). Time-frequency domain features may include the sum, mean, or maximum power of the wavelet transform, rhythm, etc. Complexity domain features may include compression ratio (e.g., determined by the encoding process), nonlinear energy, spectral entropy, sample entropy, permutation entropy, fractal dimension, line length, zero crossing, zero crossing uniformity, uncorrelated time, Hjorth complexity, polynomial approximation error, etc. In specific implementations, extracted features may include correlation measurements between two or more EEG data channels, including maximum cross-correlation, spectral cosine similarity, etc., and state dynamics.

[0042] As explained in the section "Expanding Extracted EEG Data Features" below, feature extraction block 428 may include expanding at least one extracted feature to create a metafeature (details below). A metafeature can be created by pooling two or more extracted features.

[0043] Feature extraction may include providing temporal context, such as adding delayed features, adding rolling averaging or rolling normalization of features, or extracting features using different segmentation combinations.

[0044] The seizure detection pathway proceeds to block 430, where a classifier (e.g., a classifier trained by a machine learning process) may assign a probability that a segment of the EEG data has been identified as seizure-like. This can be assigned to each segment of multiple segments using multiple extracted features associated with each segment. For example, the classifier could be an extra-tree classifier, a boosted-tree classifier, a random forest classifier, or a gradient-boosting classifier. More generally, the classifier could utilize a decision tree classifier, a neural network classifier, a nearest neighbor classifier, or a support vector machine classifier. The probability assigned by the classifier ranges from 0 to 1, where 0 represents a 0% probability that the segment contains EEG data representing a seizure, and 1 represents a 100% probability that the segment contains EEG data representing a seizure (e.g., EEG data showing seizure activity). In some cases, this probability value is a binary output, where the first value indicates that no seizure activity was detected in the segment, and the second value indicates that seizure activity was detected in the segment. For example, 0 might represent no seizure activity, and 1 might indicate that seizure activity was detected.

[0045] As described herein, the performance of the classifier can be further improved by adding one or more features associated with patient health information such as clinical symptoms, medical history, patient demographics, and medications. This additional information can be used in any other block of the seizure detection pathway 420.

[0046] The seizure detection pathway 420 proceeds to block 432, where segment probabilities may be joined. The joining of segment probabilities may be referred to herein as event identification, event segmentation, or stitching. (Event identifier (or event segmenter, event join)) The classifier (or stitcher) can combine the temporal sequences of probability values ​​determined by the classifier to output labels representing the presence of temporally extended seizure phenomena. The output event identifier can be any label derived from temporally extended phenomena that exist across individual time segments. The label may indicate, for example, that an individual seizure event has started, that an individual seizure event has been detected from start to finish, the start time of an individual seizure event, the stop time of an individual seizure event, or the occurrence rate of an EEG characteristic or pattern has been detected. The occurrence rate is, for example, ACNS Critical Care EEG Events can be categorized based on guidelines. For example, an occurrence rate of at least 10% may be classified as "Frequent," at least 50% as "Abundant," and at least 90% as "Continuous." One or more labels output by the event identifier can assist in post-event EEG data interpretation or in the rapid detection of ongoing or recent seizure events.

[0047] An event identifier can be a pipeline of one or more mathematical operations (or transformations) that use the probability values ​​of adjacent time segments to determine a new value (which is not necessarily a probability) for a given segment or set of segments within a given time. For example, transformations may include convolution or filtering (e.g., tuned filters), smoothing, thresholding, averaging (e.g., exponential or moving averages), or morphological operations. As described herein, the use of a tuned filter may include convolving a finite vector with either a probability output or a thresholded output, and optimizing the elements of the vector to increase sensitivity (or true positive rate) and decrease false positive rate.

[0048] The transformation can take a sequence of probability values ​​as input and output whether a seizure event exists and which time segments belong to that temporally consecutive event. The step sequence performed by the transformation can be defined by a set of parameters, which can ultimately determine which sequence of probability values ​​an event identifier identifies as belonging to a seizure event. A given parameterization of an event identifier can provide different interpretations of whether or not to consider probability values ​​as part of an event. A search across many different parameterizations can be performed to find an event identifier that gives an accurate interpretation that provides an optimal trade-off between detection and error. Many event identifiers can be generated with many different parameterizations, and these event identifiers can be used to generate events (or labels) from the training data. Then, it can be determined which event identifier most frequently predicts events that can be identified as true events in the data while minimizing the number of false identifications. As described in the "Event Identification and Confidence" section, many event identifiers can be generated during the training of the seizure detection pathway to select a subset that can provide an acceptable trade-off between detection and error.

[0049] This can be used to determine the confidence value associated with a seizure event or label. Additional details of event identification are described in the sections "Identifying Seizure Events" and "Event Identification and Confidence" of this specification, etc.

[0050] The seizure detection pathway 420 proceeds to block 434, where notification of a seizure event may be provided to a user, such as a patient or physician. Notification of a seizure detection event may be provided as a report, alarm, or alert. In the case of post-detection of an individual seizure event, this may include the start and end times of the seizure event, as well as / or its duration. A confidence value for the seizure event may be output as described herein.

[0051] Depending on the implementation, the seizure detection pathway 420 may be designed and adapted for different product features and / or clinical outcomes. For example, as described herein, the seizure detection pathway 420 may be designed and adapted for continuous real-time detection (rapid detection) of ongoing or recent seizure activity, which may advantageously provide real-time alerts. Seizure Events The onset of the condition can be identified and included in notifications. In some cases, this may be possible after a sufficient amount of EEG data has been acquired and processed.

[0052] EEG data collection and processing using differentiated model paths Seizure detection in EEG data across different patient types and disease states presents a technical challenge in the field of EEG monitoring. Seizures manifest differently depending on the patient, disease state, and seizure type (e.g., focal, tonic, clonic, tonic-clonic, absence, asymptomatic, flaccid, or myoclonic), making reliable detection and diagnosis difficult. Seizures may possess one or more electrical seizure characteristics that distinguish one seizure type / category from another. Detecting diverse and distinguished electrical seizure characteristics may require a sufficiently high sensitivity (or high true positive rate). However, increased sensitivity can also increase the false positive rate. The false positive rate can be measured by the detection rate of false positives (or incorrect identification of a seizure when one does not exist). If the seizure detection process is not properly designed and adjusted across diverse patient populations, detection may suffer from high error rates, for example, due to high sensitivity but a high false positive rate, or low sensitivity and a low false positive rate. Advantageously, the disclosed embodiments of seizure detection can detect seizures with high sensitivity, effectiveness, and efficiency across different patient types and disease states without exceeding an acceptable false-positive rate. In some examples, the disclosed embodiments detect seizure events with a sensitivity of at least about 80% and a false-positive rate of about 0.21 or less per hour (in some cases, about 0.08 or less per hour).

[0053] The electrographic nature of seizures (recorded using EEG) can vary considerably. Some seizures may manifest as lower frequency spikes (e.g., around 3 Hz) and slow waves, indicating a typical absence seizure. Other seizures may manifest as progressive burst activity transitioning from a continuous state to an on / off state, indicating a tonic-clonic seizure. Yet another seizure may manifest as polyspikes and slow waves, low-frequency rhythmicity (or sometimes medium-frequency rhythmicity, e.g., around 10 Hz), slowing in the frequency band, increasing in the medium-frequency band, or high-frequency spiking, indicating a focal seizure. Yet another seizure may manifest as solitary or recurrent clonus spikes, indicating a myoclonic seizure. Yet another seizure may manifest as on / off burst activity (and sometimes higher EMG activity), indicating a clonic seizure. Yet another seizure may manifest as strong burst activity (sometimes accompanied by spiking and / or high EMG activity), indicating a tonic seizure. Furthermore, another type of seizure may manifest as a quiet EEG signal, which may indicate a flaccid seizure. These differences in specific electrical seizure characteristics make it difficult to create a "universal" seizure detection process that can accurately detect all electrical seizures. The disclosed embodiments solve these problems by using a combination of multiple seizure detection pipelines / pathways that are designed and tuned (or learned) to detect one or more specific electrical seizure characteristics. As a result, the combination of pathways can detect distinct electrograph characteristics (or distinct characteristics) of various seizure types across a disparate patient population.

[0054] In some embodiments, one or more blocks or steps of a seizure detection pathway (e.g., pathway 420) may have one or more parameters tuned to detect at least one distinct electrical seizure characteristic. For example, a seizure detection pathway may be tuned to detect the presence of low-frequency spikes (e.g., about 3 Hz) and slow waves indicating an absence seizure. As another example, a seizure detection pathway may be tuned to detect progressive burst activity from a continuous state to an on / off state indicating a tonic-clonic seizure. Multiple seizure detection pathways can be used to detect multiple distinct electrograph characteristics of different types of seizures. That is, a first seizure detection pathway may be distinguished from a second seizure detection pathway by distinguishing one or more blocks or parameters designed and tuned for detecting different electrograph characteristics.

[0055] Seizure detection algorithms trained to detect seizures without such distinctions can, unfortunately, suffer from reduced sensitivity due to overgeneralization and a worsening of the false positive rate. The disclosed approach using the exit pathway advantageously facilitates the detection of seizure-type electrographic characteristics with high sensitivity and a low false-positive rate.

[0056] Figure 4B shows a seizure detection process 450 that utilizes multiple seizure detection pathways, including pathway 452, pathway 454, and one or more additional pathways 456. Each illustrated seizure detection pathway may be similar to pathway 420 in Figure 4A, but the pathways may be distinguished in that they can be designed and tuned to detect different types of seizure electrographic characteristics. For example, pathway 452 may be designed and tuned to detect a first one or more electrographic characteristics (e.g., spikes and slow waves around 3 Hz). Another example is that pathway 454 may be designed and tuned to detect a second one or more electrographic characteristics (e.g., progressive burst activity from continuous state to on / off state). Yet another example is that pathway 456 may be designed and tuned to detect a third one or more electrographic characteristics (e.g., mid-frequency rhythmicity). In some examples, another pathway (not shown) may be designed and tuned to detect a fourth one or more electrographic characteristics associated with seizure events not included in pathways 452, 454, and 456. Process 450 can be performed by one or more computing devices, for example, by the remote computing device 304 alone, or in combination with one or more portable computing devices 302 or wearable sensors 301.

[0057] Multiple seizure detection pathways can be distinguished in one or more of the following stages: preprocessing, data segmentation (or splitting), feature extraction, classification, or event identification. Since each pathway is designed and tuned to detect one or more unique electrograph characteristics, the data segmentation can also be unique. As an example of distinction in data segmentation, since absence seizures tend to be short in duration, pathway 452 may use short segments, such as 1-2 seconds. Pathways 454 and 456 may use longer segments, such as 1-4 seconds. As an example of distinction in classification, pathway 452 may use a classifier trained on data from known seizure events with approximately 3 Hz spike and slow wave morphology that may be associated with absence seizures. Pathway 454 may use a classifier trained on data from known seizure events with progressive burst activity from continuous to on / off states that may be associated with tonic-clonic seizures. Pathway 456 may use a classifier trained on data from known seizure events with approximately 10 Hz rhythmicity that may be associated with focal seizures.

[0058] As an example of distinction in event identification, a distinguished event identifier may generate seizure events with one or more unique electrograph characteristics. As described in the "Seizure Event Identification" section below, event identifiers may output such seizure events using filtering, smoothing, gap filling, and reduction. Because some distinguished electrical seizures tend to occur in closer temporal proximity than others, one or more event identifiers in path 452 (designed and tuned to detect one or more characteristics associated with low-frequency spikes and slow-wave seizures) may fill the gaps between individual events that are relatively close in time (e.g., within 15 seconds of each other) to generate a single individual seizure event (or the entire seizure event). One or more event identifiers in path 454 (designed and tuned to detect one or more characteristics associated with tonic-clonic seizures) may combine any seizure events found over a longer time range (e.g., within 30 seconds of each other) to generate a single individual seizure event. One or more event identifiers in path 456 (designed and tuned to detect one or more characteristics related to mid-frequency rhythms that are expected to occur at greater time intervals) can combine any seizure events found within a wider time range (e.g., within one minute of each other) to generate a single, distinct seizure event. More generally, event identifiers can be distinguished based on the time length between individual seizure events.

[0059] Seizure events output by seizure detection pathways 452, 454, and 456 are combined, linked, or merged by event merge block 460 (sometimes called a linking block). Events can be merged. The event merge block can analyze the confidence values ​​(and possibly durations) of seizure events determined by the event identifiers for each pathway, as described in the "Event Identification and Confidence" section. For example, in post-event detection, the event merge block 460 may select individual seizure events associated with the highest confidence value. In rapid detection, the event merge block 460 may select labels that show the occurrence rate of an EEG characteristic or pattern over a period of time, meet the occurrence rate threshold, and have the highest confidence value. If there are no event identifiers for pathways that output labels with such occurrence rates, the event merge block 460 may select the label with the highest confidence value.

[0060] Notifications of one or more detected seizure events can be provided. Notifications may be provided after the event merge block 460. Notifications may be provided to the user (e.g., a patient or clinician) via a portable computing device 302 or a remote computing device 304. Notifications may include information about the seizure events, such as the start time, stop time, and / or duration of the individual seizure event, along with its confidence level. EEG data collected by multiple individual wireless EEG sensors may also be included in the notification.

[0061] The seizure detection pathways 452, 454, and 456 are stored as instructions in computer-readable memory and can be executed by one or more processors. They can be executed concurrently or sequentially. As another example, a subset of blocks of the seizure detection pathways can be executed in parallel. Depending on the embodiment, at least two of the seizure detection pathways 452, 454, and 456 may share some blocks (which may be called undifferentiated blocks). For example, at least two pathways may share a preprocessing block 424. This may be due to similar temporal and frequency characteristics of the seizure types processed by each pathway.

[0062] Figure 5 shows an exemplary process 500 for detecting distinct electrical seizure characteristics using multiple seizure detection paths, including paths 452, 454, and one or more additional paths 456, as described in relation to Figure 4B. Process 500 may be performed by one or more computing devices, such as the remote computing device 304 alone, or in combination with one or more portable computing devices 302 or wearable sensors 301. Process 500 may include more or fewer steps. One or more steps of process 500 may be performed in a different order or simultaneously with one or more other steps. Instructions for performing process 500 may be stored in a computer-readable medium. The instructions may cause one or more processors to perform the steps of process 500. In some cases, process 500 may include additional elements, or fewer elements. For example, the process may be performed by one seizure detection path, or by three or more seizure detection paths.

[0063] Process 500 is initiated in step 502, in which multiple seizure detection pathways are configured. The multiple seizure detection pathways may be learned or tuned to detect distinct electrical seizure characteristics. The multiple seizure detection pathways may be learned or tuned using signal processing and / or statistical learning techniques on EEG data collected from a patient population. The learning data may include EEG data with seizures, non-seizure EEG data obtained from subjects who do not experience seizures, and inter-seizure data from subjects who experience seizures. The learning data may be obtained from subjects fitted with multiple individual EEG sensors (e.g., EEG sensor 301) and / or wired EEG electrodes (e.g., arranged according to a 10-10 or 10-20 electrode configuration).

[0064] Multiple seizure detection pathways can be trained to detect distinct electrical seizure characteristics as described herein. These pathways can be trained to detect absence seizures, focal-onset seizures with clinical convulsions, generalized-onset seizures without clinical convulsions, bilateral tonic-clonic events, etc. As described above, each seizure detection pathway can be learned to detect at least one specific electrical seizure characteristic.

[0065] While the specific examples disclosed relate to seizure detection, pathways can be learned to detect cortical spreading inhibition in stroke, obstructive sleep apnea, or migraine based on EEG data. Pathways can also be learned to utilize EEG data to determine sleep quality, classify sleep stages, and / or predict conditions such as Alzheimer's disease, depression, fatigue, multiple sclerosis, and Parkinson's disease. For example, a first seizure detection pathway could be learned to detect electrical seizure characteristics associated with absence seizures, and a second seizure detection pathway could be learned to detect electrographic characteristics associated with Alzheimer's disease.

[0066] A seizure detection pathway may be trained to detect at least one electrographic characteristic associated with two or more distinct types of seizures. For example, a first seizure detection pathway may be trained to detect a first characteristic associated with absence seizures, a second seizure detection pathway may be trained to detect a second characteristic associated with focal-origin seizures, and a third seizure detection pathway, as a hybrid pathway, may be trained to detect at least one characteristic associated with both absence seizures and focal-origin seizures. In another example, a third seizure detection pathway may detect characteristics associated with all types of seizures other than absence seizures and tonic-clonic seizures.

[0067] Process 500 proceeds to step 504, in which multiple individual wireless EEG sensors, for example, multiple EEG sensors 301, may be provided to collect EEG data. As described above in relation to Figures 2A-2C, the sensors may be applied to discrete locations on the patient's scalp. Multiple individual wireless EEG sensors may include four separate wireless EEG sensors, as described herein. The four EEG sensors may be placed at four discrete locations on the patient's scalp (for example, locations corresponding to the left and right frontal and temporoparietal regions). For example, the first EEG sensor may be placed on the left forehead of the patient, the second EEG sensor on the right forehead, the third EEG sensor behind the left ear, and the fourth EEG sensor behind the right ear. These locations may correspond to the F7, F8, TP9, and TP10 locations, respectively. Process 500 (or any other process described herein) may be described in relation to data obtained by an individual wireless EEG sensor, but process 500 (or any other process described herein) and the seizure detection pathway may also be implemented in a wired EEG system. As described herein, an individual EEG sensor may collect patient EEG data which may include multiple EEG data channels.

[0068] As described herein, patient health information (or patient information) may be detected or acquired. Patient information may include physiological measurements such as heart rate, body temperature, and movement, which may be collected by additional sensors of a separate wireless EEG sensor or by independent additional sensors. Patient information may include information related to the patient's environment, such as time of day, current weather, current temperature, current humidity, and patient movement. Patient information may include clinical symptoms, past diagnoses, patient demographics, patient medication, and time of day. Patient information may be provided by a portable computing device 302. For example, the portable computing device 302 may receive the patient's medical history from a remote computing device 304. As another example, the patient may provide patient information to the portable computing device 302 through a user interface. Patient information may be used in either step 508 or 510 of process 500.

[0069] Additional sensors may include one or more of the following: A photoelectric volume plethysmography (PPG) sensor can be used to detect the patient's heart rate, heart rate variability, etc. A differential PPG sensor can be used to detect blood pressure and / or vascular tone. An accelerometer can be used to detect the patient's movement and / or orientation. This information is relevant to tonic-clonic seizures. This may be useful for detecting characteristics related to the exercise effect (or muscle contraction) obtained, or for controlling the orientation of the EEG sensor. A temperature sensor can be used to detect temperature. Other sensors can be used to detect skin electrical activity, skin chemical composition, temperature, humidity, light, sound, etc. As described herein, information related to environmental conditions (e.g., temperature or humidity) can be used to standardize EEG data, for example, during preprocessing. Weather-related information can be obtained to facilitate seizure detection, as weather conditions can affect the occurrence of seizures.

[0070] Process 500 proceeds to step 506, where non-temporary computer-readable media may be provided. The non-temporary computer-readable media may store instructions for executing a seizure detection process configured to process EEG data (and, if applicable, patient information) using a seizure detection pathway. These instructions may be executed by one or more processors, for example, the processors of one or more portable computing devices 302 and / or remote computing devices 304.

[0071] Process 500 proceeds to step 508, where the EEG data is processed through multiple seizure detection paths, providing multiple distinguished outputs. As described in relation to Figure 4B, each seizure detection path may include a preprocessing step, a segmentation step, a feature extraction step, a classification step, and an event identification step. At least one of these steps may be distinguished among the multiple seizure detection paths. For example, three seizure detection paths may be used, and the steps of the seizure detection paths may not be distinguished except for the classification step. The distinguished classification step of the three seizure detection paths may be adjusted to detect distinguished electrical seizure characteristics. In some embodiments, two or more steps of the seizure detection paths may be distinguished.

[0072] As described in relation to Figures 4A and 4B, the preprocessing steps of a seizure detection pathway (e.g., step 424 of seizure detection pathway 420) may include one or more of the following: normalization, standardization, filtering, denoising, detreating, demeaning, de-articulation, or desaturation of the EEG data. One or more of these processes or parameters may be distinguished between seizure detection pathways. The EEG data as a whole may be preprocessed as individual EEG data channels or as subgroups of EEG data channels. In some cases, the preprocessing steps may be adaptive, as will be discussed later in the section titled “Adaptive Preprocessing.”

[0073] Differentiated preprocessing steps for multiple seizure detection pathways can standardize patient EEG data to account for inter-patient and intra-patient variability. Inter-patient variability may include differences between two or more patients, such as differences in skin impedance or skull thickness. Intra-patient variability may include differences between skin-electrode interfaces of different electrodes, changes in individual skin-electrode interfaces over time, or changes resulting from electrode replacement (e.g., placing a new electrode in a slightly different position or orientation from the original electrode). Differentiated preprocessing steps for multiple seizure detection pathways can maximize the ability to separate seizure-like data from non-seizure-like data in the feature space (e.g., based on power in a specific frequency band).

[0074] As described in relation to Figures 4A and 4B, the segmentation step of the seizure detection pathway (e.g., step 426 of seizure detection pathway 420) may include dividing the patient EEG data into multiple EEG data segments. Multiple seizure detection pathways may be distinguished in the segmentation step.

[0075] As described in relation to Figures 4A and 4B, the feature extraction step of the seizure detection pathway (e.g., step 428 of seizure detection pathway 420) may process one or more segments of the patient's EEG data to extract at least one EEG data feature. In this configuration, the feature extraction step may include obtaining patient information or other contextual information. Patient information may include physiological status, demographics, patient's medical history, time of day, etc. Physiological status may be measured by multiple individual wireless EEG sensors or one or more other sensors. Contextual information may include time of day, current weather, or other environmental conditions to which the patient may be exposed. In some cases, patient information (e.g., patient's medical records or seizure diary) may be obtained from a remote computing device 304 or via a user interface running on a portable computing device 302.

[0076] In some implementations, multiple seizure detection paths may be distinguished in the feature extraction step. For example, a first seizure detection path may extract at least one feature different from that extracted by a second detection path. The distinguished feature may be relevant or useful for the detection of the characteristic by the first seizure detection path, while being irrelevant to the detection of the characteristic by the second seizure detection path. In some examples, relevant features may be identified by statistical analysis. For example, a feature selection method may be used to identify one or more relevant features for detecting at least one electrical seizure characteristic. The distinguished paths may differ in the number of features selected.

[0077] In some cases, an excess of irrelevant or highly correlated features can degrade the performance of certain classifiers while allowing others to function well. Reducing the number of features can improve the performance of specific paths. Feature reduction can be implemented using dimensionality reduction techniques such as principal component analysis. For example, principal component analysis can reduce 50 features to 3 or 4 principal component dimensions. Feature reduction can also be implemented using feature selection techniques, where the number of features can be specified or automatically determined. Large scale differences between features can degrade the performance of some classifiers, and normalizing features can improve performance. Various classifiers can benefit from using contextual information as features, such as patient demographics and time of day.

[0078] For example, a low-frequency spike and slow-wave seizure detection pathway (e.g., pathway 452) may extract power spectral density features at specific frequencies such as 3 Hz, but may not extract complexity domain features. A higher-frequency spiking and / or motion effect detection pathway (e.g., pathway 454) may extract different frequency-based metrics. Advantageously, by prioritizing relevant features and / or removing unimportant features, sensitivity may increase and false positive rates may decrease. Processor load and memory consumption may also be reduced, resulting in improved performance for at least one computing device running the seizure detection pathway.

[0079] As described in relation to Figures 4A and 4B, the seizure detection pathway classification step (e.g., step 430 of seizure detection pathway 420) includes assigning a probability value to a segment of EEG data identified as seizure-like by the classifier. The probability value may be determined based on at least one extracted feature.

[0080] Multiple seizure detection pathways can be distinguished in the classification step. A first seizure detection pathway may use a classifier trained to detect one or more first electrical seizure characteristics, while a second seizure detection pathway may use a classifier trained to detect one or more second electrical seizure characteristics. As described above, the classifier used by a particular seizure detection pathway may be trained using EEG data that has seizure events containing the specific electrical seizure characteristics identified by that pathway.

[0081] As described in relation to Figures 4A and 4B, the event identification step may include identifying seizure events by applying one or more event identifiers to the segment probabilities output by the classifier. For example, for individual seizure events, the event identifier may identify the start time, stop time, and / or duration, as well as the confidence value of the seizure event. As described herein, event identifiers may be distinguishable between seizure detection pathways.

[0082] Process 500 may proceed to step 510, in which a seizure event is identified based on the outputs of multiple seizure detection pathways. This may be performed by a concatenation or merging step, such as the merge step 460 in Figure 4B. Notification of the seizure event may also be provided, as described in relation to Figure 4B.

[0083] Advantageously, the detection of one or more seizure events can assist clinicians in determining whether a seizure occurred and in identifying the type of seizure. Clinicians may use additional clinical information to make such determinations. For example, the determination of an absence seizure may be made in response to a review of information related to the notification of the seizure event and the patient's condition during the seizure event to confirm that the patient was experiencing a drowsy state (or "absence") at the time the seizure event occurred. This information may be image data (e.g., recorded by a camera) or data recorded by the patient or a third party (e.g., in a diary).

[0084] The approaches described herein for detecting one or more seizure events may facilitate accurate counting and tracking of seizure occurrences, supporting effective diagnosis and treatment. For example, patients may receive medication adjustments infrequently (e.g., every six months) based on tracking of seizure occurrences. The approaches described herein may shorten this interval (e.g., to a range of several weeks), facilitating more frequent reviews and adjustments of treatment plans.

[0085] Adaptive pretreatment When collecting EEG data using multiple independent wireless EEG sensors (e.g., sensor 301) placed at different locations on a patient's scalp, additional preprocessing may be required to account for potential differences in EEG data due to temporal differences within the patient (e.g., changes in sensor position or orientation, sweat, etc.) and differences between patients (e.g., temperature, humidity, sensor orientation, etc.). These differences can lead to signal distortion due to changes in impedance and orientation, and may require standardization to aid in seizure detection.

[0086] The implementation of adaptive preprocessing described herein provides a technical solution to address these technical challenges. An advantage is that changes in EEG data can be taken into account, allowing for effective processing of EEG data using one or more seizure detection pathway approaches described herein, as illustrated in relation to Figures 4A, 4B, and 5. Adaptive preprocessing can standardize EEG data by taking into account changes in EEG signals monitored by multiple independent wireless EEG sensors worn by the patient. Changes can be detected, and in response to these changes, one or more processing parameters can be adjusted to ensure the continuity of the EEG data. Changes can occur due to changes at the skin-sensor interface (e.g., due to adhesive replacement or degradation), changes in the patient's skin, environmental changes (e.g., temperature, humidity, etc.), changes in the EEG sensor (e.g., sensor replacement, change in sensor orientation, sensor movement, etc.), changes caused by user activity (e.g., sweating during exercise), etc.

[0087] As described above, preprocessing may include normalizing (e.g., bringing within a certain range), standardizing (e.g., scaling), filtering, denoising, desaturating, and / or modifying patient EEG data to account for inter-patient and inter-sensor differences. Preprocessing parameters may include parameters used during such operations.

[0088] Figure 6 shows an exemplary process 600 for adaptively preprocessing EEG data collected by multiple individual wireless EEG sensors, such as sensor 301. Process 600 may be part of preprocessing block 424 in Figure 4A. Process 600 may consist of more or fewer steps. One or more steps of process 600 may be performed in a different order or simultaneously with the other one or more steps.

[0089] In block 602, process 600 can receive multiple EEG signals (or data) from multiple independent wireless EEG sensors. As described herein, multiple independent EEG sensors may not share a common reference (e.g., a reference electrode). In block 604, process 600 can modify at least one preprocessing parameter in response to a change in the EEG signal.

[0090] To address the impact of intra-patient or inter-patient issues in EEG data, adaptive preprocessing may include adaptively normalizing the EEG data. This may involve normalizing the EEG data by evaluating the variance of a window of EEG data, setting a variance threshold (or scaling factor) as the minimum value of the identified variance, the mean value of the identified variance, or its square root (i.e., standard deviation), and dividing by the variance threshold. Window durations can be less than 1 second, 2 seconds, 30 seconds, 1 minute, 2 minutes, 5 minutes, 10 minutes, 30 minutes, or more than 1 hour. Since a window can serve as a best estimate of background baseline noise, it can be used to standardize the EEG data within the current window. As an example, suppose we use a 2-second window. By comparing the variance determination values ​​of each window of EEG data with each other, the window with the smallest variance can be selected as the so-called "quiet" window, assumed to contain background noise. The variance threshold can be set to the variance of such a quiet window. EEG values ​​in other windows can be scaled by the variance threshold. Zeroed or saturated EEG data may not be considered during the calculation of the variance threshold.

[0091] Because the patient's environment can be highly non-stationary, it may be advantageous to periodically adaptively update the variance threshold if a new quiescent window is not found after a certain period of time. If the variance threshold is not updated after a sufficient amount of time, a growth rate approach may be used, increasing the current variance threshold by a coefficient (e.g., 5%, 10%, 20%, 30%, 40%, 50%, etc.) until a new quiescent window is identified. The duration of this increase can range from one minute to one hour, for example, 30 seconds, one minute, two minutes, three minutes, four minutes, five minutes, ten minutes, fifteen minutes, twenty minutes, twenty-five minutes, thirty minutes, forty-five minutes, one hour, etc. The variance threshold may be increased continuously (e.g., every minute) until a new quiescent window is identified. The approach described in this paragraph is particularly applicable when the variance threshold coefficient is set to the minimum variance, as described here.

[0092] As described here, at least one preprocessing parameter may be modified in response to a change in the position or orientation of the EEG sensor (e.g., due to the replacement of the EEG sensor). The orientation or replacement of the EEG sensor may be detected based on the output of an accelerometer or other position / orientation measuring sensor. For example, an accelerometer may detect that one of several discrete wireless EEG sensors has been attached to the patient in an inverted orientation. When the accelerometer detects the change in orientation, the EEG data received from the inverted EEG sensor may be inverted (e.g., multiplied by -1) in block 604.

[0093] Process 600 proceeds to block 606, where the pre-processed EEG data may be provided to the next processing block (for example, the data segmentation block in Figure 4A).

[0094] As described here, an adaptive preprocessing approach can monitor one or more changes in EEG data and adjust one or more preprocessing parameters in response to those changes. For example, the adaptive standardization described here can monitor the occurrence of one or more changes in EEG data by adaptively updating the dispersion threshold if no new quiet window is found after a certain period of time. Therefore, as a result of not detecting any changes in the EEG data over a period of time, preprocessing parameters (or dispersion thresholds) may be adjusted. As another example, changes in the mounting position or orientation of the EEG sensor may trigger position / orientation measurement... It can be monitored by a sensor.

[0095] Figure 7 is an example timeline 700 showing how EEG data collected by multiple discrete wireless EEG sensors can be adaptively standardized using process 600. In timeline 700, the x-axis represents time and the y-axis represents the value of the preprocessing parameter. Based on the detected changes, the preprocessing parameter value may be adjusted as shown at time 702 in Figure 7.

[0096] Advantageously, the adaptive preprocessing approach described here facilitates increased freedom of movement (due to the use of independent wireless EEG sensors) and can increase the duration of EEG monitoring sessions in which the EEG sensors record EEG data substantially continuously. Rather than being limited to a hospital bed or dedicated observation facility as with wired EEG monitoring systems, patients can wear discrete wireless sensors for days or weeks.

[0097] In some approaches, EEG data can be adaptively normalized in the feature space, potentially eliminating the need for adaptive processing. The extracted features can then be normalized using methods such as rolling normalization or moving normalization, as disclosed here. In such approaches, normalization is applied to a rolling window of the data, so the normalization parameters change according to the characteristics of the data, making the normalization adaptive.

[0098] Adaptive noise reduction Depending on the implementation, preprocessing may include denoising, which may further include identifying and removing noise components from EEG data. Because noise is highly non-stationary, distinguishing it from the underlying brain signals in EEG data can be difficult. Denoising techniques that operate using overall signal statistics or by broadly attenuating specific frequencies may fail to distinguish between changes in the underlying signal statistics and the inclusion of noise. Fortunately, denoising approaches have been developed that adapt to these changing statistics.

[0099] This approach removes the linear combination of frequency components with the worst signal-to-noise ratio (SNR), and then reconstructs the signal from only the components with better SNRs. Advantageously, since the developed approach operates independently on discrete segments of the time series, these components can be selected segment by segment. If the frequency statistics of the underlying signal change, the developed approach can selectively remove different frequency combinations.

[0100] EEG data can be divided into segments. Segment durations can range from 1 second to 1 hour, e.g., 1 second, 2 seconds, 30 seconds, 1 minute, 2 minutes, 5 minutes, 10 minutes, 30 minutes, 1 hour, etc. EEG data segments can be decomposed into multiple frequency component signals using appropriate transforms such as the Short-Time Fourier Transform and / or Continuous Wavelet Transform. For a particular segment, the result can be a measure of frequency power over time (or segment duration) across many frequencies. The frequency spectral representation can be transposed and decomposed by principal component analysis (PCA). Principal components explaining the smallest variance (representing a linear combination of frequencies) can be progressively eliminated until only the principal component explaining the largest variance remains. A threshold can be used to eliminate principal components that do not meet the threshold. As a result, the dimensionality of the component representation can be reduced. This reduced representation in component space can be used to reconstruct the frequency power representation by inverse transform. This reconstruction can then be inverse frequency transformed to reconstruct a time-series segment free of noise components.

[0101] In the modified version, the number of components used in the reconstruction of principal component analysis may be selected based on the minimum description length between the spectral power distribution of the reconstruction and the spectral power distribution of the original EEG segment. Minimizing the minimum description length is essential for effectively compressing the EEG data (and therefore the basis and The best compression of the data can be selected in terms of balancing the amount of EEG data actually lost due to compression (learning patterns). Advantageously, the number of PCA components can be selected to represent a good trade-off between removing non-essential parts of the EEG signal in the frequency domain (thus learning the underlying meaningful patterns) and accurately reproducing the frequency characteristics of the original signal.

[0102] The denoising approach described in this section may be part of preprocessing block 424 in Figure 4A. In some cases, the denoising approach described in this section may be used before or after adaptively standardizing the EEG data as described in the previous section. Depending on the implementation, the denoising approach described in this section may be used instead of adaptive standardization of the EEG data.

[0103] Expansion of extracted EEG data features Learning a seizure detection pathway (e.g., seizure detection pathway 420 in Figure 4A) can be challenging, especially if the seizure detection pathway is intended to detect specific electrical seizure characteristics across diverse patient populations. Often, available training data may consist only of limited EEG datasets collected from small patient populations. Unfortunately, such limited training data may be biased towards specific regions of the patient's brain. For example, the training dataset may be biased towards seizures occurring on the right side of the patient's brain. The metafeature creation approach described herein can increase the sensitivity of the seizure detection pathway while reducing the false positive rate by generalizing patient EEG data and considering or avoiding potential biases. In particular, metafeature generation can create cross-sensor features where location is no longer important. As described herein, metafeatures can be used to learn and / or detect one or more seizure events.

[0104] Figure 8 shows an exemplary process 800 for creating metafeatures by expanding multiple features. Process 800 can be executed by one or more computing devices, for example, by a remote computing device 304 alone, or in combination with one or more portable computing devices 302 or wearable sensors 301. Process 800 may include more or fewer steps. One or more steps of process 800 may be executed in a different order or simultaneously with one or more other steps. Instructions for executing process 800 may be stored in a computer-readable medium. These instructions may cause one or more processors to execute the steps of process 800.

[0105] Process 800 can be used to train a classifier for a single seizure detection pathway (e.g., seizure detection pathway 420 in Figure 4A) or a classifier for multiple seizure detection pathways (e.g., those described in relation to Figures 4B and 5).

[0106] Process 800 begins in step 802, where the EEG data channel may be segmented into multiple segments. As described above, each of the multiple individual wireless EEG sensors may be placed on the patient's scalp and provide an EEG data channel. Each of the multiple EEG data channels may be segmented into multiple EEG data segments. The multiple data segments may be associated with the original EEG sensor. For example, a segment generated from an EEG data channel provided by a first individual wireless EEG sensor may be associated with the first sensor. As described above, the multiple segments may have segment lengths. For example, each of the multiple segments may be 2 seconds.

[0107] Step 802 may include one or more aspects of the segmentation step of a single seizure detection pathway, or distinguished seizure detection pathways, for example, related to block 426 in Figure 4A. In some cases, multiple EEG channels may be combined into patient EEG data. Patient EEG data may be segmented into multiple EEG segments.

[0108] Process 800 then moves to step 804, where multiple features may be extracted from multiple EEG data segments. Multiple features may be extracted to form at least one set of extracted data features. Step 804 may include at least some aspects of the feature extraction steps of the seizure detection pathway described above, for example, related to block 428 in Figure 4A.

[0109] Process 800 proceeds to step 806, where multiple features may be expanded to create at least one metafeature. A metafeature may be created from one or more reference features of the extracted feature dataset. A reference feature may be at least one feature extracted from multiple segments. A metafeature may contain data related to one or more extracted features and may be useful for expanding the extracted feature dataset. A metafeature may generalize data from one or more EEG channels, for example, by pooling channels. This may reduce artifacts associated with data characteristics appearing differently across multiple discrete EEG sensors.

[0110] Feature augmentation may involve pooling one or more reference features to create at least one metafeature. The metafeature may be derived from at least one corresponding or related feature associated with at least two EEG data channels (e.g., using pooling, correlation, division, etc.) or from a single EEG data channel (e.g., using trend, moving average, moving normalization, etc.). For example, when a discrete EEG sensor is used, pooling may provide a feature derived from a subset (or all) of the EEG channels associated with the discrete EEG sensor. For example, pooling may involve taking the maximum (or minimum, mean, etc.) value across all EEG channels, thereby allowing seizure events to be generalized independently of their origin and / or sensor order. Advantageously, pooling makes the classifier independent of the order and / or position of the EEG channels, and as a result, it may be more robust, even when trained on smaller EEG datasets, and in seizure event detection at use.

[0111] For example, suppose features A(FA) are determined as FA_1, FA_2, FA_3, and FA_4 for four EEG channels. These can each be independent features and can be used independently by the classifier of the seizure detection pathway. The pooled metafeatures could be the minimum, maximum, mean, variance, etc., of the four features FA_1, FA_2, FA_3, and FA_4. Assuming FA_1=1, FA_2=1, FA_3=2, and FA_4=7, the minimum pooled metafeature is 1, the maximum pooled metafeature is 7, and the mean pooled metafeature is 2.75. One or more (or all) such pooled metafeatures may be used during classification. In some cases, one or more metafeatures, along with one or more features, may be used for seizure event detection (e.g., step 808). In some cases, only metafeatures may be used and features may be omitted in order to completely generalize the detection.

[0112] One or more meta-features can be derived from time-domain, frequency-domain, or complexity-domain features. Meta-features can enable the classifier to be location-agnostic. For example, a meta-feature corresponding to the mean entropy serves this purpose. In some cases, meta-features can enable the classifier to focus on the temporal locations where seizure events are likely to occur. For example, a meta-feature corresponding to the variance of band power (which represents the difference between four EEG channels) serves this purpose.

[0113] Meta-features can be derived to function as inter-channel generalizers (such meta-features are location-independent). Meta-features can be derived to determine inter-channel statistics. Meta-features can be derived to function as channel-specific time-series generalizers (e.g., moving average of features).

[0114] Metafeatures can be determined using statistical analyses such as first-order (e.g., feature derivative or feature variance) or second-order (e.g., variance of feature derivative). In some cases, composite metafeatures (e.g., moving average of pooled features) may be determined. For example, relative bandpower features can be determined for each EEG data channel and averaged to create a pooled relative bandpower feature. Extracted features can be pooled in time across a segment window containing multiple segments for one or more EEG data channels to create at least one metafeature. For example, mean derivative voltage values ​​can be pooled in time across a segment window to generate a moving average metafeature. As yet another example, metafeatures can be created based on the moment-to-moment variance of the extracted features. This moment-to-moment variance can be determined based on the variance of the extracted features over a certain period (e.g., seconds, minutes, or hours). As yet another example, metafeatures can be created based on a long-term temporal analysis of the extracted features, which may be performed over periods such as minutes, hours, days, or weeks to identify trends. Long-term temporal analysis may include calculations of moving averages, variances, etc. Long-term temporal analysis may correlate EEG data based on time and / or patient activity levels. For example, long-term temporal analysis may determine the moving average and / or mean-variance of a patient's EEG data at bedtime.

[0115] Metafeatures can be created from one or more features extracted from a single EEG data channel (e.g., derivative, variance, maximum, minimum, etc.) or from one or more features extracted from multiple EEG data channels (e.g., inter-channel variance, inter-channel maximum, inter-channel minimum, etc.).

[0116] Process 800 may proceed to step 808, in which seizure events are detected based on the extracted feature dataset and / or at least one meta-feature. Step 808 may include the classification and event identification steps described above in relation to Figure 4A (e.g., blocks 430, 432, and 434 in Figure 4A). Similar to block 430 in Figure 4A, step 808 may include applying a classifier to identify seizure events within a segment of a group of segments. The classifier may make this decision based on the extracted EEG data features and / or at least one meta-feature.

[0117] As described herein, process 800 may include training a classifier of seizure detection pathways using at least one metafeature. As described above, one or more models may be distinguished among multiple seizure detection pathways configured to detect multiple distinct electrical seizure characteristics.

[0118] In some cases, the creation of one or more metafeatures may be distinct for each distinguished seizure detection pathway. For example, a first seizure detection pathway may create and use a first set of metafeatures, while a second seizure detection pathway may create and use a second set of metafeatures that is different from the first set. As another example, a first seizure detection pathway may create and use a first set of metafeatures, while a second seizure detection pathway may not create any metafeatures at all.

[0119] Figure 9 shows a method in which multiple features 902 can be extracted from multiple EEG data channels 904, expanded, and one or more metafeatures 906 can be created. The EEG channel 904 may be collected by four separate EEG sensors positioned at four distinct locations on the patient's scalp. For example, the locations may correspond to the left and right frontal regions (indicated as F7 and F8, respectively) and the temporoparietal region (indicated as TP9 and TP10, respectively).

[0120] Extensions for training EEG data Learning a seizure detection pathway (e.g., seizure detection pathway 420 in Figure 4A) can be challenging, especially if the seizure detection pathway is intended to detect specific electrical seizure characteristics across diverse patient populations. Ideally, the seizure detection pathway would be trained using a very large set of training data that reflects many different data acquisition scenarios for diverse patient populations. However, acquiring such a set of training data may not always be feasible. Instead, a smaller set of training data can be augmented to cover a variety of data acquisition scenarios, creating a training dataset with sufficient diversity to train the seizure detection pathway to operate with target sensitivity and false-positive rates. Furthermore, augmenting a small set of training data can be particularly important if the training data is acquired from a small number of scalp-mounted EEG sensors, such as the four separate wireless EEG sensors described herein. The channels of EEG data obtained from such a sensor arrangement can be augmented to create a training dataset containing additional features. Advantageously, this can improve the convergence of the seizure detection pathway classifier and enhance the robustness of the pathway.

[0121] The training data can be augmented by scaling, inverting (or scaling by -1), adding noise, rearranging, changing associated positions, etc., for one or more channels of EEG data in the training dataset. For example, multiplying a channel of EEG data generated by a particular EEG sensor by -1 can simulate the placement of that particular EEG sensor in the opposite direction on the patient's scalp. As another example, EEG data collected by an EEG sensor located on the left forehead can be swapped with EEG data collected by an EEG sensor located on the right forehead; this is an example of rearrangement. As a result, the training set may include a first set of EEG data with data collected by sensors located on the left and right foreheads, and a second set of EEG data with the swapped EEG data. In this way, the training set can be augmented, making the classifier of seizure detection pathways more robust and more location-independent of possible seizure events. As yet another example, EEG data collected by an EEG sensor located behind the left ear can be swapped with EEG data collected by an EEG sensor located behind the right ear; this is another example of rearrangement. In this way, the EEG dataset can be expanded and used to train a classifier for seizure detection pathways.

[0122] As yet another example, different hypothetical noise characteristics that may exist in the real world, such as 60Hz line noise, mechanical noise (e.g., from a hair dryer), or electromagnetic noise emitted from a microwave oven, can be added to the EEG data. Adding such noise characteristics to raw EEG data not only augments the training data but also creates a wider variety of real-world simulations that can make the classifier more robust.

[0123] Figure 10 shows the creation of a new EEG data channel 1212 based on multiple EEG data channels 1214. The new EEG data channel 1212 and the multiple EEG data channels are available for training. One or more seizure detection pathways may be trained to detect seizure events using the multiple EEG data channels and the new EEG data channel. A particular seizure detection pathway may be trained to detect distinguished electrical seizure characteristics. As described herein (for example in relation to Figures 4A-4B and 5), a seizure detection pathway uses the multiple EEG data channels 1214 and the new EEG data channel 1212 (or multiple such EEG data channels) It may include a classifier trained to detect seizure events.

[0124] Identification of seizure events As described herein, seizure detection can be performed on EEG data segments of a specific duration. A seizure event may include multiple segments. Some seizure detection methods rely on comparing the detected seizure probability for a segment to a threshold. If the probability of a given segment meets the threshold, that segment is determined to be part of a seizure event. The individual seizure event is then determined to begin when the first data segment exceeds the threshold and end when the first data segment falls below the threshold. Such methods are prone to errors when encountering erroneous segments with erroneous probabilities. An erroneous segment may be misinterpreted as the end (or beginning) of an individual seizure event.

[0125] Advantageously, the event identification approach described herein can effectively solve at least these technical problems associated with detecting seizure events, such as individual seizure events. Unlike conventional seizure detection methods, by using different event identifiers within and between different seizure detection pathways, target sensitivity and false positive rates can be achieved across diverse patient populations. As described herein, in some cases, a sensitivity of at least 70% and a false positive rate of approximately 0.21 or less per hour (or 0.08 or less per hour) can be achieved.

[0126] Figure 11 shows an exemplary process 1000 for detecting a seizure event. Process 1000 can be performed by one or more computing devices, for example, by a remote computing device 304 alone, or in combination with one or more portable computing devices 302 or wearable sensors 301. Process 1000 may include more or fewer steps. One or more steps of process 1000 may be performed in a different order or simultaneously with one or more other steps of process 1000. Instructions for performing process 1000 may be stored in a computer-readable medium. The instructions may cause one or more processors to perform the steps of process 1000.

[0127] Process 1000 begins in step 1002, in which patient EEG data may be segmented into multiple segments (for example, as described in relation to block 426 in Figure 4A). As described above, each sensor of a plurality of individual wireless EEG sensors may be placed on the patient's scalp and provide an EEG data channel. The plurality of EEG data channels may be combined as patient EEG data and segmented into multiple segments of segment length. For example, each of the multiple segments may be 2 seconds long. Each of the plurality of EEG data channels may be segmented independently. Step 1002 may include one or more aspects of the segmentation step of the distinguished seizure detection pathway described above (for example, as described in relation to Figures 4B and 5). In some implementations, the segmentation step of the plurality of seizure detection pathways may be distinguished as described above.

[0128] Process 1000 proceeds to step 1004, where the probability that seizure activity was identified by the classifier (sometimes called the seizure probability) may be determined for each segment of a group of segments. Step 1004 may include some or all aspects of the classification steps of the seizure detection pathway described above in relation to Figures 4A-4B and 5. In some cases, the classifier may be applied to determine the probability of a seizure event within a segment based on at least one feature extracted from that segment. The probability value may be determined on a scale ranging from 0 to 1.

[0129] Process 1000 proceeds to step 1006, where the probability determined for a segment is modified based on the probabilities determined for one or more adjacent segments. Step 1006 may include several aspects of the event identification step (or stitching) described above in relation to the distinguished seizure detection pathway (for example, in relation to Figures 4B and 5). Step 1006 may include applying one or more event identifiers to analyze the probabilities determined for one or more segments (for example, one or more segments within a segment window). In some cases, the event identifiers may evaluate the probability value of a particular segment and the probability values ​​of one or more adjacent segments.

[0130] In some implementations, one or more filtering, smoothing, thresholding, and morphological transformation processes are used to convert segment probabilities into seizure events (e.g., individual seizure events) in order to meet target sensitivity and false positive rates.

[0131] To address issues with incorrect segments and probabilities, event identifiers may be modified using one or more morphological transformations to correct the probability values ​​of segments. In some cases, one or more morphological transformations may operate on binary probability values ​​(e.g., 0 and 1) (rather than continuous values ​​between 0 and 1). The probability values ​​of segments may be transformed into binary probability values ​​based on a probability threshold. For example, if the probability value is less than the probability threshold, a binary value of 0 may be assigned to the segment (indicating that at least initially, a seizure was not identified). Conversely, if the probability value is greater than the probability threshold, a binary value of 1 may be assigned to the segment (indicating that at least initially, a seizure was identified). In some implementations, probability thresholds may be distinguished between seizure detection pathways. For example, the event identification steps of distinguished seizure detection pathways may be distinguished by applying different probability thresholds.

[0132] The binary values ​​of a particular segment can be modified based on the binary values ​​of one or more adjacent segments using one or more morphological transformations. This may include modifying the probability values ​​of a segment based on the probability values ​​of adjacent segments. For example, the probability value of a segment that does not exceed a probability threshold can be adjusted to exceed the threshold based on the probability values ​​of one or more adjacent segments exceeding the probability threshold. The probability value of a segment that exceeds a probability threshold can be adjusted so that it does not exceed the threshold based on the probability values ​​of one or more adjacent segments not exceeding the probability threshold.

[0133] One or more morphological transformations may include contraction, expansion, opening (or contraction followed by expansion), and closing (or expansion followed by contraction). In some implementations, one or more morphological transformations performed by an event identifier may include one or more of the following operations: 1) "removal" of gaps shorter than a minimum gap threshold (e.g., consecutive 2, 3, 4… segments depending on segment length) (where consecutive binary 0 segments are between consecutive binary 1 segments); 2) "removal" of binary 1 segments whose duration is less than a minimum discrete seizure duration threshold (e.g., consecutive 2, 3, 4… segments depending on segment length); and 3) padding discrete seizure events with a specific padding coefficient (e.g., 10, 20, 30… depending on segment length). The order of these morphological operations may be associated with morphological closing, as morphological operation #1 may be associated with expansion, followed by morphological operation #2, which may be associated with contraction.

[0134] Removal can be achieved by modifying the binary values ​​of one or more affected segments (e.g., modifying the binary from 0 to 1 or vice versa). For example, suppose two consecutive "no seizure" segments with a binary of 0 are sandwiched between two consecutive "seizure" segments with a binary of 1. Also, suppose the minimum gap threshold is 3. Following morphological transformation operation #1 that may be associated with expansion, the binary values ​​of the two adjacent "no seizure" segments are modified to 1 to remove the gap associated with the spurious no seizure event. As another example, consecutive... Assume that two "seizures" segments are sandwiched between "no seizure" segments, and the minimum discrete seizure duration threshold is 3. Following morphological transformation operation #2 that may be associated with contractions, the binary values ​​of the two adjacent "seizures" segments are modified to 0 to remove spurious seizure events.

[0135] As an example of padding related to morphological transformation processing #3 that may be associated with dilation, individual seizure events may be expanded to include one or more adjacent segments with a binary value of 0. This may facilitate clinicians' review of electroencephalogram (EEG) data. In some cases, padding processing may be omitted. Padding processing may be distinguished between multiple seizure detection pathways. For example, absence seizures may last at least about 10 seconds, while focal seizures may have a minimum duration of about 2 minutes. Therefore, one or more event identifiers in a seizure detection pathway designed and tuned to detect one or more electrical seizure characteristics that may be associated with absence seizures may be padded to ensure a minimum duration of about 10-15 seconds. One or more event identifiers in a seizure detection pathway designed and tuned to detect one or more electrical seizure characteristics that may be associated with focal seizures may be padded to ensure a minimum duration of about 1-2 minutes.

[0136] In some cases, the probability values ​​of a segment may be modified based on the moving average and / or the variation in the probability values ​​of one or more segments. This type of smoothing may be performed prior to one or more morphological transformations.

[0137] Process 1000 may proceed to step 1008, where the entire seizure event may be detected. The entire seizure event may span a segment with a modified probability value and one or more adjacent segments. The start time, stop time, and / or duration of the entire seizure event, as well as a confidence value, may be determined.

[0138] Using the approach described herein, seizure events (e.g., the entire seizure event) can be detected with target sensitivity and false-positive rates. Sensitivity and false-positive rates may be inversely correlated but can be adjusted to meet output standards. In some implementations, the entire seizure event can be detected with a sensitivity of approximately 70% to approximately 90% (e.g., at least 80%) and a false-positive rate of approximately 0.21 events / hour (or, in some cases, approximately 0.08 events / hour).

[0139] Channel expansion EEG data collected by one of several sensors may be lost as a result of one or more malfunctions. For example, if individual wireless EEG sensors (e.g., sensor 101 or 301) are used, EEG data collected by one or more sensors may be lost, for example, during transmission. The techniques described herein can favorably address these problems by generating missing EEG data channels (or missing data channels) from the received EEG data channels. Learning and / or seizure detection can be performed using the received EEG data channels along with the generated EEG data channels.

[0140] Figure 12 shows an exemplary process 1100 for creating a model (or synthesized) EEG data channel that can be used in place of a missing EEG data channel. Process 1100 can be performed by one or more computing devices, for example, a remote computing device 304 alone, or in combination with one or more portable computing devices 302 or wearable sensors 301. Process 1100 may include more or fewer steps. One or more steps of process 1100 may be performed in a different order or simultaneously with one or more other steps. Instructions for performing process 1100 may be stored in a computer-readable medium. These instructions may cause one or more processors to perform the steps of process 1100.

[0141] Process 1100 may begin with step 1102, in which EEG data collected by a plurality of individual wireless EEG sensors placed on the patient's scalp is processed. The EEG data collected by the plurality of individual wireless EEG sensors may be processed using one or more seizure detection pathways, as described herein in relation to Figures 4A-4B and 5. In some cases, one or more seizure detection pathways may be configured to detect distinct electrical seizure characteristics from the EEG data collected by the plurality of individual wireless EEG sensors.

[0142] As described herein, a plurality of individual wireless EEG sensors may include four EEG sensors positioned on the patient's scalp at locations corresponding to the left and right frontal and temporal-parietal regions. For example, the four EEG sensors may be positioned at the left forehead location corresponding to the F7 position, the right forehead location corresponding to the F8 position, behind the left ear corresponding to the TP9 position, and behind the right ear corresponding to the TP10 position.

[0143] Each individual wireless EEG sensor may provide an EEG data channel. An EEG data channel may be associated with the location of the corresponding EEG sensor. For example, an EEG data channel may be associated with the precise location of a discrete wireless sensor, such as the F7 or TP10 location. In some cases, an EEG data channel may be associated with the patient's cerebral hemisphere. For example, an EEG data channel may be associated with the left or right hemisphere, which may be defined in relation to the patient's midline. In another example, an EEG data channel may be associated with the anterior or posterior hemisphere, which may be defined in relation to the patient's frontal plane. Further in some cases, an EEG data channel may be associated with a lobe of the patient's brain, such as the frontal, parietal, occipital, or temporal lobe, based on the lobe closest to the location of the corresponding EEG sensor.

[0144] Next, the process may proceed to step 1104, in which missing EEG data channels may be detected. The number of EEG data channels provided or received from the EEG sensor may be compared to the expected number of EEG data channels. If the number of EEG data channels is less than the expected number, missing EEG data channels may be detected. For example, as discussed herein, EEG data channels may be transmitted independently by multiple individual wireless EEG sensors to a portable computing device (e.g., device 302) and compiled into patient EEG data. The portable computing device may receive three EEG data channels when four EEG data channels are expected, and therefore missing EEG data channels may be detected.

[0145] In some embodiments, missing EEG data channels may be detected based on the location associated with the EEG data channel. Continuing with the example of the four individual wireless EEG sensors described above, a portable computing device (or another computing device) might expect to receive EEG data channels F7, F8, TP9, and TP10. However, only EEG data channels F8, TP9, and TP10 might be received. Thus, the portable computing device might detect that the EEG data channel F7 is missing. This EEG data channel may be missing due to a failure in the individual wireless EEG sensor.

[0146] Process 1100 may proceed to step 1106 in which a model EEG data channel is created. The model EEG data channel may be created based on an EEG data channel that exists in the received EEG data. The model EEG data channel may be created based on EEG data collected from a patient population. As discussed herein, the patient population may include patients, a group of patients suffering from a particular seizure type, a group of patients suffering from any type of seizure, or a group of patients including a first subgroup suffering from seizures and a second subgroup not suffering from seizures. A model EEG data channel can be created based on EEG data channels present within the EEG data and EEG data collected from a patient population.

[0147] A model EEG data channel can be created using one or more existing EEG data channels (e.g., by modification) based on the location of a missing EEG data channel and the locations of existing EEG data channels. For example, the location of a missing EEG data channel may be on the opposite side of the midline to at least one EEG data channel. The model EEG data channel can be created by modifying that at least one EEG data channel. As an example, an EEG data channel obtained from an EEG sensor located behind the left ear may be used to generate a model EEG channel corresponding to a location behind the right ear (e.g., using shadow manifold cross-sampling as described herein). In another example, the location of a missing EEG data channel may be associated with a first EEG sensor located on the opposite side of the frontal plane to a second EEG sensor that provided at least one EEG data channel. The model EEG data channel can be created by modifying that at least one EEG data channel (e.g., using shadow manifold cross-sampling as described herein).

[0148] Shadow manifold cross-sampling (SMaCS) may be used to create model EEG data channels. Shadow manifold cross-sampling (SMaCS) is a technique that generates highly realistic synthetic EEG data that can share many important statistics with real EEG data and can be used to replace missing EEG channels or to expand the training dataset by extending existing data. SMaCS can use an existing EEG data channel as a base signal and emulate that signal by forcing a stochastic differential equation derived based on the sample-by-sample difference statistics of the original signal, thereby maintaining similar long-term temporal relationships to those seen in the base signal. The base signal can be selected as the EEG data channel opposite (contra-side) to the missing EEG data channel, as described herein. The output of SMaCS can be used to expand the training dataset by extending an EEG dataset containing all EEG data channels, either by replacing the missing EEG data channel with data from the contra-side EEG data channel, or by selectively replacing the channel with an SMaCS synthetic time series derived from the contra-side channel. SMaCS is a general-purpose method for generating synthetic EEG data and can be used in any situation where such data is needed or may be beneficial. SMaCS can be used during training and seizure detection.

[0149] SMaCS may consist of the following two parts. The first part may relate to the sampling and emulation of stochastic processes. The base EEG signal S can be treated as a stochastic process, and the probability density function estimate (P) of the sample-by-sample difference (hereinafter referred to as the first-order difference) of the original signal can be collected for each interval over the output range of S. This allows for sampling (S) of any EEG data channel at time t. t ) to, S t It is possible to map a given range of values ​​of a signal S to a probability distribution of a first difference. The range of the signal S can be divided into M non-overlapping bins that cover that range. The bins may or may not be equally spaced. For each interval R (corresponding to the M bins) of S, a set of probability density estimates P can be constructed. For any P Rk is, range Rk This can represent the probability density estimation of the derivative of S in .

[0150] A mapping function G can be defined between values ​​within the range S and probability density estimates for a suitable interval R.

number

[0151] Next, process 1100 moves to step 1108, in which seizure events may be detected based on the existing EEG data channels and model EEG data channels included in the EEG data. Step 1108 may include processing the existing EEG data channels and model EEG data channels with one or more seizure detection pathways. For example, as described herein (in relation to Figures 4A-4B and 5), the existing EEG data channels and model EEG data channels may be preprocessed and segmented into multiple segments. Multiple features may be extracted from these multiple segments. Then, by applying a classifier based on these multiple features, seizure events may be detected.

[0152] Figure 13 shows how to create a model EEG data channel 1202 based on multiple EEG data channels present in the EEG data, such as EEG data channel 1204 for F7, EEG data channel 1204 for F8, and EEG data channel 1206 for TP9.

[0153] Event identification and confidence As described herein (for example, in relation to Figures 4A, 4B, and 5), seizure detection pathways can be designed and trained for the detection of seizure events. In some embodiments, a seizure detection pathway includes multiple event identifiers. As described herein (for example, in the section titled “Seizure Detection Pathways for EEG Data Collection and Processing”), event identifiers may be configured to evaluate EEG data segment probabilities (e.g., 0 to 1) determined by the seizure detection pathway's classifier and output labels representing temporally extended seizure phenomena (e.g., start and end times, and / or duration, of individual seizure events). Parameterizing some event identifiers or interpreting their probabilities may detect more true events at the cost of more errors, while others may detect fewer true events but also produce fewer errors. Therefore, as described herein, a subset of event identifiers that offer a beneficial trade-off between detecting events and generating errors may be selected during training. Confidence values ​​may be derived from which event identifiers detect events and which do not in the segment probability data. Parameterizing event identifiers, which can lead to many errors, can result in lower certainty in positive detection, leading to lower confidence when predicting events. On the other hand, parameterizing event identifiers, which produces fewer false detections, can result in higher certainty in the positive detection it performs, leading to higher confidence when predicting events.

[0154] Many different parameterizations of the event identification pipeline can be created. It can be determined how often they detect true events and how often they produce false positives. From all of those tested during training, a subset that provides a useful trade-off between these two measures can be selected to interpret EEG data collected from patients. When used with patients, one of the subsets of parameterizations for that event identifier can be selected. Confidence can potentially be derived based on whether or not one or more event identification parameterizations detect events, and how frequently such one or more event identification parameterizations produced false detections on the training data.

[0155] Figure 14 shows plot 1350 with event identifier parameterizations 1362, 1364, and 1366. In plot 1350, the x-axis corresponds to false detections and the y-axis corresponds to true detections. Parameterization 1362 of the event identifier detects many true events but also many errors, and can be considered to have high true detection but low confidence. Parameterization 1364 of the event identifier detects fewer true events than 1362 but fewer true events, and can be considered to have moderate confidence. Parameterization 1366 of the event identifier detects even fewer true events than 1362 and 1364, but has very few errors and can be considered to have high confidence.

[0156] If an event is detected by parameterizations of multiple event identifiers, the confidence value of that event can be determined based on the confidence of the parameterization of the event identifier that detected the event, which has the "highest" confidence level. For example, if only parameterization 1362 of event identifiers detected the event, a low confidence value would be assigned to that event. As another example, if parameterization 1364 of event identifiers detected the event (and possibly parameterization 1362 as well), a moderate confidence value would be assigned to that event. As yet another example, if parameterization 1366 of event identifiers detected the event (and possibly parameterizations 1362 and / or 1364 as well), a high confidence value would be assigned to that event. In other words, the confidence value for an event may be assigned based on the determination that the parameterization of the event identifier with the highest confidence level detected the event, regardless of whether other parameterizations of event identifiers with lower confidence levels also detected the event. One or more events detected by parameterizations of event identifiers with lower confidence levels may be discarded.

[0157] In some cases, the confidence value can represent precision and can be determined as follows:

[0158] Confidence = 1 - False detection rate = 1 - False positives / (True positives + False positives)

[0159] Figure 15 shows an exemplary process 1400 that outputs seizure events along with confidence levels. Process 1400 can be performed by one or more computing devices, either by the remote computing device 304 alone or in combination with one or more portable computing devices 302 or wearable sensors 301. Process 1400 may consist of more or fewer steps. One or more steps of process 1400 may be performed in a different order or concurrently with one or more other steps. Instructions for performing process 1400 may be stored in a computer-readable medium. These instructions may cause one or more processors to perform the steps of process 1400.

[0160] Process 1400 begins in step 1402, in which EEG data collected by multiple individual wireless EEG sensors is processed by multiple classifiers (e.g., Figure 4B) in a seizure detection pathway (e.g., Figure 4A) and / or a distinguished pipeline. The output of step 1402 may be a probability value associated with an EEG data segment.

[0161] Process 1400 moves to step 1404, where a subset of the parameterized event identifiers selected during learning is converted into probability values ​​for seizure events.

[0162] Process 1400 then moves to step 1406, where confidence values ​​are determined based on which of the parameterized subsets of event identifiers detected events and how frequently they resulted in false positives in the training data.

[0163] Process 1400 proceeds to an optional step 1408, where, if multiple pipelines are used, the seizure events and confidence values ​​output by the parameterization of event identifiers associated with the distinct pipelines are merged. This may be similar to the event merge block 460 in Figure 4B. As described herein, step 1408 can analyze the confidence values ​​of seizure events determined by the parameterization of each event identifier. For example, in post-detection, the individual seizure event associated with the highest confidence value among all detected events may be selected in step 1408. As another example, in rapid detection, the event with the highest confidence value that shows an occurrence rate of an EEG characteristic or pattern that satisfies an occurrence rate threshold may be selected in step 1408. If there is no parameterization of event identifiers that outputs an event with such an occurrence rate, the event with the highest confidence value may be selected in step 1408.

[0164] Process 1400 proceeds to step 1410, where notification of a seizure event may be provided along with a confidence value (e.g., Figure 16). The confidence value can be reported as a numerical value (e.g., 68%) or as a level. The latter may be a more useful tool for clinicians. For example, as described in this section, confidence levels may include three levels or ratings (high, medium, low). As another example, confidence levels may include four levels or ranges (very high, high, medium, low). In some cases, low confidence may correspond to less than 50%, medium confidence to 50%-79%, high confidence to 80%-94%, and very high confidence to 95% or higher. More generally, low confidence may correspond to less than A%, medium confidence to A%-B% (where B is greater than A), high confidence to C%-D% (where C is greater than B and less than D), and very high confidence to E% (where E is greater than D).

[0165] Notifications may include visual, auditory, or tactile outputs. For example, notifications may be generated on a portable computing device. Notifications may include elements of a graphical user interface that include indications of possible seizures and confidence levels. Indications may include information about possible seizure events, including labels (e.g., onset time, termination time, duration, or incidence rate of individual seizure events) and / or patient EEG data associated with the seizure event.

[0166] Figure 16 shows the seizure detection output 1500, which can be presented as a graphical user interface on the display. Traces of EEG data obtained by individual EEG sensors (e.g., four sensors 301 located on the left forehead, right forehead, behind the left ear, and behind the right ear) as well as differential EEG data are presented in a chart or graph 1502. Differential EEG data can be obtained by combining EEG data obtained by individual EEG sensors (e.g., F8-TP10, F7-F8, TP9-TP10, F7-TP10, and F8-TP9). The illustrated traces of EEG data can be associated with vertical and horizontal montages.

[0167] The detection of potential seizure events is presented in region 1504 along with their confidence levels. For example, potential seizure event 1510 is associated with a high confidence level. As another example, potential seizure event 1512 is associated with a moderate confidence level. As described herein, the confidence levels can assist clinicians in determining whether a detected seizure event is a true seizure. In some cases, confidence values ​​may be output additionally or alternatively.

[0168] As described herein (see, for example, Figures 4A and 4B), one or more seizure detection pathways may output seizure events. Events detected by multiple seizure detection pathways that overlap with each other may be merged into a single seizure event. The outputs of seizure detection pathways may be merged into a single seizure event. This may be done, for example, by the event merging block 460 in Figure 4B. Assume that a first seizure event output by a first seizure detection pathway overlaps in time with a second seizure event output by a second seizure detection pathway. The first and second seizure events may be concatenated as a single seizure event for output. For example, the first seizure event (which may be separate seizure events) may be associated with a start time of 10:00:00, an end time of 10:00:35, and a confidence level of 50%. The second seizure event may be associated with a start time of 10:00:20, an end time of 10:00:48, and a confidence level of 30%. A concatenated seizure event associated with a start time of 10:00:00, an end time of 10:00:48, and a confidence level of 30% may be output.

[0169] In some cases, one of the first or second seizure events may be retained, while the other is discarded. In some cases, the seizure event with the highest confidence level may be retained. Continuing the above example, the first seizure event may be retained, and the second seizure event may be discarded. If both the first and second seizure events are associated with the same confidence level, the event with the shortest duration may be retained.

[0170] In some implementations, retaining one of the seizure events may be done based on a confidence value determined over time. For example, suppose seizure detection path A outputs high-confidence seizure events over a period of time, and seizure detection path B outputs lower-confidence seizure events over the same period. At some point later, if paths A and B output seizure events with the same confidence level, the seizure events provided by path A may be retained because path A had previously output seizure events with a higher confidence level.

[0171] Additional examples The following are exemplary systems, methods, and computer-readable media for detecting individual seizure events. These examples are intended to illustrate various embodiments and are not intended to limit the embodiments described herein. Any feature in one example can be combined with one or more features in one or more other examples.

[0172] 1. Differentiated routes In some embodiments, the techniques described herein relate to a method for detecting seizure events using electroencephalogram (EEG) signals, the method comprising: by one or more first processors: configuring a plurality of seizure detection paths based on EEG data collected from a patient population by a plurality of first EEG sensors, each seizure detection path comprising one or more distinguished steps, the plurality of seizure detection paths being configured to detect distinguished electrical seizure characteristics; and by one or more second processors: processing EEG data collected from patients evaluated using the plurality of seizure detection paths by a plurality of second EEG sensors to provide a plurality of distinguished outputs relating to seizure events associated with distinguished electrical seizure characteristics; and outputting labels associated with seizure events from the plurality of distinguished outputs relating to possible seizure events.

[0173] In some embodiments, the techniques described herein relate to methods in which labels indicate the onset time and duration of a seizure event.

[0174] In some embodiments, the techniques described herein relate to methods, and labels are used for a certain period. This shows the occurrence rate of electrical seizure characteristics over time.

[0175] In some embodiments, the techniques described herein relate to a method and include a plurality of seizure detection pathways configured to detect a first seizure detection pathway configured to detect a first electrical seizure characteristic indicating an absence seizure; a second seizure detection pathway configured to detect a second electrical seizure characteristic indicating a tonic-clonic seizure; and a third seizure detection pathway configured to detect a third electrical seizure characteristic indicating a combination of two or more seizure types.

[0176] In some embodiments, the techniques described herein relate to a method in which each of a plurality of seizure detection pathways includes a preprocessing step, a segmentation step, a feature extraction step, a classification step, and an event identification step.

[0177] In some embodiments, the techniques described herein relate to methods in which multiple seizure detection pathways are distinguished in a preprocessing step that includes standardizing EEG data to account for differences between patients and within patients.

[0178] In some embodiments, the techniques described herein relate to a method in which multiple seizure detection pathways are distinguished in a segmentation step.

[0179] In some embodiments, the techniques described herein relate to methods and are distinguished in a feature extraction step which includes identifying multiple features based on electrical seizure characteristics detected by the seizure detection paths, and extracting the multiple features.

[0180] In some embodiments, the techniques described herein relate to methods and are distinguished in a classification step in which multiple seizure detection pathways assign a probability of the occurrence of an electrical seizure characteristic by applying a classifier tuned to detect electrical seizure characteristics.

[0181] In some embodiments, the techniques described herein relate to a method in which multiple seizure detection pathways are distinguished in an event identification step that includes converting one or more probabilities determined by a classifier into labels.

[0182] In some embodiments, the techniques described herein relate to methods, the methods further comprising: receiving one or more patient physiological data acquired by one or more first sensors and environmental data acquired by one or more second sensors; and outputting labels based on one or more of the patient physiological data and environmental data by one or more second processors.

[0183] In some embodiments, the techniques described herein relate to a non-temporary computer-readable medium, the computer-readable medium storing instructions, and when the instructions are executed by one or more first and second processors, the first and second processors cause one or more first and second processors to perform a method for detecting seizure events using electroencephalogram (EEG) signals, the method being: by one or more first processors: configuring a plurality of seizure detection paths based on EEG data collected from a patient population by a plurality of first EEG sensors, each seizure detection path comprising one or more distinguished steps, the plurality of seizure detection paths being configured to detect distinguished electrical seizure characteristics; and by one or more second processors: processing EEG data collected from patients evaluated using the plurality of seizure detection paths by a plurality of second EEG sensors to provide a plurality of distinguished outputs relating to seizure events associated with distinguished electrical seizure characteristics; and from the plurality of distinguished outputs relating to possible seizure events, extracting labels associated with the seizure events. This includes the act of exerting force.

[0184] In some embodiments, the techniques described herein relate to non-temporary computer-readable media, where the labels indicate the onset time and duration of a seizure event.

[0185] In some embodiments, the techniques described herein relate to non-transient computer-readable media, where the label indicates the occurrence rate of electrical seizure characteristics over a period of time.

[0186] In some embodiments, the techniques described herein relate to a non-transient computer-readable medium, and the seizure detection pathways include a first seizure detection pathway configured to detect a first electrical seizure characteristic indicating an absence seizure, a second seizure detection pathway configured to detect a second electrical seizure characteristic indicating a tonic-clonic seizure, and a third seizure detection pathway configured to detect a third electrical seizure characteristic indicating a combination of two or more seizure types.

[0187] In some embodiments, the techniques described herein relate to non-temporary computer-readable media, and each of the multiple seizure detection paths includes a preprocessing step, a segmentation step, a feature extraction step, a classification step, and an event identification step.

[0188] In some embodiments, the techniques described herein relate to non-temporal computer-readable media, and multiple seizure detection pathways are distinguished in a preprocessing step that includes standardizing EEG data to account for inter-patient and intra-patient differences.

[0189] In some embodiments, the techniques described herein relate to non-temporary computer-readable media, and multiple seizure detection pathways are distinguished in the segmentation step.

[0190] In some embodiments, the techniques described herein relate to non-temporary computer-readable media, and the multiple seizure detection paths are distinguished in a feature extraction step which includes: identifying multiple features based on electrical seizure characteristics detected by the seizure detection paths; and extracting multiple features.

[0191] In some embodiments, the techniques described herein relate to non-temporary computer-readable media, and multiple seizure detection pathways are distinguished in a classification step which includes assigning probabilities of the occurrence of electrical seizure characteristics by applying a classifier tuned to detect electrical seizure characteristics.

[0192] In some embodiments, the techniques described herein relate to non-temporary computer-readable media, and multiple seizure detection pathways are distinguished in an event identification step which includes converting one or more probabilities determined by a classifier into labels.

[0193] In some embodiments, the technology described herein relates to a non-temporary computer-readable medium, which further includes: receiving one or more of a patient's physiological data acquired by one or more first sensors and environmental data acquired by one or more second sensors; and outputting a label based further on one or more of the patient's physiological data and environmental data.

[0194] In some embodiments, the technology described herein relates to a system for detecting seizure events using electroencephalogram (EEG) signals, the system comprising: a plurality of individual wireless EEG sensors configured to be placed on the patient's scalp, each discrete wireless EEG sensor configured to collect EEG data channels; and a non-temporary computer-readable medium for storing instructions. The instruction, when executed by one or more processors, causes one or more processors to perform any of the methods of the preceding embodiments, and includes the following:

[0195] 2. Event identification and confidence level In some embodiments, the techniques described herein relate to a method for detecting seizure events using electroencephalogram (EEG) signals, the method comprising: by a first one or more processor: selecting a set of event identifiers configured to translate from a plurality of event identifiers into a first set of labels associated with a first seizure event, where the first plurality of probabilities are determined by a classifier, and each event identifier in the set of event identifiers has a confidence value indicating the accuracy with which the event identifier determines the label; and by a second one or more processors: using the set of event identifiers to translate a second set of labels associated with a second seizure event, where the second plurality of probabilities are determined by a classifier; and outputting from the second set of labels the label determined by the event identifier having a higher confidence value than any other event identifier in the set of event identifiers; and providing notification of the second seizure event, along with the confidence level associated with the highest confidence value.

[0196] In some embodiments, the techniques described herein relate to methods, and a second set of labels corresponds to the onset time and duration of individual seizure events.

[0197] In some embodiments, the techniques described herein relate to a method in which a second set of labels corresponds to the occurrence rate of electrical seizure characteristics over a period of time, and the output labels further satisfy an occurrence rate threshold.

[0198] In some embodiments, the techniques described herein relate to a method and provide notification of a second seizure event along with a confidence level associated with the highest confidence value, which includes displaying the notification along with a confidence level, along with EEG data segments collected from the patient.

[0199] In some embodiments, the techniques described herein relate to methods, and the confidence levels include high, medium, or low.

[0200] In some embodiments, the techniques described herein relate to a method in which the EEG data segment is displayed as a vertical and horizontal montage including four channels of EEG data collected by four separate wireless EEG sensors located on the left forehead, right forehead, behind the left ear, and behind the right ear of the patient, and multiple channels derived by subtracting one pair of channels from the four channels of EEG data.

[0201] In some embodiments, the techniques described herein relate to a method in which each event identifier in a set of event identifiers has different parameterizations for translating a second set of probabilities into labels in a second set of labels.

[0202] In some embodiments, the techniques described herein relate to methods, and the highest confidence value corresponds to the highest precision.

[0203] In some embodiments, the techniques described herein relate to a method that outputs from a second set of labels a label determined by an event identifier having the highest confidence value of any other event identifier from the set of event identifiers, and discards labels determined by another event identifier having a lower confidence value than the event identifier. nothing.

[0204] In some embodiments, the techniques described herein relate to a method, which further includes: using a second set of one or more processors: converting a set of different probabilities of the occurrence of a set of different electrical seizure characteristics in an EEG data segment collected from a patient into a set of different labels associated with a set of different seizure events, the set of different probabilities being determined by a set of different classifiers; outputting a set of different labels determined by an event identifier having a higher confidence value than any other event identifier of the set of different labels; determining that the confidence value of the set of different labels exceeds the confidence value of the label; and providing notification of a set of different seizure events along with a set of different confidence levels associated with the set of different labels in response to the determination that the confidence value of the set of different labels exceeds the confidence value of the label.

[0205] In some embodiments, the techniques described herein relate to a method, the method further comprising: using a second one or more processors: converting a set of different probabilities of the occurrence of a set of different electrical seizure characteristics in an EEG data segment collected from a patient into a set of different labels associated with a set of different seizure events, the set of different probabilities being determined by a set of different classifiers; merging the second seizure event and confidence level with another seizure event and another confidence level to form a merged seizure event and merged confidence level; and providing notification of the merged seizure event along with the merged confidence level.

[0206] In some embodiments, the techniques described herein relate to a non-temporary computer-readable medium, and instructions are executed by one or more first and second processors, causing one or more first and second processors to perform a method for detecting seizure events using electroencephalogram (EEG) signals, the method being: by one or more first processors: selecting a set of event identifiers from a plurality of event identifiers, configured to translate a first plurality of probabilities of the occurrence of one or more electrical seizure characteristics in an EEG data segment of training data into a first set of labels associated with a first seizure event, the first plurality of probabilities being determined by a classifier, each of the set of event identifiers An event identifier has a confidence value indicating the accuracy with which the event identifier determines a label; and by one or more second processors: using the set of event identifiers, convert a second set of multiple probabilities of the occurrence of one or more electrical seizure characteristics in EEG data segments collected from a patient into a second set of labels associated with the second seizure event, the second set of multiple probabilities being determined by a classifier; and outputting from the second set of labels the label determined by the event identifier having a higher confidence value than any other event identifier in the set of event identifiers; and providing notification of the second seizure event along with the confidence level associated with the highest confidence value.

[0207] In some embodiments, the techniques described herein relate to a non-temporary computer-readable medium, where a second set of labels corresponds to the onset time and duration of individual seizure events.

[0208] In some embodiments, the techniques described herein relate to a non-temporary computer-readable medium, where a second set of labels corresponds to the occurrence rate of electrical seizure characteristics over a period of time, and the output labels further satisfy an occurrence rate threshold.

[0209] In some embodiments, the technology described herein relates to a non-temporary computer-readable medium and provides notification of a second seizure event along with a confidence level associated with the highest confidence value, along with EEG data segments collected from the patient, accompanied by a confidence level. This includes displaying knowledge.

[0210] In some embodiments, the techniques described herein relate to non-temporary computer-readable media, and the confidence level includes high, medium, or low.

[0211] In some embodiments, the techniques described herein relate to a non-temporary computer-readable medium, and the EEG data segment is displayed as a vertical and horizontal montage including four channels of EEG data collected by four separate wireless EEG sensors located on the left forehead, right forehead, behind the left ear, and behind the right ear of the patient, and multiple channels derived by subtracting one pair of channels from the four channels of EEG data.

[0212] In some embodiments, the techniques described herein relate to a non-temporary computer-readable medium, where each event identifier in a set of event identifiers has different parameterizations for translating a second set of probabilities into labels in a second set of labels.

[0213] In some embodiments, the techniques described herein relate to non-temporary computer-readable media, where the highest confidence value corresponds to the highest precision.

[0214] In some embodiments, the techniques described herein relate to a non-temporary computer-readable medium and include outputting from a second set of labels a label determined by an event identifier having a higher confidence value than any other event identifier in the set of event identifiers, and discarding labels determined by another event identifier having a lower confidence value than the event identifier.

[0215] In some embodiments, the techniques described herein relate to a non-temporary computer-readable medium, which further includes, by one or more second processors: converting a set of different probabilities of the occurrence of one or more different electrical seizure characteristics in an EEG data segment collected from a patient into a set of different labels associated with a different seizure event, using a set of different event identifiers, the different probabilities being determined by a different classifier; outputting a different label from the set of different labels, determined by an event identifier having a higher confidence value than any other event identifier of the other event identifier; determining that the confidence value of the other label exceeds the confidence value of the label; and providing notification of the other seizure event along with a different confidence level associated with the other label in response to determining that the confidence value of the other label exceeds the confidence value of the label.

[0216] In some embodiments, the techniques described herein relate to a non-temporary computer-readable medium, which further includes, by one or more second processors: converting a set of different probabilities of the occurrence of one or more different electrical seizure characteristics in an EEG data segment collected from a patient into a set of different labels associated with a different seizure event, where the different probabilities are determined by a different classifier; merging seizure events and confidence levels with other seizure events and confidence levels to form merged seizure events and merged confidence levels; and providing notification of the merged seizure events along with the merged confidence levels.

[0217] In some embodiments, the technology described herein relates to a system for detecting seizure events using electroencephalogram (EEG) signals, the system comprising: a plurality of individual wireless EEG sensors configured to be placed on the scalp of a patient, each discrete wireless EEG sensor configured to collect EEG data channels; and instructions A non-temporary computer-readable medium for storing instructions, the non-temporary computer-readable medium, when executed by one or more processors, causes one or more processors to perform the method of any of the preceding embodiments.

[0218] 3. Adaptive pretreatment In some embodiments, the techniques described herein relate to a method for detecting seizure events from electroencephalogram (EEG) signals, the method comprising: preprocessing patient EEG data collected from the scalp of a particular patient by a plurality of individual wireless EEG sensors to obtain preprocessed patient EEG data, wherein at least one parameter of the preprocessing changes over time in response to changes in the patient EEG data; and detecting seizure events from the preprocessed patient EEG data.

[0219] In some embodiments, the techniques described herein relate to a method in which patient EEG data are independently collected by each individual wireless EEG sensor of a plurality of individual wireless EEG sensors.

[0220] In some embodiments, the techniques described herein relate to methods, and the preprocessing includes standardizing patient EEG data, which is done by determining the variance of patient EEG data of multiple segments of patient EEG data; and correcting one or more segments of patient EEG data that have variances that do not meet a variance threshold which varies based on the variance of the multiple segments.

[0221] In some embodiments, the techniques described herein relate to a method in which the current value of the variance threshold is greater than the variance of a segment of patient EEG data, and the variance threshold is set to the variance of that segment of patient EEG data.

[0222] In some embodiments, the techniques described herein relate to a method in which, during the threshold period, the variance threshold is increased by a multiplier in response to not encountering any segments of patient EEG data where the variance is less than the current value of the variance threshold.

[0223] In some embodiments, the techniques described herein relate to a method which further includes monitoring changes in patient EEG data and adjusting at least one parameter of preprocessing in response to the changes.

[0224] In some embodiments, the techniques described herein relate to methods, and variations are due to the position or orientation of the individual wireless EEG sensors of a plurality of individual wireless EEG sensors.

[0225] In some embodiments, the techniques described herein relate to a method, and adjustment of at least one parameter of the pretreatment is performed in response to the failure to detect changes in patient EEG data over a certain period of time.

[0226] In some embodiments, the techniques described herein relate to a method in which multiple individual wireless EEG sensors are configured to collect patient EEG data without a common reference electrode.

[0227] In some embodiments, the preprocessing further includes denoising the patient EEG data to remove at least one frequency component in the patient EEG data.

[0228] In some embodiments, the techniques described herein relate to methods, where noise reduction includes: identifying one or more noise signals in a patient's electroencephalogram (EEG) data; and removing one or more noise signals from the patient's EEG data.

[0229] In some embodiments, the techniques described herein relate to a method for identifying one or more noise signals, which includes: dividing a patient's EEG data into multiple segments; and identifying one or more noise signals in response to determining that a linear combination of frequency components does not meet a threshold associated with frequency dispersion.

[0230] In some embodiments, the technology described herein relates to a non-temporary computer-readable medium, which, when executed by one or more processors, stores instructions causing one or more processors to: preprocess patient EEG data collected from a plurality of individual wireless EEG sensors from the scalp of a particular patient to obtain preprocessed patient EEG data, wherein at least one parameter of the preprocessing changes over time in response to changes in the patient's EEG data; and detect seizure events from the preprocessed patient EEG data.

[0231] In some embodiments, the techniques described herein relate to a non-temporary computer-readable medium, and patient EEG data are collected independently by each individual wireless EEG sensor of a plurality of individual wireless EEG sensors.

[0232] In some embodiments, the techniques described herein relate to non-temporary computer-readable media, and the preprocessing includes standardizing patient EEG data, the standardization including determining the variance of multiple segments of the patient's EEG data, and correcting one or more segments of the patient's EEG data that have variances that do not meet a variance threshold which changes based on the variance of the multiple segments.

[0233] In some embodiments, the techniques described herein relate to a non-temporary computer-readable medium, and the instructions are further configured to cause one or more processors to set a variance threshold to the variance of a segment in response to the current value of the variance threshold being greater than the variance of a segment of the patient's EEG data.

[0234] In some embodiments, the techniques described herein relate to non-temporal computer-readable media, where the variance threshold is increased by a multiplier in response to not encountering any segments of patient EEG data having a variance smaller than the current value of the variance threshold during the threshold period.

[0235] In some embodiments, the techniques described herein relate to a non-temporary computer-readable medium, and instructions are further configured to cause one or more processors to monitor a patient's EEG for changes in EEG data and to adjust at least one parameter of preprocessing in response to changes in EEG data.

[0236] In some embodiments, the techniques described herein relate to non-temporary computer-readable media, and the changes are due to the position or orientation of individual wireless EEG sensors among a plurality of individual wireless EEG sensors.

[0237] In some embodiments, the techniques described herein relate to non-temporary computer-readable media, and adjustment of at least one parameter of preprocessing is performed in response to the failure to detect changes in the patient's EEG data over a period of time.

[0238] In some embodiments, the technology described herein relates to a non-temporary computer-readable medium, and a plurality of individual wireless EEG sensors are configured to collect patient EEG data without a common reference electrode.

[0239] In some embodiments, the techniques described herein relate to a non-temporal computer-readable medium, and instructions are further configured to cause one or more processors to preprocess the patient's EEG data by denoising the patient's EEG data in order to remove at least one frequency component in the patient's EEG data.

[0240] In some embodiments, the techniques described herein relate to a non-temporary computer-readable medium, and instructions are further configured to cause one or more processors to denoise the patient's EEG data by identifying one or more noise signals in the patient's EEG data and removing one or more noise signals from the patient's EEG data.

[0241] In some embodiments, the techniques described herein relate to a non-temporary computer-readable medium, and instructions are further configured to cause one or more processors to identify one or more noise signals by: dividing a patient's EEG data into multiple segments; and identifying one or more noise signals in response to determining that a linear combination of frequency components does not meet a threshold associated with frequency dispersion.

[0242] In some embodiments, the technology described herein relates to a system for detecting seizure events from electroencephalogram (EEG) signals, the system comprising: a plurality of EEG sensors placed on the scalp of a particular patient and configured to collect the patient's EEG data; and a non-temporary computer-readable memory for storing instructions, the instructions of which, when executed by one or more processors, cause one or more processors to: preprocess the patient's EEG data collected by a plurality of individual wireless EEG sensors from the scalp of a particular patient to obtain preprocessed patient EEG data, wherein at least one parameter of the preprocessing changes over time in response to changes in the patient's EEG data; and detect seizure events from the preprocessed patient EEG data.

[0243] In some embodiments, the techniques described herein relate to a system in which patient EEG data is independently collected by each individual wireless EEG sensor of a plurality of individual wireless EEG sensors.

[0244] In some embodiments, the techniques described herein relate to a system, and the instructions further cause one or more processors to: determine the variance of the patient's EEG data in a plurality of segments of the patient's EEG data; and modify one or more segments of the patient's EEG data that have a variance that does not meet a variance threshold that changes based on the variances of the plurality of segments, thereby standardizing the patient's EEG data.

[0245] In some embodiments, the techniques described herein relate to a system, and the instructions further cause one or more processors to set the variance threshold to the variance of a segment of the patient's EEG data in response to the current value of the variance threshold being greater than the variance of the segment of the patient's EEG data.

[0246] In some embodiments, the techniques described herein relate to a system, and the variance threshold is increased by a multiplier in response to not encountering a segment of the patient's EEG data having a variance less than the current value of the variance threshold during a threshold period.

[0247] In some embodiments, the techniques described herein relate to a system, and the instructions further cause one or more processors to monitor the patient's EEG for changes in the EEG data and adjust at least one parameter of the preprocessing in response to the changes in the EEG data.

[0248] In some embodiments, the techniques described herein relate to a system, and the changes are due to the position or orientation of an individual wireless EEG sensor among a plurality of individual wireless EEG sensors.

[0249] In some embodiments, the techniques described herein relate to a system, and adjusting at least one parameter of the preprocessing is performed in response to not detecting a change in the EEG data over a period of time.

[0250] In some embodiments, the technology described herein relates to a system in which multiple individual wireless EEG sensors are further configured to collect patient EEG data without a common reference electrode.

[0251] In some embodiments, the techniques described herein relate to a system, and instructions are further configured to cause one or more processors to preprocess the patient's EEG data by denoising the patient's EEG data to remove at least one frequency component in the patient's EEG data.

[0252] In some embodiments, the techniques described herein relate to a system and instructions are further configured to cause one or more processors to denoise the patient's EEG data by: identifying one or more noise signals in the patient's EEG data; and removing one or more noise signals from the patient's EEG data.

[0253] In some embodiments, the techniques described herein relate to a system and instructions are further configured to cause one or more processors to identify one or more noise signals by: dividing patient EEG data into multiple segments; and identifying one or more noise signals in response to a determination that a linear combination of frequency components does not satisfy a threshold associated with frequency dispersion.

[0254] 4. Expansion of EEG data features (meta-features) and channel expansion (synthetic EEG data channels) In some embodiments, the techniques described herein relate to a method for detecting seizure events from electroencephalogram (EEG) signals, the method comprising: segmenting data collected from multiple EEG data channels collected by each discrete radio-EEG sensor of multiple individual radio-EEG sensors placed on a patient's scalp to form multiple EEG data segments; extracting multiple features from the multiple EEG data segments to form an extracted feature dataset associated with each of the multiple EEG data channels from which data was collected; creating a metafeature from at least one feature from the multiple features; and detecting seizure events based at least on the metafeature, the method being performed by one or more processors.

[0255] In some embodiments, the techniques described herein relate to a method, and creating metafeatures involves pooling values ​​for at least one feature.

[0256] In some embodiments, the techniques described herein relate to a method in which the multiple EEG data channels include four EEG data channels collected by four separate wireless EEG sensors located on the left forehead, right forehead, behind the left ear, and behind the right ear of the patient.

[0257] In some embodiments, the techniques described herein relate to a method for detecting seizure events, which involves using a classifier; metafeatures are used to train the classifier. It can be done.

[0258] In some embodiments, the techniques described herein relate to a method in which each discrete wireless EEG sensor of a plurality of individual wireless EEG sensors collects EEG data channels independently of other discrete wireless EEG sensors; and creating metafeatures involves generalizing the EEG data across two or more EEG data channels of the plurality of EEG data channels, thereby pooling the EEG data of two or more EEG data channels together for seizure event detection.

[0259] In some embodiments, the techniques described herein relate to a method for detecting seizure events, which further relies on one or more features from a plurality of features.

[0260] In some embodiments, the techniques described herein relate to a method in which metafeatures remove bias in multiple EEG data channels directed toward specific locations on the scalp for the detection of seizure events.

[0261] In some embodiments, the techniques described herein relate to an electroencephalogram (EEG) system, which includes: a plurality of individual wireless EEG sensors configured to be placed on the scalp of a patient, each discrete wireless EEG sensor configured to collect EEG data channels of a plurality of EEG data channels; and a non-temporary computer-readable medium storing instructions, which, when executed by one or more processors, causes one or more processors to: form a plurality of EEG data segments by segmenting the plurality of EEG data channels; extract a plurality of features from the plurality of EEG data segments to form an extracted feature dataset associated with each of the plurality of EEG data channels from which data has been collected; create a metafeature from at least one of the plurality of features; and detect a seizure event based at least on the metafeature.

[0262] In some embodiments, the techniques described herein are related to a system, and a meta-feature is created by pooling values for at least one feature.

[0263] In some embodiments, the techniques described herein are related to a system, and a plurality of individual wireless EEG sensors include four distinct wireless EEG sensors configured to be placed on a patient's left front, right front, behind the left ear, and behind the right ear.

[0264] In some embodiments, the techniques described herein are related to a system, and the instructions further configure one or more processors to detect seizure events using a classifier, and the meta-features are used to train the classifier.

[0265] In some embodiments, the techniques described herein are related to a system, and each discrete wireless EEG sensor of a plurality of individual wireless EEG sensors collects an EEG data channel independently of other discrete wireless EEG sensors; creating a meta-feature includes generalizing EEG data across two or more of the plurality of EEG data channels and thereby pooling the EEG data of the two or more EEG data channels together for the detection of seizure events.

[0266] In some embodiments, the techniques described herein are related to a system, and seizure events are further detected based on one or more of a plurality of features.

[0267] In some embodiments, the techniques described herein are related to a system, and the meta-feature removes bias to specific locations on the scalp in a plurality of electroencephalogram (EEG) data channels for the detection of seizure events.

[0268] In some embodiments, the techniques described herein relate to a method for detecting seizure events in a patient, the method comprising: receiving EEG data from a plurality of individual wireless EEG sensors by a first processor, each discrete wireless EEG sensor providing a plurality of EEG data channels; generating at least one composite EEG data channel based on the plurality of EEG data channels; training a seizure detection model based on the at least one composite EEG data channel and the plurality of EEG data channels; and detecting a seizure event by a second processor by applying the seizure detection model to at least four data channels provided by at least four discrete wireless EEG sensors located on the scalp of a particular patient.

[0269] In some embodiments, the techniques described herein relate to a method in which at least one synthetic EEG data channel is created by extending at least one EEG data channel of a plurality of EEG data channels.

[0270] In some embodiments, the techniques described herein relate to methods, and extensions include scaling at least one EEG data channel.

[0271] In some embodiments, the techniques described herein relate to a method, and scaling at least one EEG data channel includes inverting at least one EEG data channel.

[0272] In some embodiments, the techniques described herein relate to methods, and extensions include adding noise to at least one EEG data channel.

[0273] In some embodiments, the techniques described herein relate to an electroencephalogram (EEG) system, which includes: a plurality of individual radio-EEG sensors configured to be placed on a patient's scalp, each discrete radio-EEG sensor configured to collect EEG data channels; and a non-temporary computer-readable medium storing instructions, which, when an instruction is executed by one or more first and second processors, causes one or more first processors to: receive EEG data from the plurality of individual radio-EEG sensors, each discrete radio-EEG sensor providing an EEG data channel; generate at least one composite EEG data channel based on the plurality of EEG data channels; learn a seizure detection model based on the at least one composite EEG data channel and the plurality of EEG data channels; and cause one or more second processors to detect seizure events by applying the seizure detection model to at least four data channels provided by at least four discrete radio-EEG sensors located on the scalp of a particular patient.

[0274] In some embodiments, the techniques described herein relate to a system in which at least one synthetic EEG data channel is created by extending at least one EEG data channel of a plurality of EEG data channels.

[0275] In some embodiments, the techniques described herein relate to a system, and the extensions include scaling at least one EEG data channel.

[0276] In some embodiments, the techniques described herein relate to a system, and scaling at least one EEG data channel includes inverting at least one EEG data channel.

[0277] In some embodiments, the techniques described herein relate to a system, and the extensions include adding noise to at least one EEG data channel.

[0278] In some embodiments, the techniques described herein relate to a non-temporary computer-readable medium that stores instructions, and when an instruction is executed by one or more processors, the one or more processors are caused to implement one of the methods of the preceding embodiments.

[0279] 5. Identification of seizure events In some embodiments, the techniques described herein relate to a method for detecting seizure events from electroencephalogram (EEG) signals, the method comprising: segmenting EEG data collected by a plurality of separate wireless EEG sensors placed on the scalp of a particular patient to form a plurality of segments; determining the probability of occurrence of one or more electrical seizure characteristics for the plurality of segments; modifying the probability of a segment among the plurality of segments using one or more probabilities of one or more adjacent segments among the plurality of segments; and detecting individual seizure events spanning segments of the plurality of segments and one or more adjacent segments, the method being performed by one or more processors.

[0280] In some embodiments, the techniques described herein relate to methods, and modifying the probability of a segment involves performing at least one morphological transformation using the probability of the segment and one or more probabilities of one or more adjacent segments.

[0281] In some embodiments, the techniques described herein relate to methods in which at least one morphological transformation includes a morphological closure.

[0282] In some embodiments, the techniques described herein relate to a method which further includes performing an exponential moving average of the probabilities of a plurality of segments before modifying the probabilities of the segments.

[0283] In some embodiments, the techniques described herein relate to a method in which the detection of individual seizure events is performed with a sensitivity of at least 80% and a false positive rate of 0.08 or less per hour.

[0284] In some embodiments, the techniques described herein relate to a method in which individual seizure events have a duration between 10 seconds and 15 minutes.

[0285] In some embodiments, the techniques described herein relate to a method and further include discarding an individual seizure event in response to determining that the duration of the individual seizure event exceeds 15 minutes.

[0286] In some embodiments, the techniques described herein relate to a method which further includes combining a first individual seizure event and a second individual seizure event that begins within a threshold time from the end of the first individual seizure event into a single individual seizure event.

[0287] In some embodiments, the technique described herein relates to a method, and the threshold time includes 2 minutes.

[0288] In some embodiments, the technology described herein relates to an electroencephalogram (EEG) system, the system comprising: a plurality of individual wireless EEG sensors configured to be placed on the scalp of a patient, each discrete wireless EEG sensor configured to collect EEG data channels; and a non-temporary computer-readable medium for storing instructions, the instructions, when executed by one or more processors, cause one or more processors to: segment the EEG data collected by the plurality of individual wireless EEG sensors to form a plurality of segments; determine the probability of the occurrence of one or more electrical seizure characteristics for the plurality of segments; modify the probability of one of the plurality of segments using one or more probabilities of one or more adjacent segments; and detect individual seizure events spanning a segment and one or more adjacent segments.

[0289] In some embodiments, the techniques described herein relate to a system, and the instructions are further configured to cause one or more processors to modify the probability of a segment by performing at least one morphological transformation using the probability of the segment and one or more probabilities of one or more adjacent segments.

[0290] In some embodiments, the techniques described herein relate to a system in which at least one morphological transformation includes a morphological closure.

[0291] In some embodiments, the techniques described herein relate to a system, and the instructions are further configured to detect individual seizure events with a sensitivity of at least 80% and a false positive rate of no more than 0.08 per hour.

[0292] In some embodiments, the techniques described herein relate to a system in which individual seizure events have a duration between 10 seconds and 15 minutes.

[0293] In some embodiments, the techniques described herein relate to a system and are configured to cause an individual seizure event to be discarded in response to determining that the duration of the individual seizure event exceeds 15 minutes.

[0294] In some embodiments, the techniques described herein relate to a system, and instructions are further configured to cause one or more processors to combine a first separate seizure event and a second separate seizure event that begins within a threshold time from the end of the first separate seizure event into a single separate seizure event.

[0295] In some embodiments, the techniques described herein relate to the system according to Embodiment 16, and the threshold time includes 2 minutes.

[0296] In some embodiments, the techniques described herein relate to a non-temporary computer-readable medium that, when executed by one or more processors, stores instructions causing one or more processors to perform a method according to any of the preceding embodiments.

[0297] 6. Channel expansion In some embodiments, the techniques described herein can detect seizure events in patients. In relation to a method, the method includes: processing EEG data collected by a plurality of individual wireless EEG sensors, each of which provides an EEG data channel; detecting missing EEG data channels in the EEG data; creating a model EEG data channel using one or more EEG data channels present in the EEG data; and detecting seizure events based on one or more EEG data channels present in the EEG data and the model EEG data channel, and the method is performed by one or more processors.

[0298] In some embodiments, the techniques described herein relate to methods in which multiple individual wireless EEG sensors include four discrete wireless sensors.

[0299] In some embodiments, the techniques described herein relate to methods in which model EEG data channels are created based on probabilistic sampling and emulation of one or more EEG data channels present in the EEG data.

[0300] In some embodiments, the techniques described herein relate to a method in which a model EEG data channel is created using shadow manifold cross-sampling of one or more EEG data channels present in the EEG data.

[0301] In some embodiments, the techniques described herein relate to methods, and the model EEG data channel is created based on EEG data collected from a patient population.

[0302] In some embodiments, the techniques described herein relate to a method in which a model EEG data channel is created based on at least one EEG data channel collected by at least one discrete radio EEG sensor of a plurality of separate radio EEG sensors that spatially correspond to the discrete radio EEG sensor associated with the missing EEG data channel.

[0303] In some embodiments, the techniques described herein relate to a method in which at least one discrete radio EEG sensor is positioned opposite the discrete radio EEG sensor associated with the missing EEG data channel in the midline.

[0304] In some embodiments, the techniques described herein relate to a method in which at least one discrete radio EEG sensor is positioned opposite the discrete radio EEG sensor associated with the missing EEG data channel in the frontal plane.

[0305] In some embodiments, the techniques described herein relate to a method in which missing EEG data channels in the EEG data are detected as a result of a failure of a discrete radio EEG sensor of multiple individual radio EEG sensors.

[0306] In some embodiments, the technology described herein relates to an electroencephalogram (EEG) system, the system comprising: a plurality of individual wireless EEG sensors configured to be placed on a patient's scalp, each discrete wireless EEG sensor configured to collect EEG data channels; and a non-temporary computer-readable medium for storing instructions, the instructions, when executed by one or more processors, causing one or more processors to: process EEG data collected by a plurality of individual wireless EEG sensors placed on the scalp of a particular patient, each discrete wireless EEG sensor providing an EEG data channel; detect missing EEG data channels in the EEG data; create a model EEG data channel using one or more EEG data channels present in the EEG data; and detect seizure events based on one or more EEG data channels present in the EEG data and the model EEG data channel, the non-temporary computer-readable medium for storing instructions, the instructions, when executed by one or more processors, cause one or more processors to: process EEG data collected by a plurality of individual wireless EEG sensors placed on the scalp of a particular patient, each discrete wireless EEG sensor providing an EEG data channel; detect missing EEG data channels in the EEG data; create a model EEG data channel using one or more EEG data channels present in the EEG data and the model EEG data channel, Includes the medium.

[0307] In some embodiments, the technology described herein relates to a system in which multiple individual wireless EEG sensors include four discrete wireless sensors.

[0308] In some embodiments, the techniques described herein relate to a system in which the model EEG data channel is created based on probabilistic sampling and emulation of one or more EEG data channels present in the EEG data.

[0309] In some embodiments, the techniques described herein relate to a system in which a model EEG data channel is created using shadow manifold cross-sampling of one or more EEG data channels present in the EEG data.

[0310] In some embodiments, the techniques described herein relate to a system, and the model EEG data channel is created based on EEG data collected from a patient population.

[0311] In some embodiments, the techniques described herein relate to a system in which a model EEG data channel is created based on at least one EEG data channel collected by at least one discrete radio EEG sensor of a plurality of separate radio EEG sensors that spatially correspond to the discrete radio EEG sensor associated with the missing EEG data channel.

[0312] In some embodiments, the techniques described herein relate to a system in which at least one discrete radio EEG sensor is positioned opposite the discrete radio EEG sensor associated with the missing EEG data channel in the midline.

[0313] In some embodiments, the techniques described herein relate to a system in which at least one discrete radio EEG sensor is positioned opposite the discrete radio EEG sensor associated with the missing EEG data channel in the frontal plane.

[0314] In some embodiments, the techniques described herein relate to a system in which missing EEG data channels in the EEG data are detected as a result of a failure of a discrete radio EEG sensor of multiple individual radio EEG sensors.

[0315] In some embodiments, the techniques described herein relate to a non-temporary computer-readable medium that, when executed by one or more processors, stores instructions causing one or more processors to perform any of the methods of the preceding embodiments.

[0316] Other variations Further details of EEG monitoring systems and methods are described in U.S. Patent No. 11,020,035, U.S. Patent Publication No. 2021 / 0307672, and U.S. Patent Application No. 18 / 067611 filed December 16, 2022, all of which are incorporated herein by reference in their entirety.

[0317] The general principles described herein can be extended to other scenarios. For example, in pediatric and adult intensive care, two sensors, four sensors, eight sensors, or various combinations of sensors may be used.

[0318] Various other configurations are also possible, and certain elements described as being implemented as hardware may instead be implemented in software, firmware, or a combination thereof. Good. Those skilled in the art will recognize various alternatives to the specific embodiments described herein.

[0319] This specification and the drawings describe specific embodiments provided for the sake of clarity and illustration, and are not intended to limit them. Embodiments may be implemented for use in a variety of environments without departing from the spirit and scope of this disclosure.

[0320] At least some elements of the implementation of one or more devices disclosed are controllable, and at least some steps of the implementation of one or more methods disclosed are operationally operable in cooperation with a programmable processor controlled by instructions stored in memory. The memory may be RAM, ROM, flash memory, any other memory, or any combination thereof suitable for storing control software or other instructions and data. Those skilled in the art will also readily understand that instructions or programs defining the functionality of the disclosed implementation may be provided to the processor in many forms, such as information permanently stored in a non-writable storage medium (such as ROM in a computer, or a medium readable via a computer I / O connection such as a CD-ROM or DVD), information variably stored in a writable storage medium (such as a floppy disk, removable flash memory, or hard drive), or information transmitted to a computer through a communication medium, including a wired or wireless computer network. Furthermore, while some implementations may be embodied as software, the functions necessary to implement the disclosed features may, at their discretion or as an alternative, be embodied in part or in whole using combinational logic, ASICs, FPGAs, other hardware, or firmware and / or hardware components such as any combination of hardware, software, and firmware.

[0321] In various embodiments, user input may be required. Examples of methods for receiving user input, such as receiving a user button press, are illustrative and not limiting. Alternative methods for receiving user input may include button presses on a touchscreen, physical button presses on a device, swipes, taps, other touch gestures, and speech (voice) input.

[0322] Various modifications to the implementations described herein will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other implementations without departing from the spirit or scope of this disclosure. Accordingly, the claims are not intended to be limited to the implementations shown herein, but should be granted the broadest scope consistent with this disclosure, the principles, and the novel features disclosed herein.

[0323] Certain features described herein in the context of separate implementations can also be implemented in combination within a single implementation. Conversely, various features described in the context of a single implementation can also be implemented separately in multiple implementations or in any suitable subcombination. Furthermore, even if features are described above as operating in a particular combination and initially claimed to do so, one or more features may be removed from the claimed combination, and the claimed combination may be directed towards a subcombination or a variation thereof.

[0324] Depending on the embodiment, any part of any operation, event, or function of any process or algorithm described herein may be executed in a different order, and may be added, merged, or omitted entirely. Furthermore, in certain embodiments, the operation or event may be executed in parallel, rather than sequentially, for example, by multithreading, interrupt handling, multiple processors / processor cores, or other parallel architectures.

[0325] The various exemplary logic blocks, modules, routines, and algorithmic steps described in connection with embodiments of this disclosure can be implemented as electronic hardware or as a combination of electronic hardware and executable software. To clearly demonstrate this interchangeability, the various exemplary components, blocks, modules, and steps have generally been described above in terms of their functionality. Whether such functionality is implemented as hardware or as software running on hardware depends on the specific application and design constraints imposed on the overall system. The described functionality can be implemented in various ways depending on the application, but such implementation decisions should not be construed as resulting in a departure from the scope of this disclosure.

[0326] Furthermore, various exemplary logic blocks and modules described in connection with embodiments of this disclosure can be implemented or executed by machines (machine learning service servers, DSPs, ASICs, FPGAs, other programmable logic devices, individual gate / transistor logic, individual hardware components, or any combination thereof) to perform the functions described herein. A machine learning service server may be a microprocessor, or alternatively, a controller, microcontroller, state machine, or a combination thereof configured to generate and expose machine learning services backed by machine learning models. A machine learning service server may include electrical circuits configured to process computer executable instructions. Although this specification has primarily described digital technologies, a machine learning service server may also primarily include analog components. For example, some or all of the modeling, simulation, or service algorithms described herein may be implemented in analog circuits or mixed analog-digital circuits. A computing environment may include any type of computer system, including but not limited to microprocessor-based computer systems, mainframe computers, DSPs, portable computing devices, device controllers, or in-device computing engines.

[0327] Elements of methods, processes, routines, or algorithms described in connection with embodiments disclosed herein may be embodied directly in hardware, in software modules executed by a machine learning service server, or in a combination of both. The software modules may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disks, removable disks, CD-ROMs, or any other form of non-temporary computer-readable storage medium. An exemplary storage medium may be coupled to the machine learning service server so that the machine learning service server can read information from and write information to the storage medium. Alternatively, the storage medium may be integrated with the machine learning service server. The machine learning service server and storage medium may reside within an ASIC. The ASIC may reside within a user terminal. Alternatively, the machine learning service server and storage medium may reside as separate components within a user terminal (e.g., an access device or a client device for a network service).

[0328] Conditional language used herein, such as “can,” “could,” “may,” “might,” and “for example,” is generally intended to convey that certain features, elements, and / or steps are included in one embodiment, while others are not, unless otherwise stated or understood from the context in which they are used. Therefore, such conditional language does not necessarily imply that features, elements, and / or steps are required in any embodiment, or that any embodiment necessarily includes logic for determining whether these features, elements, and / or steps are included or performed in a particular embodiment, with or without other inputs or prompts. Terms such as “equip,” “include,” and “have” are synonymous and inclusive (open-ended). It is used in a way that does not exclude additional elements, features, actions, or behaviors. Also, "or" is used in an inclusive sense (not exclusive), for example, when used to connect a list of elements, "or" means one, some, or all of the elements in that list.

[0329] Disjunctive language such as "at least one of X, Y, or Z" is generally understood, in the context in which it is used, to indicate that an item, term, etc., may be X, Y, or Z, or any combination thereof (e.g., X, Y, and / or Z), unless otherwise specified. Therefore, such disjunctive language does not, and should not, mean that a particular embodiment requires the presence of at least one X, at least one Y, or at least one Z, respectively.

[0330] Unless otherwise explicitly stated, articles such as "a" or "an" should generally be interpreted as including one or more of the items described. Therefore, phrases such as "device configured to..." are intended to include one or more of the devices described. One or more such devices may be collectively configured to perform the described items. For example, "processor configured to perform items A, B and C" may include a first processor configured to perform item A working in conjunction with a second processor configured to perform items B and C.

[0331] While the above detailed description illustrates, explains, and points out novel features applied to various embodiments, it should be understood that various omissions, substitutions, and modifications are possible in the form and details of the illustrated apparatus or algorithm without departing from the spirit of the disclosure. As recognized, certain embodiments described herein may be embodied in a form that does not provide all of the features and benefits listed herein, as some of the features described herein may be used or implemented separately from others. The scope of certain embodiments disclosed herein is indicated by the appended claims rather than the description. All modifications that fall within the meaning and equivalent scope of the claims shall be included within that scope.

Claims

1. A method for detecting seizure events using electroencephalogram (EEG) signals, One or more processors Using a set of event identifiers, the process involves transforming multiple probabilities of the occurrence of one or more electrical seizure characteristics in EEG data segments collected from a patient into a set of labels associated with seizure events, wherein each event identifier in the set of event identifiers has a confidence value indicating the accuracy with which the event identifier determines the label, and the process involves transforming the multiple probabilities determined by the classifier. Outputting a label from the set of labels that is determined by the event identifier having a higher confidence value than any other event identifier in the set of event identifiers, To provide notification of the seizure event along with the confidence level associated with the highest confidence value, A method characterized by including the following.

2. The method according to claim 1, characterized in that the set of labels corresponds to the onset time and duration of individual seizure events.

3. The method according to claim 1, characterized in that the set of labels corresponds to the occurrence rate of electrical seizure characteristics over a certain period of time, and the output labels further satisfy an occurrence rate threshold.

4. The method according to claim 1, wherein providing the notification of the seizure event together with the confidence level associated with the highest confidence value includes displaying the notification together with the confidence level along with the EEG data segment collected from the patient.

5. The method according to 4, characterized in that the confidence level includes high, medium, or low.

6. The method according to 4, characterized in that the EEG data segment is displayed as a vertical and horizontal montage including four channels of EEG data collected by four separate wireless EEG sensors located on the left forehead, right forehead, behind the left ear, and behind the right ear of the patient, and a plurality of channels derived by subtracting one pair of channels from the four channels of EEG data.

7. The method according to claim 1, characterized in that each event identifier in the set of event identifiers has a different parameterization for converting the plurality of probabilities into labels in the set of labels.

8. The method according to claim 1, characterized in that the highest confidence value corresponds to the highest precision.

9. The method according to claim 1, wherein outputting the label determined by the event identifier having the highest confidence value of any other event identifier from the set of event identifiers from the set of event identifiers includes discarding labels determined by other event identifiers having a lower confidence value than the event identifier.

10. Translating, using a different set of event identifiers, into a different set of labels associated with a different seizure event, the probabilities of the occurrence of one or more different electrical seizure characteristics in the EEG data segment collected from the patient, the probabilities being determined by a different classifier, Outputting another label from the aforementioned set of labels, determined by the event identifier having a higher confidence value than any other event identifier of the aforementioned other event identifier, It is determined that the confidence value of the aforementioned other label exceeds the confidence value of the aforementioned label, In response to determining that the confidence value of the other label exceeds the confidence value of the label, notification of the other seizure event is provided along with the other confidence level associated with the other label. The method according to claim 1, further comprising:

11. Translating, using a different set of event identifiers, into a different set of labels associated with a different seizure event, the probabilities of the occurrence of one or more different electrical seizure characteristics in the EEG data segment collected from the patient, the probabilities being determined by a different classifier, The seizure event and the confidence level are merged with the other seizure event and the other confidence level to obtain a merged seizure event and a merged confidence level. The notification of the merged seizure event is provided along with the merged confidence level. The method according to claim 1, further comprising:

12. A computer-readable medium that, when executed by one or more processors, stores instructions causing the one or more processors to perform a method for detecting seizure events using electroencephalogram (EEG) signals, The aforementioned method, Using a set of event identifiers, the process involves transforming multiple probabilities of the occurrence of one or more electrical seizure characteristics in EEG data segments collected from a patient into a set of labels associated with seizure events, wherein each event identifier in the set of event identifiers has a confidence value indicating the accuracy with which the event identifier determines the label, and the process involves transforming the multiple probabilities determined by the classifier. Outputting a label from the set of labels that is determined by the event identifier having a higher confidence value than any other event identifier in the set of event identifiers, To provide notification of the seizure event along with the confidence level associated with the highest confidence value, A computer-readable medium characterized by containing [a certain element].

13. The computer-readable medium according to claim 12, characterized in that the set of labels corresponds to the onset time and duration of individual seizure events.

14. The computer-readable medium according to claim 12, characterized in that the set of labels corresponds to the occurrence rate of electrical seizure characteristics over a certain period, and the output labels further satisfy an occurrence rate threshold.

15. The computer-readable medium according to claim 12, wherein providing the notification of the seizure event together with the confidence level associated with the highest confidence value includes displaying the notification together with the confidence level along with the EEG data segment collected from the patient.

16. The computer-readable medium according to claim 15, characterized in that the confidence level includes high, medium, or low.

17. The computer-readable medium according to claim 15, characterized in that the EEG data segment is displayed as a vertical and horizontal montage including four channels of EEG data collected by four separate wireless EEG sensors located on the left forehead, right forehead, behind the left ear, and behind the right ear of the patient, and a plurality of channels derived by subtracting one pair of channels from the four channels of EEG data.

18. The computer-readable medium according to claim 12, characterized in that each event identifier in the set of event identifiers has a different parameterization for converting the plurality of probabilities into labels in the set of labels.

19. The computer-readable medium according to claim 12, characterized in that the highest confidence value corresponds to the highest precision.

20. The computer-readable medium according to claim 12, wherein outputting the label determined by the event identifier having the highest confidence value of any other event identifier from the set of event identifiers includes discarding labels determined by other event identifiers having a lower confidence value than the event identifier.

21. The above method is performed by a second or more processors, Using a different set of event identifiers, transforming a set of different probabilities of the occurrence of one or more different electrical seizure characteristics in the EEG data segment collected from the patient into a different set of labels associated with a different seizure event, and transforming the said set of probabilities determined by a different classifier, Outputting another label from the aforementioned set of labels, determined by the event identifier having a higher confidence value than any other event identifier of the aforementioned other event identifier, It is determined that the confidence value of the aforementioned other label exceeds the confidence value of the aforementioned label, In response to determining that the confidence value of the other label exceeds the confidence value of the label, notification of the other seizure event is provided along with the other confidence level associated with the other label. The computer-readable medium according to claim 12, further comprising the above.

22. The above method is performed by a second or more processors, Using a different set of event identifiers, transforming a set of different probabilities of the occurrence of one or more different electrical seizure characteristics in the EEG data segment collected from the patient into a different set of labels associated with a different seizure event, and transforming the said set of probabilities determined by a different classifier, The seizure event and the confidence level are merged with the other seizure event and the other confidence level to obtain a merged seizure event and a merged confidence level. The notification of the merged seizure event is provided along with the merged confidence level. The computer-readable medium according to claim 12, further comprising the above.

23. A system for detecting seizure events using electroencephalogram (EEG) signals, A plurality of individual wireless EEG sensors configured to be placed on the scalp of a patient, wherein each individual wireless EEG sensor is configured to collect EEG data channels, A computer-readable medium for storing instructions, wherein, when the instructions are executed by one or more processors, the computer-readable medium causes the one or more processors to perform the method according to any one of claims 1 to 22. A system characterized by comprising the following features.

Images