A personalized prediction system and device for temporal lobe epilepsy based on multi-site cross-frequency coupling

Abstract

The present invention discloses a personalized prediction system and device for temporal lobe epilepsy based on multi-site cross-frequency coupling, including: a signal acquisition module for synchronously acquiring electroencephalogram signals at multiple sites in the brain; a signal preprocessing module for preprocessing the acquired electroencephalogram signals, drawing a cross-frequency coupling co-modulation map, and segmenting frequency bands; a feature extraction module for extracting phase-amplitude coupling features, including the average coupling center phase frequency, amplitude frequency, and average coupling center intensity; a prediction module for inputting the phase-amplitude coupling features into a pre-trained machine learning model and outputting a prediction result of temporal lobe epilepsy; and an early warning module for generating an early warning signal according to the prediction result. The present invention breaks through the traditional fixed frequency band analysis paradigm, captures phase / amplitude migration features through energy distribution weighting; constructs a personalized adaptation mechanism to identify atypical cross-frequency coupling features of epilepsy; and has stronger interpretability for the mechanism of neural oscillation desynchronization.

Description

Technical Field

[0001] The present invention belongs to the field of electroencephalogram signal processing, and in particular relates to a personalized prediction system and device for temporal lobe epilepsy based on multi-site cross-frequency coupling. Background Art

[0002] Epilepsy is a chronic neurological disease caused by abnormal discharges of brain neurons, and its clinical manifestations are recurrent neurological disorders. This disease not only leads to a progressive decline in the cognitive function of patients, but may also induce psychological disorders such as anxiety and depression, posing a serious threat to the quality of life of patients. Current clinical treatments mainly rely on anti-epileptic drugs to control the condition, but about 30%-40% of patients will experience drug tolerance and develop into drug-resistant epilepsy (DRE). For the DRE patient group, surgical resection of the epileptogenic focus is an important treatment method. However, limited by factors such as the positioning accuracy of the epileptogenic focus, surgical risks, and postoperative complications, a considerable proportion of patients still cannot obtain effective treatment through surgery. Against this background, the research value of epilepsy seizure prediction technology has become increasingly prominent - by accurately predicting the onset time of epileptic seizures, patients can take preventive measures in advance (such as adjusting body position, activating a neuromodulation device), thereby significantly reducing the risk of accidental injuries during the seizure period and alleviating the psychological burden caused by the uncertainty of seizures. More importantly, reliable prediction results can provide a basis for clinicians to optimize treatment plans, avoid excessive medical interventions, and overall improve the level of disease management.

[0003] Existing epilepsy prediction methods mainly rely on the analysis of electroencephalogram (EEG) signals. Methods such as linear analysis, non-linear dynamics analysis, and time-frequency analysis have been widely used in epilepsy prediction research. Traditional epilepsy prediction methods often have difficulty capturing the dynamic coupling relationship between different frequency bands in the brain. In particular, the interaction between neural oscillations in different frequency bands often undergoes significant changes before epileptic seizures, and such changes are often not effectively captured by traditional features. For this reason, in recent years, research on phase amplitude coupling (PAC) features has gradually received attention. PAC features can effectively reveal the coupling relationship between neural oscillations in different frequency bands, and thus provide a new idea for the early prediction of epileptic seizures.

[0004] Compared with traditional EEG features, the PAC technology system has multi-dimensional advantages: First, in terms of feature sensitivity, PAC can dynamically reflect the evolution process of the coupling strength between different frequency bands, especially within dozens of minutes before epileptic seizures. A significant enhancement phenomenon of the frequency band coupling relationship, this time-varying characteristic is extremely easy to be ignored through conventional power spectrum analysis; secondly, in terms of time resolution, the eigenvalues calculated by PAC based on instantaneous phase and amplitude can be updated on a second-by-second basis, which is conducive to capturing subtle electrophysiological changes before seizures; secondly, in terms of anti-interference ability, the PAC feature has strong robustness to common electromyogram artifacts and motion noise in EEG signals, which is crucial for improving the practicality of bedside monitoring devices. On the contrary, traditional time-domain features (such as signal slope and peak-to-peak value) are easily affected by noise interference and produce feature drift, directly affecting the stability of the prediction model.

[0005] Although PAC technology shows unique value in the field of epilepsy prediction, existing methods still have limitations: 1) The problem of single-dimensional feature extraction. Current research generally uses the maximum value of the modulation index (MI) as the core feature. Although this method can reflect the coupling strength of specific frequency band combinations, it cannot describe the dynamic migration law of the phase center frequency and amplitude center frequency (such as in the pre-ictal period The phase of the frequency band shifts to the frequency band, the amplitude of the frequency band expands to the ripple frequency band). The strategy relying on fixed thresholds or global statistics results in insufficient sensitivity to changes in the pre-ictal network state; 2) Lack of individual adaptation ability. Existing methods mostly preset a single frequency band combination (such as θ-phase-γ-amplitude coupling), ignoring the atypical coupling patterns caused by different epilepsy types (such as frontal lobe epilepsy and temporal lobe epilepsy) or pathological mechanisms (such as structural lesions and gene mutations). For example, clinical data show that some DRE patients have abnormal coupling between the phase and the amplitude of the ripple frequency band, and the traditional preset frequency band analysis method has a detection blind spot for such variant patterns; 3) Limitations in feature interpretability. As a mathematical statistic, the MI index has no clear association between its calculation results and neurophysiological mechanisms, and it is difficult to explain the specific change pattern of the excitation-inhibition balance in the epileptogenic network, which seriously restricts the trust of clinicians in the prediction results. To address the above problems, this patent proposes three core technological breakthroughs: breaking through the traditional fixed frequency band analysis paradigm; constructing an individualized epilepsy prediction model to analyze individualized PAC patterns, effectively capturing non-atypical cross-frequency interaction features such as frontal lobe epilepsy -Ripple coupling, drug-resistant epilepsy -ultra-high frequency coupling, etc.; innovatively proposing the average coupling strength feature, and establishing a pathological association between the eigenvalue and the desynchronization mechanism of neural oscillations through the weight fusion of phase / amplitude contribution. This solution constructs an epilepsy prediction feature system that is more adaptable to clinical complex scenarios through the collaborative innovation of three dimensions: dynamic frequency band tracking, individualized pattern adaptation, and enhanced physiological interpretability, providing a new methodological framework for the development of a precise early warning system. Summary of the Invention

[0006] In view of the deficiencies of the prior art, the present invention provides a personalized prediction system and device for temporal lobe epilepsy based on multi-site cross-frequency coupling, which solves the problems raised in the above background art.

[0007] A personalized prediction system for temporal lobe epilepsy based on multi-site cross-frequency coupling, comprising:

[0008] A signal acquisition module for synchronously acquiring electroencephalogram signals at multiple sites in the brain;

[0009] A signal preprocessing module for preprocessing the acquired electroencephalogram signals, drawing a co-modulation map of cross-frequency coupling, and segmenting frequency bands;

[0010] A feature extraction module for extracting phase-amplitude coupling features, including average coupling center phase frequency, average coupling center amplitude frequency, and average coupling strength;

[0011] A prediction module for inputting the phase-amplitude coupling features into a pre-trained machine learning model and outputting a prediction result of temporal lobe epilepsy;

[0012] An early warning module for generating an early warning signal according to the prediction result.

[0013] Further, the preprocessing of the acquired electroencephalogram signals includes:

[0014] Using notch filtering to remove power frequency noise and mean filtering;

[0015] The continuous electrical signals after notch filtering to remove power frequency noise and mean filtering are divided into processing windows according to a fixed division time length, and at the same time, adjacent windows are overlapped by a certain overlapping time length, and the overlapping time length is less than the fixed division time length.

[0016] Further, the process of drawing the co-modulation map of cross-frequency coupling is as follows:

[0017] For the electrical signals in each processing window, the instantaneous phase of low-frequency oscillation and the instantaneous amplitude of high-frequency oscillation are extracted by using complex wavelet transform, and the MI is calculated using the modulation index calculation formula; with the phase frequency as the abscissa and the amplitude frequency as the ordinate, a co-modulation map (Co-modulation map) of phase frequency-amplitude frequency cross-frequency coupling between different frequency bands within the processing window is obtained; this co-modulation map is used to describe the phase-amplitude coupling strength of multi-site signals;

[0018] Further, the modulation index MI calculation formula is:

[0019] ;

[0020] Wherein, The instantaneous phase of low-frequency oscillation at time is defined as , the instantaneous amplitude of high-frequency oscillation is defined as ; is the instantaneous phase, the frequency is , the instantaneous amplitude frequency is when the modulation index; is the imaginary unit, used to construct a rotating vector on the complex plane.

[0021] Further, the divided frequency bands include:

[0022] The low-frequency oscillation is divided into 4 frequency bands, namely wave, wave, wave, the overall low-frequency band, and the high-frequency oscillation is divided into 4 frequency bands, namely low wave, high wave, ripple wave and the overall high-frequency band.

[0023] Further, the calculation process of the phase-amplitude coupling feature is as follows:

[0024] The 4 low-frequency bands and 4 high-frequency bands obtained by dividing the frequency bands form 16 cross-frequency coupling modes, and further calculate the average coupling center phase frequency, average coupling center amplitude frequency and average coupling strength feature; where:

[0025] The average coupling center phase frequency :

[0026] ;

[0027] The average coupling center amplitude frequency :

[0028] ;

[0029] The average coupling strength feature :

[0030] ;

[0031] Among them, and respectively represent the row and column indices of the modulation index matrix, represents time when the value at position ( ) in the matrix; and are the phase frequency dimension and amplitude frequency dimension respectively.

[0032] The machine learning model described above is a support vector machine, which is used to input 16 combinations of features corresponding to all channels for classification, verify the performance to screen the optimal individual prediction channels and frequency band combinations, save the model parameters and output the prediction results of temporal lobe epilepsy.

[0033] Furthermore, under 16 frequency band combinations, all channels' , and MCCV features are input into the SVM model, and the model evaluation metrics are obtained, including accuracy, sensitivity, specificity, precision and F1 score, to determine the optimal channel and frequency band combination with the highest classification accuracy, and save the optimal frequency band combination and the model parameters corresponding to the channels.

[0034] The personalized prediction system for temporal lobe epilepsy described above further includes a post-processing module for optimizing the prediction results. The optimization process of the post-processing module is as follows:

[0035] Use the k-of-n module to post-process the output of the classifier in the machine learning model on the prospective study dataset and mark the pre-epileptic seizure state;

[0036] Calculate the false alarm rate per hour, prediction sensitivity and prediction lead time according to the post-processing results.

[0037] A computer device includes a memory and a processor. The memory stores a computer program. It is characterized in that when the processor executes the computer program, the functions of the above-mentioned temporal lobe epilepsy prediction system are realized.

[0038] Furthermore, the system adopts an embedded processor and low-power design, is compatible with the long-term monitoring requirements of implantable devices, and supports the integration of a closed-loop neuromodulation system.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. The personalized prediction system for temporal lobe epilepsy of the present invention combines the phase-frequency and amplitude-frequency couplings between different frequency bands at multiple sites to predict temporal lobe epilepsy, improving the accuracy and efficiency of temporal lobe epilepsy prediction.

[0041] 2. Compared with the traditional multi-channel feature extraction method, the present invention uses single-channel EEG signals to extract PAC features, which can not only reduce the complexity of data acquisition, but also fully capture the cross-frequency band coupling information of neural oscillations. The system of the present invention not only has a high time resolution and can capture the subtle changes before epileptic seizures, but also has strong anti-noise ability and can adapt to the noise interference in the clinical environment. In addition, PAC features can provide a more biologically meaningful interpretation of neural mechanisms and can effectively realize the early prediction of epileptic seizures, having high clinical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is the implementation flowchart of the personalized prediction system for temporal lobe epilepsy in the embodiments of the present invention.

[0043] Figure 2 This is the feature description of the single-channel phase-amplitude co-modulation map in the embodiments of the present invention.

[0044] Figure 3 This is the training result of the SVM model corresponding to the coupling results of different sub-bands in the embodiments of the present invention.

[0045] Figure 4 This is an example of the actual prediction result of the personalized prediction system for temporal lobe epilepsy in the embodiments of the present invention. Detailed implementation manners

[0046] The following further describes the present invention in detail with reference to the drawings and embodiments. It should be noted that the following embodiments are intended to facilitate the understanding of the present invention and do not limit it in any way.

[0047] In the process of studying the changes in phase-amplitude coupling from before epilepsy seizure to the end of epilepsy seizure, it is found that: from before epilepsy seizure to the end of seizure, the phase-amplitude coupling characteristics of each site have significant changes. The study also finds that there are differences in the results of predicting epilepsy using the phase-amplitude coupling characteristics of different frequency bands for the same individual, and there is a frequency band combination with the best epilepsy prediction ability. At the same time, the best frequency band combinations are also different among different individuals.

[0048] Based on this research result, the embodiments of the present invention provide a personalized prediction system for temporal lobe epilepsy based on multi-site cross-frequency coupling. The specific steps of the implementation process of this system are as Figure 1 shown, and the detailed steps are as follows:

[0049] Step 1, obtain abnormal electrical signals at multiple sites synchronously recorded on the hippocampus and limbic system, and prepare a data set.

[0050] In the embodiment, electrodes are implanted at multiple sites on the hippocampus and limbic system of the brain. The limbic system refers to the structures around the hippocampus, such as the anterior thalamic nucleus, amygdala, etc., which are clear to those skilled in the art. After the electrodes are implanted, start the electroencephalogram (EEG) signal acquisition device and continuously acquire the EEG signals at multiple sites. There are abnormal discharge signals in the EEG signals, and the abnormal discharge signals refer to the EEG signals during temporal lobe epilepsy seizures. The acquired EEG signals are input to and received by the temporal lobe epilepsy prediction system.

[0051] Step 2, perform preprocessing and feature calculation on the data set.

[0052] The purpose of preprocessing is to improve the data quality and eliminate noise. Specifically, perform preprocessing on the received EEG signals to determine the parameter combination for the prediction model, including:

[0053] (a) Notch filter the dataset at 50 Hz to remove power frequency noise and perform 8-point mean filtering on each acquisition channel.

[0054] (b) Plot a co-modulation map with the phase frequency on the abscissa and the amplitude frequency on the ordinate in continuous time for each recording site. This co-modulation map is used to describe the changes in cross-frequency coupling characteristics between different frequency bands at each site.

[0055] The constructed co-modulation map is mainly used for cross-frequency analysis of continuous signals. Specifically, the construction process of the co-modulation map is as follows:

[0056] Divide the collected continuous electrical signal into processing windows according to a fixed time length, and at the same time ensure that adjacent windows overlap by a certain time length. The overlapping time length is less than the fixed division time length. Use complex wavelet transform to extract the instantaneous phase of the low-frequency signal and the instantaneous amplitude of the high-frequency signal in each processing window to calculate MI, and obtain a co-modulation map that reflects the coupling characteristics between the phase of the low-frequency oscillation and the amplitude of the high-frequency oscillation within the processing window.

[0057] For example: Divide the continuous EEG signal into processing windows every 5 s. At the same time, in order to improve the resolution of the co-modulation map in continuous time, adjacent processing windows overlap by 1 s. In each window, the calculation formula of complex wavelet transform is as follows:

[0058] ;

[0059] where is the complex coefficient, is the scale parameter, is the time shift parameter, is the conjugate of the complex wavelet. The phase of the low-frequency oscillation is defined as and the amplitude of the high-frequency oscillation is defined as . The specific calculation formulas are as follows:

[0060] ;

[0061] ;

[0062] MI is a phase-amplitude coupling quantization index based on information entropy. Its core lies in reflecting the phase modulation effect through the statistical characteristics of the amplitude distribution. The calculation formula is as follows:

[0063] ;

[0064] After calculating the modulation index, plot with the phase frequency on the abscissa and the amplitude frequency on the ordinate, with a resolution of 0.1 Hz and 1 Hz, to obtain the co-modulation map.

[0065] (c) Calculate features based on the obtained phase-amplitude co-modulation map.

[0066] Specifically, the calculation formula for the phase amplitude coupling characteristics within the window is as follows:

[0067] Average coupling center phase frequency :

[0068] ;

[0069] Average coupling center amplitude frequency :

[0070] ;

[0071] Average coupling strength characteristics :

[0072] ;

[0073] in, and Represent the row and column indices of the modulation index matrix respectively. Indicates time hour Position in the matrix ( ) value, and are the phase-frequency dimension and the amplitude-frequency dimension respectively, is the average coupling center phase frequency, is the average coupling center amplitude frequency, is the average coupling strength.

[0074] Step 3: Divide the EEG signal into different periods, build a prediction model, and perform channel selection and frequency band optimization to determine the model parameters based on the model classification results of the phase-amplitude coupling characteristics obtained from the multi-site phase-amplitude co-modulation map.

[0075] The start and end time of the epileptic EEG signals in the collected data set are annotated, and the inter-epileptic period and pre-epileptic period are mainly extracted based on the start and end time of the epileptic seizure.

[0076] In the embodiment, the pre-epileptic period is defined as the time range from 180 seconds before the epileptic onset to the epileptic onset; correspondingly, the inter-epileptic period is the time range from the end of the epileptic onset to 180 seconds before the next epileptic onset.

[0077] Calculation data set EEG signal different recording sites and 4 low-frequency oscillation bands ( Wave, Wave, Wave and overall low frequency band) and 4 high frequency oscillation bands (low wave, high wave, ripple wave and overall high frequency band) under 16 frequency band combinations. The calculated features are divided into a training set and a test set and put into the SVM model for training; the model parameters after training are saved, and the test set is used to test the model performance to obtain an epilepsy prediction model based on cross-frequency coupling.

[0078] Specifically, training and testing the epilepsy prediction model according to the multi-site phase-amplitude coupling features includes:

[0079] The data set is randomly divided into 10 subsets. Each round, 9 subsets are used to train the model, and the remaining 1 subset is used to test the model, repeating 10 times in total. The classification results are evaluated through the classification confusion matrix of the model, and 5 model performance indicators, namely accuracy, sensitivity, specificity, precision, and F1 score, are calculated.

[0080] In the embodiment, the , and features of 10 channels and 16 frequency band combinations are put into the SVM model for training. After ten-fold cross-validation, 5 model performance indicators are obtained, and the calculation methods of the model performance indicators are as follows:

[0081] ;

[0082] ;

[0083] ;

[0084] ;

[0085] ;

[0086] Among them, is the true positive sample, that is, the pre-epileptic sample classified as pre-epileptic, is the true negative sample, that is, the inter-ictal sample classified as inter-ictal, is the false positive sample, is the false negative sample.

[0087] In the SVM classification results, there are differences in the classification performance indicators for different channel and different frequency band combinations. The channel and frequency band combination with the highest classification accuracy is selected as the optimal parameter combination of the model. The model parameters of the optimal parameter combination are saved for the subsequent personalized epilepsy prediction system. Different from the traditional single frequency band or fixed frequency band combination, the present invention proposes a dynamic screening mechanism for 16 cross-frequency band combinations, and for the first time introduces the full frequency band coupling analysis of 4 low frequencies ( / / / 1 - 12Hz) and 4 high frequencies (low / high / ripple / 30 - 200Hz) into the prediction of temporal lobe epilepsy, revealing the non-linear correlation between different frequency bands.

[0088] Step 4, use the prospective dataset, calculate the features of the preprocessed signal and input them into the pre-trained SVM model for classification.

[0089] Specifically, post-processing is a crucial step in the epilepsy prediction process. The main purpose is to optimize and adjust the original prediction results output by the model to improve the accuracy, reliability and practicality of the prediction. It usually includes data smoothing, threshold selection and other processing methods to reduce false alarms and missed alarms. Use the "k-of-n" module proposed by Parvez et al. to post-process the labels obtained by the real-time classification of the classifier. Mark 0 as the interictal state of epilepsy, and 1 as the pre-ictal state. If there are more than or equal to k output sample labels as 1 in n consecutive windows, then mark it as the pre-ictal state at this time and issue a warning, otherwise mark it as no warning in the interictal period.

[0090] In this embodiment, use the "8-of-10" module to post-process the labels output by the model based on the prospective dataset to obtain the false positive rate per hour (FPR / h), the successful prediction rate (True Predicted Percent, TPP) and the seizure prediction lead time (Prediction Horizon, PH). The calculation formulas for the false positive rate per hour and the successful prediction rate are as follows:

[0091] ;

[0092] .

[0093] The embodiment of the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it realizes the functions of the temporal lobe epilepsy prediction system in the above embodiment.

[0094] To verify the effect of the present invention, the temporal lobe epilepsy prediction system of the present invention is tested through the following experiments.

[0095] Six Sprague-Dawley rat models with lithium chloride-pilocarpine abdominal injection were used in this experiment. After inducing the rats into a continuous epileptic state, 45-μm nickel-chromium alloy electrodes were implanted at 10 sites in the bilateral hippocampus (CA1, CA3, DG, SUB) and the left and right amygdala (AMD) of the rats, and continuous monitoring was carried out for 6 to 8 weeks. The signal sampling frequency was 1000 Hz, and the resolution was 16 bits.

[0096] Pretreatment stage: Time annotation was performed on the collected EEG signals, which were divided into interictal, pre-epileptic, and epileptic seizure periods. 50-Hz power frequency notch filtering and 8-point mean filtering were performed on each channel. Figure 2 It shows the EEG signal of channel 5 (R-CA3) after filtering, which contains an epileptic seizure. The dataset was segmented into continuous windows with a window length of 5 s and an overlapping time of 1 s. The instantaneous phase in the low-frequency band and the instantaneous amplitude in the high-frequency band were extracted using complex wavelet transform for each window, and then the degree of phase-amplitude coupling in each window was calculated using the MI calculation formula to obtain the co-modulation map (as shown in the matrix diagram in Figure 2 ). The co-modulation map was segmented into sub-bands of δ wave (1-3 Hz), θ wave (3-8 Hz), α wave (8-12 Hz), and the sum of the overall low-frequency band (1-12 Hz), and low γ wave (30-50 Hz), high γ wave (50-120 Hz), ripple wave (120-200 Hz), and the overall high-frequency band (30-200 Hz). Then, through , and the feature calculation formula, the eigenvalue corresponding to 16 frequency band combinations can be obtained. Figure 2 It is the co-modulation map of channel 5 in experimental example 1 with a phase frequency range of 1-12 Hz and an amplitude frequency range of 120-200 Hz. The average coupling strength of the co-modulation map is the MCCV. The pentagram in the figure represents the average coupling center of the co-modulation map, and its corresponding horizontal and vertical coordinates are and .

[0097] The phase-amplitude coupling features of the 10 channels in the 1-12 / 30-200 Hz frequency band combination obtained after pretreatment , and were classified using the SVM model. Table 1 gives an example of the classification results of the 10 channels under the aforementioned experimental conditions. For experimental example 1, the channel with the highest accuracy (R-CA3) was selected to calculate the , and features of this channel for 16 frequency band combinations. Figure 3It shows the classification results of 16 frequency band combinations obtained by using the SVM classifier for this channel. For Experimental Example 1, the frequency band combination with the highest classification accuracy is 1 - 12 Hz / Ripple. According to the above experimental results, the parameter combination of the 1 - 12 Hz / Ripple frequency band combination using R - CA3 is saved for the online epilepsy prediction system of Experimental Example 1.

[0098] Table 1

[0099]

[0100] On the prospective dataset, each experimental example uses the personalized prediction model saved after pre - training for prediction, and the "8 - of - 10" module post - processing is performed on the prediction results.

[0101] Through the verification of this experiment, the specific prediction results of each experimental example are shown in Table 2. The prediction system of the present invention can achieve a maximum successful prediction of 87.5% of epileptic seizures, with an average of 5.68 ± 2.81 false alarms per hour, and an average prediction of epileptic seizures 63.47 ± 14.36 s in advance. Figure 4 A schematic diagram of the successful prediction of an epileptic seizure by the prediction model of the present invention is given.

[0102] Table 2

[0103]

[0104] The above - described embodiments have detailed the technical solutions and beneficial effects of the present invention. It should be understood that the above - described are only specific embodiments of the present invention and are not used to limit the present invention. Any modifications, supplements, and equivalent replacements made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A personalized prediction system for temporal lobe epilepsy based on multi-site cross-frequency coupling, characterized in that: include: A signal acquisition module, used to synchronously acquire EEG signals at multiple sites in the brain; The signal preprocessing module is used to synchronously preprocess the collected EEG signals, draw the co-modulation diagram of cross-frequency coupling, and divide the frequency bands; The feature extraction module is used to extract the phase amplitude coupling feature, including the average coupling center phase frequency, the average coupling center amplitude frequency and the average coupling strength; the calculation process of the phase amplitude coupling feature is: The four low-frequency bands and four high-frequency bands obtained by dividing the frequency bands are combined into 16 cross-frequency coupling modes, and the average coupling center phase frequency, average coupling center amplitude frequency and average coupling strength characteristics are further calculated, where: Average coupling center phase frequency : ； Average coupling center amplitude frequency : ； Average coupling strength characteristics : ； in, and represent the row and column indices of the modulation index matrix, respectively, Indicates time hour Position in the matrix ( ) value; and They are phase-frequency dimension and amplitude-frequency dimension respectively; A prediction module, which is used to input the phase amplitude coupling features into a pre-trained machine learning model and output the prediction results of epilepsy; The early warning module is used to generate early warning signals according to the prediction results.

2. The personalized prediction system for temporal lobe epilepsy based on multi-site cross-frequency coupling according to claim 1, characterized in that: Preprocess the collected EEG signals, including: Power frequency noise filtering, mean filtering and segmentation processing; The continuous EEG signal after notch filtering, power frequency noise removal and mean filtering is divided into processing windows according to a fixed time length, while ensuring that adjacent windows overlap for a certain time length, and the overlapping time length is less than the fixed division time length.

3. The personalized prediction system for temporal lobe epilepsy based on multi-site cross-frequency coupling according to claim 1, characterized in that: The process of drawing the co-modulation diagram of the cross-frequency coupling is as follows: The instantaneous phase of low-frequency oscillation and the instantaneous amplitude of high-frequency oscillation are extracted from the EEG signal in each processing window by using complex wavelet transform, and the modulation index is calculated using the modulation index calculation formula; the horizontal axis is the phase frequency and the vertical axis is the amplitude frequency, and the co-modulation diagram of the cross-frequency coupling of phase frequency-amplitude frequency between different frequency bands in the processing window is obtained; Wherein, the modulation index calculation formula is: ； in, The instantaneous phase of the low-frequency oscillation at time is defined as , the instantaneous amplitude of high-frequency oscillation is defined as ; The instantaneous phase frequency is , the instantaneous amplitude frequency is The modulation index at ; is an imaginary unit used to construct rotation vectors on the complex plane.

4. The personalized prediction system for temporal lobe epilepsy based on multi-site cross-frequency coupling according to claim 1, characterized in that: The divided frequency bands include: The low frequency oscillation is divided into 4 frequency bands, namely Wave, Wave, The overall low frequency band, high frequency oscillation is divided into 4 frequency bands, namely low wave, high waves, ripple waves and overall high frequency band.

5. The personalized prediction system for temporal lobe epilepsy based on multi-site cross-frequency coupling according to claim 1, characterized in that: The machine learning model is a support vector machine, which is used to classify the 16 combined feature inputs corresponding to all channels, verify the performance of screening the individual optimal prediction channel and frequency band combination, save the model parameters and output the prediction results of temporal lobe epilepsy.

6. The personalized prediction system for temporal lobe epilepsy based on multi-site cross-frequency coupling according to claim 1, characterized in that: The personalized prediction system for temporal lobe epilepsy also includes a post-processing module for optimizing the prediction results. The optimization process of the post-processing module is as follows: Post-process the output of the classifier in the machine learning model using a k-of-n module to label the pre-ictal state; According to the post-processing results, the hourly false alarm rate, prediction sensitivity and prediction lead time are obtained.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the functions of the personalized prediction system for temporal lobe epilepsy according to any one of claims 1 to 6 are realized.

Images