Invasive brain-computer electroencephalogram high-frequency oscillation detection device

Abstract

The application provides an invasive brain-machine brain electro signal high-frequency oscillation detection device, which 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 realized, efficiency is greatly improved, and manual deviation is eliminated. First-level structure is extracted through time-domain-frequency-domain double-branch feature extraction and gate fusion, second-level structure is modeled through lead diagram modeling and multi-head attention analysis, multi-dimensional features and multi-lead synchronization characteristics of HFO are fully captured, false detection and missed detection rates are significantly reduced, quantitative basis is provided for epilepsy epileptogenic zone positioning and surgical plan, and the dual goals of technology cost reduction and efficiency increase and diagnosis and treatment quality improvement are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical treatment, in particular to an invasive brain-machine electroencephalogram high-frequency oscillation detection device. BACKGROUND

[0002] High-frequency oscillation (HFO) is a key marker of abnormal electrical activity in the brain, especially closely related to the localization of epileptogenic zone. Pathological HFO is often concentrated in the epileptogenic zone, and its distribution range and frequency directly guide the development of surgical plans for epilepsy, so accurate identification of HFO from invasive brain electroencephalogram signals is a core requirement for clinical diagnosis and treatment of neurological diseases such as epilepsy.

[0003] Traditional detection relies on manual labeling of HFO by medical staff, which requires analysis of massive electroencephalogram data segment by segment, consuming time and effort. Moreover, the labeling results are greatly influenced by the experience of personnel, and the consistency between different labelers is poor, which is prone to miss weak signal HFO or misjudge noise as HFO. SUMMARY

[0004] Therefore, the purpose of the present application is to provide an invasive brain-machine electroencephalogram high-frequency oscillation detection device to overcome the problems in the prior art.

[0005] In a first aspect, the embodiments of the present application provide an invasive brain-machine electroencephalogram high-frequency oscillation detection device, which comprises a collection device, a processing device and a display device.

[0006] The collection device is configured to collect an initial electroencephalogram signal according to a collection requirement and a preset sampling rate, and send the initial electroencephalogram signal to the processing device.

[0007] The processing device is configured to perform segmented processing on a target electroencephalogram signal and a corresponding target time-frequency map to obtain a processed basic data set. The target electroencephalogram signal is the initial electroencephalogram signal after preprocessing.

[0008] Based on different preset screening dimensions, a candidate data set with potential high-frequency oscillation characteristics is screened from the basic data set.

[0009] The candidate data set is input into a preset two-level architecture deep learning model, a single-lead high-frequency oscillation probability of the candidate data set is obtained using a first-level structure submodel, a window-level high-frequency oscillation probability of the candidate data set is obtained using a second-level structure submodel, and a high-frequency oscillation discharge segment and a lead-level pathological high-frequency oscillation probability are obtained according to the single-lead high-frequency oscillation probability and the window-level high-frequency oscillation probability.

[0010] The display device is configured to display the high-frequency oscillation discharge segment and the lead-level pathological high-frequency oscillation probability, thereby providing a treatment basis for medical staff.

[0011] In some technical solutions of the present application, the processing device preprocesses the initial electroencephalogram signal in the following manner:

[0012] According to a preset reference point or reference method, the initial electroencephalogram signal is re-referenced to obtain a first processed electroencephalogram signal;

[0013] A first filter is used to eliminate power frequency noise in the first electroencephalogram signal to obtain a second electroencephalogram signal; wherein the first filter is a notch filter based on a first filter frequency;

[0014] A second filter is used to eliminate baseline drift in the second electroencephalogram signal to obtain the target electroencephalogram signal; wherein the second filter is a high-pass filter based on a second filter frequency.

[0015] In some technical solutions of the present application, the processing device is used to segment the target electroencephalogram signal and the corresponding target time-frequency map to obtain a processed basic data set; wherein the target electroencephalogram signal is the preprocessed initial electroencephalogram signal, including:

[0016] A preset wavelet transform method is used to process the target electroencephalogram signal to obtain an initial time-frequency map;

[0017] The effective signal features in the initial time-frequency map that have a correlation with high-frequency oscillations are enhanced, while noise and redundancy in the initial time-frequency map that are irrelevant to high-frequency oscillations are suppressed to obtain the target time-frequency map;

[0018] According to a preset time window, the target electroencephalogram signal and the target time-frequency map are segmented to obtain a plurality of data units; wherein the data unit includes a single-lead waveform segment and a time-frequency map corresponding to the waveform segment, and all data units constitute the basic data set.

[0019] In some technical solutions of the present application, the preset different screening dimensions include frequency band filtering, envelope threshold screening, synchronicity constraint, and spatial density screening;

[0020] The processing device obtains the candidate data set in the following manner:

[0021] According to a preset order, the screening dimensions are used in sequence to process the basic data set to obtain the candidate data set.

[0022] In some technical solutions of the present application, the first-level structure sub-model includes a time-domain feature extraction component, a time-frequency feature extraction component, and a gating fusion component; the processing device obtains the single-lead high-frequency oscillation probability in the following manner:

[0023] The time domain feature extraction component is configured to extract a first high-frequency oscillation feature of a waveform segment of a single lead in a time domain waveform in the candidate data set;

[0024] The time-frequency feature extraction component is configured to extract a second high-frequency oscillation feature of a target time-frequency map corresponding to the same lead in the candidate data set in a frequency domain dimension;

[0025] The gating fusion component is configured to filter and fuse the first high-frequency oscillation feature and the second high-frequency oscillation feature to obtain a target high-frequency oscillation feature.

[0026] In some technical solutions of the present 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 full connection layer;

[0027] The processing device obtains the window-level high-frequency oscillation probability in the following manner:

[0028] The construction module is configured to convert the target high-frequency oscillation feature into a lead graph data structure; wherein the lead graph data structure includes node features and edge weights;

[0029] The multi-head attention graph convolution module is configured to determine a global correlation feature vector according to the node features, and determine a node attention weight according to the edge weights;

[0030] The context fusion module is configured to obtain a fusion feature vector according to the global correlation feature vector and the target high-frequency oscillation feature;

[0031] The first full connection layer calculates a single-lead high-frequency oscillation probability corresponding to the target high-frequency oscillation feature, and a window-level high-frequency oscillation probability corresponding to the fusion feature vector.

[0032] In some technical solutions of the present application, the processing device obtains a high-frequency oscillation discharge segment and a lead-level pathological high-frequency oscillation probability in the following manner:

[0033] According to the window-level high-frequency oscillation probability and a preset high-frequency oscillation threshold, an effective time window is screened from a candidate time window in the candidate data set;

[0034] The effective time window is merged according to a preset time interval to obtain a merged high-frequency oscillation discharge segment;

[0035] According to the single-lead high-frequency oscillation probability, the node attention weight, and the high-frequency oscillation discharge segment, a lead-level pathological high-frequency oscillation probability is calculated.

[0036] In some technical solutions of the application, the first-level structure sub-model comprises a single-lead Local-CNN encoder, and the second-level structure sub-model comprises a variable-lead Group-GNN determinator.

[0037] In some technical solutions of the application, the time-domain feature extraction component comprises a Wave-branch model, the time-frequency feature extraction component comprises a TF-branch component, and the gating fusion component comprises a GLU fusion module.

[0038] In some technical solutions of the application, the context fusion module comprises a Context-Gate fusion module.

[0039] The technical solutions provided by the embodiments of the application can have the following beneficial effects.

[0040] The detection device for high-frequency oscillation of the invasive brain-computer electroencephalogram in the application comprises a collection device, a processing device and a display device. The collection device is configured to collect an initial electroencephalogram signal according to a collection requirement and a preset sampling rate, and send the initial electroencephalogram signal to the processing device. The processing device is configured to perform segmentation processing on a target electroencephalogram signal and a corresponding target time-frequency map to obtain a processed basic data set. The target electroencephalogram signal is the initial electroencephalogram signal after preprocessing. A candidate data set having a potential high-frequency oscillation feature is screened out from the basic data set based on different preset screening dimensions. The candidate data set is input into a preset two-level architecture deep learning model, a single-lead high-frequency oscillation probability of the candidate data set is obtained by using a first-level structure sub-model, a window-level high-frequency oscillation probability of the candidate data set is obtained by using a second-level structure sub-model, and a high-frequency oscillation discharge segment and a lead-level pathological high-frequency oscillation probability are obtained according to the single-lead high-frequency oscillation probability and the window-level high-frequency oscillation probability. The display device is configured to display the high-frequency oscillation discharge segment and the lead-level pathological high-frequency oscillation probability, and provide a treatment basis for medical staff.

[0041] The application scheme realizes full-process automatic processing by cooperation of the collection device, the processing device and the display device, greatly improves efficiency and eliminates manual bias; the first-level structure realizes time-domain-frequency-domain double-branch feature extraction and gating fusion, the second-level structure realizes lead graph modeling and multi-head attention analysis, fully captures HFO multi-dimensional features and multi-lead synchronization characteristics, significantly reduces the false detection / missed detection rate, provides a quantitative basis for epilepsy epileptogenic zone positioning and surgical plan, and realizes the dual goals of technology cost reduction and efficiency improvement and diagnosis and treatment quality improvement.

[0042] In order to make the above-mentioned purposes, features and advantages of the application more obvious and easy to understand, the following preferred embodiments are described in detail below, and the accompanying drawings are described as follows. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some of the embodiments of the present application, and therefore should not be regarded as limiting the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor.

[0044] Figure 1 A schematic diagram of an invasive brain-machine brain electrical signal high-frequency oscillation detection device provided by an embodiment of the present application is shown.

[0045] Figure 2 A schematic diagram of a primary structure sub-model provided by an embodiment of the present application is shown.

[0046] Figure 3 A schematic diagram of a secondary structure sub-model provided by an embodiment of the present application is shown.

[0047] Figure 4 A specific flowchart provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. It should be understood that the drawings in the present application only play the purpose of illustration and description, and do not limit the protection scope of the present application. In addition, it should be understood that the schematic drawings are not drawn according to the actual proportion. The flowchart used in the present application shows the operations realized according to some embodiments of the present application. It should be understood that the operations of the flowchart can not be implemented in sequence, and the steps without logical context relationship can be reversed in sequence or implemented simultaneously. In addition, one or more other operations can be added to the flowchart or removed from the flowchart by those skilled in the art under the guidance of the content of the present application.

[0049] In addition, the described embodiments are only some of the embodiments of the present application, not all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0050] It should be noted that the term "comprising" will be used in the embodiments of the present application to specify the presence of stated features, but does not exclude the presence of other features.

[0051] High frequency oscillation (HFO) is a key marker of abnormal electrical activity in the brain, especially closely related to the localization of epileptogenic zone. Pathological HFOs are often concentrated in the epileptogenic zone, and their distribution range and frequency directly guide the development of surgical plans for epilepsy. Therefore, accurately identifying HFOs from invasive brain signals is a core requirement for the clinical diagnosis and treatment of neurological diseases such as epilepsy.

[0052] At present, the detection of high frequency oscillation in invasive brain signals mainly adopts the process of "band-pass filtering + threshold judgment" or "time-frequency analysis + artificial features": first, IIR / FIR filtering or continuous wavelet transform (CWT) is performed on the frequency band above 80 Hz, and then the envelope mean, RMS, Line-length and other statistical quantities are set to fixed or semi-automatic threshold. Some studies introduce EMD, STFT or simple convolution network to extract features from fixed-length multi-lead segments and then use SVM / RandomForest classification. In order to avoid too many false positives, manual review or combined with heuristic rules (such as event duration < 250ms, spike exclusion) is often needed to complete the final determination.

[0053] The above scheme relies on manual threshold setting, fixed number of leads and is not sensitive to changes in electrode layout, making it difficult to adapt to iEEG data in which both the number and position of leads can vary in the clinic. It is not robust enough to muscle artifact and stimulation artifact, resulting in high false positives and the need for a lot of manual review. At the same time, it can only give a binary conclusion of whether a segment contains HFOs, and cannot quantify the discharge probability in each lead, limiting the precise localization of the epileptogenic zone.

[0054] Based on this, the embodiments of the present application provide a detection device for high frequency oscillation of invasive brain signals. The following embodiments are described. In the case of no conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0055] Figure 1 A schematic diagram of a detection device for high frequency oscillation of invasive brain signals provided by the embodiments of the present application is shown, wherein the detection device in the embodiments of the present application comprises a collection device, a processing device and a display device, the collection device is in communication connection with the processing device, and the processing device is in communication connection with the display device.

[0056] The acquisition device here is a professional invasive electroencephalogram acquisition device for directly acquiring intracranial brain electrical signals. The core function is to acquire the acquisition requirement and collect the invasive brain-computer electroencephalogram signal (iEEG, initial electroencephalogram signal) according to the preset sampling rate (which needs to meet the high sampling requirement of HFO detection, usually ≥2000 Hz, the mainstream is 2000 Hz-5000 Hz, for example, 2000 Hz, 2500 Hz, 2700 Hz, 3650 Hz, 4587 Hz, etc.) according to the clinical standard. The acquisition requirement here is mainly the difference of brain region positioning requirement, patient individual disease difference and detection target precision difference of HFO detection in the clinical scene, etc. The structure characteristics of the required sampling electrode, the implantation method, etc. are different according to the different acquisition requirements. The sampling electrode here can use SEEG, ECoG grid / band signal, etc.

[0057] After the acquisition device acquires the initial electroencephalogram signal, the initial electroencephalogram signal is sent to the processing device. The processing device is a device with data transmission and data processing functions, which can be a desktop computer, a laptop computer, etc. After the processing device processes the initial electroencephalogram signal, the processing result is obtained and sent to the display device for display. The display device is a device with data receiving and display functions. The specific structure and model of the processing device and the display device are not limited in the present application, as long as they have the above functions.

[0058] In an optional embodiment, after the processing device receives the initial electroencephalogram signal, considering that interference data will be collected during data acquisition, in order to ensure the accuracy of the detection result, the present application embodiment needs to preprocess the initial electroencephalogram signal to obtain a target electroencephalogram signal after preprocessing. Detection is carried out based on the target electroencephalogram signal.

[0059] The preprocessing of the initial electroencephalogram signal in the present application embodiment mainly includes re-reference, removal of power frequency noise and baseline drift. Specifically: according to the preset reference point or reference method, the initial electroencephalogram signal is re-referenced to obtain a processed first electroencephalogram signal; a first filter is used to eliminate power frequency noise in the first electroencephalogram signal to obtain a second electroencephalogram signal; wherein the first filter is a notch filter based on a first filter frequency; a second filter is used to eliminate baseline drift in the second electroencephalogram signal to obtain the target electroencephalogram signal; wherein the second filter is a high-pass filter based on a second filter frequency.

[0060] In the process of re-referencing, the reference points or reference methods of re-referencing are different considering the diversity of sampling electrodes, so the re-referencing needs to use the reference points or reference methods corresponding to the acquisition electrodes. For example, the SEEG signal can adopt bipolar reference, the ECoG grid signal can adopt local reference, etc. After re-referencing the initial electroencephalogram signal, a first electroencephalogram signal is obtained, and then the first electroencephalogram signal is filtered twice. After the first electroencephalogram signal is filtered for the first time, a second electroencephalogram signal is obtained, and the second electroencephalogram signal is filtered again to obtain a target electroencephalogram signal. In order to be more targeted to the power frequency noise and baseline drift, different filters (a first filter and a second filter) are set in the above two filtering processes. 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.

[0061] The notch filter can only deeply suppress the preset first filtering frequency (such as 50Hz), and has almost no effect on other frequency signals; the high-pass filter can set the second filtering frequency (such as 1Hz), attenuate all signals (including baseline drift) below the frequency, and completely retain the signals (especially the 80-500Hz frequency band of HFO) above the frequency; the notch filter is only for narrowband fixed frequency and cannot process wideband low-frequency interference such as baseline drift; and the low-frequency blocking characteristic of the high-pass filter, the combination of the two can cover the two most core interferences in the electroencephalogram signal. The first filter and the second filter can both select the FIR filter.

[0062] For example, through SEEG electrodes (stereotactic electroencephalogram electrodes, minimally invasive insertion into the deep part of the patient's brain) or ECoG strip electrodes (placed on the surface of the cerebral cortex), invasive electroencephalogram signals (initial electroencephalogram signals) of a patient with a sampling rate of 2000Hz are collected, and a total of 12 SEEG leads and 8 ECoG lead signals are obtained. The SEEG signal adopts bipolar reference (the difference between adjacent SEEG electrode signals is the effective signal), and the ECoG strip signal adopts local reference (the edge electrode of the ECoG strip is used as a reference to calculate the signal difference between the target electrode and the reference). The signal is filtered by a FIR notch filter (50Hz frequency band) to eliminate power supply interference; the signal is processed by a 1Hz high-pass FIR filter (zero phase) to avoid the influence of overall signal offset on subsequent analysis. Output: 12 SEEG leads, 8 ECoG leads, clean iEEG signals (target electroencephalogram signals) without power frequency interference and baseline drift.

[0063] After the processing device obtains the target electroencephalogram signal through the above preprocessing, the embodiment of the present application also needs to obtain a target time-frequency graph of the target electroencephalogram signal. The target electroencephalogram signal can be transformed through a preset wavelet transform method, and a time-frequency graph corresponding to the target electroencephalogram signal can be obtained. However, the time-frequency graph obtained through direct wavelet transform may have problems of effective signals being covered and interference of redundant information. Therefore, the embodiment of the present application takes the time-frequency graph obtained through wavelet transform as an initial time-frequency graph, and the time-frequency graph after overcoming the problems of effective signals being covered and interference of redundant information in the initial time-frequency graph is called a target time-frequency graph. The wavelet transform here can use Morlet wavelet or CMOR wavelet, etc. The way to overcome the above problems is to enhance the effective signal characteristics in the initial time-frequency graph that have a correlation with high-frequency oscillation, and suppress the noise and redundancy in the initial time-frequency graph that have no correlation with high-frequency oscillation. Specifically, the log of the power of the initial time-frequency graph can be calculated, the z-score of each frequency band can be calculated, and Gaussian smoothing can be performed.

[0064] In order to ensure the feasibility of HFO detection, after obtaining the target electroencephalogram signal and the corresponding target time-frequency graph, the embodiment of the present application also needs to segment the target electroencephalogram signal and the target time-frequency graph to obtain a plurality of data units; wherein the data unit includes a single-lead waveform segment and a time-frequency graph corresponding to the waveform segment, and all the data units constitute the basic data set. The segmentation here can use a sliding window segmentation method, and the segmentation is performed in turn according to a preset time window. Considering that the data unit needs to ensure to contain HFO and part of the background, the sliding window here can be set to 120ms-200ms, and the step length can be set to 10-20ms, so as to ensure coverage and control the amount of calculation.

[0065] For example, for the above target electroencephalogram signal (20-lead clean iEEG signal), Morlet wavelet is used to transform each lead clean signal to generate a time-frequency graph; the log power of the time-frequency graph is calculated (to highlight the high-frequency feature), the z-score of each frequency band is calculated (to standardize the signals of each frequency band), and Gaussian smoothing is performed (to reduce noise); the clean signal of each lead and the time-frequency graph are segmented with a window width of 150ms (to ensure that HFO and part of the background signal are contained) and a step length of 15ms (to balance the coverage range and the amount of calculation), and a total of 1000 waveform segments and 1000 corresponding time-frequency graph segments are obtained. A waveform segment and a corresponding time-frequency graph segment data unit constitute the basic data set.

[0066] After obtaining the basic data set, since the basic data set contains a large number of redundant data units that have no correlation with high-frequency oscillation (HFO), direct use will cause waste of computing resources and low efficiency, and may also interfere with reasoning due to noise data, thereby reducing the HFO detection accuracy. Therefore, the embodiment of the present application also needs to screen the basic data set to obtain a candidate data set with potential high-frequency oscillation characteristics after screening.

[0067] 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.

[0068] 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.

[0069] 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.

[0070] 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.

[0071] Spatial density screening: true HFOs will appear in high frequency at the same trigger focus; random artifacts are usually scattered, so screen out the leads with less high frequency event occurrence. Count the number of candidate segments of each lead in a period of time (such as 5-10 min) , set a threshold, such as or the first 80 percentile of all lead candidates, retain number of leads, and remove the other lead candidates. Among them is the average of the number of all lead candidate segments, is the standard deviation of the number of all lead candidate segments.

[0072] For example, after the above 1000 waveform segments and time-frequency picture segments are screened in the order of frequency band filtering, envelope threshold screening, synchronicity constraint, and spatial density screening, 180 waveform segments and time-frequency picture segments are obtained as a candidate data set.

[0073] Then, in order to accurately capture local features in single-lead signals and exclude isolated noise (global correlation) by using the physiological characteristics of multi-lead synchronization, the embodiment of the application inputs the candidate data set into a pre-set two-level architecture deep learning model for analysis, to obtain high frequency oscillation discharge segments 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 structure sub-model and a second-level structure sub-model; the first-level structure sub-model focuses on local feature extraction of "single lead-single window" to solve the problem of detail capture, and the second-level structure sub-model focuses on global correlation analysis of "multi-lead-single window" to solve the problem of noise exclusion. Through phased processing, the goal of not losing details, excluding noise, and adapting to scenes is achieved. In specific implementation, the first-level structure sub-model includes a single-lead Local-CNN encoder, and the second-level structure sub-model includes a Group-GNN determinator with variable leads.

[0074] As shown in Figure 2 , for the single-lead Local-CNN encoder: it includes a time domain feature extraction component, a time-frequency feature extraction component, and a gating fusion component; the single-lead high frequency oscillation probability is obtained by the following way: the time domain feature extraction component is used to extract first high frequency oscillation features of a waveform segment of a single lead in the time domain waveform in the candidate data set; the time-frequency feature extraction component is used to extract second high frequency oscillation features of a target time-frequency map corresponding to the same lead in the frequency domain dimension in the candidate data set; and the gating fusion component is used to screen and fuse the first high frequency oscillation features and the second high frequency oscillation features to obtain target high frequency oscillation features.

[0075] 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.

[0076] Wave-branch model: the input is a waveform segment in the candidate data set, and the main structure of the model includes a one-dimensional convolution block, a residual block, a pooling layer, a CBAM attention module, and a second full connection layer. The one-dimensional convolution block structure can include one-dimensional convolution, BatchNorm, and SiLU activation layer; 1-2 residual modules are accessed to fuse the features before and after coding for next step calculation, thereby reducing the risk of overfitting; the CBAM attention module is first performed on the feature map dimension C generated by convolution, and attention is paid to which type of waveform feature has the best diagnostic value, and then the time dimension is performed, and attention is paid to the time period with more HFO features in the waveform segment; and finally, the second full connection layer encodes the CBAM attention layer output to a specific dimension, such as 128 dimensions.

[0077] TF-branch: the input is a time-frequency graph in the candidate data set, and the main structure of the model includes a plurality of Depthwise-Separable convolution layers, a global pooling layer, and a third full connection layer. The Depthwise-Separable Conv first performs independent 3x3 convolution on each input channel, and then uses 1x1 Point-wise convolution to mix channels, which can reduce the training parameter amount; the global pooling layer combines the time dimension and the frequency band dimension after convolution, and then performs pooling after flattening; and finally, the third full connection layer encodes the features to a specific dimension, such as 128 dimensions.

[0078] GLU fusion module: the outputs (first high-frequency oscillation feature and second high-frequency oscillation feature) of the Wave-branch and the TF-branch are spliced to obtain a merged dimension feature (target high-frequency oscillation feature). Then, the target high-frequency oscillation feature is gated and fused through the following formula.

[0079]

[0080] wherein is a Sigmoid activation function, which maps the gating vector to the range of 0-1, is the spliced target high-frequency oscillation feature, and the dimension is the superposition of the output dimensions of the two branches, and are trainable matrices, and the dimension is , and are bias vectors, and the dimension is , a gating vector is generated, a content vector is generated, The gating signal.

[0081] As shown in Figure 3 , for the variable lead Group-GNN determinator: including a construction module of lead graph structure, a multi-head attention graph convolution module, a context fusion module and a first full connection layer; the window-level high frequency oscillation probability is obtained in the following manner: the construction module is used for converting 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 is used for determining a global correlation feature vector according to the node features, and determining node attention weights according to the edge weights; the context fusion module is used for obtaining a fusion feature vector according to the global correlation feature vector and the target high frequency oscillation feature; the first full connection 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 fusion feature vector. The context fusion module includes a Context-Gate fusion module.

[0082] The variable lead Group-GNN determinator is based on the Local-CNN and increases the group constraint, that is, embedding the Local-CNN in the same time window into a window-level vector according to the physical / functional relationship, and then outputting the probability of “whether HFO exists in this window”. The model allows the number of nodes to change with the surgical layout.

[0083] Construction module: receiving target high frequency oscillation features, the number of vertices V is the number of leads in the candidate data set, the feature of each vertex is the feature of the single-lead Local-CNN encoder after gating fusion . The construction of the edge can be obtained by the anatomical connection strength obtained by image acquisition such as DTI, or the functional connection strength obtained by fMRI; or by the physical distance of the electrode, such as the grid adjacency of ECoG or the spatial distance of SEEG electrode; or by iEEG signal, such as the amplitude correlation , phase correlation or their weighted fusion of envelope signals of two leads.

[0084] Multi-head attention graph convolution network: set multiple attention heads, calculate attention coefficients of each attention head module , and update at each iteration:

[0085] where k is the kth attention head, and are trainable matrices, and are the feature vectors of nodes i and j, respectively, is the edge weight of node i and node j, is the trainable weight vector. DropEdge and Normalize structure are added to prevent overfitting.

[0086] Then, the access attention pooling module is used to compress the variable node set into a fixed-dimensional global vector while highlighting key leads.

[0087]

[0088] wherein, is a single-lead feature mapping vector, is a trainable matrix that realizes node mapping, is a trainable "query" vector, and V is a node set.

[0089] Context-Gate fusion and classification head: In order to deal with the situation that the lead may be less and the attention mechanism cannot work, the Context-Gate fusion module is added to make the output probability more stable.

[0090]

[0091] wherein is a trainable matrix used to generate a gate vector.

[0092] Finally, the first fully connected layer and the activation layer are used. Through the first fully connected layer, the probability value of the current segment being HFO (single-lead high frequency oscillation probability) and the probability value of the i-th lead appearing HFO (single-lead high frequency oscillation probability) can be obtained, and the probability of whether the final segment contains HFO .

[0093] After obtaining the single-lead high frequency oscillation probability and the window-level high frequency oscillation probability, according to the window-level high frequency oscillation probability and a preset high frequency oscillation threshold, an effective time window is screened out from the candidate time window in the candidate data set; the effective time window is merged according to a preset time interval, to obtain a merged high frequency oscillation discharge segment; and according to the single-lead high frequency oscillation probability, the node attention weight and the high frequency oscillation discharge segment, the lead-level pathological high frequency oscillation probability is calculated.

[0094] For the discharge segment: according to the Group-GNN model output, a threshold can be set for the probability, and 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.

[0095] 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:

[0096]

[0097]

[0098] 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.

[0099] 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.

[0100] 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.

[0101] Finally, it should be noted that the above-described embodiments are merely specific embodiments of the present application, which are used to illustrate the technical solutions of the present application, but not to limit the same. The protection scope of the present application is not limited thereto. Although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that any skilled person in the art can still modify or easily think of changes to the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some of the technical features, within the technical scope disclosed by the present application. The modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application. All should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An apparatus for detecting invasive brain-machine brain electrical signals high frequency oscillations, characterized in that, The detection device comprises a collection device, a processing device and a display device; The collection device is used for collecting an initial electroencephalogram signal according to a collection requirement and a preset sampling rate, and sending the initial electroencephalogram signal to the processing device; The processing device is used for segmenting a target electroencephalogram signal and a corresponding target time-frequency graph to obtain a processed basic data set; Based on four screening dimensions of frequency band filtering, envelope threshold screening, synchronism constraint and spatial density screening, a candidate data set with potential high-frequency oscillation characteristics is screened from the basic data set; The candidate data set is input into a preset two-level architecture deep learning model, a single-lead high-frequency oscillation probability of the candidate data set is obtained by using a first-level structure submodel, a window-level high-frequency oscillation probability of the candidate data set is obtained by using a second-level structure submodel, and a high-frequency oscillation discharge segment and a lead-level pathological high-frequency oscillation probability are obtained according to the single-lead high-frequency oscillation probability and the window-level high-frequency oscillation probability; wherein the target electroencephalogram signal is a preprocessed initial electroencephalogram signal; the first-level structure submodel comprises a time domain feature extraction component for extracting a first high-frequency oscillation feature of a waveform segment of a single lead in the candidate data set in a time domain waveform, a time-frequency feature extraction component for extracting a second high-frequency oscillation feature of a target time-frequency graph corresponding to the same lead in the candidate data set in a frequency domain dimension, and a gate fusion component for screening and fusing the first high-frequency oscillation feature and the second high-frequency oscillation feature to obtain a target high-frequency oscillation feature; the second-level structure submodel comprises a lead graph structure construction module for converting the target high-frequency oscillation feature into a lead graph data structure, a multi-head attention graph convolution module for determining a global correlation feature vector according to a node feature and determining a node attention weight according to an edge weight, a context fusion module for obtaining a fusion feature vector according to the global correlation feature vector and the target high-frequency oscillation feature, and a first full connection layer for calculating a single-lead high-frequency oscillation probability corresponding to the target high-frequency oscillation feature and a window-level high-frequency oscillation probability corresponding to the fusion feature vector; wherein the lead graph data structure comprises a node feature and an edge weight; The display device is used for displaying the high-frequency oscillation discharge segment and the lead-level pathological high-frequency oscillation probability to provide a treatment basis for medical staff.

2. The detection device of claim 1, wherein, The processing device preprocesses the initial electroencephalogram signal in the following manner: The initial electroencephalogram signal is re-referenced to obtain a first electroencephalogram signal; A first filter is used to eliminate power frequency noise in the first electroencephalogram signal to obtain a second electroencephalogram signal; wherein the first filter is a notch filter based on a first filter frequency; A second filter is used to eliminate baseline drift in the second electroencephalogram signal to obtain the target electroencephalogram signal; wherein the second filter is a high-pass filter based on a second filter frequency.

3. The detection device of claim 1, wherein, The processing device is used for segmenting a target electroencephalogram signal and a corresponding target time-frequency graph to obtain a processed basic data set; comprising: The target electroencephalogram signal is processed by using a preset wavelet transform method to obtain an initial time-frequency graph; The effective signal features associated with the high-frequency oscillation in the initial time-frequency map are enhanced, while noise and redundancy unrelated to the high-frequency oscillation in the initial time-frequency map are suppressed, to obtain the target time-frequency map; According to a preset time window, the target electroencephalogram signal and the target time-frequency map are slidingly segmented to obtain a plurality of data units; wherein the data unit includes a waveform segment of a single lead and a time-frequency map corresponding to the waveform segment, and all data units constitute the basic data set.

4. The detection device according to claim 1, characterized in that, The processing device obtains the high-frequency oscillation discharge segment and the lead-level pathological high-frequency oscillation probability by: According to the window-level high-frequency oscillation probability and a preset high-frequency oscillation threshold, an effective time window is screened from the candidate time window in the candidate data set; According to a preset time interval, the effective time window is merged to obtain a merged high-frequency oscillation discharge segment; According to the single-lead high-frequency oscillation probability, the node attention weight and the high-frequency oscillation discharge segment, the lead-level pathological high-frequency oscillation probability is calculated.

5. The detection device of claim 1, wherein, The primary structure sub-model includes a single-lead Local-CNN encoder, and the secondary structure sub-model includes a Group-GNN determinator of variable leads.

6. The detection device of claim 1, wherein, 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.

7. The detection device of claim 1, wherein, The context fusion module includes a Context-Gate fusion module.

Images