Epilepsy automatic detection method and system based on EEG

Abstract

The invention discloses an EEG-based epilepsy automatic detection method and system, and belongs to the field of disease automatic detection. The method comprises the steps of collecting multi-channel electroencephalogram EEG signals in real time, conducting signal preprocessing, time-frequency feature extraction, deep learning analysis and multi-modal prediction, achieving accurate detection and prediction of epileptic seizure, generating real-time risk assessment of epileptic seizure through a prediction module in combination with multi-modal feature information, and achieving real-time risk assessment of epileptic seizure. The epilepsy early warning system sends an early warning signal to a user or a medical worker, starts corresponding intervention measures, has high detection precision and real-time response capability, and can provide effective early warning and management for epilepsy patients.

Description

Technical Field

[0001] The present application relates to the field of automatic disease detection, and in particular to an EEG-based automatic epilepsy detection method and system. Background Art

[0002] Epilepsy is a common neurological disease characterized by recurrent epileptic seizures, typically manifested by symptoms such as loss of consciousness and movement disorders. Epileptic seizures often occur suddenly, significantly impacting patients' quality of life. Early prediction of epileptic seizures is crucial for effective prevention and treatment of epilepsy. However, existing epilepsy detection methods rely primarily on traditional neuroelectrophysiological diagnostic techniques, such as electroencephalogram (EEG) monitoring. While these techniques can accurately record brain activity, they are often limited by the subjectivity and real-time nature of manual analysis, and are unable to rapidly predict the onset of an epileptic seizure in its early stages.

[0003] Traditional automated epilepsy detection methods typically rely on EEG feature extraction and machine learning algorithms. However, these methods often suffer from inaccuracies in processing complex EEG signals, making it difficult to guarantee efficient and stable performance, particularly in the presence of noise and artifacts. Furthermore, existing technologies lack effective integration of time-frequency feature fusion with deep learning, hindering the real-time and accuracy of detection to meet clinical needs.

[0004] Therefore, how to improve the accuracy and real-time performance of automatic epilepsy detection through more precise feature extraction and deep learning models, especially in complex environments, has become an urgent problem to be solved in the current technical field. Summary of the Invention

[0005] The present invention provides an EEG-based automatic epilepsy detection method and system. This method utilizes a time-frequency analysis method combining wavelet transform and short-time Fourier transform, and uses a deep learning model to comprehensively analyze EEG signals. This method addresses the shortcomings of existing technologies, enabling more accurate prediction of epileptic seizures and providing real-time warnings. The technical solution is as follows:

[0006] According to one aspect of the present application, a method for automatic epilepsy detection based on EEG is provided, wherein the method is used in an automatic epilepsy detection system based on EEG, and the method comprises:

[0007] Collecting multi-channel electroencephalogram (EEG) signals, wherein the EEG signals include EEG signals of epileptic patients and EEG signals of a normal control group;

[0008] The EEG signal is processed by a signal preprocessing module, including bandpass filtering to remove low-frequency noise, remove electrooculographic artifacts, electromyographic artifacts and power frequency interference, and standardize the signal;

[0009] Extracting the time-frequency features of the EEG signal by a time-frequency analysis method that combines multi-scale wavelet transform with adaptive short-time Fourier transform;

[0010] Based on the hybrid deep learning model of deep convolutional neural network (CNN) and long short-term memory network (LSTM), the extracted time-frequency features are fused and classified at multiple levels.

[0011] Combining the fusion features of time domain, frequency domain and time-frequency domain, a multimodal prediction model is used to detect and predict epileptic seizures;

[0012] Notifications are sent to users or medical personnel through the alarm system based on the detection results.

[0013] Optionally, the comprehensive denoising formula for removing low-frequency noise, electrooculogram artifacts, electromyography artifacts and power frequency interference is:

[0014] , Among them, F LP (y k ) is the filtering operation of low-frequency noise, λ LF is the removal gain of low-frequency noise;

[0015] s eye (y k ) is the component of electrooculographic artifact, λ eye is the gain for removing electrooculogram artifacts;

[0016] s EMG (y k ) is the component of myoelectric artifact, λ EMG is the EMG artifact removal gain;

[0017] F Notch (y k ) is the band-stop filter operation of power frequency interference, λ power It is the gain of removing power frequency interference.

[0018] Optionally, after the EEG signal is processed by the signal preprocessing module, the method further includes:

[0019] The processed EEG signal is segmented using an adaptive sliding time window method, wherein the length and step size of the sliding time window are dynamically adjusted according to the individual characteristics of the patient and the specific pattern of epileptic seizures. The length of the time window is represented by Tw, the step size is represented by Ts, and the signal segmentation formula is:

[0020] X(t)={x(t),t∈[t0,t0+T w ]}, t0=nT s, where n=0, 1, 2, 3…, the signal segmentation formula dynamically adjusts the window length according to the seizure pattern of each patient.

[0021] Optionally, the time-frequency feature extraction formula of the EEG signal is:

[0022]

[0023] Among them, T f (t) is the time-frequency feature, F AMFT (t, f) is the time-frequency spectrum extracted by adaptive multi-scale Fourier transform, and M is the number of scales.

[0024] Optionally, the hybrid deep learning model based on deep convolutional neural network (CNN) and long short-term memory network (LSTM) performs multi-level feature fusion and classification on the extracted time-frequency features. The model formula after fusion of CNN and LSTM is:

[0025] H l+1 =ResNet(LSTM(H l )), where H l+1 =H l +△H, where H l+1 is the output of the l+1th layer, ResNet and LSTM are convolutional networks and long short-term memory networks respectively, and ΔH is the residual term, indicating the nonlinear optimization process of the network.

[0026] Optionally, in the case of combining the fusion features of the time domain, frequency domain, and time-frequency domain and using a multimodal prediction model to detect and predict epileptic seizures, the multimodal feature weighted fusion formula is:

[0027]

[0028] The weighting coefficient formula is:

[0029]

[0030] Among them, F final is the final fusion feature, ω i is the weighting coefficient of each mode, σ i is the standard deviation of the feature, and λ is the adjustment coefficient that controls the weighting factor.

[0031] Optionally, the EEG signal is collected by electrode arrays arranged at different parts of the patient's head.

[0032] In another aspect, an EEG-based automatic epilepsy detection system is provided. The system is used in the above-mentioned EEG-based automatic epilepsy detection method, comprising:

[0033] EEG signal acquisition module, used for real-time acquisition of multi-channel EEG signals;

[0034] Signal preprocessing module, used to filter, remove artifacts and standardize the collected EEG signals;

[0035] Feature extraction module, used to extract time-frequency features from the preprocessed EEG signal;

[0036] Deep learning module, including deep convolutional neural network (CNN) and long short-term memory (LSTM) network, is used to perform multi-level feature fusion and classification on the extracted time-frequency features;

[0037] A prediction module, used to generate epileptic seizure detection and prediction results through a multimodal feature fusion method;

[0038] The alarm module is used to output early warning information of epileptic seizures based on the detection and prediction results.

[0039] Optionally, the system further comprises:

[0040] A data storage module is used to store EEG signals, preprocessed data, extracted time-frequency features, classification results, and prediction information. The data storage module supports query and backtracking of historical data;

[0041] The user interface module is used to interact with the user, display real-time epilepsy detection results, and provide operation options to start or stop the detection function. The user interface module is connected to the alarm module and the data storage module to provide information feedback and management functions.

[0042] On the other hand, a computer-readable storage medium is provided, wherein a computer program is stored in the computer-readable storage medium. The computer program is loaded and executed by a processor to implement the above-mentioned EEG-based automatic epilepsy detection method.

[0043] The present invention provides an EEG-based automatic epilepsy detection method, which includes real-time acquisition of multi-channel electroencephalogram (EEG) signals, and through signal preprocessing, time-frequency feature extraction, deep learning analysis, and multimodal prediction, etc., it can achieve accurate detection and prediction of epileptic seizures. Specifically, first, EEG signals of epilepsy patients and normal controls are collected, and the signals are bandpass filtered, artifacts are removed, and normalized to ensure data quality. Then, a time-frequency analysis method combining wavelet transform and short-time Fourier transform is used to extract the time-frequency features in the EEG signals. Then, based on the deep convolutional neural network (CNN) and the long short-term memory network (LSTM) model, the extracted features are subjected to multi-level feature fusion and classification to achieve the detection of epileptic seizures. Finally, combined with multimodal feature information, a prediction module is used to generate a real-time risk assessment of epileptic seizures, and an early warning signal is sent to the user or medical personnel to initiate corresponding intervention measures. This method has high detection accuracy and real-time response capability, and can provide effective early warning and management for epilepsy patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A flowchart of an EEG-based automatic epilepsy detection method provided by an exemplary embodiment of the present application is shown;

[0045] Figure 2 To correspond Figure 1 Figure 2 shows the test scenario implementation diagram of the EEG-based automatic epilepsy detection method. DETAILED DESCRIPTION

[0046] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0047] In this document, "plurality" refers to two or more. "And / or" describes a relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can mean: A exists alone, A and B exist simultaneously, or B exists alone. The character " / " generally indicates an "or" relationship between the associated objects.

[0048] Example 1

[0049] Please refer to Figure 1 , which shows a flow chart of an EEG-based automatic epilepsy detection method provided by an exemplary embodiment of the present application. This embodiment is described using the method for an EEG-based automatic epilepsy detection system as an example. The method includes:

[0050] Step 101 : collecting multi-channel EEG signals, wherein the EEG signals include EEG signals of epileptic patients and EEG signals of a normal control group.

[0051] Among them, EEG signals are collected through electrode arrays set at different parts of the patient's head.

[0052] Traditional EEG signal acquisition is affected by a variety of factors, such as poor electrode contact, motion artifacts, and environmental noise, all of which can compromise signal quality. This is particularly true for monitoring epilepsy patients, where noise and artifacts can interfere significantly and potentially lead to misinterpretations of seizure signals. Furthermore, seizure signals often exhibit complex spatial and temporal characteristics, requiring efficient fusion and analysis of multi-channel EEG signals. Current technology presents challenges in processing and utilizing signals from diverse brain regions in a timely and efficient manner.

[0053] The present invention addresses this issue by acquiring multi-channel EEG signals using an array of up to 64 electrodes positioned at different locations on the patient's head, covering key brain regions. This high-density electrode configuration improves spatial resolution of the epileptic seizure zone, ensuring comprehensive monitoring of epileptic activity. The electrode array design considers the characteristics of electrical activity transmission between the epileptogenic region and the opposite hemispheres of the brain, ensuring comprehensive signal acquisition.

[0054] To further enhance signal quality, this invention optimizes electrode selection and placement, employing improved, highly sensitive electrodes that can stably detect EEG signals even at lower current intensities. Furthermore, a signal amplification module and a noise suppression module amplify and purify EEG signals in real time, reducing errors caused by motion artifacts and ambient noise.

[0055] Furthermore, to effectively address interference in EEG signals, the present invention employs an adaptive filtering algorithm for noise suppression during the preprocessing phase after signal acquisition. This method can adaptively adjust the filtering parameters based on the real-time signal characteristics, effectively removing different types of noise (such as electrooculographic artifacts, electromyographic artifacts, and power frequency interference). The advantage of this dynamic filtering technique is that it can identify and automatically adjust the filtering parameters in real time, ensuring maximum retention of signals related to epileptic seizures.

[0056] In one possible implementation, EEG signal acquisition utilizes synchronous sampling technology to ensure temporal consistency across all channel signals. Each electrode array is connected to the main control unit via a dedicated synchronous clock signal, ensuring time synchronization during data acquisition and avoiding feature extraction errors caused by timing deviations. Furthermore, during the signal preprocessing stage, a time-frequency analysis method combining wavelet transform and short-time Fourier transform is used to extract the time-frequency features of the EEG signal, enhancing the ability to capture changes in EEG activity before and after an epileptic seizure.

[0057] This method also includes a real-time data transmission module that transmits the collected EEG signals via high-bandwidth wireless or wired channels to a data processing platform. This platform utilizes an efficient parallel computing framework combined with deep learning algorithms for data analysis. To reduce data transmission latency and improve processing efficiency, data is segmented and preprocessed locally. This allows for preliminary denoising and feature extraction at the signal acquisition end, reducing the computational burden on subsequent analysis modules and improving real-time responsiveness.

[0058] It can be seen that step 101 can comprehensively monitor the EEG activities in different areas of the brain, especially the starting area of epileptic seizures, through the configuration of a high-density electrode array, thereby improving the accuracy of detecting epileptic seizures; the adaptive signal preprocessing technology can identify and remove different types of noise and artifacts in real time, so that the quality of EEG signals is significantly improved, especially in an environment with large noise interference, and can provide reliable signal data; by adopting efficient parallel computing and adaptive filtering technology, the present invention can process multi-channel EEG signals in real time, quickly respond to the early warning needs of epileptic seizures, and provide timely intervention opportunities for the clinic. In the design of the high-density electrode array and the signal acquisition method, the combination of adaptive preprocessing and time-frequency feature extraction methods solves the problems of poor signal quality, large interference and poor real-time performance in traditional EEG acquisition methods. Through this technical means, the early detection capability of epileptic seizures can be significantly improved, and the stability and accuracy of the system in complex environments can be guaranteed.

[0059] Step 102 : Processing the EEG signal through a signal preprocessing module, including bandpass filtering to remove low-frequency noise, electrooculographic artifacts, electromyographic artifacts and power frequency interference, and standardizing the signal.

[0060] In a possible implementation, step 102 includes the following content.

[0061] Content 1: Use a bandpass filter to filter out low-frequency 0.5Hz-1Hz and high-frequency 50Hz-60Hz noise in the EEG signal;

[0062] Based on traditional bandpass filters, this implementation uses adaptive bandpass filtering technology, dynamically adjusting the filter parameters based on the frequency characteristics of the real-time EEG signal. This adaptive filter not only effectively removes low-frequency (0.5Hz-1Hz) and high-frequency (50Hz-60Hz) noise, but also adjusts the bandwidth in real time based on the EEG activity characteristics of different patients to avoid affecting the effective EEG signal.

[0063] It should be noted that the frequency range of EEG signals in certain brain regions may overlap with noise (for example, some EEG activity contains important information in the low-frequency band), and traditional fixed-band filtering methods may mistakenly remove these signals. To this end, the adaptive filtering technology of the present invention uses dynamic frequency adjustment to accurately set the filter bandwidth and frequency response based on the spectral information of the real-time signal, thereby effectively retaining the valid EEG signal.

[0064] Content 2: Use independent component analysis (ICA) or principal component analysis (PCA) to remove electrooculographic and myoelectric artifacts;

[0065] This embodiment of the present application combines ICA and PCA methods to adopt a hybrid artifact removal model. Traditional ICA or PCA methods are susceptible to the complexity of noise sources when removing artifacts, and cannot completely and effectively eliminate artifacts such as electrooculography and electromyography. This embodiment introduces an artifact signal prediction algorithm that enhances the accuracy of artifact removal by combining the time and frequency domain characteristics of the signal.

[0066] Because electrooculographic and electromyographic artifacts in EEG signals often overlap in different frequency ranges, ICA or PCA alone cannot completely separate these artifacts. In this embodiment, the artifact signal prediction algorithm first performs time-frequency analysis on the EEG signal, then uses a prediction model to determine the dynamic characteristics of the artifact signal. Finally, ICA and PCA are used to optimize and remove the artifacts, ensuring more accurate artifact removal.

[0067] Content three: Standardize the EEG signal and normalize the signal amplitude to a preset standard.

[0068] The standardization process not only normalizes the EEG signal amplitude to a preset standard but also introduces an adaptive standardization method that automatically adjusts the standardization range based on the characteristics of each patient's EEG activity. Specifically, standardization parameters (such as mean and variance) are automatically generated based on each patient's real-time signal statistics, making the standardized EEG signal more accurate to the individual patient's actual situation.

[0069] Traditional normalization methods use fixed parameters (such as mean and variance) to process signals, which may not fully account for differences in EEG activity between different patients, thereby affecting the accuracy of subsequent analysis. To overcome this technical bottleneck, the embodiments of the present application introduce adaptive normalization technology to automatically evaluate the EEG activity characteristics of each patient at the initial stage of signal acquisition and generate personalized normalization parameters, thereby ensuring that the normalization process can better adapt to individual differences and improve the accuracy of subsequent epilepsy detection.

[0070] As can be seen, in step 102, the bandpass filter used combines the real-time spectral changes of the signal and uses a spectrum adaptive algorithm to dynamically select the filter bandwidth. This method enhances the adaptability to the signal on the basis of conventional bandpass filtering, and can retain the effective EEG signal to the greatest extent while ensuring noise removal. The artifact signal prediction algorithm is based on a multi-layer neural network model. The training set simulates the interference pattern of electrooculography and electromyography artifacts to identify the typical characteristics of the artifact signal. Then, the algorithm uses a prediction method to identify artifacts that may affect the quality of the EEG signal in advance and automatically adjusts the artifact removal parameters of ICA or PCA to further improve the accuracy of artifact removal. The personalized normalization algorithm calculates personalized normalization parameters by analyzing the statistical characteristics of each patient's EEG signal (such as the standard deviation and mean of the signal amplitude) in real time. In particular, the normalization process is no longer uniform, but is flexibly adjusted according to the patient's actual EEG activity characteristics, so that the standardized signal can effectively reflect the individual's EEG characteristics.

[0071] In addition, in order to provide a more detailed description of the calculation process, the comprehensive denoising formula for removing low-frequency noise, electrooculographic artifacts, electromyographic artifacts, and power frequency interference is:

[0072]

[0073] , where F LP (y k ) is the filtering operation of low-frequency noise, λ LF is the removal gain of low-frequency noise; s eye (y k ) is the component of electrooculographic artifact, λ eye is the gain for removing electrooculogram artifacts; s EMG (y k ) is the component of myoelectric artifact, λ EMG is the gain for removing electromyographic artifacts; F Notch (y k ) is the band-stop filter operation of power frequency interference, λ power It is the gain of removing power frequency interference.

[0074] In addition, after the EEG signal is processed by the signal preprocessing module, the method further includes:

[0075] The processed EEG signal is segmented using an adaptive sliding time window method, wherein the length and step size of the sliding time window are dynamically adjusted according to the individual characteristics of the patient and the specific pattern of epileptic seizures. The length of the time window is represented by Tw, the step size is represented by Ts, and the signal segmentation formula is:

[0076] X(t)={x(t),t∈[t0,t0+T w ]}, t0=nTs , where n=0, 1, 2, 3…, the signal segmentation formula dynamically adjusts the window length according to the seizure pattern of each patient.

[0077] Step 103 : extracting the time-frequency features of the EEG signal by using a time-frequency analysis method that combines multi-scale wavelet transform with adaptive short-time Fourier transform.

[0078] The extracted time-frequency features are obtained by combining wavelet transform with short-time Fourier transform to fuse time-frequency information of different scales.

[0079] Traditional wavelet transform methods typically rely on fixed scales to analyze signals, but EEG signals have different local characteristics in different frequency bands. This method uses a multiscale wavelet transform to assign an adaptive scale to each frequency component, thereby simultaneously capturing both transient changes and long-term patterns in EEG signals at different levels in the time and frequency domains. This method can effectively improve sensitivity to dynamic signals during and before epileptic seizures.

[0080] In EEG signals, epileptic seizures often exhibit time-frequency localization, meaning that certain frequency components in the spectrum change significantly within a specific timeframe. To accurately extract these localized changes, traditional wavelet transforms suffer from a fixed scale selection and difficulty adapting to varying signal characteristics. This embodiment employs a multi-scale wavelet transform based on an adaptive algorithm, dynamically adjusting the scale of the wavelet basis functions to accommodate the characteristics of different EEG activities, particularly the dynamic variations of epileptic seizures.

[0081] Regarding adaptive short-time Fourier transform (STFT) enhancement of time-frequency features, the traditional STFT method uses a fixed window function to analyze the signal, which may not capture the subtle features of instantaneous frequency changes in the signal. In this embodiment, the short-time Fourier transform (STFT) technology with adaptive window adjustment automatically adjusts the length of the time window according to the spectral characteristics of the EEG signal, ensuring that the frequency distribution of the signal can be accurately extracted within a short period of time, thereby avoiding the information loss problem caused by a fixed window.

[0082] Choosing the window length for the short-time Fourier transform (STFT) is a significant technical challenge. A window that is too long loses high-frequency transient information, while one that is too short reduces frequency domain resolution, affecting the accurate capture of signal features. To overcome this problem, the present invention incorporates an adaptive time window algorithm that automatically adjusts the window length based on the signal's changing characteristics to achieve optimal time-frequency resolution. This algorithm dynamically adjusts the width of the time window based on the instantaneous frequency variations of the EEG signal, enabling more accurate extraction of transient frequency spectrum characteristics during epileptic seizures.

[0083] To address the limitations of both the multi-scale wavelet transform and the short-time Fourier transform (STFT) for fusing time-frequency information, this implementation proposes a multi-scale fusion time-frequency feature extraction method. This method combines the advantages of both methods, effectively fusing the high-frequency time-domain features extracted by the wavelet transform with the low-frequency frequency-domain features extracted by the STFT to generate a high-dimensional time-frequency feature map that encompasses the multi-level characteristics of epileptic signals in both the time and frequency domains.

[0084] The time-frequency features of wavelet transforms and short-time Fourier transforms are often difficult to directly fuse because they extract signal information from different perspectives. To overcome this problem, the present invention introduces a weighted fusion algorithm that assigns weights to each transform result based on the signal's contribution at different scales and frequencies, thereby achieving multi-scale, full-band information fusion. This fusion method can improve the stability and comprehensiveness of feature extraction, particularly in the detection of premonitory signals and abnormal waveforms of epileptic seizures, enhancing detection accuracy.

[0085] In one example, the time-frequency feature extraction formula of the EEG signal is:

[0086]

[0087] Among them, T f (t) is the time-frequency feature, F AMFT (t, f) is the time-frequency spectrum extracted by adaptive multi-scale Fourier transform, and M is the number of scales.

[0088] Furthermore, to better adapt to the characteristics of EEG signals, the present embodiment introduces an adaptive selection algorithm into the multi-scale wavelet transform process, automatically selecting the wavelet basis type (such as Morlet wavelet, Haar wavelet, etc.) based on the local characteristics of the signal. This method can automatically adjust the type and scale of the wavelet basis based on the characteristics of the EEG signal, so that the transformed characteristics more accurately reflect the time-frequency characteristics of the signal.

[0089] Furthermore, a weighting method based on information entropy is used during feature fusion. By calculating the information content of the time-frequency features extracted by wavelet transform and STFT at each scale, the features at each scale are weighted. Scales with high information content receive higher weights, ensuring that the most representative feature information is retained during the fusion process.

[0090] Furthermore, considering the real-time requirements of EEG signals, the present embodiment optimizes the efficiency of time-frequency analysis through a parallel computing architecture. By distributing the computational tasks of the multi-scale wavelet transform and adaptive STFT to multi-core processors and combining them with GPU acceleration technology, efficient time-frequency feature extraction is achieved under large-scale data processing.

[0091] It can be seen that in step 103, by combining the short-time Fourier transform of the adaptive window with the multi-scale wavelet transform, it is possible to avoid the feature loss problem caused by the fixed window length in the traditional method while ensuring the time-frequency resolution, especially to more accurately capture the frequency mutation and waveform changes during the epileptic seizure process; by the multi-scale fusion method, signal features are extracted from different angles of the time domain and frequency domain, effectively enhancing the sensitivity to epileptic signals, especially in the early warning of epileptic seizures, and more accurately identifying abnormal EEG activity; by dynamically adjusting the time window and wavelet scale, combined with the information weighted fusion strategy, this method can reduce the interference of noise on signal feature extraction, improve the stability and identifiability of epileptic signals in complex environments, and ensure high diagnostic accuracy; the multi-core parallel computing architecture and GPU acceleration enable the processing of large-scale EEG data to be completed under real-time requirements, ensuring the timely response and accuracy of epileptic seizure detection, and meeting the needs of clinical applications.

[0092] Step 104: Based on a hybrid deep learning model of a deep convolutional neural network (CNN) and a long short-term memory (LSTM) network, the extracted time-frequency features are subjected to multi-level feature fusion and classification.

[0093] The deep convolutional neural network (CNN) is used to extract the spatial features of the EEG signal, and the long short-term memory (LSTM) network is used to capture the time series dependencies of the signal.

[0094] When extracting spatial features from EEG signals using convolutional neural networks (CNNs), traditional CNNs, when used for image processing, primarily rely on convolution kernels to extract local spatial features of images. However, these methods struggle to capture the multidimensional, nonlinear spatial information of EEG signals' time-frequency characteristics. The present invention designs a spatial convolution module that incorporates multi-scale convolution kernels to extract spatial features of EEG signals at different scales, particularly capable of identifying spatial variations in signals during the pre-ictal, ictal, and recovery phases of epileptic seizures.

[0095] Epileptic EEG signals have strong spatial localization characteristics, which are difficult to directly identify with traditional CNN architectures. This paper designs a multi-scale convolutional module, enabling the network to learn different levels of spatial features in different convolutional layers. In particular, the design based on depthwise separable convolution operations further enhances the network's ability to capture spatial localization characteristics and reduces computational complexity.

[0096] Long-short-term memory (LSTM) networks are used to capture time series dependencies in EEG signals. LSTM networks, through their specialized memory units, are capable of capturing long-term dependencies in time series data. To enhance their ability to perceive the temporal dynamics of epileptic seizures in EEG signals, this implementation incorporates an adaptive learning rate mechanism into the LSTM training process. This allows the network to adaptively adjust the time window based on signal characteristics at different stages, accurately capturing early changes in epileptic seizures.

[0097] Time series dependencies in EEG signals are crucial for early warning of epileptic seizures, but traditional LSTM networks are prone to vanishing or exploding gradients in long time series. This paper employs adaptive learning rate adjustment and a residual connection strategy to enable LSTM to effectively overcome the gradient problem in long-sequence learning during training, enhancing its ability to capture time series features, particularly the short-term changes before and after an epileptic seizure.

[0098] For the feature fusion and multi-level classification of CNN and LSTM, the CNN extracts spatial features, while the LSTM captures the dependencies between time series. The key to this technical solution is how to integrate these two types of information. This paper adopts a multi-level feature fusion strategy, fusing the spatial features extracted by the CNN with the time series features extracted by the LSTM. This fusion is performed layer by layer at multiple network levels, ensuring that spatial and temporal information can be fully interacted and integrated within the deep network.

[0099] The fusion of spatial and temporal features often leads to information redundancy and feature conflicts, resulting in unstable model training. This implementation introduces a bidirectional feature fusion module into the network, adaptively adjusting the fusion strategy based on the importance of features at different levels. This ensures that spatial and temporal features complement each other effectively, improving the network's classification capabilities.

[0100] Furthermore, a multi-scale convolution module introduces convolution kernels of multiple scales (such as 3x3, 5x5, and 7x7 kernels) into the CNN convolutional layers. This extracts local spatial features of the EEG signal at different convolutional levels, enabling the identification of subtle changes in the signal. Batch normalization and ReLU activation functions are introduced after each convolutional layer to further improve the stability of feature extraction and nonlinear modeling capabilities.

[0101] Furthermore, the LSTM network uses an adaptive learning rate and momentum update strategy, trained using an adaptive optimization algorithm (such as Adam or RMSprop). During training, the LSTM memory unit adjusts its learning rate based on changes in the EEG signal, enabling long-term dependencies to be effectively learned without exploding or vanishing gradients.

[0102] Furthermore, at different layers of the network, spatial and temporal features are combined using a weighted fusion strategy. This fusion method takes into account the variability of signals at different scales and time periods, weighting features based on metrics such as information entropy to ensure that important information is preserved during the fusion process.

[0103] Furthermore, in order to solve the problem of gradient vanishing and information loss in deep networks, residual connections are used in feature fusion networks, which enables the network to maintain smooth information flow and reduce the problem of gradient vanishing or exploding during training.

[0104] It can be seen that in step 104, through the combination of multi-scale convolution and LSTM, the spatial local features and time series dependent features in the EEG signal can be fully exploited, especially in the early stage and dynamic change process of epileptic seizures, and epilepsy-related signal changes can be captured more accurately; through the multi-level feature fusion strategy, spatial features and temporal features can be efficiently interactively fused in the deep network, improving the accuracy and stability of classification, especially in complex epilepsy signal detection tasks; by adopting an adaptive learning rate mechanism, residual connection and efficient optimization algorithm, the model can converge quickly, while improving the robustness of the network in real-time applications and reducing the delay problem caused by large data volume and high computational complexity; after fusing CNN and LSTM, the model can suppress the influence of noise on signal analysis without losing important information, thereby improving the reliability of epilepsy detection, especially when the patient signal is more complex or the noise is larger, and still maintain a high detection accuracy.

[0105] In an example, in the multi-level feature fusion and classification of the extracted time-frequency features, the model formula after fusion of CNN and LSTM is H l+1 =ResNet(LSTM(H l )) Among them, H l+1 =H l +△H, where H l+1 is the output of the l+1th layer, ResNet and LSTM are convolutional networks and long short-term memory networks respectively, and ΔH is the residual term, indicating the nonlinear optimization process of the network.

[0106] Step 105 , combining the time domain, frequency domain and time-frequency domain fusion features, and using a multimodal prediction model to detect and predict epileptic seizures.

[0107] In one example, a multimodal prediction model is used to detect and predict epileptic seizures by combining fusion features of the time domain, frequency domain, and time-frequency domain. The weighted fusion formula of the multimodal features is:

[0108]

[0109] The weighting coefficient formula is:

[0110]

[0111] Among them, F final is the final fusion feature, ω i is the weighting coefficient of each mode, σ i is the standard deviation of the feature, and λ is the adjustment coefficient that controls the weighting factor.

[0112] The deep learning model uses a multimodal feature fusion method to perform a weighted combination of time-frequency features, time series features, and spatial features. The multimodal prediction model combines the deep learning model with a traditional machine learning regression model to process the fused feature data and generate epileptic seizure prediction results.

[0113] For the weighted combination of time-frequency features, time series features and spatial features, the multidimensional features of epileptic seizure signals include time domain information (such as the waveform of the EEG signal), frequency domain information (such as spectrum changes), and time-frequency domain information (such as instantaneous changes in time-frequency features). Traditional methods can often only process a certain type of feature separately and cannot fully utilize all information. The present invention proposes a weighted fusion mechanism to perform a weighted combination of the features of the time domain, frequency domain and time-frequency domain, so that each feature can fully play its role at different levels.

[0114] During the fusion of time-frequency features, the relative importance of different features is often difficult to quantify, and the scales of these features vary greatly. Direct fusion can lead to information redundancy or loss. Therefore, a weighted fusion strategy based on information gain was adopted. Based on the importance of each feature at different stages (such as the epileptic seizure phase and the premonitory phase), the contribution of each feature was quantified using information gain. This dynamic weight was then assigned to each feature, ensuring that the model retains the most critical information when fusing features.

[0115] This application combines deep learning models with traditional machine learning regression models. Deep learning models (such as convolutional neural networks (CNNs) and LSTMs) are used to process complex time-frequency features and extract high-level abstract information, while traditional machine learning regression models (such as support vector machines (SVMs) and random forests) are used to more accurately model and regress the fused features to generate epileptic seizure prediction results.

[0116] Deep learning models possess powerful feature learning capabilities, but when faced with multimodal features, they are prone to overfitting and vanishing gradients during training. Traditional machine learning models, while stable, are less capable of extracting complex features. To address this issue, a phased training strategy was adopted, first extracting high-order features using a deep learning model, then passing these features to a traditional machine learning regression model for fine-tuning. Furthermore, cross-validation was used to optimize the integration of deep learning and traditional machine learning models, ensuring robust and accurate predictions.

[0117] For multimodal feature fusion and multi-level classification, after feature fusion, a multi-level classification network is used to predict epileptic seizures. This network designs a multi-stage classifier based on the processing depth and level of different features. Feature selection and fusion are performed separately at each stage, ensuring that the classification results at each stage are based on the most relevant features.

[0118] After feature fusion, selecting the optimal decision path among multiple classifiers to achieve the most accurate prediction results becomes a key challenge in this technical solution. To address this problem, the present invention introduces a multi-stage convolutional regression network. This network extracts richer features through multiple layers of convolution and pooling operations, predicting epileptic seizures at each level. Ultimately, the results of these multiple levels are weighted and fused to generate the final prediction output.

[0119] Furthermore, to ensure a more rational weighted fusion of features, a dynamic weighting algorithm based on information gain was employed. During model training, the weights of each feature were automatically adjusted according to the stage of signal change, ensuring a rational fusion of time-domain, frequency-domain, and time-frequency-domain features. For example, in the early stages of an epileptic seizure, time-domain features may be more important, while at the peak of an attack, time-frequency-domain features may be more representative.

[0120] Furthermore, the combination of deep learning and traditional machine learning models employs a hierarchical training strategy: first, the deep learning model is trained to extract time-frequency features, and then the traditional machine learning regression model is trained to model and predict these features. This process uses a pre-training + fine-tuning approach, ensuring that the deep learning model fully learns the in-depth information of the time-frequency features, while the traditional machine learning model can obtain efficient regression capabilities from the trained deep learning features.

[0121] Furthermore, the network employs a multi-layer convolutional module, progressively extracting features through multiple layers of convolution and pooling operations. Each convolutional layer uses weighted convolution kernels to optimize feature extraction. In the final regression stage, a regression layer is used to predict the occurrence of epileptic seizures, ensuring that the model can make precise predictions based on the importance of features at different classification levels.

[0122] Furthermore, to further improve the generalization ability of the model and reduce overfitting, the present invention introduces the k-fold cross-validation technique. This method ensures the stability of the model and improves the generalization ability of the model on different data sets by dividing the data set into multiple training and validation steps.

[0123] It can be seen that in step 105, through the weighted fusion strategy, key information can be fully retained in multiple feature spaces, avoiding the information loss problem that may occur in a single feature model, and improving the accuracy and stability of epileptic seizure detection; by combining deep learning with traditional machine learning regression models, the present invention can extract higher-order features from complex epileptic EEG signals, and make precise predictions through regression models, thereby improving the model's predictive ability for epileptic seizures, especially in real-time monitoring and early warning; through the design of a multi-stage convolutional regression network, the model's perception of different feature levels is further enhanced, ensuring that the prediction results of each level can be classified based on the most relevant features, and finally fusing multi-level results to improve the overall prediction accuracy; adopting a staged training and cross-validation strategy, the overfitting phenomenon in the training process is reduced, the stability and generalization ability of the model are improved, and it is ensured that in the actual application of different epileptic signals, the model can run stably and efficiently.

[0124] Step 106: Send a notification to the user or medical personnel through the alarm system based on the detection result.

[0125] Regarding the real-time notifications delivered via an intelligent alarm system, the present invention also designs an intelligent alarm system that combines epileptic seizure detection results with the patient's real-time status to generate personalized notifications, which are automatically sent to the user or medical staff via multiple communication channels (such as mobile phone apps, text messages, phone calls, and emails). The key to this system lies in its real-time and intelligent nature, enabling customized notification methods and content based on different seizure phases (e.g., warning period, onset period, and recovery period), ensuring timely and accurate delivery of critical information.

[0126] The occurrence of an epileptic seizure is a rapidly changing dynamic process. How to promptly notify relevant personnel through the alarm system at the moment of occurrence while avoiding errors or delays is a technical challenge. To solve this problem, the present invention adopts a multi-level information processing mechanism, including:

[0127] Real-time monitoring of epileptic seizure signals through deep learning models ensures that changes in each signal can be quickly fed back to the alarm system;

[0128] The notification content automatically adjusts according to the different stages of the attack. For example, during the warning period, the system will send a reminder notification; during the attack period, the system will push an emergency alert and notify medical personnel for timely intervention.

[0129] To enhance the accuracy and reliability of the alarm system, this invention employs a multi-tiered alarm mechanism and precise notification strategy. Different alarm levels trigger notifications based on signal strength, seizure duration, and the patient's current status (e.g., whether they are in a medical facility). The system not only supports personalized notifications but also automatically selects appropriate notification content and methods based on the patient's clinical information (e.g., epilepsy history, seizure frequency, etc.).

[0130] A key challenge is reducing unnecessary false alarms and missed alerts. By introducing a multi-level decision tree algorithm, the alarm system can make comprehensive judgments based on different data sources (such as real-time analysis of EEG signals, patient health status, and historical data), ensuring that notifications are sent only when necessary, thereby reducing the probability of false alarms and missed alerts.

[0131] For the notification system integrating wearable devices and telemedicine services, the present invention further combines wearable devices and telemedicine services. Through integration with wearable devices such as smart bracelets and smart watches, the alarm system can monitor the patient's physical condition (such as heart rate, blood pressure, etc.) in real time and connect to the alarm system via Bluetooth or Wi-Fi. If it is detected that the patient is in the epileptic seizure period and the physical indicators are abnormal, the system will automatically activate the alarm and send detailed information to medical staff or relatives.

[0132] Another challenge in this technical solution is how to ensure the real-time nature of telemedicine services and the rapid response of medical personnel. To this end, the present invention has designed a multi-channel response system that connects with medical personnel's terminal devices (such as mobile phones, tablets, etc.) in real time through the network to ensure that each medical staff member can receive the patient's seizure information in the first place. In addition, the system also supports the dynamic selection of the most appropriate responder based on the medical staff's availability, ensuring the speed and accuracy of the response.

[0133] Furthermore, the system supports personalized adjustments to notification content based on individual patient differences. For example, for patients with long-term epilepsy, the system will use historical data to determine the severity of the attack and choose whether to send a simple reminder, emergency alert, or detailed medical guidance based on the patient's medical needs. The system will generate corresponding notification templates based on the medical staff's preset parameters (such as the patient's age, disease progression, etc.), ensuring that the notification content is accurate, clear, and easy to understand.

[0134] Furthermore, this implementation ensures timely notifications through multiple methods. In addition to traditional SMS and phone calls, the system also uses modern tools like smartwatches, smart headphones, and mobile devices to deliver notifications to multiple platforms for patients, their families, and medical staff. Furthermore, the system can also utilize voice notifications to convey emergency messages, helping patients quickly access critical information in critical situations.

[0135] Furthermore, when medical staff receive an alert notification, the system provides follow-up tracking, ensuring they can initiate immediate intervention based on the notification and provide real-time updates on the patient's status. Following an attack, the system automatically records all relevant data and provides medical staff with detailed information, including time, duration, and frequency, allowing doctors to analyze the patient's progress.

[0136] It can be seen that in the present invention, the alarm system can promptly notify patients and medical personnel based on the real-time monitoring results of epileptic seizures, provide rapid response and intervention, and reduce the potential risks brought by epileptic seizures; by introducing multi-level decision-making algorithms and dynamic push mechanisms, the system can ensure that alarms are sent only when necessary, avoid false alarms and missed alarms, and ensure the safety of patients; the alarm system can adjust the notification content and notification method according to the patient's personalized needs and health status, provide customized alarm services, and improve the adaptability and accuracy of the system; through the integration of telemedicine services and wearable devices, medical personnel can obtain the patient's real-time data at the first time, make judgments and responses quickly, and avoid delays caused by insufficient medical resources.

[0137] Further, such as Figure 2 As shown, Figure 2 To correspond Figure 1 Figure 2 shows the test scenario implementation diagram of the EEG-based automatic epilepsy detection method. Figure 2 This paper demonstrates a test scenario for an EEG-based automated epilepsy detection system, using a mouse as the experimental subject to verify the system's functionality and performance. In this scenario, EEG signals are acquired using a multi-channel electrode array placed on the subject's head. The acquired signals are first processed by a signal amplification module and a noise suppression module to improve signal quality and remove artifacts and ambient noise. Next, the EEG signals undergo time-frequency analysis to extract their time-frequency features, which are then analyzed using a deep learning model to identify the precursors and progression of epileptic seizures. In the "Model Parameters" module (shown in the figure), the system adjusts the parameters of the deep learning model based on the subject's actual data to ensure the precision and accuracy of signal analysis and epileptic seizure detection. Finally, the detection results are transmitted to medical personnel in real time via an alarm module, ensuring a timely response. Despite the use of mice as the experimental subject, this test scenario demonstrates the effectiveness of the system and suggests that this technology could be further applied to clinical monitoring and early warning of epilepsy in humans.

[0138] The following is a further analysis of the processes involved in the test diagram.

[0139] Experimental subjects (the mouse icon shown in the figure) use mice as examples in the experimental subject section. This is because, in many biomedical studies, especially in the study of neurological diseases such as epilepsy, mice are often used as an ideal animal model to preliminarily verify the effectiveness of technologies and methods. Although this figure uses mice as experimental subjects, in actual applications, all signal acquisition, processing and analysis steps are also applicable to human patients, especially in clinical trials. Specifically, the system is started by the experimental subject (such as an animal model or an epilepsy patient) providing an EEG signal. EEG signals are collected through electrode arrays set at different parts of the experimental subject's head. These electrode arrays are responsible for collecting multi-channel EEG signals, including signals from epilepsy patients and normal controls. This part optimizes signal quality through the "signal amplification module" and "noise suppression module" to ensure the validity of the data.

[0140] EEG signal acquisition. The EEG signal acquisition module (configured by a computer) is also applicable to humans. In animal experiments, EEG signals are collected via an electrode array placed on the animal's head. Depending on the experimental requirements, the computer system can adjust the position and number of the electrode arrays based on the activity of different brain regions. In a clinical setting, the configuration of the electrode array will take into account the anatomical characteristics of the human brain.

[0141] Signal amplification and noise suppression. In the "signal amplification module" and "noise suppression module" shown in the figure, although the experimental subjects in the figure are mice, they are technically also applicable to humans. In human patients, EEG signals also need to be amplified and noise processed to ensure the quality and accuracy of the signal. These steps have been verified in animal experiments and can be directly extended to clinical applications. Specifically, after signal acquisition, the "signal amplification module" in the figure is responsible for amplifying the collected EEG signal, while the "noise suppression module" purifies the signal in real time. The material mentions that interference from noise and artifacts, especially in the monitoring of epileptic patients, may lead to misjudgment of epileptic seizure signals. Through high-sensitivity electrode configuration and improved noise suppression methods, the signal quality is significantly improved.

[0142] Time-Frequency Feature Extraction and Analysis. The "Time-Frequency Analysis Method" module in the figure is primarily used to extract the time-frequency features of EEG signals. These techniques have been validated in animal experiments and can be applied to human EEG signal processing. In both animal models and human clinical trials, these techniques have effectively identified EEG activity patterns before and after epileptic seizures. Specifically, after signal acquisition and preprocessing, the "Learning Model" module in the figure is responsible for processing the signal and extracting its time-frequency features. In the above embodiment, a time-frequency analysis method combining wavelet transform and short-time Fourier transform is used to extract the time-frequency features of the EEG signal. This method enables the system to extract the underlying patterns of epileptic seizures from complex spatial and temporal features. Furthermore, the "EEG signal" in the figure is collected simultaneously from multiple signal sources, overcoming the limitations of single-channel signals and ensuring the comprehensiveness and accuracy of the data. This corresponds to the 64-electrode array configuration mentioned in the materials, enhancing the signal's spatial resolution. The "Signal Preprocessing" module in the figure optimizes the EEG signal using various methods, including bandpass filtering to remove low- and high-frequency noise and artifact removal. The materials mention the use of adaptive filtering techniques, independent component analysis (ICA), and principal component analysis (PCA) to remove electrooculogram and electromyographic artifacts. "Time-frequency analysis" extracts the time-frequency characteristics of the signal, integrating information from the time, frequency, and time-frequency domains shown in the figure. By combining multi-scale wavelet transforms with short-time Fourier transforms, the system can effectively capture dynamic changes before and after an epileptic seizure.

[0143] The "Multimodal Prediction Model" in the figure demonstrates the application of deep learning in signal analysis, combining a convolutional neural network (CNN) and a long short-term memory network (LSTM). CNN is used to extract spatial features from EEG signals, while LSTM focuses on capturing dependencies within time series. This is precisely the method mentioned in the material for extracting and fusing spatiotemporal features using a hybrid CNN and LSTM deep learning model. This module not only extracts spatial features but also effectively combines them with temporal features using a multi-level feature fusion strategy, ensuring more accurate and efficient detection of epileptic seizures.

[0144] In the figure, the "model parameters" part is a key module in the system, which directly affects the performance and accuracy of the deep learning model. According to the description in the material, this embodiment uses a hybrid deep learning model that combines a deep convolutional neural network (CNN) and a long short-term memory network (LSTM) to process the time-frequency characteristics of EEG signals. In this process, the "model parameters" module is responsible for adjusting and optimizing the various parameters in the network to ensure the system's ability to recognize and analyze different EEG activity signals. Based on data collected from experimental subjects (such as mouse models), the model parameters can be dynamically adjusted to adapt to different types of epileptic seizure signals and ensure the robustness and accuracy of the model.

[0145] The "Model Parameters" in the figure indicate how the system adjusts parameters based on training data, including the size of the convolution kernel, the configuration of memory cells in the LSTM network, and adaptive learning rate adjustments. This ensures the model operates stably in various environments and accurately predicts epileptic seizures. This approach improves the system's ability to provide early warning of epileptic seizures and ensures its reliability in clinical applications.

[0146] Example 2

[0147] On the other hand, the present application also provides an EEG-based automatic epilepsy detection system, which is used in the above-mentioned automatic epilepsy detection method, including:

[0148] EEG signal acquisition module, used for real-time acquisition of multi-channel EEG signals;

[0149] Signal preprocessing module, used to filter, remove artifacts and standardize the collected EEG signals;

[0150] Feature extraction module, used to extract time-frequency features from the preprocessed EEG signal;

[0151] Deep learning module, including deep convolutional neural network (CNN) and long short-term memory (LSTM) network, is used to perform multi-level feature fusion and classification on the extracted time-frequency features;

[0152] A prediction module, used to generate epileptic seizure detection and prediction results through a multimodal feature fusion method;

[0153] The alarm module is used to output early warning information of epileptic seizures based on the detection and prediction results.

[0154] Furthermore, the system further includes the following contents.

[0155] A data storage module is used to store EEG signals, preprocessed data, extracted time-frequency features, classification results, and prediction information. The data storage module supports query and backtracking of historical data;

[0156] The user interface module is used to interact with the user, display real-time epilepsy detection results, and provide operation options to start or stop the detection function. The user interface module is connected to the alarm module and the data storage module to provide information feedback and management functions.

[0157] The EEG-based automatic epilepsy detection system provided in the embodiment of the present application combines multi-channel EEG acquisition, signal preprocessing, time-frequency feature extraction, deep learning model and alarm system, which improves the accuracy and real-time performance of epileptic seizure detection as a whole. Through multi-level processing of EEG signals, including filtering, artifact removal and standardization, high-quality input of signals is ensured. The combination of time-frequency analysis method and deep convolutional neural network (CNN) and long short-term memory network (LSTM) effectively extracts the spatial and temporal features of EEG signals and improves detection accuracy through multimodal feature fusion. Combined with the multimodal prediction model, the system can quickly and accurately detect epileptic seizures and notify users or medical personnel in real time through the alarm module to ensure timely intervention. In addition, the data storage module supports historical data query and backtracking, enhances the ability of long-term monitoring and clinical analysis, and the user interface module provides an intuitive interactive interface, which improves the convenience of operation and response efficiency. Through the combination of these technical means, this system has obvious advantages in epileptic seizure detection and early warning.

[0158] An embodiment of the present application further provides a computer-readable medium storing at least one instruction, wherein the at least one instruction is loaded and executed by the processor to implement the EEG-based automatic epilepsy detection method as described in the above embodiments.

[0159] The above description is merely an optional embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. An automatic epilepsy detection method based on EEG, characterized in that: The method is used in an EEG-based automatic epilepsy detection system, and the method comprises: Collecting multi-channel electroencephalogram (EEG) signals, wherein the EEG signals include EEG signals of epileptic patients and EEG signals of a normal control group; The EEG signal is processed by a signal preprocessing module, including bandpass filtering to remove low-frequency noise, remove electrooculographic artifacts, electromyographic artifacts and power frequency interference, and standardize the signal; Extracting the time-frequency features of the EEG signal by a time-frequency analysis method that combines multi-scale wavelet transform with adaptive short-time Fourier transform; Based on the hybrid deep learning model of deep convolutional neural network (CNN) and long short-term memory network (LSTM), the extracted time-frequency features are fused and classified at multiple levels. Combining the fusion features of time domain, frequency domain and time-frequency domain, a multimodal prediction model is used to detect and predict epileptic seizures; Notifications are sent to users or medical personnel through the alarm system based on the detection results.

2. The method according to claim 1, characterized in that The comprehensive denoising formula for removing low-frequency noise, electrooculogram artifacts, electromyography artifacts and power frequency interference is: Among them, F LP (y k ) is the filtering operation of low-frequency noise, λ LF is the removal gain of low-frequency noise; s eye (y k ) is the component of electrooculographic artifact, λ eye is the gain for removing electrooculogram artifacts; s EMG (y k ) is the component of myoelectric artifact, λ EMG is the EMG artifact removal gain; F Notch (y k ) is the band-stop filter operation of power frequency interference, λ power It is the gain of removing power frequency interference.

3. The method according to claim 1, characterized in that After the EEG signal is processed by the signal preprocessing module, the method further includes: The processed EEG signal is segmented using an adaptive sliding time window method, wherein the length and step size of the sliding time window are dynamically adjusted according to the individual characteristics of the patient and the specific pattern of epileptic seizures. The length of the time window is represented by Tw, the step size is represented by Ts, and the signal segmentation formula is: X(t)={x(t),t∈[t0,t0+T w ]},t0=nT s , where n=0, 1, 2, 3…, the signal segmentation formula dynamically adjusts the window length according to the seizure pattern of each patient.

4. The method according to claim 1, wherein The time-frequency feature extraction formula of the EEG signal is: Among them, T f (t) is the time-frequency feature, F AMFT (t, f) is the time-frequency spectrum extracted by adaptive multi-scale Fourier transform, and M is the number of scales.

5. The method according to claim 1, wherein The hybrid deep learning model based on deep convolutional neural network (CNN) and long short-term memory (LSTM) performs multi-level feature fusion and classification on the extracted time-frequency features. The model formula after fusion of CNN and LSTM is: H l+1 =ResNet(LSTM(H l )), where H l+1 =H l +△H, where H l+1 is the output of the l+1th layer, ResNet and LSTM are convolutional networks and long short-term memory networks respectively, and ΔH is the residual term, indicating the nonlinear optimization process of the network.

6. The method according to claim 1, characterized in that In the above-mentioned multimodal prediction model for detecting and predicting epileptic seizures by combining the fusion features of the time domain, frequency domain, and time-frequency domain, the weighted fusion formula of the multimodal features is: The weighting coefficient formula is: Among them, F final is the final fusion feature, ω i is the weighting coefficient of each mode, σ i is the standard deviation of the feature, and λ is the adjustment coefficient that controls the weighting factor.

7. The method according to any one of claims 1 to 6, characterized in that: The EEG signals are collected by electrode arrays arranged at different parts of the patient's head.

8. An EEG-based automatic epilepsy detection system, characterized in that: The system is used for the EEG-based automatic epilepsy detection method according to any one of claims 1 to 7, comprising: EEG signal acquisition module, used for real-time acquisition of multi-channel EEG signals; Signal preprocessing module, used to filter, remove artifacts and standardize the collected EEG signals; Feature extraction module, used to extract time-frequency features from the preprocessed EEG signal; Deep learning module, including deep convolutional neural network (CNN) and long short-term memory (LSTM) network, is used to perform multi-level feature fusion and classification on the extracted time-frequency features; A prediction module, used to generate epileptic seizure detection and prediction results through a multimodal feature fusion method; The alarm module is used to output early warning information of epileptic seizures based on the detection and prediction results.

9. The EEG-based automatic epilepsy detection system according to claim 8, characterized in that: The system further comprises: A data storage module is used to store EEG signals, preprocessed data, extracted time-frequency features, classification results, and prediction information. The data storage module supports query and backtracking of historical data; The user interface module is used to interact with the user, display real-time epilepsy detection results, and provide operation options to start or stop the detection function. The user interface module is connected to the alarm module and the data storage module to provide information feedback and management functions.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which is loaded and executed by a processor to implement the EEG-based automatic epilepsy detection method according to any one of claims 1 to 7.

Images