Detection device for high-frequency oscillation of electroencephalogram signals of intrusive brain machine

Abstract

The invention provides an invasive brain-computer electroencephalogram signal high-frequency oscillation detection device, the detection device comprises a collection device, a processing device and a display device, through cooperation of the collection device, the processing device and the display device, full-process automatic processing is achieved, efficiency is greatly improved, and manual deviation is eliminated; the first-level structure is subjected to time domain-frequency domain double-branch feature extraction and gating fusion, the second-level structure is subjected to lead graph modeling and multi-head attention analysis, HFO multi-dimensional features and multi-lead synchronization characteristics are fully captured, the false detection / omission ratio is remarkably reduced, a quantitative basis is provided for epilepsy induced area positioning and an operation scheme, and the epilepsy induced area positioning and multi-lead attention analysis method is applied to epilepsy induced area positioning. The dual goals of technology cost reduction, efficiency improvement and diagnosis and treatment quality improvement are achieved.

Description

Technical Field

[0001] This application relates to the field of medical technology, and more specifically, to an invasive brain-computer interface device for detecting high-frequency oscillations of electroencephalogram (EEG) signals. Background Technology

[0002] High-frequency oscillating foci (HFOs) are key biomarkers of abnormal electrical activity in the brain, especially closely related to the localization of epileptogenic zones. Pathological HFOs are often concentrated in epileptogenic zones, and their distribution and frequency directly guide the formulation of epilepsy surgical plans. Therefore, accurately identifying HFOs from invasive EEG signals is a core requirement for the clinical diagnosis and treatment of epilepsy and other neurological diseases.

[0003] Traditional detection relies on medical staff to manually annotate HFOs, which requires analyzing massive amounts of EEG data segment by segment, which is time-consuming and labor-intensive. Moreover, the annotation results are greatly affected by the experience of the personnel, and the consistency between different annotators is poor, which easily leads to the omission of weak signal HFOs or misjudgment of noise as HFOs. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide an invasive brain-computer interface for detecting high-frequency oscillations of brain-computer signals, in order to overcome the problems in the prior art.

[0005] In a first aspect, embodiments of this application provide an invasive brain-computer interface (BCI) device for detecting high-frequency oscillations of EEG signals, the device comprising: a data acquisition device, a processing device, and a display device; The acquisition device is used to acquire initial electroencephalogram (EEG) signals according to acquisition requirements and a preset sampling rate, and send the initial EEG signals to the processing device. The processing device is used to segment the target EEG signal and the corresponding target time-frequency map to obtain a processed basic dataset; wherein, the target EEG signal is the preprocessed initial EEG signal; Based on different preset screening dimensions, candidate datasets with potential high-frequency oscillation characteristics are selected from the basic dataset; The candidate dataset is input into a preset two-level architecture deep learning model. The first-level structure sub-model is used to obtain the single-lead high-frequency oscillation probability of the candidate dataset. The second-level structure sub-model is used to obtain the window-level high-frequency oscillation probability of the candidate dataset. Based on the single-lead high-frequency oscillation probability and the window-level high-frequency oscillation probability, the high-frequency oscillation discharge fragment and the lead-level pathological high-frequency oscillation probability are obtained. The display device is used to show the high-frequency oscillating discharge fragments and the probability of pathological high-frequency oscillations at the lead level, providing medical staff with a basis for treatment.

[0006] In some technical solutions of this application, the above-mentioned processing device preprocesses the initial electroencephalogram (EEG) signal in the following manner: The initial EEG signal is subjected to rereference processing according to a preset reference point or reference method to obtain the processed first EEG signal. The first filter is used to eliminate power frequency noise in the first EEG signal to obtain the second EEG signal; wherein, the first filter is a notch filter based on a first filtering frequency; The target EEG signal is obtained by using a second filter to eliminate baseline drift in the second EEG signal; wherein the second filter is a high-pass filter based on a second filtering frequency.

[0007] In some technical solutions of this application, the aforementioned processing device is used to segment the target EEG signal and the corresponding target time-frequency map to obtain a processed basic dataset; wherein, the target EEG signal is the preprocessed initial EEG signal, including: The target EEG signal is processed using a preset wavelet transform method to obtain an initial time-frequency map; The effective signal features related to high-frequency oscillations in the initial time-frequency graph are enhanced, while noise and redundancy unrelated to high-frequency oscillations in the initial time-frequency graph are suppressed to obtain the target time-frequency graph. According to a preset time window, the target EEG signal and the target time-frequency map are slide-segmented to obtain multiple data units; wherein, each data unit includes a waveform segment of a single lead and the corresponding time-frequency map of the waveform segment, and all data units constitute the basic dataset.

[0008] In some technical solutions of this application, the above-mentioned preset different screening dimensions include: frequency band filtering, envelope threshold screening, synchronization constraint and spatial density screening; The processing device obtains the candidate dataset in the following manner: The basic dataset is processed sequentially using the preset sorting dimensions to obtain the candidate dataset.

[0009] In some technical solutions of this application, the above-mentioned primary structure sub-model includes a time-domain feature extraction component, a time-frequency feature extraction component, and a gated fusion component; the processing device obtains the single-lead high-frequency oscillation probability in the following manner: The time-domain feature extraction component is used to extract the first high-frequency oscillation feature of the waveform segment of a single lead in the candidate dataset in the time-domain waveform; The time-frequency feature extraction component is used to extract the second high-frequency oscillation feature in the frequency domain dimension of the target time-frequency map corresponding to the same lead in the candidate dataset. The gated fusion component is used to filter and fuse the first high-frequency oscillation feature and the second high-frequency oscillation feature to obtain the target high-frequency oscillation feature.

[0010] In some technical solutions of this application, the above-mentioned secondary structure sub-model includes a lead graph structure construction module, a multi-head attention graph convolution module, a context fusion module, and a first fully connected layer; The processing device obtains the window-level high-frequency oscillation probability in the following manner: The construction module is used to convert the target high-frequency oscillation characteristics into a lead graph data structure; wherein, the lead graph data structure includes node features and edge weights; The multi-head attention graph convolution module is used to determine the global association feature vector based on the node features and to determine the node attention weight based on the edge weights. The context fusion module is used to obtain a fused feature vector based on the global correlation feature vector and the target high-frequency oscillation feature; The first fully connected layer calculates the single-lead high-frequency oscillation probability corresponding to the target high-frequency oscillation feature and the window-level high-frequency oscillation probability corresponding to the fused feature vector.

[0011] In some technical solutions of this application, the above-mentioned processing device obtains high-frequency oscillating discharge fragments and lead-level pathological high-frequency oscillation probabilities in the following ways: Based on the window-level high-frequency oscillation probability and the preset high-frequency oscillation threshold, an effective time window is selected from the candidate time windows in the candidate dataset; The effective time windows are merged according to a preset time interval to obtain a merged high-frequency oscillation discharge segment. The probability of pathological high-frequency oscillation at the lead level is calculated based on the single-lead high-frequency oscillation probability, the node attention weight, and the high-frequency oscillation discharge segment.

[0012] In some technical solutions of this application, the above-mentioned primary structural sub-model includes a single-lead Local-CNN encoder, and the secondary structural sub-model includes a variable-lead Group-GNN determiner.

[0013] In some technical solutions of this application, the aforementioned time-domain feature extraction component includes a Wave-branch model; the time-frequency feature extraction component includes a TF-branch component; and the gated fusion component includes a GLU fusion module.

[0014] In some technical solutions of this application, the aforementioned context fusion module includes a Context-Gate fusion module.

[0015] The technical solutions provided by the embodiments of this application may include the following beneficial effects: The invasive brain-computer interface (BCI) device for detecting high-frequency oscillations in EEG signals disclosed in this application comprises: a data acquisition device, a processing device, and a display device. The data acquisition device acquires initial EEG signals according to acquisition requirements and a preset sampling rate, and sends the initial EEG signals to the processing device. The processing device segments the target EEG signal and its corresponding target time-frequency map to obtain a processed basic dataset. The target EEG signal is the preprocessed initial EEG signal. Based on preset different screening dimensions, candidate datasets with potential high-frequency oscillation characteristics are selected from the basic dataset. The candidate datasets are input into a preset two-level architecture deep learning model. A first-level structural sub-model is used to obtain the single-lead high-frequency oscillation probability of the candidate datasets. A second-level structural sub-model is used to obtain the window-level high-frequency oscillation probability of the candidate datasets. Based on the single-lead high-frequency oscillation probability and the window-level high-frequency oscillation probability, high-frequency oscillating discharge fragments and lead-level pathological high-frequency oscillation probabilities are obtained. The display device displays the high-frequency oscillating discharge fragments and lead-level pathological high-frequency oscillation probabilities, providing treatment guidance for medical personnel.

[0016] This application solution utilizes the combined use of acquisition, processing, and display equipment: fully automated processing significantly improves efficiency and eliminates human error; the primary structure employs time-domain and frequency-domain dual-branch feature extraction and gating fusion, while the secondary structure utilizes lead graph modeling and multi-head attention analysis to fully capture the multi-dimensional features and multi-lead synchronization characteristics of HFO, significantly reducing false positive / false negative rates. This provides quantitative evidence for epileptogenic zone localization and surgical planning, achieving the dual goals of cost reduction and efficiency improvement, and enhanced diagnostic and treatment quality.

[0017] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A schematic diagram of an invasive brain-computer interface device for detecting high-frequency oscillations of electroencephalogram (EEG) signals, provided in an embodiment of this application, is shown. Figure 2 A schematic diagram of a primary structure sub-model provided in an embodiment of this application is shown; Figure 3A schematic diagram of a secondary structure sub-model provided in an embodiment of this application is shown; Figure 4 A specific flowchart provided by an embodiment of this application is shown. Detailed Implementation

[0020] 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. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0021] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0022] It should be noted that the term "comprising" will be used in the embodiments of this application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0023] High-frequency oscillating foci (HFOs) are key biomarkers of abnormal electrical activity in the brain, especially closely related to the localization of epileptogenic zones. Pathological HFOs are often concentrated in epileptogenic zones, and their distribution and frequency directly guide the formulation of epilepsy surgical plans. Therefore, accurately identifying HFOs from invasive EEG signals is a core requirement for the clinical diagnosis and treatment of epilepsy and other neurological diseases.

[0024] Currently, invasive brain-computer interface (BCI) signal high-frequency oscillation detection often employs a "bandpass filtering + threshold discrimination" or "time-frequency analysis + manual feature" process: first, IIR / FIR filtering or continuous wavelet transform (CWT) is performed on frequencies above 80Hz; then, fixed or semi-automatic thresholds are set using statistical measures such as envelope mean, RMS, and line-length. Some studies introduce EMD, STFT, or simple convolutional networks to extract features from fixed-length multi-lead segments and classify them using SVM / RandomForest. To avoid excessive false alarms, manual review or the use of heuristic rules (such as event duration <250ms, spike exclusion) is often required to complete the final judgment.

[0025] The above-mentioned approach relies on manual threshold setting, a fixed number of leads, and is insensitive to changes in electrode layout, making it difficult to adapt to iEEG data where the number and position of leads can vary in clinical practice. It also lacks robustness against electromyography artifacts and stimulation artifacts, resulting in a high false positive rate and requiring extensive manual verification. Furthermore, it can only provide a binary conclusion on whether HFO is present at the fragment level, and cannot quantify the probability of discharge in each lead, thus limiting the precise localization of the epileptogenic zone.

[0026] Based on this, this application provides an invasive brain-computer interface (BCI) device for detecting high-frequency oscillations of EEG signals, which is described below through embodiments. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0027] Figure 1 The illustration shows a schematic diagram of an invasive brain-computer interface (BCI) high-frequency oscillation detection device provided in an embodiment of this application. The detection device in this embodiment includes: a data acquisition device, a processing device, and a display device. The data acquisition device is communicatively connected to the processing device, and the processing device is communicatively connected to the display device.

[0028] The acquisition equipment here is a professional invasive EEG acquisition device used to directly acquire intracranial EEG signals. Its core function is to acquire invasive brain-computer interface (iEEG) signals (initial EEG signals) according to clinical standards after the acquisition requirements are met, based on a preset sampling rate (which must meet the high sampling requirements of HFO detection, typically ≥2000Hz, with the mainstream being 2000Hz-5000Hz, such as 2000Hz, 2500Hz, 2700Hz, 3650Hz, 4587Hz, etc.). The acquisition requirements here are mainly determined by differences in brain region localization needs for HFO detection in clinical scenarios, individual patient conditions, and differences in the accuracy of the detection target. The structural characteristics and implantation methods of the sampling electrodes vary depending on the acquisition requirements. The sampling electrodes here can use SEEG, ECoG grid / strip signals, etc.

[0029] After acquiring the initial EEG signal, the acquisition device sends the signal to the processing device. This processing device is a device with data transmission and processing capabilities, such as a desktop computer or laptop. The processing device processes the initial EEG signal, obtains the processing result, and sends the result to the display device for presentation. The display device is a device with data receiving and display capabilities. This application does not limit the specific structure or model of the processing and display devices; as long as they possess the aforementioned functions, they are acceptable.

[0030] In an optional implementation, after receiving the initial EEG signal, the processing device, considering the interference data collected during data acquisition, needs to preprocess the initial EEG signal to obtain a preprocessed target EEG signal in order to ensure the accuracy of the detection results. Detection is then performed based on the target EEG signal.

[0031] In this embodiment, the preprocessing of the initial EEG signal mainly includes rereference, removal of power frequency noise, and baseline drift. Specifically: the initial EEG signal is rereferenced according to a preset reference point or reference method to obtain a processed first EEG signal; a first filter is used to eliminate power frequency noise in the first EEG signal to obtain a second EEG signal; wherein the first filter is a notch filter based on a first filtering frequency; a second filter is used to eliminate baseline drift in the second EEG signal to obtain the target EEG signal; wherein the second filter is a high-pass filter based on a second filtering frequency.

[0032] When performing rereference, considering the diversity of sampling electrodes, the reference point or reference method for rereference will differ. Therefore, the rereference here needs to use a reference point or reference method corresponding to the sampling electrode. For example, the SEEG signal can use bipolar reference, and the ECoG grid signal can use local reference, etc. After rereference of the initial EEG signal, a first EEG signal is obtained. Then, the first EEG signal is filtered twice: after the first filtering of the first EEG signal, a second EEG signal is obtained, and the second EEG signal is filtered again to obtain the target EEG signal. In order to be more targeted at power frequency noise and baseline drift, the embodiments of this application set different filters (first filter and second filter) in the above two filtering. The first filter is a notch filter based on a first filtering frequency, and the second filter is a high-pass filter based on a second filtering frequency.

[0033] Notch filters can deeply suppress only a preset first filtering frequency (e.g., 50Hz), while having almost no effect on other frequency signals; high-pass filters can be set to a second filtering frequency (e.g., 1Hz), attenuating all signals below that frequency (including baseline drift), while completely preserving signals above that frequency (especially the 80-500Hz band of HFO); notch filters are only for narrowband fixed frequencies and cannot handle broadband low-frequency interference such as baseline drift; while the low-frequency blocking characteristics of high-pass filters, combined with the high-pass filter, can cover the two most critical types of interference in EEG signals. Both the first and second filters can be selected with FIR filters.

[0034] For example, invasive EEG signals (initial EEG signals) with a sampling rate of 2000 Hz were acquired using SEEG electrodes (stereotactic EEG electrodes, minimally invasively inserted deep into the patient's brain) or ECoG strip electrodes (placed on the surface of the cerebral cortex), obtaining signals from 12 SEEG leads and 8 ECoG leads. The SEEG signals used a bipolar reference (the difference between signals from adjacent SEEG electrodes was considered the effective signal), and the ECoG strip signals used a local reference (the signal difference between the target electrode and the reference was calculated using the edge electrode of the ECoG strip as a reference). A FIR notch filter (50 Hz band) was used to filter the signals to eliminate power supply interference; a 1 Hz high-pass FIR filter (zero phase) was used to process the signals to avoid overall signal shift affecting subsequent analysis. Output: Clean iEEG signals (target EEG signals) from 12 SEEG leads and 8 ECoG leads, free from power supply interference and baseline drift.

[0035] After the processing device obtains the target EEG signal through the above preprocessing, this embodiment of the application also needs to obtain the target time-frequency map of the target EEG signal. The target EEG signal can be transformed using a preset wavelet transform method to obtain the corresponding time-frequency map. However, the time-frequency map after direct wavelet transform suffers from problems such as effective signal masking and redundant information interference. Therefore, this embodiment of the application uses the time-frequency map obtained by wavelet transform as the initial time-frequency map, and the time-frequency map after overcoming the problems of effective signal masking and redundant information interference in the initial time-frequency map is called the target time-frequency map. The wavelet transform here can be Morlet wavelet or CMOR wavelet, etc. The way to overcome the above problems is to enhance the effective signal features in the initial time-frequency map that are related to high-frequency oscillations, and suppress noise and redundancy unrelated to high-frequency oscillations in the initial time-frequency map. Specifically, the logarithm of the initial time-frequency map power can be calculated, and z-scores can be calculated band-by-band, Gaussian smoothing, etc.

[0036] To ensure the feasibility of HFO detection, this embodiment of the application, after obtaining the target EEG signal and the corresponding target time-frequency map, needs to segment the target EEG signal and the target time-frequency map to obtain multiple data units. Each data unit includes a waveform segment of a single lead and the corresponding time-frequency map of that waveform segment. All data units constitute the basic dataset. The segmentation can be performed using a sliding window method, sequentially sliding according to a preset time window. Considering that the data units need to include both the HFO and some background, the sliding window can be set to 120ms-200ms, and the step size can be set to 10-20ms to ensure coverage while controlling the computational load.

[0037] For example, for the target EEG signal (clean iEEG signal from 20 leads) mentioned above, Morlet wavelet transform is used to transform the clean signal of each lead to generate a time-frequency plot. Log power (highlighting high-frequency characteristics), band-by-band z-score (standardizing signals in each frequency band), and Gaussian smoothing (reducing noise) are calculated on the time-frequency plot. Using a window width of 150ms (ensuring inclusion of the HFO and some background signal) and a step size of 15ms (balancing coverage and computational load), the clean signal and time-frequency plot of each lead are segmented, resulting in 1000 waveform segments and 1000 corresponding time-frequency image segments. A basic dataset is constructed using a waveform segment and its corresponding time-frequency image segment as data units.

[0038] After obtaining the basic dataset, it is found that the dataset contains a large number of redundant data units unrelated to high-frequency oscillations (HFOs). Direct use of these data would lead to wasted computational resources and low efficiency, and may also cause noise data to interfere with inference, reducing the accuracy of HFO detection. Therefore, this embodiment of the application also requires screening the basic dataset to obtain candidate datasets with potential high-frequency oscillation characteristics.

[0039] To reduce errors when filtering the basic dataset, this embodiment sets multiple different filtering dimensions, and the processing device filters the basic dataset sequentially based on these different dimensions. These filtering dimensions include: band filtering, envelope threshold filtering, synchronization constraints, and spatial density filtering. The specific order of these filtering dimensions is not limited in this embodiment. For example, the order could be: band filtering, envelope threshold filtering, synchronization constraints, and spatial density filtering; or the order could be: band filtering, synchronization constraints, spatial density filtering, and envelope threshold filtering; or synchronization constraints, band filtering, envelope threshold filtering, and spatial density filtering; or spatial density filtering, synchronization constraints, band filtering, and envelope threshold filtering, etc.

[0040] Bandwidth filtering: High-frequency oscillations include ripple and fastripple. Design two FIR filters (a third filter based on a third filtering frequency and a fourth filter based on a fourth filtering frequency), filtering corresponding frequency bands: the ripple band is generally considered to be 80-250Hz, and the fastripple band is generally considered to be 250-500Hz. The filters are FIR filters, and the window function is chosen from Hamming windows or Blackman windows, etc., with appropriate order. The zero-phase filtering function `filtfilt` is used for filtering, where `fs` is the sampling rate.

[0041] Envelope Calculation and Threshold Statistics: The signal envelope is calculated using Hilbert transform, an envelope threshold is set, and time segments exceeding the threshold are filtered out. Threshold setting can employ more localized methods, such as robust-z, for windowed envelope signals of 1 second. The median m is calculated as follows: The MAD value d is: Then the normalized value z is: Then to Set thresholds for filtering. Constrain the duration of the detection window, such as merging segments with intervals less than 2ms and retaining segments with durations between 6-200ms.

[0042] Synchronization constraints: Pathological HFOs are often confined to local networks, but frequently occur synchronously between adjacent contacts of the same stereo electrode; single-channel bursts are more like hardware or electromyography artifacts. Therefore, for each time window, the number of leads in which candidate high-frequency oscillation segments appear within the time window is calculated. If the number of leads is less than a threshold (e.g., 3), all candidate segments within that time window are removed.

[0043] Spatial density screening: True HFOs will occur frequently in the same trigger zone; random artifacts are usually scattered, so leads with fewer high-frequency events are screened out. The number of candidate segments in each lead is counted over a period of time (e.g., 5-10 minutes). Set a threshold, such as Or the top 80 percentile of global leads, retain The number of leads was determined, and other candidate leads were removed. The average number of candidate segments across all leads. The standard deviation of the number of candidate segments across all leads.

[0044] For example, after filtering the 1000 waveform segments and time-frequency image segments mentioned above in the order of frequency band filtering, envelope thresholding, synchronization constraint, and spatial density filtering, a candidate dataset of 180 waveform segments and time-frequency image segments was obtained.

[0045] Subsequently, in order to accurately capture local features in single-lead signals while utilizing the physiological characteristics of multi-lead synchronization to eliminate isolated noise (global correlation), this embodiment inputs the candidate dataset into a pre-defined two-level architecture deep learning model for analysis, obtaining the high-frequency oscillating discharge fragments and lead-level pathological high-frequency oscillation probabilities output by the two-level architecture deep learning model. The two-level architecture deep learning model includes a first-level structural sub-model and a second-level structural sub-model. The first-level structural sub-model focuses on local feature extraction in a "single-lead-single-window" manner to solve the detail capture problem, while the second-level structural sub-model focuses on global correlation analysis in a "multi-lead-single-window" manner to solve the noise elimination problem. Through phased processing, the goals of preserving details, eliminating noise, and adapting to different scenarios are achieved. In specific implementation, the first-level structural sub-model includes a single-lead Local-CNN encoder, and the second-level structural sub-model includes a variable-lead Group-GNN determiner.

[0046] like Figure 2 As shown, for a single-lead Local-CNN encoder, it includes a temporal feature extraction component, a time-frequency feature extraction component, and a gated fusion component. The single-lead high-frequency oscillation probability is obtained in the following ways: the temporal feature extraction component is used to extract the first high-frequency oscillation feature of the waveform segment of a single lead in the candidate dataset in the time domain waveform; the time-frequency feature extraction component is used to extract the second high-frequency oscillation feature of the target time-frequency map corresponding to the same lead in the candidate dataset in the frequency domain dimension; the gated fusion component is used to filter and fuse the first high-frequency oscillation feature and the second high-frequency oscillation feature to obtain the target high-frequency oscillation feature.

[0047] The time-domain feature extraction component includes a Wave-branch model; the time-frequency feature extraction component includes a TF-branch component; and the gated fusion component includes a GLU fusion module.

[0048] The Wave-branch model takes waveform segments from a candidate dataset as input. Its main structure includes a one-dimensional convolutional block, residual blocks, pooling layers, a CBAM attention module, and a second fully connected layer. The one-dimensional convolutional block can consist of one-dimensional convolution, BatchNorm, and SiLU activation layers. One or two residual modules are connected to fuse pre-encoded and post-encoded features for further computation, reducing the risk of overfitting. The CBAM attention module first focuses on the feature map dimension C generated by the convolution, identifying which waveform features have the best diagnostic value, and then on the time dimension, focusing on time periods with higher HFO (Hard Forward-Looking) characteristics within the waveform segments. Finally, the second fully connected layer encodes the output of the CBAM attention layer into a specific dimension, such as 128 dimensions.

[0049] TF-branch: The input is the time-frequency plot from the candidate dataset. The main structure of the model includes: multiple Depthwise-Separable Convolutional Layers, Global Pooling, and a third fully connected layer. The Depthwise-Separable Convolutional Layer first performs independent 3×3 convolutions on each input channel, and then uses 1×1 Point-wise convolutions to mix the channels, which can reduce the number of training parameters; the global pooling layer merges and flattens the time dimension and frequency band dimension after convolution before pooling; finally, the third fully connected layer encodes the features into a specific dimension, such as 128 dimensions.

[0050] The GLU fusion module concatenates the outputs of the Wave-branch and TF-branch (first and second high-frequency oscillation features) to obtain the merged dimension feature (target high-frequency oscillation feature). Then, the target high-frequency oscillation feature is gated and fused using the following formula.

[0051]

[0052] in The sigmoid activation function maps the gate vector to the range of 0-1. The spliced ​​target high-frequency oscillation characteristics are represented by the superposition of the two-branch output dimensions. and The matrix is ​​trainable and has a dimension of . , and Let be the bias vector, with dimension . , Generate gate vectors, Generate content vectors. This is a gating signal.

[0053] like Figure 3 As shown, the Group-GNN determiner for variable leads includes a lead graph structure construction module, a multi-head attention graph convolution module, a context fusion module, and a first fully connected layer. The window-level high-frequency oscillation probability is obtained as follows: the construction module transforms the target high-frequency oscillation feature into a lead graph data structure, wherein the lead graph data structure includes node features and edge weights; the multi-head attention graph convolution module determines a global correlation feature vector based on the node features and determines node attention weights based on the edge weights; the context fusion module obtains a fused feature vector based on the global correlation feature vector and the target high-frequency oscillation feature; the first fully connected layer calculates the single-lead high-frequency oscillation probability corresponding to the target high-frequency oscillation feature and the window-level high-frequency oscillation probability corresponding to the fused feature vector. The context fusion module includes a Context-Gate fusion module.

[0054] The variable-lead Group-GNN determiner adds group constraints to Local-CNN, that is, it embeds the physical / functional relationships of Local-CNNs within the same time window into a window-level vector, and then outputs the probability of "whether HFO exists in this window". The model allows the number of nodes to vary with the surgical layout.

[0055] Construction module: Receives the target high-frequency oscillation features, where the number of vertices V is the number of leads in the candidate dataset, and the features of each vertex... Features after gating fusion in a single-lead Local-CNN encoder The construction of boundaries can be obtained through imaging, such as anatomical connectivity strength obtained by DTI, or functional connectivity strength obtained by fMRI; it can also be obtained through the physical distance of electrodes, such as the grid adjacency of ECoG or the spatial distance of SEEG electrodes; or it can be obtained through iEEG signals, such as the amplitude correlation of the envelope signals of two leads. Phase correlation Or a weighted fusion of them.

[0056] Multi-head attention graph convolutional network: This involves setting up multiple attention heads, with each attention head module calculating attention coefficients. and update in each iteration : Where k is the kth attention head. and For trainable matrices, and These are the feature vectors of nodes i and j, respectively. Let the edge weights be those between nodes i and j. These are trainable weight vectors. DropEdge and Normalize structures are added to prevent overfitting.

[0057] Then, the attention pooling module is connected to process the variable node set. Compressed into a fixed-dimensional global vector At the same time, key leads are highlighted.

[0058]

[0059] in, For single-lead feature mapping vectors, Node mapping is achieved using trainable matrices. V is the trainable "query" vector, and V is the node set.

[0060] Context-Gate fusion and classification head: To address situations where there may be fewer leads and the attention mechanism cannot function, a Context-Gate fusion module is added to make the output probability more stable.

[0061]

[0062] in This is a trainable matrix used to generate gate vectors.

[0063] Finally, a first fully connected layer and an activation layer are used. The probability value of the current segment being an HFO (high-frequency oscillation probability in a single lead) can be obtained through the first fully connected layer. And the probability value of HFO appearing in the i-th lead (probability of high-frequency oscillation in a single lead). The probability of whether the final fragment contains an HFO .

[0064] After obtaining the single-lead high-frequency oscillation probability and the window-level high-frequency oscillation probability, effective time windows are selected from the candidate time windows in the candidate dataset based on the window-level high-frequency oscillation probability and the preset high-frequency oscillation threshold; the effective time windows are merged according to the preset time interval to obtain the merged high-frequency oscillation discharge segment; the lead-level pathological high-frequency oscillation probability is calculated based on the single-lead high-frequency oscillation probability, the node attention weight, and the high-frequency oscillation discharge segment.

[0065] For the discharge segment: based on the output of the Group-GNN model, a threshold can be set for the probability to select... Within the effective time window exceeding the threshold, HFO discharge segments (high-frequency oscillating discharge segments) are spliced ​​together according to the preset time interval. The detection single-lead HFO detection segments within the HFO discharge segments can be displayed in conjunction with image reading software to assist in clinical image interpretation.

[0066] For discharge leads: the output of the Local-CNN model can be weighted and fused. The node attention weights of group-GNN are obtained, and one fusion method is as follows:

[0067]

[0068] in Let be the weight value of the contribution of the i-th lead to the HFO group determination within the current time window. For all leads The mean. Leads with higher readings are those with a higher probability of pathological HFO, which may indicate that the HFO is located in the epileptogenic zone.

[0069] The specific implementation process of this application is as follows: Figure 4 As shown, through the coordination of acquisition, processing, and display equipment, the entire process is automated, significantly improving efficiency and eliminating human error. The primary structure utilizes time-domain and frequency-domain dual-branch feature extraction and gating fusion, while the secondary structure employs lead graph modeling and multi-head attention analysis to fully capture the multi-dimensional features and multi-lead synchronization characteristics of HFO, significantly reducing the false detection / false detection rate. This provides a quantitative basis for epileptogenic zone localization and surgical planning, achieving the dual goals of cost reduction and efficiency improvement, as well as enhanced diagnostic and treatment quality.

[0070] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In addition, the terms "first", "second", "third", etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0071] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application; and these modifications, changes, 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. All should be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. An invasive brain-computer interface (BCI) device for detecting high-frequency oscillations of brain-computer interfaces (BCIs), characterized in that, The detection device includes: a data acquisition device, a processing device, and a display device; The acquisition device is used to acquire initial electroencephalogram (EEG) signals according to acquisition requirements and a preset sampling rate, and send the initial EEG signals to the processing device. The processing device is used to segment the target EEG signal and the corresponding target time-frequency map to obtain a processed basic dataset; wherein, the target EEG signal is the preprocessed initial EEG signal; Based on different preset screening dimensions, candidate datasets with potential high-frequency oscillation characteristics are selected from the basic dataset; The candidate dataset is input into a preset two-level architecture deep learning model. The first-level structure sub-model is used to obtain the single-lead high-frequency oscillation probability of the candidate dataset. The second-level structure sub-model is used to obtain the window-level high-frequency oscillation probability of the candidate dataset. Based on the single-lead high-frequency oscillation probability and the window-level high-frequency oscillation probability, the high-frequency oscillation discharge fragment and the lead-level pathological high-frequency oscillation probability are obtained. The display device is used to show the high-frequency oscillating discharge fragments and the probability of pathological high-frequency oscillations at the lead level, providing medical staff with a basis for treatment.

2. The detection device according to claim 1, characterized in that, The processing device preprocesses the initial electroencephalogram (EEG) signal in the following manner: The initial EEG signal is subjected to rereference processing according to a preset reference point or reference method to obtain the processed first EEG signal. The first filter is used to eliminate power frequency noise in the first EEG signal to obtain the second EEG signal; wherein, the first filter is a notch filter based on a first filtering frequency; The target EEG signal is obtained by using a second filter to eliminate baseline drift in the second EEG signal; wherein the second filter is a high-pass filter based on a second filtering frequency.

3. The detection device according to claim 1, characterized in that, The processing device is used to segment the target EEG signal and the corresponding target time-frequency map to obtain a processed basic dataset; wherein, the target EEG signal is the preprocessed initial EEG signal, including: The target EEG signal is processed using a preset wavelet transform method to obtain an initial time-frequency map; The effective signal features related to high-frequency oscillations in the initial time-frequency graph are enhanced, while noise and redundancy unrelated to high-frequency oscillations in the initial time-frequency graph are suppressed to obtain the target time-frequency graph. According to a preset time window, the target EEG signal and the target time-frequency map are slide-segmented to obtain multiple data units; wherein, each data unit includes a waveform segment of a single lead and the corresponding time-frequency map of the waveform segment, and all data units constitute the basic dataset.

4. The detection device according to claim 1, characterized in that, Different preset filtering dimensions include: frequency band filtering, envelope threshold filtering, synchronization constraint, and spatial density filtering; The processing device obtains the candidate dataset in the following manner: The basic dataset is processed sequentially using the preset sorting dimensions to obtain the candidate dataset.

5. The detection device according to claim 1, characterized in that, The primary structural sub-model includes a time-domain feature extraction component, a time-frequency feature extraction component, and a gated fusion component; the processing device obtains the single-lead high-frequency oscillation probability in the following manner: The time-domain feature extraction component is used to extract the first high-frequency oscillation feature of the waveform segment of a single lead in the candidate dataset in the time-domain waveform; The time-frequency feature extraction component is used to extract the second high-frequency oscillation feature in the frequency domain dimension of the target time-frequency map corresponding to the same lead in the candidate dataset. The gated fusion component is used to filter and fuse the first high-frequency oscillation feature and the second high-frequency oscillation feature to obtain the target high-frequency oscillation feature.

6. The detection device according to claim 5, characterized in that, The secondary structure sub-model includes a lead graph structure construction module, a multi-head attention graph convolution module, a context fusion module, and a first fully connected layer; The processing device obtains the window-level high-frequency oscillation probability in the following manner: The construction module is used to convert the target high-frequency oscillation characteristics into a lead graph data structure; wherein, the lead graph data structure includes node features and edge weights; The multi-head attention graph convolution module is used to determine the global association feature vector based on the node features and to determine the node attention weight based on the edge weights. The context fusion module is used to obtain a fused feature vector based on the global correlation feature vector and the target high-frequency oscillation feature; The first fully connected layer calculates the single-lead high-frequency oscillation probability corresponding to the target high-frequency oscillation feature and the window-level high-frequency oscillation probability corresponding to the fused feature vector.

7. The detection device according to claim 6, characterized in that, The processing device obtains the high-frequency oscillating discharge fragment and the lead-level pathological high-frequency oscillation probability in the following manner: Based on the window-level high-frequency oscillation probability and the preset high-frequency oscillation threshold, an effective time window is selected from the candidate time windows in the candidate dataset; The effective time windows are merged according to a preset time interval to obtain a merged high-frequency oscillation discharge segment. The probability of pathological high-frequency oscillation at the lead level is calculated based on the single-lead high-frequency oscillation probability, the node attention weight, and the high-frequency oscillation discharge segment.

8. The detection device according to claim 1, characterized in that, The first-level structural sub-model includes a single-lead Local-CNN encoder, and the second-level structural sub-model includes a variable-lead Group-GNN determiner.

9. The detection device according to claim 5, characterized in that, The time-domain feature extraction component includes a Wave-branch model; the time-frequency feature extraction component includes a TF-branch component; and the gated fusion component includes a GLU fusion module.

10. The detection device according to claim 6, characterized in that, The context fusion module includes the Context-Gate fusion module.

Images