A spindle wave detection method, device and system combined with frequency attention mechanism

Abstract

The present application relates to a kind of spindle wave detection method, device and system combined with frequency attention mechanism, wherein spindle wave detection method includes obtaining partial electroencephalogram, and partial electroencephalogram is initialized, obtain a default event group;Feature extraction is carried out to default event group, and feature tensor is obtained;Feature tensor is convoluted, and first latent event group is obtained, and each latent event in first latent event group includes its corresponding center position and duration;According to the center position and duration of each latent event in first latent event group, original electroencephalogram is positioned and frequency attention processing is carried out, and a set of latent event one-to-one corresponding confidence array in first latent event group is obtained;And according to first latent event group, probability vector is calculated;According to probability vector and corresponding confidence value, the classification of latent event is completed.The detection method can more accurately capture the event of frequency falling in the range, improve its identification and prediction event.

Description

Technical Field

[0001] This invention relates to the field of spindle wave detection, and in particular to a spindle wave detection method and apparatus that incorporates a frequency attention mechanism. Background Technology

[0002] Sleep characteristic waves are an important component of sleep electroencephalography (EEG) and are significant markers for distinguishing different sleep stages. Spindle waves are an important member of sleep characteristic waves.

[0003] Sleep spindles are brief neural oscillations (11-16 Hz, 0.5-3 seconds) generated by the interaction between the N2 nucleus and thalamic nuclei during sleep. They are associated with memory consolidation and information processing during sleep. Some medical studies have shown that the presence of sleep spindles is of significant predictive value for the prognosis of patients with acute disorders of consciousness (ADOC). This is because sleep spindles appear to be related to cortical plasticity, and their generation depends on the functional and structural integrity of the thalamo-reticular ascending activating system or thalamic-cortical circuits that maintain consciousness. The presence of spindles indicates that the patient has a high level of residual consciousness.

[0004] Currently, the standard for spindle wave detection mainly relies on visual recognition and labeling by trained experts. However, this process is time-consuming, and the subjective judgment standards among different experts can affect the accuracy of the results. Therefore, automatic sleep spindle wave detection technology plays an increasingly important role in spindle wave detection due to its advantages of high efficiency, accuracy, and objectivity. Existing automatic spindle wave detection technologies typically process a segment of the EEG signal, extract the corresponding features of this segment, and predict the potential EEG events matched by this segment based on the extracted features, thereby completing the automatic detection of sleep characteristic waves. In recent years, several automatic spindle wave detection methods based on mathematical calculations and machine learning have been proposed. L Ray et al. proposed a method using polyphony detection to adjust the threshold and extract the sigma phase change signal associated with the spindle wave, achieving a detection accuracy of 73%. Fazil Duman et al. proposed an efficient sleep spindle wave detection algorithm based on decision trees, which combines a multi-signal classification algorithm and the Teager energy operator algorithm, achieving a sensitivity of 96.17%. Patti et al. proposed a method for sleep spindle wave detection using a random forest classifier and a Gaussian mixture model (GMM), which can reduce the probability of overfitting during model training. This method achieved a detection sensitivity of 71.2%. However, these methods still have some limitations, such as relying on defined parameters and pre-prepared manually labeled data, making them unsuitable for spindle wave detection tasks in patients with acute disturbances of consciousness, exhibiting poor generalization ability, and resulting in inaccurate predictions. Summary of the Invention

[0005] Therefore, the purpose of this invention is to provide a spindle wave detection method incorporating a frequency attention mechanism to solve the above-mentioned problems. The method specifically includes:

[0006] Acquire a portion of the EEG signals and initialize the portion of the EEG signals to obtain a default event group;

[0007] Feature extraction is performed on the default event group to obtain a feature tensor;

[0008] The feature tensor is convolved to obtain a first potential event group, where each potential event in the first potential event group includes its corresponding center position and duration.

[0009] Based on the center location and duration of each potential event in the first potential event group, the original EEG signal is located and frequency attention processing is performed to obtain a set of confidence arrays that correspond one-to-one with the potential events in the first potential event group; and a probability vector is calculated based on the first potential event group; and each potential event is classified based on the probability vector and the corresponding confidence value.

[0010] Compared to existing technologies, this invention introduces a frequency attention mechanism to reverse-localize the initial results detected by the deep learning model, thereby influencing the generation and adjustment of its final probability vector. This enables the model to more accurately capture events whose frequencies fall within the target range, improving its ability to identify and predict these events.

[0011] Furthermore, after obtaining the first potential event group, it specifically includes:

[0012] Based on the center position and duration of each potential event in the first potential event group, the original EEG signal represented by the potential event is located in reverse, and frequency attention processing is performed on the original EEG signal to obtain a set of confidence arrays that correspond one-to-one with the potential events in the first potential event group.

[0013] Determine whether the corresponding confidence value of the potential event in the first potential event group meets the preset condition. If the corresponding confidence value of the potential event in the first potential event group meets the preset condition, then the potential event is directly identified as a spindle wave. If the corresponding confidence value of the potential event in the first potential event group does not meet the preset condition, then proceed to the next step.

[0014] Convolve the potential events in the first potential event group to obtain the probability vector of the potential event, and output the classification result of the potential event based on the probability vector.

[0015] This invention reduces the convolutional classification of potential events by performing prior checks on frequency attention confidence processing, thereby saving system resources and improving detection efficiency.

[0016] Furthermore, after obtaining the first potential event group, it specifically includes:

[0017] Convolve the potential events in the first potential event group to obtain the probability vector of the potential event, and output the classification result of the potential event based on the probability vector;

[0018] Determine whether the classification result is a spindle wave. If the classification result is a spindle wave, output the classification result directly; if the classification result is not a spindle wave, proceed to the next step.

[0019] Based on the center location and duration of each potential event in the first potential event group, the original EEG signal represented by the potential event is located in reverse. Frequency attention processing is performed on the original EEG signal to obtain the confidence value of the original EEG signal. The classification result of the potential event is output based on the confidence value of the original EEG signal.

[0020] This invention reduces frequency attention processing for potential events by prior judgment of model classification results, saving system resources and improving detection efficiency.

[0021] Furthermore, the frequency attention processing includes:

[0022] Wavelet transform was performed on the original EEG signal to obtain wavelet coefficients;

[0023] Calculate the amplitude spectrum of the wavelet coefficients, calculate the dominant frequency based on the amplitude spectrum, and obtain the confidence value based on the magnitude of the dominant frequency.

[0024] By processing the raw EEG signals of the potential event, the problem of information loss in the raw signals caused by a series of convolution operations is avoided. The classification results of the classification module are further verified by the main frequency of the raw data to see if the classification is missed because the frequency exceeds the preset range.

[0025] Furthermore, the confidence value is calculated using the following formula:

[0026] T(r)=J(F(wt(C(ri.center_time,riduration))),frequency_range

[0027] Where T represents the confidence result of the event (with a value of 0 or 1); J represents determining whether the dominant frequency falls within the expected frequency range; F represents obtaining the dominant frequency in the EEG (electroencephalogram) segment; Wt represents wavelet transform; and C represents extracting the EEG segment represented by the two parameters .center_time.

[0028] On the other hand, the present invention also provides a spindle wave detection device incorporating a frequency attention mechanism, comprising:

[0029] Initialization module: used to acquire partial EEG signals and initialize the partial EEG signals to obtain a default event group;

[0030] Feature extraction module: used to extract features from the default event group to obtain a feature tensor;

[0031] The localization module is used to perform convolution processing on the feature tensor to obtain a first potential event group, wherein each potential event in the first potential event group includes its corresponding center position and duration.

[0032] Frequency attention classification module: It is used to locate the original EEG signal and perform frequency attention processing based on the center position and duration of each potential event in the first potential event group to obtain a set of confidence arrays that correspond one-to-one with the potential events in the first potential event group; and calculate the probability vector based on the first potential event group; and complete the classification of each potential event based on the probability vector and the corresponding confidence value.

[0033] Furthermore, the frequency attention classification module includes:

[0034] Frequency attention submodule: used to reverse locate the original EEG signal represented by each potential event in the first potential event group based on the center position and duration of each potential event, and perform frequency attention processing on the original EEG signal to obtain a set of confidence arrays that correspond one-to-one with the potential events in the first potential event group;

[0035] First Judgment Submodule: Used to determine if the corresponding confidence value of the potential event in the first potential event group meets the preset condition. If the corresponding confidence value of the potential event in the first potential event group meets the preset condition, the potential event is directly output as a spindle wave. If the corresponding confidence value of the potential event in the first potential event group does not meet the preset condition, the classification submodule is triggered.

[0036] The classification submodule is used to perform convolution on the potential events in the first potential event group to obtain the probability vector of the potential event, and output the classification result of the potential event based on the probability vector.

[0037] Furthermore, the frequency attention classification module includes:

[0038] The classification submodule is used to perform convolution on each potential event in the first potential event group to obtain the probability vector of the potential event, and to obtain the classification result of the potential event based on the probability vector.

[0039] The second judgment submodule is used to determine whether the classification result is a spindle wave. If the classification result is a spindle wave, the classification result is output directly. If the classification result is not a spindle wave, the frequency attention submodule is triggered.

[0040] Frequency attention submodule: It is used to reverse locate the original EEG signal represented by each potential event in the first potential event group based on the center position and duration of each potential event, perform frequency attention processing on the original EEG signal to obtain the confidence value of the original EEG signal, and output the classification result of the potential event based on the confidence value of the original EEG signal.

[0041] Furthermore, the frequency attention submodule includes:

[0042] Wavelet transform processing layer: used to perform wavelet transform on the original EEG signal to obtain wavelet coefficients;

[0043] Confidence value calculation processing layer: Calculates the amplitude spectrum of wavelet coefficients, calculates the dominant frequency based on the amplitude spectrum, and obtains the confidence value based on the magnitude of the dominant frequency.

[0044] On the other hand, the present invention also provides a spindle wave detection system incorporating a frequency attention mechanism, which includes a brain-computer interface and a spindle wave detection device incorporating a frequency attention mechanism as described in any of the above claims.

[0045] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0046] Figure 1 A structural block diagram of a spindle wave detection device incorporating a frequency attention mechanism is provided for this invention.

[0047] Figure 2 A flowchart of the spindle wave detection method performed by the spindle wave detection device;

[0048] Figure 3 The flowchart shows an improved spindle wave detection method incorporating a frequency attention mechanism according to the present invention.

[0049] Figure 4 This is a flowchart of another improved spindle wave detection method incorporating a frequency attention mechanism, as described in this invention. Detailed Implementation

[0050] This invention, through research, has discovered that the frequency of spindle waves in patients with acute consciousness disorder may decrease from 11-16 Hz to 8-15 Hz compared to normal sleep spindle waves. However, the preset frequency range of standard spindle waves in existing sleep spindle detection models is usually pre-set as a model hyperparameter during training. Since the spindle wave frequency in patients with acute consciousness disorder deviates from that of normal individuals, whether the frequency range (usually used as a model hyperparameter) remains constant or is increased, it will reduce the performance of the method in the patient spindle wave detection task. Keeping it constant will miss the 8-11 Hz spindle wave signal, while expanding the detection range will lead to misjudgments when detecting normal sleep spindle waves. Based on this, this invention proposes introducing a frequency attention mechanism to back-locate the preliminary results detected by the deep learning model. Simultaneously, wavelet transform and energy density analysis are performed on the segment containing this result to verify whether the frequency of the result matches expectations, thereby affecting the generation and adjustment of its final probability vector. This enables the model to more accurately capture events with frequencies falling within the target range, improving its ability to identify and predict these events.

[0051] This invention is based on an experimental spindle wave detection system that incorporates a frequency attention mechanism. It includes a brain-computer interface (BCI) device and a spindle wave detection device connected to the BCI device. The BCI device is used to acquire and transmit electroencephalogram (EEG) signals, and the spindle wave detection device performs analysis of the EEG signals and filters out spindle wave signals.

[0052] Please see Figure 1 and Figure 2 ,in Figure 1 This is a block diagram of the spindle wave detection device that incorporates the frequency attention mechanism described above. Figure 2 A flowchart of the spindle wave detection method performed by the spindle wave detection device. The spindle wave detection method incorporating a frequency attention mechanism specifically includes the following steps:

[0053] Initialization module 10 is used to execute step S10: acquire partial EEG signals and initialize the partial EEG signals to obtain a set of default events;

[0054] In spindle wave detection, "event" is a general term. A spindle wave can be an event, a K-complex wave can be an event, and awakening can also be an event.

[0055] Since the duration of a spindle wave is typically 0.5-3 seconds, this invention performs two segments on a long EEG signal. First, a 30-second segment of the EEG signal is extracted. Then, within that 30-second segment, it is divided into 5-second intervals, with each 5-second segment serving as a default event. The center time of that segment and a default duration (e.g., 2.5 seconds) provided by the system are used as the initialization for that segment. Therefore, a set of default events can be obtained from a portion of the EEG signal through parameterization.

[0056] The feature extraction module 20 is used to perform S20: extract features from the default event group to obtain a feature tensor.

[0057] In this invention, feature extraction includes two steps: spatial filtering and temporal feature extraction, specifically including:

[0058] Spatial filtering is performed on the default event group to obtain the first tensor; this step is performed by a spatial subprocessing module.

[0059] In spindle wave detection, spatial filtering is crucial because it enhances the signal-to-noise ratio, highlights the spatial patterns of characteristic waves, and improves the overall sensitivity and specificity of the detection process. Therefore, we designed a spatial sub-processing module. This module is constructed using a two-dimensional convolutional layer consisting of c convolutional kernels of size (c,1) (c∈N), followed by a transpose operation. It performs c linear combinations on the c input time series, ultimately generating the first tensor.

[0060] This module synthesizes the results of filtering signals from multiple channels into a tensor, utilizing the signal information from multiple channels to improve the detection accuracy to a certain extent.

[0061] Temporal feature extraction is performed on the first tensor to obtain the feature tensor; this step is executed by a temporal processing submodule.

[0062] This temporal subprocessing module consists of several blocks. Each block comprises a two-dimensional convolutional layer with batch normalization, ReLU activation, and temporal max pooling. For each block, a series of operations are performed.

[0063] First, use a 4×2 matrix with size (1,3) and stride of 1. k Each kernel performs an inverse convolution on the first tensor. This convolution extracts spatial and temporal features with zero padding while keeping the tensor dimension unchanged.

[0064] Next, the ReLU activation function is used to introduce nonlinearity, thereby enhancing the network's expressive power.

[0065] Finally, a time-based max-pooling operation with a kernel size of (1,2) and a stride of 2 is applied to halve the time dimension.

[0066] It is important to note that within a block, since the convolution operation only focuses on the temporal dimension, the spatial dimension remains unchanged. The final output of the temporal sub-processing module is a feature tensor of shape (f, c, t), where f = 4 × 2k and t = t / 2k.

[0067] Extracting the temporal features of spindle waves can determine their occurrence time during sleep and help assess their duration. This extraction provides crucial information about the temporal characteristics of spindle waves, aiding in the analysis of their dynamic properties during detection.

[0068] The positioning module 30 is used to perform step S30: convolving the feature tensor to obtain a first potential event group, wherein each potential event in the first potential event group includes its corresponding center position and duration.

[0069] The feature tensor is an array representation of the feature map after spatial and temporal filtering of the original EEG signal map. It contains some information stored in the EEG signal. This invention uses a convolutional layer with 2×Kn kernels of size (c,t / 2k-1) to extract information. This layer uses a linear activation function to obtain information such as the center position and duration of each potential event in the feature tensor.

[0070] The frequency attention classification module 40 is used to perform step S40: based on the center position and duration of each potential event in the first potential event group, locate the original EEG signal and perform frequency attention processing to obtain a set of confidence arrays that correspond one-to-one with the potential events in the first potential event group; calculate the probability vector based on the first potential event group; and complete the classification of each potential event based on the probability vector and the corresponding confidence value.

[0071] The localization module generates the center location and duration of potential events. The frequency attention module then performs reverse localization on these potential events, extracting the EEG signal represented by the center location and duration. A wavelet transform processing layer performs a wavelet transform on the EEG signal to obtain wavelet coefficients. A confidence calculation processing layer calculates the amplitude spectrum of the wavelet coefficients, finding the largest peak in the amplitude spectrum as the dominant frequency of the potential event, and calculating the confidence level based on the dominant frequency. If the dominant frequency of the potential event falls within the expected frequency range, the confidence level of the potential event is set to 1; otherwise, it is set to 0. All confidence values ​​of the first group of potential events are combined into a confidence array. The specific formula for calculating the confidence value obtained by the frequency attention processing is as follows:

[0072] T(r i )=J(F(wt(C(r i .center_time,r iduration), frequency_range)

[0073] Where T represents the confidence result of the event (value is 0 / 1); J represents determining whether the dominant frequency falls within the expected frequency range; F represents obtaining the dominant frequency in the EEG segment; Wt represents wavelet transform; and C represents truncating the EEG segment represented by the two parameters .center_time.

[0074] By processing the raw EEG signals of the potential event, the problem of information loss in the raw signals caused by a series of convolution operations is avoided. The classification results of the classification module are further verified by the main frequency of the raw data to see if the classification is missed because the frequency exceeds the preset range.

[0075] The classification part uses a dimension of (c, t / 2) k-1 The system uses a 2D convolutional layer with (L+1)×Kn kernels to process the center position and duration of each event. A softmax activation function is applied to generate a probability vector representing each potential event. This vector is then combined with the confidence value for each potential event. If the confidence value is 1, the potential event is considered a spindle wave, and the "spindle wave probability" in the probability vector is set to 1, with the output being the final result. If the confidence value is 0, the maximum value in the classification probability vector is used as the predicted event. For example, "(0.2, 0.4, 0.3) represents the probability of the event being a spindle wave (0.2), a K-composite wave (0.4), and other events (0.3)," and the highest probability of 0.4 is selected as the final classification result.

[0076] In the above embodiments, frequency attention processing is performed before classifying each potential event in the first potential event group, i.e., a pre-confidence assessment is performed on the potential events. However, in practice, existing models are only accurate in predicting spindle wave frequencies in patients with acute consciousness disorders when they drop to the 8-11 Hz range, but are relatively accurate in predicting frequencies in the 11-16 Hz range. Therefore, the pre-confidence assessment for frequencies falling between 11-16 Hz is somewhat wasteful of resources. Therefore, the inventors have improved the above-mentioned spindle wave detection method incorporating a frequency attention mechanism, resulting in the following method, which is described in detail below. Figure 3 , Figure 3 This is a flowchart of an improved spindle wave detection method incorporating a frequency attention mechanism according to the present invention. The improved step S40 is executed by the improved frequency attention classification module, which includes a frequency attention submodule, a first judgment submodule, and a classification submodule. Specifically, the improved step S40 includes:

[0077] The frequency attention submodule is used to perform step S41.A: based on the center position and duration of each potential event in the first potential event group, the original EEG signal represented by the potential event is located in reverse, and frequency attention processing is performed on the original EEG signal to obtain a set of confidence arrays that correspond one-to-one with the potential events in the first potential event group.

[0078] The first judgment submodule is used to execute step S42.A: determine whether the corresponding confidence value of the potential event in the first potential event group meets the preset condition. If the corresponding confidence value of the potential event in the first potential event group meets the preset condition, then the potential event is directly identified as a spindle wave. If the corresponding confidence value of the potential event in the first potential event group does not meet the preset condition, then step S43.A is executed.

[0079] The classification submodule is used to execute step S43.A: convolve the potential events in the first potential event group to obtain the probability vector of the potential event, and output the classification result of the potential event according to the probability vector; step S43.A is executed by the classification submodule.

[0080] If the potential event is a spindle wave, then based on the center position and duration of each potential event in the first potential event group, the original EEG signal represented by the potential event is located in reverse, and frequency attention processing is performed on the original EEG signal to obtain the confidence value of the EEG signal; if the confidence value of the EEG signal meets the preset conditions, then the potential event is a spindle wave.

[0081] Or, please see [link to relevant information] Figure 4 , Figure 4 The flowchart below shows another improved spindle wave detection method incorporating a frequency attention mechanism according to the present invention. The improved spindle wave detection method based on the frequency attention mechanism of the present invention is further improved as follows: step S40 is executed by the improved frequency attention classification module, which includes a classification submodule, a first judgment submodule, and a frequency attention submodule. Specifically, step S40 includes:

[0082] The classification submodule is used to perform step S41.B: convolve the potential events in the first potential event group to obtain the probability vector of the potential event, and output the classification result of the potential event based on the probability vector.

[0083] The second judgment submodule is used to execute step S42.B: determine whether the classification result is a spindle wave. If the classification result is a spindle wave, the classification result is directly output. If the classification result is not a spindle wave, the frequency attention submodule is triggered.

[0084] The frequency attention submodule is used to perform step S43.B: based on the center position and duration of each potential event in the first potential event group, reverse locate the original EEG signal represented by the potential event, perform frequency attention processing on the original EEG signal to obtain the confidence value of the original EEG signal, and output the classification result of the potential event based on the confidence value of the original EEG signal.

[0085] Any of the above improvements—such as first classifying potential events, and then performing reverse localization and frequency attention processing on the potential event when the classification result is not a spindle wave, calculating the dominant frequency of the corresponding EEG signal and calculating the confidence value; or first performing frequency attention processing on the potential event, calculating the confidence value, and then classifying the potential event through convolution when the confidence value does not meet the confidence condition—can both achieve the effect of avoiding missed predictions of spindle wave events due to frequency decrease, and saving time wasted on pre-confidence calculations for each potential event. Ultimately, without affecting the accuracy of the detection method, it saves system resources and improves detection efficiency.

[0086] This invention addresses the issue of insufficient accuracy in existing spindle wave detection methods. It introduces a frequency attention mechanism to perform simultaneous confidence processing on classification predictions, correcting missed classification results from the classification submodule. This solves the difficulty of detecting spindle waves in patients with acute impairment of consciousness without affecting the detection accuracy of existing models. Furthermore, the frequency attention classification module is structurally optimized. By prior judgment of the model's classification results, it reduces the frequency attention processing of some potential events; or by prior verification of the frequency attention confidence processing, it reduces the convolutional classification of some potential events. This saves system resources and improves detection efficiency.

[0087] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A spindle wave detection method incorporating a frequency attention mechanism, characterized in that, include: Acquire a portion of the EEG signals and initialize the portion of the EEG signals to obtain a default event group; Feature extraction is performed on the default event group to obtain a feature tensor; The feature tensor is convolved to obtain a first potential event group, where each potential event in the first potential event group includes its corresponding center position and duration. Based on the center location and duration of each potential event in the first potential event group, the original EEG signal is located and frequency attention processing is performed to obtain a set of confidence arrays that correspond one-to-one with the potential events in the first potential event group; and a probability vector is calculated based on the first potential event group; and each potential event is classified based on the probability vector and the corresponding confidence value.

2. The spindle wave detection method combining frequency attention mechanism according to claim 1, characterized in that, After obtaining the first potential event group, it specifically includes: Based on the center position and duration of each potential event in the first potential event group, the original EEG signal represented by the potential event is located in reverse, and frequency attention processing is performed on the original EEG signal to obtain a set of confidence arrays that correspond one-to-one with the potential events in the first potential event group. Determine whether the corresponding confidence value of the potential event in the first potential event group meets the preset condition. If the corresponding confidence value of the potential event in the first potential event group meets the preset condition, then the potential event is directly identified as a spindle wave. If the corresponding confidence value of the potential event in the first potential event group does not meet the preset condition, then proceed to the next step. Convolve the potential events in the first potential event group to obtain the probability vector of the potential event, and output the classification result of the potential event based on the probability vector.

3. The spindle wave detection method combining frequency attention mechanism according to claim 1, characterized in that, After obtaining the first potential event group, it specifically includes: Convolve the potential events in the first potential event group to obtain the probability vector of the potential event, and output the classification result of the potential event based on the probability vector; Determine whether the classification result is a spindle wave. If the classification result is a spindle wave, output the classification result directly; if the classification result is not a spindle wave, proceed to the next step. Based on the center location and duration of each potential event in the first potential event group, the original EEG signal represented by the potential event is located in reverse. Frequency attention processing is performed on the original EEG signal to obtain the confidence value of the original EEG signal. The classification result of the potential event is output based on the confidence value of the original EEG signal.

4. A spindle wave detection method combining a frequency attention mechanism according to any one of claims 1-3, characterized in that, The frequency attention processing includes: Wavelet transform was performed on the original EEG signal to obtain wavelet coefficients; Calculate the amplitude spectrum of the wavelet coefficients, calculate the dominant frequency based on the amplitude spectrum, and obtain the confidence value based on the magnitude of the dominant frequency.

5. The spindle wave detection method combining frequency attention mechanism according to claim 4, characterized in that, The confidence value is calculated using the following formula: T(r i )=J(F(Wt(C(r i .center_time,r i duration))),frequency_range) Where T represents the confidence result of the event (with a value of 0 or 1); J represents determining whether the dominant frequency falls within the expected frequency range; F represents obtaining the dominant frequency in the EEG (electroencephalogram) segment; Wt represents wavelet transform; and C represents extracting the EEG segment represented by the two parameters .center_time.

6. A spindle wave detection device incorporating a frequency attention mechanism, characterized in that, include: Initialization module: used to acquire partial EEG signals and initialize the partial EEG signals to obtain a default event group; Feature extraction module: used to extract features from the default event group to obtain a feature tensor; The localization module is used to perform convolution processing on the feature tensor to obtain a first potential event group, wherein each potential event in the first potential event group includes its corresponding center position and duration. Frequency attention classification module: It is used to locate the original EEG signal and perform frequency attention processing based on the center position and duration of each potential event in the first potential event group to obtain a set of confidence arrays that correspond one-to-one with the potential events in the first potential event group; and calculate the probability vector based on the first potential event group; and complete the classification of each potential event based on the probability vector and the corresponding confidence value.

7. A spindle wave detection device incorporating a frequency attention mechanism according to claim 6, characterized in that, The frequency attention classification module includes: Frequency attention submodule: used to reverse locate the original EEG signal represented by each potential event in the first potential event group based on the center position and duration of each potential event, and perform frequency attention processing on the original EEG signal to obtain a set of confidence arrays that correspond one-to-one with the potential events in the first potential event group; First Judgment Submodule: Used to determine if the corresponding confidence value of the potential event in the first potential event group meets the preset condition. If the corresponding confidence value of the potential event in the first potential event group meets the preset condition, the potential event is directly output as a spindle wave. If the corresponding confidence value of the potential event in the first potential event group does not meet the preset condition, the classification submodule is triggered. The classification submodule is used to perform convolution on the potential events in the first potential event group to obtain the probability vector of the potential event, and output the classification result of the potential event based on the probability vector.

8. A spindle wave detection device incorporating a frequency attention mechanism according to claim 6, characterized in that, The frequency attention classification module includes: The classification submodule is used to perform convolution on each potential event in the first potential event group to obtain the probability vector of the potential event, and to obtain the classification result of the potential event based on the probability vector. The second judgment submodule is used to determine whether the classification result is a spindle wave. If the classification result is a spindle wave, the classification result is output directly. If the classification result is not a spindle wave, the frequency attention submodule is triggered. Frequency attention submodule: It is used to reverse locate the original EEG signal represented by each potential event in the first potential event group based on the center position and duration of each potential event, perform frequency attention processing on the original EEG signal to obtain the confidence value of the original EEG signal, and output the classification result of the potential event based on the confidence value of the original EEG signal.

9. A spindle wave detection device incorporating a frequency attention mechanism according to any one of claims 6-8, characterized in that, The frequency attention submodule includes: Wavelet transform processing layer: used to perform wavelet transform on the original EEG signal to obtain wavelet coefficients; Confidence value calculation processing layer: used to calculate the amplitude spectrum of wavelet coefficients, calculate the main frequency based on the amplitude spectrum, and obtain the confidence value based on the magnitude of the main frequency.

10. A spindle wave detection system incorporating a frequency attention mechanism, characterized in that: Brain-computer interface device and a spindle wave detection device incorporating a frequency attention mechanism as described in any one of claims 6-9.

Images