A Streaming Epilepsy Prediction Method Based on Feature Screening

Abstract

The present invention relates to a streaming epilepsy prediction method based on feature screening. The method includes: collecting and annotating electroencephalogram signals; normalizing the data; extracting features from the normalized data and splicing the extracted features; obtaining the contour curve of the time series by using the recursive average filtering method, and calculating the trend ranking of the features in combination with the dynamic time warping algorithm; selecting the optimal features by using the recursive feature elimination method; constructing a time-domain convolutional network as a training model, putting the extracted optimal features into the model for model training, constructing a streaming training mode, and stopping training when the model output reaches the expected accuracy rate to obtain a trained model; using the trained model to give an alarm in the pre-epileptic or epileptic seizure period, evaluating the epilepsy monitoring effect by using specificity and sensitivity, and improving the probability of correct early warning and reducing the number of false early warnings in the daily monitoring of brain waves.

Description

Technical Field

[0001] The present invention relates to the field of epilepsy prediction in electroencephalogram (EEG) signal processing, and particularly to a streaming epilepsy prediction method based on feature screening. Background Art

[0002] Epilepsy is a chronic brain disease caused by excessive discharge of brain neurons. There are approximately 50 million epilepsy patients worldwide, and epileptic seizures are characterized by repeatability and suddenness. A sudden epileptic seizure poses a high risk and can cause various injuries to epilepsy patients, such as falls, drowning, asphyxiation, etc. Repeated epileptic seizures can also damage the normal brain function of patients, causing them to suffer from unpredictable chronic diseases, losing the autonomy of daily life, and even leading to premature death in severe cases. Timely detection of epileptic seizure phenomena in patients and providing treatment have become crucial for epilepsy patients. By performing artificial intervention on epilepsy patients during or before seizures, on the one hand, it can relieve the pain suffered by patients and slow down their symptoms, and on the other hand, it can also reduce the high-risk injuries caused by epileptic seizures.

[0003] EEG brain wave signals are a feature expressing the activities of brain neurons, which contain rich and complex activity signals in the brain. By observing the EEG signals of patients and detecting whether there is repetitive neural discharge, thereby detecting epileptic seizures, has become the standard technique for diagnosing epilepsy in current clinical practice. Using EEG signals for epilepsy detection has the advantages of universality, safety, and all-weather availability.

[0004] Currently, the research on EEG-based epilepsy detection or epilepsy warning has been increasing day by day. The main goal is to improve the classification accuracy of the model by more precisely reducing signal noise, selecting more appropriate EEG features, and training better classification models to achieve this goal. In the research on EEG-based epilepsy warning, there are many feature extraction methods, and how to select effective features is a great challenge. In addition, the existing EEG signal acquisition basically adopts the 10-20 international electrode standard system, and how to screen channels to obtain better prediction effects is also the difficulty of the research. Summary of the Invention

[0005] The object of the present invention is to provide a streaming epilepsy prediction method for EEG trend feature screening and channel adaptive selection, taking into account both individual differences of subjects and the relationships between EEG signal channels, screening effective features, thereby reducing the complexity of the network model and improving the epilepsy prediction accuracy. In the actual life of patients, daily epilepsy monitoring can be carried out according to brain waves. If it is found that the wave band enters the pre-epileptic seizure stage, a warning will be given, so as to provide better treatment and protection for patients. The epilepsy prediction method provided by the present invention includes three processes: EEG signal acquisition and obtaining, feature extraction and screening, and epilepsy streaming prediction. It mainly includes the following steps:

[0006] Step 1: Collect EEG signals and label the interictal, pre-ictal, and ictal stages;

[0007] Step 2: Preprocess the data, that is, perform normalization processing;

[0008] Step 3: Extract features from the normalized data and splice the extracted features;

[0009] Step 4: Obtain the contour curve of the time series and calculate the trend ranking of the features;

[0010] Step 5: Select the optimal features by means of the recursive feature elimination method;

[0011] Step 6: Construct a time-domain convolutional network as a training model, put the optimal features extracted in Step 5 into the model for model training, optimize the training data, use the training result of the previous moment as the feature of the next moment to construct a streaming training mode, and stop training when the model output reaches the expected accuracy to obtain the trained model;

[0012] Step 7: In EEG epilepsy monitoring, use the trained model. When the specified number of positive times is reached, it is determined as an alarm for a pre-epileptic or epileptic seizure period.

[0013] Furthermore, it may further include:

[0014] Step 8: Evaluate the epilepsy monitoring effect using specificity and sensitivity.

[0015] Furthermore, Step 1 includes the following steps:

[0016] Step 1-1: Collect EEG signals according to the 10-20 international electrode standard.

[0017] Step 1-2: Use EdfReader in Python to read out the data, and improve the labels in the way that the 30 minutes before the seizure is the pre-ictal stage, the 4 hours before the seizure or 4 hours after the seizure is the interictal stage, and the epileptic seizure is the ictal stage.

[0018] Step 1-3: Segment the pre-ictal, inter-ictal, and ictal periods of each patient into 20s data segments.

[0019] Further, Step 2 includes the following steps: data normalization processing.

[0020] Perform data normalization processing. Denote the original data as Z = [z1, z2, z3, z4......, z n Perform normalization on each dimension, and the operation is where, Z i_new is the data after normalization, z i is the original data, n is the data dimension, i ∈ (1, n), i new ∈ (1, n).

[0021] Further, the feature extraction in Step 3 includes: time-domain feature extraction, frequency-domain feature extraction, and time-frequency spectrum feature extraction.

[0022] Further, Step 3 includes the following steps:

[0023] Step 3-1: Extract time series features, and the time series features include: mean, variance, peak-to-peak value, skewness, root mean square amplitude, approximate entropy, and sample entropy.

[0024] Step 3-2: Extract frequency-domain features using the power spectral density.

[0025] Step 3-3: Extract time-frequency spectrum features using the short-time Fourier transform.

[0026] Further, Step 4 includes the following steps:

[0027] Step 4-1: Arrange the extracted features in time series, and smooth the curve using the recursive average filter method.

[0028] Step 4-2: Calculate the distance between the smoothed features and the labels using the dynamic time warping algorithm, so as to obtain the trend ranking of the features.

[0029] Further, Step 5 includes the following steps:

[0030] Step 5-1: Use the common convolutional neural network model Alexnet, and select the learning rate and loss function of the model.

[0031] Step 5-2: According to the feature ranking obtained in Step 4, use the recursive feature elimination method to perform model training, and finally select the optimal feature combination.

[0032] Further, Step 6 includes the following steps:

[0033] Step 6-1: Organize the sample matrices of the training set and the test set according to the feature combinations obtained in Step 5.

[0034] Step 6-2: Construct a temporal convolutional network, add a channel adaptive selection module based on the attention mechanism at the input end of the network, and send the weighted combined feature matrix into the temporal convolutional network.

[0035] In Step 6-3, the focal loss is used as the loss function to correct the problem of unbalanced positive and negative samples in the training set.

[0036] Step 6-4: Construct a streaming training strategy, and use the prediction result of the previous time step as the feature of the current time step to participate in the training process.

[0037] Furthermore, the number of positive times specified in Step 7 is that among 5 consecutive predictions, 4 times are determined to be positive.

[0038] According to the above concept, the technical solutions for implementing the present invention mainly have the following characteristics:

[0039] A trend feature screening method based on the dynamic time warping algorithm is proposed, and the characteristic trend change from the interictal period to the seizure period is fitted by calculation. And through the way of recursive elimination, the Alexnet network is selected to select the optimal feature combination.

[0040] Based on the features extracted by the above method, an epilepsy streaming prediction method is proposed, which makes full use of the selected effective features to improve the prediction effect of the model; uses the channel self-selection algorithm based on the attention mechanism to complete the automatic screening of data channels; finally adopts the streaming training and prediction method to improve the probability of correct early warning and reduce the number of false early warnings in the daily monitoring of brain waves.

[0041] Based on the above technical solutions, the beneficial effects of the present invention are mainly reflected in:

[0042] The present invention obtains a richer feature matrix by extracting the time domain features, frequency domain features and time-frequency domain features of electroencephalogram data, and can better represent the changes of electroencephalogram signals.

[0043] The present invention selects effective feature combinations by combining the dynamic time warping method and the recursive elimination method, reduces the complexity of the network model and improves the classification ability of the model at the same time.

[0044] On the basis of the above feature screening, the present invention establishes an epilepsy prediction model, completes the automatic selection of data channels, and through the streaming training and testing methods, improves the probability of correct early warning and reduces the number of false early warnings in the daily monitoring of brain waves. Description of the Drawings

[0045] The specific implementation manners of the present invention will be further described in detail below with reference to the accompanying drawings.

[0046] Figure 1 It is the overall architecture diagram of the present invention

[0047] Figure 2 It is the model architecture diagram proposed by the present invention

[0048] Figure 3 It is the structural diagram of the AlexNet model Specific implementation manners

[0049] The present invention will be further described below in conjunction with specific embodiments, but the protection scope of the present invention is not limited thereto:

[0050] In Embodiment 1, there are problems in electroencephalogram (EEG) monitoring technology, such as a large number of features, large computational complexity, difficult fitting, many EEG acquisition channels for epilepsy and the main effective channels vary from person to person, resulting in low accuracy of epilepsy early warning and high false alarm probability in daily monitoring algorithms. In view of this problem, the present invention proposes a streaming epilepsy prediction method based on feature screening. As Figure 1 shown, the streaming epilepsy prediction method based on feature screening provided by the embodiment of the present invention includes the following steps:

[0051] Step 1: First, prepare the data of real patients. The data in this embodiment uses the epilepsy EEG data collected by Boston Children's Hospital and is labeled according to the interictal, preictal, and ictal periods. The above EEG data is the data obtained by collecting EEG signals according to the 10-20 international electrode standard;

[0052] Step 2: Preprocess the data, that is, perform normalization processing.

[0053] Step 3: Extract features from the normalized data. The present invention uses three methods to extract features. The first is time-domain feature extraction, the second is frequency-domain feature extraction, and the third is time-frequency spectrum feature extraction, and the three extracted features are concatenated.

[0054] Step 4: Use the recursive average filtering method to obtain the contour curve of the time series, and calculate the trend ranking of the features in combination with the dynamic time warping algorithm.

[0055] Step 5: Select the best features by means of the recursive feature elimination method.

[0056] Step 6: Construct a time-domain convolutional network as the training model, allocate channel weights and perform weighted summation using the attention mechanism, use focal loss to solve the problem of sample imbalance in the training data, and additionally use the training result of the previous moment as the feature of the next moment to construct a streaming training mode. Put the optimal features extracted in Step 5 into this model for model training. The overall architecture of the model is as Figure 2 shown. When the model output reaches the expected accuracy rate, stop training to obtain the trained model.

[0057] Step 7: In EEG epilepsy monitoring, use the trained model. When the preset number of positive times is reached, it is determined as an alarm for a pre-epileptic or epileptic seizure period. The method adopted in this embodiment is that when 4 out of 5 consecutive predictions are determined to be positive, it is determined as an alarm for a pre-epileptic or epileptic seizure period.

[0058] Step 8: Evaluate the epilepsy monitoring effect using specificity and sensitivity.

[0059] Furthermore, Step 1 further includes the following sub-steps:

[0060] Step 1-1: First, prepare the data of real patients. The data uses the epilepsy EEG data collected by Boston Children's Hospital. The dataset includes the records of 23 cases collected from 22 subjects (5 males, aged 3 to 22 years; 17 females, aged 1.5 to 19 years). The specific seizure conditions of each patient are shown in Table 1.

[0061] Step 1-2: Use EdfReader in Python to read out the data, and improve the labels in the way that 30 minutes before the seizure is the pre-epileptic period, 4 hours before the seizure or 4 hours after the seizure is the inter-ictal period, and the epileptic seizure is the seizure period.

[0062] Step 1-3: Segment the pre-epileptic period, inter-ictal period, and seizure period of each patient into 2s data segments.

[0063] Step 1-4: Screen appropriate data as the training set and prediction set. Use the data of a single seizure period, pre-epileptic period, and inter-ictal period of the patient as the test set, and the remaining seizures as the training set.

[0064] Table 1: Information of Epilepsy Patients in the Dataset of Boston Children's Hospital

[0065] Patient Gender Age Signal intensity (microvolt) Number of attacks 01 Female 11 27.78 7 02 Male 11 30.92 3 03 Female 14 21.79 7 04 Male 22 31.47 4 05 Female 7 49.5 5 06 Female 1.5 61.02 10 07 Female 14.5 29.38 3 08 Male 3.5 44.74 5 09 Female 10 49.3 4 10 Male 3 54.27 7 11 Female 12 20.79 3 12 Female 2 25.16 27 13 Female 3 35.88 12 14 Female 9 35.16 7 15 Male 16 13.13 20 16 Female 7 27.25 10 17 Female 12 19.34 3 18 Female 18 13.53 6 19 Female 19 13.44 3 20 Female 6 22.17 8 21 Female 13 21.52 4 22 Female 9 26.45 3 23 Female 6 25.66 7

[0066] Furthermore, Step 2 specifically includes:

[0067] Perform normalization processing on the data. Denote the original data as Z = [z1, z2, z3, z4......, z nNormalize each dimension, and the operation is where Z i_new is the data after normalization, z i is the original data, n is the data dimension, i ∈ (1, 512), i new ∈ (1, 512).

[0068] Furthermore, step 3 includes the following sub-steps:

[0069] Step 3-1: Extract time series features. The time series features extracted in this embodiment specifically include mean, variance, peak-to-peak value, skewness, root mean square amplitude, approximate entropy, and sample entropy.

[0070] Furthermore, step 3-1 includes the following sub-steps:

[0071] Step 3-1-1: Read all the standardized training set data in step 2, and extract features from each sample among them. The data is a matrix of 18 rows and 512 columns;

[0072] Step 3-1-2: Extract the mean feature from each row of the 18 rows of data, that is, mean(x);

[0073] Step 3-1-3: Extract the variance feature, that is

[0074] Step 3-1-4: Extract the peak-to-peak value feature, that is, max(x) - min(x);

[0075] Step 3-1-5: Extract the skewness feature, that is where μ is mean(x), and σ is

[0076] Step 3-1-6: Extract the root mean square amplitude feature, that is where abs(·) is to find the absolute value;

[0077] Step 3-1-7: Extract the approximate entropy feature, that is where is Take m = 2 and r = 2 during the extraction process;

[0078] Step 3-1-8: Extract the sample entropy feature.

[0079] Furthermore, step 3-1-8 specifically includes:

[0080] Step 3-1-8-1: First, construct the time series X into an m-dimensional vector, that is, X(i) = {x(i), x(i + 1,..., x(i + m - 1)}, where i = 1, 2,..., N - m + 1.

[0081] Step 3-1-8-2: Define the distance between X(i) and X(j) as the largest difference among their corresponding elements, that is

[0082] Step 3-1-8-3: Given a threshold r (r > 0), count the number of distances between X(i) and X(j) that are < r and calculate the ratio to the total number of vectors N - m, that is: Denote it as

[0083] Step 3-1-8-4: Take the average of the results obtained in Step 3-1-8-3 to get

[0084] Step 3-2: Extract frequency-domain features using the power spectral density, which is where X(k) is

[0085] Step 3-3: Extract time-frequency spectrum features using the short-time Fourier transform, which is

[0086] Furthermore, Step 4 includes the following steps:

[0087] Step 4-1: After completing the feature extraction in Step 3, each sample becomes a feature matrix with 18 rows and 134 columns, where 18 is the number of channels. Arrange the samples in time series. For each feature of each channel, use the recursive average filtering method to smooth the curve. Specifically: where p t is the value at the current moment after smoothing, and g t is the value before smoothing.

[0088] Step 4-2: Use the dynamic time warping algorithm to calculate the distance between the smoothed features and the labels, so as to obtain the trend ranking of the features. Denote the smoothed feature sequence as x(t), and generate a sequence y(t) of the same length as x(t) for the labels of the samples.

[0089] Step 4-3: Use the dynamic time warping algorithm to calculate the distance between the smoothed features and the labels.

[0090] Furthermore, Step 4-3 includes the following sub-steps:

[0091] Step 4-3-1: Calculate the distance matrix between each point of x(t) and y(t).

[0092] Step 4-3-2: Find a path from the upper left corner to the lower right corner of the matrix such that the sum of the elements on the path is the smallest.

[0093] Furthermore, Step 4-3-2 specifically includes:

[0094] Step 4-3-2-1: Set the starting condition, the minimum distance L min (1, 1) = M(1, 1).

[0095] Step 4-3-2-2: According to the recurrence rule L min (i, j) = min{L min (i, j - 1), L min (i - 1, j, Lmini - 1, j - 1 + M(i, j) Recursively find the path with the smallest sum of elements on the path and obtain the shortest distance. Calculate the feature distance for each channel, then sum the results of 18 channels, and finally sort them in ascending order of distance to obtain the trend feature sorting of the current dataset.

[0096] Further, Step 5 includes the following steps:

[0097] Step 5-1: Build the currently widely used convolutional neural network model Alexnet, which includes five convolutional layers, three max pooling layers and two fully connected layers. The learning rate is set to 0.01, and the Adam optimizer is used.

[0098] Step 5-2: According to the feature sorting obtained in Step 4, use the method of recursive feature elimination to train the Alexnet model, that is, starting from using all feature dimensions, reducing 10 feature dimensions each time, and finally selecting the best feature combination. The total number of feature dimensions in the present invention is 134 dimensions. Training is carried out respectively for features of 134, 120, 110, 100, 90, 80, 70, 60, 50, 40, 30, 20, 10 dimensions. The final results show that when using the first 40 dimensions of the trend feature sorting, the best classification effect can be obtained.

[0099] Further, Step 6 includes the following steps:

[0100] Step 6-1: Organize the sample matrices of the training set and the test set according to the feature combination obtained in Step 5. For each sample, the channel dimension is 18 and the feature dimension is 40. Combine 19 consecutive samples into a time series sample, and regard the data of a single seizure as the test set, and the remaining seizure times as the training set.

[0101] Step 6-2: Build a time series convolutional network with channel adaptive selection.

[0102] Further, Step 6-2 includes the following steps:

[0103] Step 6-2-1: Build a time series convolutional network, where the convolutional kernel size is 4 and the number of time series convolutional layers is 3.

[0104] Step 6-2-2: Build a channel adaptive selection module at the input end to change the original input dimension of 18*19*41 to 18*779, denoted as H.

[0105] Step 6-2-3: Let H pass through a tanh activation function to obtain M = tanh(H). The dimension of matrix M is 18*779.

[0106] Step 6-2-4: Multiply M by a 779*1 matrix, and then pass through softmax to obtain the weight α of each channel, α = softmax(w T M). The dimension of α is 18*1.

[0107] Step 6-2-5: Multiply the original feature H by the channel weight α to obtain a weighted feature matrix, and change its feature dimension to 19*41 and send it into the temporal convolutional network.

[0108] Step 6-3: Adopt the focal loss function to solve the problem of imbalance between positive and negative samples in the training set.

[0109] Step 6-4: Build a streaming training strategy, and use the prediction result of the previous time step as the feature of the current time step to participate in the training process. That is, the feature dimension is 41, where the first dimension is the prediction result of the previous time step, and the remaining 40 dimensions are the extracted features, and use the result of the previous time step to correct the prediction result of the current time step.

[0110] Furthermore, step 7 includes the following steps:

[0111] Step 7-1: In electroencephalogram epilepsy monitoring, use the trained model. When 4 out of 5 consecutive predictions are judged as positive, it is regarded as an alarm for a pre-epileptic or epileptic seizure period.

[0112] Step 7-2: Use specificity and sensitivity to evaluate the epilepsy monitoring effect. Specificity is the proportion of correctly predicted negative samples in the actual negative samples, that is Sensitivity is the proportion of the number of correctly predicted positive samples to the actual positive samples, that is

[0113] In order to verify the effectiveness of the present invention and reduce the impact of different data on the final analysis results, 23 patients in Boston Children's Hospital were used for training and testing, and the final accuracy was the average of the 23 patients. Five schemes were selected for comparison: the first method was a time series convolutional neural network without feature screening, the second method was a time series convolutional neural network without adaptive channel selection, the third method was a time series convolutional neural network without focal loss, the fourth method was a time series convolutional neural network without a streaming processing mechanism, and the fifth method was the above method. Then, for the test results of the above five methods, the sensitivity and specificity were statistically analyzed, and the final results are shown in Table 2.

[0114] Table 2: Experimental results

[0115] Method Method 1 Method 2 Method 3 Method 4 Method 5 Sensitivity 91.08％ 90.25％ 93.24％ 94.24％ 98.18％ Specificity 93.1％ 97.82％ 94.22％ 95.69％ 98.82％

[0116] It is not difficult to see from the results that the present invention improves the fitting performance of the model by screening features. In addition, the effectiveness of EEG epilepsy monitoring is enhanced through adaptive screening channels and streaming training strategies. It can basically detect every attack and pre-attack of the patient, and the false alarm rate is low, which can achieve better daily monitoring effects on patients.

[0117] Finally, it should be noted that the above examples are only some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments, and there are many variations. All variations that can be directly derived or associated with the content disclosed by a person skilled in the art should be considered as the protection scope of the present invention.

Claims

1. A streaming epilepsy prediction method based on feature screening, the method comprising: Step 1: Collect electroencephalogram (EEG) signals and label the interictal, preictal, and ictal periods; Step 2: Preprocess the data, i.e., perform normalization processing; Step 3: Extract features from the normalized data and splice the extracted features; wherein, the feature extraction includes: time-domain feature extraction, frequency-domain feature extraction, and time-frequency spectrum feature extraction; and further includes: Step 3-1: Extract temporal features, the temporal features including: mean, variance, peak-to-peak value, skewness, root mean square amplitude, approximate entropy, and sample entropy; Step 3-2: Extract frequency-domain features using power spectral density; Step 3-3: Extract time-frequency spectrum features using short-time Fourier transform; Step 4: Obtain the contour curve of the time series and calculate the trend ranking of the features; further includes: Step 4-1: Arrange the extracted features in time sequence and smooth the curve using a recursive average filter; Step 4-2: Calculate the distance between the smoothed features and the labels using the dynamic time warping algorithm to obtain the trend ranking of the features; Step 5: Select the optimal features using the recursive feature elimination method; Step 6: Construct a time-domain convolutional network as a training model, put the optimal features extracted in Step 5 into the model for model training, and optimize the training data. Use the training result of the previous moment as the feature of the next moment to construct a streaming training mode. When the model output reaches the expected accuracy rate, stop training to obtain the trained model; Step 7: In EEG epilepsy monitoring, use the trained model. When the specified number of positive times is reached, it is determined as an alarm for a pre-epileptic or epileptic seizure period.

2. The epilepsy prediction method according to claim 1, characterized in that, Further includes: Step 8: Evaluate the epilepsy monitoring effect using specificity and sensitivity.

3. The epilepsy prediction method according to claim 1 or 2, characterized in that, The said Step 1 further includes: Step 1-1: Collect EEG signals according to the 10-20 international electrode standard; Step 1-2: Use EdfReader in Python to read out the data, and improve the labels in the way that the 30 minutes before the seizure is the preictal period, the 4 hours before or 4 hours after the seizure is the interictal period, and the epileptic seizure is the ictal period; Step 1-3: Segment the preictal, interictal, and ictal periods of each patient into 20s data segments.

4. The epilepsy prediction method according to claim 1 or 2, characterized in that The said normalization processing in Step 2 is: Denote the original data as Z = [z1, z2, z3, z4......, zn], and perform normalization on each dimension. The operation is Among them, Z i_new is the data after normalization, z i is the original data, n is the data dimension, i ∈ (1, n), i new ∈ (1, n).

5. The epilepsy prediction method according to claim 1 or 2, characterized in that, Step 5 further includes the following steps: Step 5-1: Use the common convolutional neural network model Alexnet, and select the learning rate and loss function of the model; Step 5-2: According to the feature ranking obtained in Step 4, use the recursive feature elimination method to perform model training and finally select the optimal feature combination.

6. The epilepsy prediction method according to claim 1 or 2, characterized in that Step 6 further includes the following steps: Step 6-1: Organize the sample matrices of the training set and the test set according to the feature combination obtained in Step 5; Step 6-2: Construct a temporal convolutional network, add a channel adaptive selection module based on the attention mechanism at the network input end, and send the weighted and combined feature matrix into the temporal convolutional network; Step 6-3: The loss function uses focal loss to correct the problem of imbalance between positive and negative samples in the training set; Step 6-4: Construct a streaming training strategy, and use the prediction result of the previous time step as the feature of the current time step to participate in the training process.

7. The epilepsy prediction method according to claim 1 or 2, characterized in that, The number of positive times specified in Step 7 is: among 5 consecutive predictions, 4 times are determined to be positive.

Images