A method, system and apparatus for epilepsy monitoring and closed loop intervention

Abstract

The application discloses a method and system for epilepsy monitoring and closed-loop intervention and equipment thereof, and belongs to the field of medical electronics and health monitoring. The application collects scalp electroencephalogram signals and motion, optical plethysmography multi-modal physiological signals, obtains high-quality signals through dynamic impedance compensation, multi-source artifact removal and signal quality screening, extracts three-dimensional features of time-space-spectrum-connectivity and deep learning feature vectors, inputs a multi-level abnormality detection and classification model, and completes subclinical to seizure period full-cycle risk assessment in combination with individualized dynamic baseline. Finally, a graded early warning and intervention decision suggestion is output. Early warning, individualized assessment and closed-loop intervention can be realized, the problems of poor home adaptability, low early warning precision and lack of decision support of traditional equipment are effectively solved, and the application is suitable for long-term home intelligent monitoring and clinical auxiliary management of epilepsy patients.

Description

Technical Field

[0001] This application belongs to the field of medical electronics and health monitoring, and specifically relates to a method, system and device for epilepsy monitoring and closed-loop intervention. Background Technology

[0002] Epilepsy, a chronic neurological disorder caused by abnormal synchronous discharges of neurons in the brain, relies heavily on continuous monitoring and analysis of the patient's brain electrical activity for diagnosis and management. Electroencephalography (EEG) is the core tool for capturing these abnormal discharges. Traditional clinical EEG monitoring is typically performed in a controlled hospital environment, using wet electrodes with conductive gel to ensure signal quality. However, this approach cannot meet the needs for monitoring patients' daily, long-term, and especially natural brain electrical activity in environments such as at home, which is crucial for capturing sporadic or nocturnal seizures, assessing treatment effectiveness, and exploring seizure triggers.

[0003] In recent years, with the development of wearable technology, head-mounted EEG monitoring devices based on dry electrodes have become a possible path to achieve long-term, home-based monitoring. These devices typically integrate electrodes and signal acquisition circuitry into reusable elastic headbands or helmets, aiming to provide a convenient means of continuous monitoring. Among existing epilepsy monitoring technologies, patent CN111616682A discloses an epilepsy seizure early warning system based on a portable EEG acquisition device, achieving epilepsy signal identification and early warning through a multi-layered complex brain network combined with a deep learning model; patent CN116999070A proposes an epilepsy seizure prediction system integrating intelligent wearable devices and modality transfer networks, relying on modality transfer networks and deep learning to complete EEG signal classification and prediction; and patent CN121080913A proposes an epilepsy signal recognition method based on multimodal information from wearable devices, reducing the false alarm rate caused by motion artifacts through a serial integrated learning framework. While these existing technologies achieve portable acquisition, multimodal / multi-feature fusion, and early warning functions, they still have the following problems: Traditional wearable electrodes are mostly reusable structures, resulting in low wearing comfort and the risk of cross-infection. The crude electrode positioning leads to significant signal attenuation due to movement and impedance changes in home environments, resulting in insufficient long-term monitoring stability. They only remove artifacts through conventional filtering and do not achieve dynamic monitoring and active compensation of electrode-skin interface impedance, making them unable to adapt to signal fluctuations in uncontrolled home environments. The early warning models and thresholds adopt a one-size-fits-all approach, failing to build dynamic baseline models based on the patient's diurnal rhythm, sleep status, and medication history, making the accuracy of early warnings highly susceptible to individual heterogeneity. Most of them provide a single early warning output that "monitors but does not control," failing to form a full-cycle intervention decision support based on risk stratification, thus lacking clinical applicability. Summary of the Invention

[0004] To address the aforementioned problems, firstly, this application proposes a method for epilepsy monitoring and closed-loop intervention, comprising the following steps: The system collects the user's scalp electroencephalogram (EEG) signals and multimodal physiological signals, wherein the multimodal physiological signals include at least motion signals and optical volumetric imaging (OCI) signals. Adaptive preprocessing is performed on the scalp EEG signals and the multimodal physiological signals to obtain high-quality signals; Multi-scale feature extraction is performed on the high-quality signal to obtain a three-dimensional feature set representing spatiotemporal-spectral-connectivity and a deep learning feature vector; The three-dimensional feature set and the deep learning feature vector are input into a pre-trained multi-level anomaly detection and classification model to perform full-cycle risk assessment from subclinical state to seizure event, and output seizure warning signals and seizure classification results. Based on risk assessment results, outbreak classification results, and pre-set individualized risk stratification strategies, the system generates and outputs tiered early warning information and intervention decision suggestions to users or guardians.

[0005] Furthermore, the motion signals are acquired using a triaxial accelerometer and a gyroscope; The optical volumetric imaging signal was acquired by a photoelectric sensor located at the superficial temporal artery.

[0006] Furthermore, adaptive preprocessing is performed on the scalp EEG signals and the multimodal physiological signals, including the following steps: The electrode-skin interface impedance of each acquisition channel is dynamically monitored. When the impedance exceeds a preset threshold, an active impedance compensation algorithm is triggered or a backup electrode switching protocol is initiated. The multi-modal physiological signals are used to jointly remove multi-source artifacts from the scalp EEG signal to obtain a processed scalp EEG signal; the multi-source artifacts include at least motion artifacts reflected by the motion signal and electrocardiogram artifacts synchronously identified by the optical volumetric imaging signal. The signal quality index of the processed scalp EEG signal is calculated in real time. The signal quality index is a weighted sum of signal-to-noise ratio, channel failure rate, and artifact percentage. High-quality EEG signals are obtained by screening based on the signal quality index, specifically by identifying the processed scalp EEG signals whose signal quality index meets the preset threshold condition as the high-quality signals.

[0007] Furthermore, the active impedance compensation algorithm is implemented in the following way: A constant current test signal with known frequency and amplitude is injected into the electrode, and the feedback voltage is measured to calculate the real-time electrode-skin interface impedance. ; When the real-time electrode-skin interface impedance is higher than a preset threshold, the original scalp EEG signal is compensated using the following formula;

[0008] in, The compensated scalp EEG signal; This represents the raw scalp EEG signal; The reference impedance; Real-time electrode-skin interface impedance; The baseline offset is estimated using the least squares method; This is the index for discrete-time sampling points.

[0009] Furthermore, multi-scale feature extraction is performed on the high-quality signal to obtain a three-dimensional feature set and deep learning feature vectors representing spatiotemporal-spectrum-connectivity, including the following steps: Features in four dimensions—time domain, frequency domain, spatial domain, and brain network connectivity—are extracted from the high-quality signal to form a three-dimensional feature set characterizing spatiotemporal-spectrum-connectivity; wherein the brain network connectivity features include a dynamic functional connectivity density matrix and network topology parameters constructed based on scalp EEG signals. The high-quality signal is input into a deep learning model composed of a convolutional recurrent neural network and a graph convolutional network to automatically extract deep learning feature vectors that characterize the spatiotemporal evolution pattern of epilepsy abnormalities.

[0010] Furthermore, the three-dimensional feature set and the deep learning feature vector are input into a pre-trained multi-level anomaly detection and classification model to perform a full-cycle risk assessment from subclinical state to seizure event, and output seizure warning signals and seizure classification results, including the following steps: Based on the user's circadian rhythm, sleep status, and medication time window, a personalized dynamic baseline model is constructed and dynamically updated. The three-dimensional feature set and the deep learning feature vector are input into the multi-level anomaly detection and classification model and compared with the individualized dynamic baseline model. Anomaly determination is made based on the degree of feature deviation. The multi-level abnormality detection and classification model is used to perform subclinical detection, pre-seizure detection and seizure detection in sequence; wherein, the subclinical detection is used to detect epilepsy-related abnormalities 30-60 minutes in advance, the pre-seizure detection is used to detect epilepsy-related abnormalities 5-10 minutes in advance, and the seizure detection is used to detect epilepsy-related abnormalities in real time. The multi-level anomaly detection and classification model uses a multi-task learning framework for synchronous classification, outputting classification results for seizure type, seizure manifestation, state of consciousness and seizure origin, and generating corresponding seizure warning signals based on the step-by-step detection results.

[0011] Furthermore, the tiered early warning information includes five levels: monitoring, attention, early warning, emergency, and cluster. Each level corresponds to different probability of occurrence, duration, or frequency of occurrence, and triggers differentiated response strategies ranging from device vibration and sound alarms to automatically dialing emergency numbers.

[0012] Secondly, this application proposes a system for epilepsy monitoring and closed-loop intervention, comprising: The signal acquisition module is used to acquire the user's scalp electroencephalogram (EEG) signals and multimodal physiological signals, wherein the multimodal physiological signals include at least motion signals and optical volumetric imaging (OPI) signals. The signal preprocessing module is used to adaptively preprocess the scalp EEG signal and the multimodal physiological signal to obtain a high-quality signal; The feature extraction module is used to perform multi-scale feature extraction on the high-quality signal to obtain a three-dimensional feature set representing spatiotemporal-spectrum-connectivity and a deep learning feature vector. The intelligent analysis module is used to input the three-dimensional feature set and the deep learning feature vector into a pre-trained multi-level anomaly detection and classification model to perform full-cycle risk assessment from subclinical state to seizure event, and output seizure warning signals and seizure classification results. The decision-making and early warning module is used to generate and output graded early warning information and intervention decision suggestions to users or guardians based on risk assessment results, outbreak classification results and preset individualized risk stratification strategies.

[0013] Thirdly, this application proposes a signal acquisition device for epilepsy monitoring, used to acquire scalp electroencephalogram (EEG) signals and multimodal physiological signals, including: The disposable EEG cap is made of elastic non-woven fabric and has multiple electrode mounting positions corresponding to the international 10-20 system standard positions. Each electrode mounting position is equipped with a positioning structure. Multiple reusable electronic acquisition modules are detachably installed on the electrode mounting positions of the disposable EEG cap and interconnected via flexible circuitry. A main processing unit is electrically connected to the plurality of electronic acquisition modules; Each of the electronic acquisition modules includes a housing, and a flexible microneedle-hydrogel composite dry electrode, an inertial measurement unit, and an optical volumetric sensor integrated within the housing; the flexible microneedle-hydrogel composite dry electrode, the inertial measurement unit, and the optical volumetric sensor are respectively used for the scalp electroencephalogram signal, motion signal, and optical volumetric signal.

[0014] Furthermore, the disposable EEG cap has a three-layer composite structure, comprising: The surface layer is provided with electrode markings and the positioning structure; The inner layer has a perforated area for electrodes to contact the scalp; An interlayer, located between the surface layer and the inner layer, has openings for electrodes to pass through.

[0015] Compared with the prior art, this application has the following advantages: This application constructs a three-level integrated intelligent monitoring system of "signal-network-clinical", which organically combines high-fidelity multimodal signal acquisition, multi-scale spatiotemporal feature analysis, individualized dynamic modeling and closed-loop decision support, and realizes full-cycle management from abnormal EEG microstate to clinical onset. This application employs an adaptive preprocessing approach combining dynamic impedance monitoring, active impedance compensation, and backup electrode switching. It also utilizes multimodal physiological signals to remove motion and ECG artifacts, and employs signal quality index quantification for screening. This approach enables stable output of high-quality EEG signals, overcoming the bottleneck of insufficient signal reliability in long-term home monitoring using existing technologies. By constructing a fusion feature system combining spatiotemporal-spectral-connectivity three-dimensional features with deep learning feature vectors, and integrating convolutional recurrent neural networks and graph convolutional networks to automatically capture the abnormal evolution patterns of epilepsy, this approach offers more comprehensive feature representation and more sensitive anomaly identification compared to single features or single models. It can effectively capture subclinical microstate abnormalities from several hours to several minutes before a seizure.

[0016] This application introduces a personalized dynamic baseline model that adaptively updates based on the user's circadian rhythm, sleep state, and medication time window. It achieves personalized risk assessment based on the degree of feature deviation, abandoning the traditional "one-size-fits-all" warning threshold. This significantly reduces the false alarm rate and improves warning accuracy, better adapting to individual differences among patients of different ages, etiologies, and medication histories. It employs a three-level cascaded detection architecture—subclinical, pre-seizure, and seizure-level—to achieve full-cycle risk identification 30-60 minutes in advance, 5-10 minutes in advance, and in real-time. Combined with a multi-task learning framework, it synchronously outputs seizure classification results, resulting in superior warning lead time and identification accuracy compared to existing technologies.

[0017] Establish a five-level hierarchical early warning and closed-loop intervention system, automatically execute differentiated responses based on risk level, and output individualized intervention decision suggestions to realize closed-loop management from "simple monitoring" to "monitoring-early warning-intervention", providing real-time and actionable clinical decision support for patients, guardians and medical staff.

[0018] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart of a method for epilepsy monitoring and closed-loop intervention proposed in an embodiment of this application is shown; Figure 2 A schematic diagram of a system for epilepsy monitoring and closed-loop intervention proposed in an embodiment of this application is shown. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] This application proposes a method for epilepsy monitoring and closed-loop intervention, such as... Figure 1 As shown, it includes the following steps: S1. Collect the user's scalp electroencephalogram (EEG) signals and multimodal physiological signals, wherein the multimodal physiological signals include at least motion signals and optical volumetric imaging (OPI) signals; S2. Adaptive preprocessing is performed on the collected scalp EEG signals and the multimodal physiological signals to obtain high-quality signals; S3. Perform multi-scale feature extraction on the high-quality signal to obtain a three-dimensional feature set of spatiotemporal-spectrum-connectivity and a deep learning feature vector; S4. Input the three-dimensional feature set and the deep learning feature vector into a pre-trained multi-level anomaly detection and classification model to perform a full-cycle risk assessment from subclinical state to epileptic event, and output epileptic warning signals and epileptic classification results. S5. Based on risk assessment results, outbreak classification results, and preset individualized risk stratification strategies, generate and output graded early warning information and intervention decision suggestions to users or guardians.

[0023] The above method will be illustrated by example below with reference to specific embodiments.

[0024] Step S1: After the user puts on the dedicated wearable device, start the synchronous acquisition of multimodal signals.

[0025] Scalp EEG signals: acquired through 64 channels according to the international 10-20 system standard location, sampling rate... , bandwidth 0.5-500Hz.

[0026] Motion signals (IMU): Acquired via a triaxial accelerometer and gyroscope integrated into the device, with a sampling rate of [missing information]. Dynamic range ±16g.

[0027] Optical volumetric plethysmography (PPG) signal: acquired via a photoelectric sensor located at the superficial temporal artery, sampling rate... It uses a 530nm green LED light source.

[0028] All signals are strictly synchronized via hardware timestamps, with a time synchronization error of <1 ms.

[0029] Step S2, specifically, involves the following: Dynamic impedance monitoring and compensation: Injecting 1kHz at a frequency of 10Hz into each electrode with an amplitude of [missing value] The constant current test signal is used to measure the feedback voltage. .

[0030] The electrode-skin interface impedance is calculated as follows:

[0031] In the formula, Real-time electrode-skin interface impedance; This is the feedback voltage; , is the reference impedance; if And if it lasts for 3 seconds, then compensation will be activated:

[0032] In the formula, The compensated scalp EEG signal; This represents the raw scalp EEG signal; The baseline offset is estimated using the least squares method; This is the index for discrete-time sampling points.

[0033] Motion artifact removal: Principal component analysis (PCA) was performed on the IMU signal to extract the principal motion components. For each EEG channel Motion artifacts are estimated and removed using a Normalized Least Mean Square (NLMS) adaptive filter:

[0034]

[0035]

[0036] in .

[0037] In the formula, For the predicted motion artifact signal; For the first The first-order filter is at the 1st order. Weighting coefficients for each sampling point; For the main motion components extracted based on IMU, This is the discrete-time sampling index, corresponding to the first (i) of the signal. One sampling point; This is the filter order index, corresponding to the delay of the reference signal, with a value range of 0 ≤ ≤M 1. M is the filter order, which is set to 32 in this embodiment; The error signal is the clean scalp EEG signal after removing motion artifacts. Indicates the first The original scalp electroencephalogram (EEG) signals from each sampling point; The adaptive learning rate is set to 0.01 in this application; The regularization coefficient is set to [value] in this application. ; Reference signal The square norm (energy).

[0038] ECG artifact removal: Using the heartbeat time extracted by PPG, an electrocardiogram template window is cropped from the EEG signal. [200, 400] ms. The template is obtained through coherent averaging. The scaling factor is determined by maximizing the cross-correlation. and latency Post-removal:

[0039] In the formula, Clean scalp EEG signal after removing ECG artifacts; Indicates the first The original scalp electroencephalogram (EEG) signals from each sampling point; This is the scaling factor; This represents the time-delay aligned ECG template, where... This is a template signal for electrocardiogram (ECG). This refers to time delay.

[0040] Signal Quality Index (SQI) )calculate: Calculate every 30 seconds, using the following formula:

[0041] In the formula, Signal quality index; Indicates the signal-to-noise ratio. Calculations are performed in the 1-45Hz frequency band; The proportion of bad channels, This represents the proportion of time periods marked as artifacts. This is considered a high-quality signal and allows us to proceed to the next step.

[0042] Step S3, specifically, is implemented as follows.

[0043] Traditional 3D feature set extraction: Using a time window of 30 seconds and a sliding step of 5 seconds, the following features were extracted: Time domain: Spike / spike occurrence rate (events with amplitude > 3 standard deviations and duration < 200 ms).

[0044] Frequency domain: The power spectrum was estimated using the Welch method, and the relative power ratios of the δ (0.5-4Hz), θ (4-8Hz), α (8-13Hz), β (13-30Hz), and γ (30-80Hz) bands were calculated.

[0045] Spatial Domain and Connectivity: Based on Weighted Phase Lag Index ( Construct a dynamic functional connection matrix :

[0046] In the formula, Represents frequency Below, brainwave channels With channel The weighted phase lag index between the two channels is used to quantify the phase synchronization of the scalp EEG signals and is a core indicator for constructing a dynamic functional connectivity matrix. Indicates brainwave channels With channel In frequency Cross power spectral density at; Represents the mathematical expectation; This indicates taking the imaginary part of a complex number.

[0047] Calculate the graph theory characteristics of this matrix: clustering coefficients. Feature path length And the resulting small-world index .

[0048] Deep learning feature vector extraction: A 5-second duration of 64-channel EEG data is fed into a pre-trained convolutional recurrent neural network (CRNN).

[0049] Convolutional part: Three layers of 1D convolution with 32, 64 and 128 filters respectively, used to extract local spatiotemporal features.

[0050] Recurrent part: Bidirectional LSTM layers (64 units each) capture long-term dependencies.

[0051] Attention mechanism: assigning weights to different time points :

[0052] In the formula, Indicates the first Attention weights for each time step are used to represent the importance of the features at that time step in the final output; This represents the trainable parameter vector of the attention mechanism, used to calculate the weight scores of features.

[0053] These represent the feature transformation weight matrix and the global feature transformation weight matrix, respectively, used to weight the output of the recurrent layer. and global features Perform a linear transformation.

[0054] : Output features of the recurrent layer (bidirectional LSTM) at time step t.

[0055] : Global feature vector, used to capture the overall information of the input sequence.

[0056] tanh: Hyperbolic tangent activation function, used to perform nonlinear mapping on the transformed features.

[0057] The exponential function is used to convert scores into non-negative values, which facilitates subsequent normalization.

[0058] The scores at all time steps are summed to normalize the attention weights.

[0059] The final output is a 128-dimensional deep feature vector. .

[0060] Multi-level anomaly detection and classification (step S4) specifically implements the following content.

[0061] Individualized dynamic baseline modeling: The system establishes independent baseline models for three states during the initial period of user wear (e.g., the first 72 hours): awake and resting. ), non-rapid eye movement sleep ( REM sleep ( Baseline feature vector Recursive update with exponential forgetting:

[0062] In the formula, and They represent the first The and the first Individualized dynamic baseline feature vector updated at each time step; For the current feature vector, The forgetting factor is 0.9 in this embodiment.

[0063] Subclinical testing module: The Isolation Forest algorithm is used to calculate the anomaly score of the current feature relative to the current state baseline. ;like If it lasts for more than 10 minutes, it is judged as "subclinical abnormality".

[0064] Pre-seizure detection module: A Long Short-Term Memory (LSTM) network is used as the prediction model. The input is a feature sequence of the past 30 minutes (one point per minute), and the output is the probability of an attack in the next 5-10 minutes. .like If the density of high-frequency oscillation (HFO) in the 80-250Hz range increases by more than 100% within the last 2 minutes, it is considered a "pre-seizure state".

[0065] Seizure-level detection module: A lightweight convolutional neural network (based on an improvement of MobileNetV3-small) is used for real-time classification. This network takes a 128ms EEG data window as input, has a latency of less than 200ms, and employs a multi-task learning framework to simultaneously output classification results across four dimensions. Seizure types: focal, generalized, and of unknown origin.

[0066] State of consciousness: preserved, damaged.

[0067] Motor manifestations: rigidity, clonic contractions, automatisms, no significant motor activity.

[0068] Origin regions: temporal lobe, frontal lobe, parietal lobe, occipital lobe, multifocal.

[0069] The total loss function is the weighted sum of the losses from each task:

[0070] The meanings of each parameter in the formula are as follows: The total loss function is the core optimization objective of the multi-task epilepsy classification model in this application, and is used to comprehensively evaluate the overall prediction error of the model in multi-dimensional classification tasks. This represents the summation operator, used to accumulate the loss terms of all subtasks, enabling unified calculation of multi-task losses; Indicates the first The loss weight coefficients for each subtask are used to balance the contribution of different subtasks to the total loss and avoid a single task dominating model training. : No. For each sub-task, the sub-loss function is used. This solution adopts Focal Loss to specifically address the class imbalance problem in clinical EEG data. [0.3,0.2,0.3,0.2]: Preset weight coefficient vector, corresponding to the weight allocation of the 4 sub-tasks, which can be flexibly adjusted according to clinical needs.

[0071] Each sub-loss Focal loss is used to address class imbalance.

[0072] Step S5, specifically, implements the following: The system executes a tiered response according to the strategy shown in Table 1.

[0073] Table 1. Five-Level Early Warning and Response Strategies

[0074] Meanwhile, the decision-making module will provide personalized intervention decision suggestions based on long-term data analysis. For example, "Data shows that 80% of the attacks in the past week occurred the day after less than 6 hours of sleep. It is recommended to maintain a regular schedule." In another embodiment of this application, a system for epilepsy monitoring and closed-loop intervention is proposed to implement the above-mentioned method, such as... Figure 2 As shown, it includes: The signal acquisition module is used to acquire the user's scalp electroencephalogram (EEG) signals and multimodal physiological signals, wherein the multimodal physiological signals include at least motion signals and optical volumetric imaging (OPI) signals. The signal preprocessing module is used to adaptively preprocess the scalp EEG signal and the multimodal physiological signal to obtain a high-quality signal; The feature extraction module is used to perform multi-scale feature extraction on the high-quality signal to obtain a three-dimensional feature set representing spatiotemporal-spectrum-connectivity and a deep learning feature vector. The intelligent analysis module is used to input the three-dimensional feature set and the deep learning feature vector into a pre-trained multi-level anomaly detection and classification model to perform full-cycle risk assessment from subclinical state to seizure event, and output seizure warning signals and seizure classification results. The decision-making and early warning module is used to generate and output graded early warning information and intervention decision suggestions to users or guardians based on risk assessment results, outbreak classification results and preset individualized risk stratification strategies.

[0075] The system is deployed in a cloud-edge-device collaborative architecture: End-side (device side): Runs signal acquisition module and lightweight preprocessing algorithm.

[0076] Side-side (user mobile phone / gateway): Runs the core signal preprocessing module, feature extraction module, intelligent analysis module, and decision and early warning module.

[0077] Cloud-based (remote server): Stores long-term data, trains and optimizes individualized models, and provides an interactive interface for doctors.

[0078] The specific implementation of each module is as follows: Signal acquisition module: Calls the device hardware API to synchronously read EEG, IMU, and PPG data streams in a multi-threaded manner.

[0079] Signal preprocessing module: contains the implementation of all the above algorithms, with the core algorithm written in C++ to ensure real-time performance.

[0080] Feature extraction module: Includes a traditional feature calculation library and a CRNN model loaded with pre-trained weights.

[0081] Intelligent analysis module: Integrates and schedules the above-mentioned multi-level detection models (Isolation Forest, LSTM, Lightweight CNN).

[0082] Decision and Early Warning Module: Executes early warning and communication tasks based on the preset strategy table (Table 1) and user-personalized configurations (such as guardian list, emergency phone number).

[0083] In another embodiment of this application, a signal acquisition device for intelligent epilepsy monitoring is proposed. The device consists of a disposable positioning structure and a reusable electronic module to solve the problem of balancing high signal quality, low skin irritation, and low cost of use in long-term home monitoring.

[0084] To ensure precise electrode positioning and stable adhesion, the device includes a disposable EEG cap. Specifically: Structure and Materials: The EEG cap is made of three layers of composite elastic non-woven fabric, formed by hot pressing. The outer surface of the outer layer is clearly printed with the international 10-20 system electrode markings; the interlayer has 5mm diameter circular holes corresponding to each electrode marking; and the inner layer has 4mm diameter circular cutouts corresponding to the electrodes.

[0085] Positioning structure: At each electrode marking position on the surface layer, an annular micro-protrusion structure with a height of about 1mm is formed by hot pressing to form a positioning groove.

[0086] Fixation structure: The EEG cap has integrally molded strip-shaped fixation straps on both sides, each 2.5cm wide and twice the thickness of the cap body. The free ends of the fixation straps have areas with medical pressure-sensitive adhesive (such as acrylic adhesive), and the surface is covered with release paper.

[0087] Size specifications: Available in four sizes (extra small, small, medium, and large). The electrode position coordinates of the international 10-20 system are scaled proportionally to the head size to ensure the accuracy of anatomical positioning.

[0088] Multiple reusable electronic modules are detachably mounted on the disposable EEG cap. Each module corresponds to one EEG recording point and integrates multiple sensors. Module housing: Made of medical-grade ABS plastic, with a bottom shape that matches the positioning groove, allowing it to be embedded in the micro-protrusion structure for physical positioning and anti-rotation.

[0089] Flexible microneedle-hydrogel composite dry electrode: The electrode is embedded in the center of the bottom of the module. Its structure and performance are consistent with the above embodiment, with a contact impedance of <10kΩ and an integrated active shielding layer. When the module is embedded in the EEG cap, the electrode passes through the interlayer circular hole, and its conductive surface directly contacts the scalp through the inner layer perforation to collect scalp EEG signals.

[0090] Inertial Measurement Unit (IMU): Integrated inside the module, using the BMI160 chip, to acquire motion signals.

[0091] Optical volumetric plethysmography (PPG) sensor: Integrated into the side wall of the module, when the module is installed in the temporal region, its photosensitive window faces the skin area of ​​the superficial temporal artery to collect optical volumetric plethysmography signals.

[0092] Inter-module connection: Each electronic module is connected in series via flexible printed circuit (FPC) cables and finally connected to the main processing unit located at the rear of the cap.

[0093] The main processing unit is encapsulated in a separate box that snaps onto the back of the EEG cap, and includes: Microcontroller: Responsible for controlling the synchronous data acquisition and initial data packaging of all electronic modules.

[0094] Wireless communication module: adopts Bluetooth Low Energy 5.2, used to send the raw data stream to a smartphone or local gateway for deep processing as described in Example 1.

[0095] Battery: Rechargeable lithium battery, ensuring continuous operation for more than 24 hours.

[0096] When using it, select the appropriate EEG cap size according to the user's head circumference.

[0097] Remove the release paper from the fixation straps, put on the EEG cap, and glue the two fixation straps together below the chin.

[0098] Press each reusable electronic module into the corresponding positioning groove on the surface of the EEG cap. A "click" sound indicates that it is locked.

[0099] Check that all module indicator lights are on, and start monitoring via the mobile app.

[0100] After a single test is completed, all electronic modules are removed for future use, and the used disposable EEG caps are disposed of as medical waste.

[0101] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for epilepsy monitoring and closed-loop intervention, characterized in that, Includes the following steps: The system collects the user's scalp electroencephalogram (EEG) signals and multimodal physiological signals, wherein the multimodal physiological signals include at least motion signals and optical volumetric imaging (OCI) signals. Adaptive preprocessing is performed on the scalp EEG signals and the multimodal physiological signals to obtain high-quality signals; Multi-scale feature extraction is performed on the high-quality signal to obtain a three-dimensional feature set representing spatiotemporal-spectral-connectivity and a deep learning feature vector; The three-dimensional feature set and the deep learning feature vector are input into a pre-trained multi-level anomaly detection and classification model to perform full-cycle risk assessment from subclinical state to seizure event, and output seizure warning signals and seizure classification results. Based on risk assessment results, outbreak classification results, and pre-set individualized risk stratification strategies, the system generates and outputs tiered early warning information and intervention decision suggestions to users or guardians.

2. The method for epilepsy monitoring and closed-loop intervention according to claim 1, characterized in that, The motion signals are acquired via a triaxial accelerometer and a gyroscope. The optical volumetric imaging signal was acquired by a photoelectric sensor located at the superficial temporal artery.

3. The method for epilepsy monitoring and closed-loop intervention according to claim 1, characterized in that, Adaptive preprocessing of the scalp EEG signals and the multimodal physiological signals includes the following steps: The electrode-skin interface impedance of each acquisition channel is dynamically monitored. When the impedance exceeds a preset threshold, an active impedance compensation algorithm is triggered or a backup electrode switching protocol is initiated. The multi-modal physiological signals are used to jointly remove multi-source artifacts from the scalp EEG signal to obtain a processed scalp EEG signal; the multi-source artifacts include at least motion artifacts reflected by the motion signal and electrocardiogram artifacts synchronously identified by the optical volumetric imaging signal. The signal quality index of the processed scalp EEG signal is calculated in real time. The signal quality index is a weighted sum of signal-to-noise ratio, channel failure rate, and artifact percentage. High-quality EEG signals are obtained by screening based on the signal quality index, specifically by identifying the processed scalp EEG signals whose signal quality index meets the preset threshold condition as the high-quality signals.

4. The method for epilepsy monitoring and closed-loop intervention according to claim 3, characterized in that, The active impedance compensation algorithm is implemented in the following way: A constant current test signal with known frequency and amplitude is injected into the electrode, and the feedback voltage is measured to calculate the real-time electrode-skin interface impedance. When the real-time electrode-skin interface impedance is higher than a preset threshold, the original scalp EEG signal is compensated using the following formula; in, The compensated scalp EEG signal; This represents the raw scalp EEG signal; The reference impedance; Real-time electrode-skin interface impedance; The baseline offset is estimated using the least squares method; This is the index for discrete-time sampling points.

5. The method for epilepsy monitoring and closed-loop intervention according to claim 1, characterized in that, Multi-scale feature extraction is performed on the high-quality signal to obtain a three-dimensional feature set and deep learning feature vectors representing spatiotemporal-spectral-connectivity, including the following steps: Features in four dimensions—time domain, frequency domain, spatial domain, and brain network connectivity—are extracted from the high-quality signal to form a three-dimensional feature set characterizing spatiotemporal-spectrum-connectivity; wherein the brain network connectivity features include a dynamic functional connectivity density matrix and network topology parameters constructed based on scalp EEG signals. The high-quality signal is input into a deep learning model composed of a convolutional recurrent neural network and a graph convolutional network to automatically extract deep learning feature vectors that characterize the spatiotemporal evolution pattern of epilepsy abnormalities.

6. The method for epilepsy monitoring and closed-loop intervention according to claim 1, characterized in that, The three-dimensional feature set and the deep learning feature vector are input into a pre-trained multi-level anomaly detection and classification model to perform a full-cycle risk assessment from subclinical state to seizure event, and output seizure warning signals and seizure classification results, including the following steps: Based on the user's circadian rhythm, sleep status, and medication time window, a personalized dynamic baseline model is constructed and dynamically updated. The three-dimensional feature set and the deep learning feature vector are input into the multi-level anomaly detection and classification model and compared with the individualized dynamic baseline model. Anomaly determination is made based on the degree of feature deviation. The multi-level abnormality detection and classification model is used to perform subclinical detection, pre-seizure detection and seizure detection in sequence; wherein, the subclinical detection is used to detect epilepsy-related abnormalities 30-60 minutes in advance, the pre-seizure detection is used to detect epilepsy-related abnormalities 5-10 minutes in advance, and the seizure detection is used to detect epilepsy-related abnormalities in real time. The multi-level anomaly detection and classification model uses a multi-task learning framework for synchronous classification, outputting classification results for seizure type, seizure manifestation, state of consciousness and seizure origin, and generating corresponding seizure warning signals based on the step-by-step detection results.

7. The method for epilepsy monitoring and closed-loop intervention according to claim 1, characterized in that, The tiered early warning information includes five levels: monitoring, attention, early warning, emergency, and cluster. Each level corresponds to different risk probabilities, durations, or frequencies of outbreaks and triggers differentiated response strategies ranging from device vibration and sound alarms to automatically dialing emergency numbers.

8. A system for epilepsy monitoring and closed-loop intervention, characterized in that, include: The signal acquisition module is used to acquire the user's scalp electroencephalogram (EEG) signals and multimodal physiological signals, wherein the multimodal physiological signals include at least motion signals and optical volumetric imaging (OPI) signals. The signal preprocessing module is used to adaptively preprocess the acquired scalp EEG signals and the multimodal physiological signals to obtain high-quality signals. The feature extraction module is used to perform multi-scale feature extraction on the high-quality signal to obtain a three-dimensional feature set of spatiotemporal-spectrum-connectivity and a deep learning feature vector. The intelligent analysis module is used to input the three-dimensional feature set and the deep learning feature vector into a pre-trained multi-level anomaly detection and classification model to perform full-cycle risk assessment from subclinical state to seizure event, and output seizure warning signals and seizure classification results. The decision-making and early warning module is used to generate and output graded early warning information and intervention decision suggestions to users or guardians based on risk assessment results, outbreak classification results and preset individualized risk stratification strategies.

9. A signal acquisition device for epilepsy monitoring, characterized in that, include: A disposable EEG cap is made of elastic non-woven fabric and has multiple electrode mounting positions, each of which is equipped with a positioning structure. Multiple reusable electronic acquisition modules are detachably installed on the electrode mounting positions of the disposable EEG cap and interconnected via flexible circuitry. A main processing unit is electrically connected to the plurality of electronic acquisition modules; Each of the electronic acquisition modules includes a housing, and a flexible microneedle-hydrogel composite dry electrode, an inertial measurement unit, and an optical volumetric sensor integrated within the housing; the flexible microneedle-hydrogel composite dry electrode, the inertial measurement unit, and the optical volumetric sensor are respectively used to acquire the scalp electroencephalogram signal, motion signal, and optical volumetric signal as described in claim 1.

10. The signal acquisition device according to claim 9, characterized in that, The disposable EEG cap has a three-layer composite structure, including: The surface layer is provided with electrode markings and the positioning structure; The inner layer has a perforated area for electrodes to contact the scalp; An interlayer, located between the surface layer and the inner layer, has openings for electrodes to pass through.

Images