Epileptic seizure early-stage prediction method and system based on electrocardio and deterministic learning

Abstract

The invention discloses an epileptic seizure early-stage prediction method and system based on electrocardio and deterministic learning, and the method comprises the steps: 1, carrying out the preprocessing of the electrocardio data of the epileptic seizure early-stage, and dividing the data into three normal groups and a symptom group; 2, aiming at the original electrocardiosignals of different groups, adopting a discrete high-gain observer to obtain an electrocardiogram state, constructing a neural network according to a determined learning theory, extracting an electrocardiogram dynamic mode, and storing to form a normalized initial mode library; 3, designing a specific judgment rule and an abrupt change prompt signal based on a dynamic estimator, comparing the electrocardio state of the original electrocardio signal with the difference of different electrocardio state modes in the mode library, outputting a point-to-point signal change diagram, and then establishing inspection of the abrupt change prompt signal; and 4, carrying out combined experiment on the modes in different groups, and analyzing and visualizing the result difference of different mode selection. According to the method, the change of the electrocardiosignal before the epileptic seizure is well described, and the accurate time node of the mutation can be obtained.

Description

Technical Field

[0001] The invention belongs to the field of analysis of electrocardiogram signal changes before an epileptic seizure, and relates to using a discrete high-gain observer and deterministic learning theory to visualize and analyze changes in electrocardiogram signals before an epileptic seizure. Background Art

[0002] Epilepsy is a neurological disease that affects approximately 50 million people worldwide. Epilepsy often disrupts patients' lives, employment, and social life because of its sudden and unpredictable nature. In particular, patients with uncontrolled generalized tonic-clonic seizures (GTCS, a type of epileptic seizure) face a higher risk of sudden death from epilepsy. Studies have shown that changes in the autonomic nervous system during epileptic seizures also affect the ECG signals. Heart rate variability (HRV) is currently widely used to analyze ECG changes before and after epileptic seizures, and a large number of literatures have pointed out that there are significant differences in ECG signals before and during epileptic seizures. However, research on the evolution of ECG signals before epileptic seizures is still insufficient. Whether there is a mutation signal in the ECG signal before the onset is a question worth exploring.

[0003] As a typical nonlinear system, the heart has inherent nonlinear dynamics that deserve in-depth analysis in addition to HRV characteristics. Deterministic learning theory performs well in identifying recursive nonlinear systems with unknown dynamics, and can effectively obtain the internal dynamics of ECG. At the same time, dynamic pattern recognition methods for unknown system dynamics are widely used in fault detection, but there has been no relevant research in the field of ECG change analysis before an epileptic seizure. Based on discrete high-gain observers and deterministic learning theory, the present invention deeply explores the changes in electrocardiographic dynamics before an epileptic seizure, and finds that there are mutation signals in the ECG signal before an epileptic seizure, which provides a new research perspective and method for the field of epilepsy analysis, and also provides important reference information for clinicians. The present invention can also accurately capture the exact time of the mutation signal, which helps to optimize the management and intervention strategies for patients. Summary of the invention

[0004] The present invention provides a method and system for predicting the pre-epilepsy attack based on electrocardiogram and deterministic learning, especially a study on analyzing the electrocardiogram changes before an epileptic attack from the perspective of electrocardiodynamics, which is helpful to solve the specific application problems of epileptic attacks in clinical practice.

[0005] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0006] Step 1: pre-processing the ECG data in the pre-epilepsy period and dividing them into three normal groups and one symptom group; the pre-epilepsy period is the period from 20 minutes before the onset of the epilepsy to the onset of the epilepsy;

[0007] Step 2: For different groups of original ECG signals, a discrete high-gain observer is used to obtain the ECG state, and a neural network is constructed based on deterministic learning theory to extract the ECG dynamics pattern, which is finally saved uniformly to form a standardized initial pattern library;

[0008] Step 3: Based on the dynamic estimator, a specific judgment rule and mutation prompt signal are designed to compare the ECG state of the original ECG signal with the difference between different ECG state patterns in the pattern library, and output a point-to-point signal change graph, and then establish a verification link to ensure the accuracy of the mutation prompt signal;

[0009] Step 4: Randomly select any one ECG state pattern from each of the four groups of patterns for combined experiments, and analyze the differences in the changes in visualized ECG signals caused by different ECG state pattern selections; select different episodes of the same case to build a pattern library for repeated experiments, analyze the commonality and specificity of the experimental results, and thus develop an exclusive pattern library for each case data.

[0010] Furthermore, the specific implementation of step 1 is as follows:

[0011] 1-1. Prepare original ECG data: For patients with at least 2 attacks and an interval between attacks greater than 1 hour, obtain ECG data from 20 minutes before the attack to the end of the attack;

[0012] 1-2. Preprocess the original ECG signal using multiple filters: first, remove the power frequency interference through a 50Hz notch filter, then remove the baseline drift through a median filter, and then use a 1-70Hz bandpass filter to remove the interference of the electromyographic signal and noise; finally, perform normalization processing;

[0013] 1-3. The pre-processed data were divided into three normal groups and one seizure symptom group; the normal groups corresponded to the time periods of 15-20 minutes, 10-15 minutes, and 5-10 minutes before the seizure, respectively, while the seizure symptom group corresponded to the time period of 0-5 minutes before the seizure;

[0014] 1-4. Each group is divided into five 1-minute segments as ECG analysis units.

[0015] Furthermore, the specific implementation of step 2 is as follows:

[0016] 2-1. For each group of ECG analysis units, a discrete high-gain observer is used to estimate the ECG state; the observer is defined as:

[0017]

[0018] Among them, x(k) is the original single-lead ECG data, is the ECG state estimated by the observer, O(k) is the state vector of the observer; A d ,Bd ,C d ,D d is the designed observer parameter; k represents the kth moment;

[0019] 2-2. Normalize the ECG state estimated by the observer and combine it with the original ECG signal to form a two-dimensional signal;

[0020] 2-3. According to the deterministic learning theory, a radial basis function neural network is constructed for two-dimensional signals to identify the intrinsic dynamics of the ECG. The ECG dynamics obtained by the network are stored in the initial pattern library in the form of constants, thus obtaining three groups of normal patterns and one group of seizure symptom patterns, and each group contains 5 patterns, so there are 25 patterns in total in the four groups.

[0021] Furthermore, the specific implementation of step 3 is as follows:

[0022] 3-1. Prepare a new set of ECG data as test data. The data comes from the period before and after the attack of the same case. Also use a high-gain observer to obtain the ECG state of the test data and normalize it.

[0023] 3-2. Select one pattern from each of the four groups, and compare the ECG state of the test data with the four selected patterns through the dynamic estimator. The difference is expressed as the residual The expression is finally expressed as the mean residual norm; the smaller the mean residual norm is, the higher the similarity between the test data and the pattern is; Represents the ECG states of the four modes at time k;

[0024] 3-3. Input the ECG state of the test data into the dynamic estimator, compare the average residual norm of the test data with that of the four modes, and record the moment when the residual of the test data with the seizure symptom mode is lower than the residual of all normal modes with the test data. This moment is the signal mutation moment;

[0025] The normal mode refers to the ECG state of a mode randomly selected from the normal group;

[0026] The seizure symptom pattern refers to the electrocardiographic state of a pattern randomly selected from the seizure symptom group;

[0027] 3-4. Formulate a mutation prompt signal to show the changes in the ECG signal before an epileptic seizure; when the residual corresponding to any normal mode is lower than the seizure symptom mode, the mutation prompt signal is 0; if and only if the residual corresponding to the seizure symptom mode is lower than the residuals of all normal modes, the mutation prompt signal becomes 1;

[0028] 3-5. Establish a mutation warning signal inspection mechanism; for the mutation warning signal caused by the residual change caused by the short-term abnormal ECG fluctuation, the inspection mechanism shall be used to determine whether the previous mutation warning signal is a false alarm;

[0029] 3-6. Experiment with other episodes of the same case in sequence, and output the residual change graph and mutation signal prompt graph of the test data point by point;

[0030] Furthermore, the inspection mechanism of steps 3-5 is specifically inspected as follows:

[0031] The moment when the mutation prompt signal changes to 1 for the first time is taken as the initial mutation point; then, as time goes on, it is determined whether there is a situation where the mutation prompt signal changes to 0, that is, the residual of the seizure symptom mode is higher than the residual of at least one normal mode; if not, the initial mutation point is maintained; if so, the ratio of the duration of the mutation prompt signal changing to 0 to the duration from the current moment to the initial mutation point is calculated; if the ratio is greater than or equal to 0.5, it means that the mutation prompt signal is a false alarm, and a new mutation point needs to be found; if the ratio is less than 0.5, it indicates that it is only a short-term residual fluctuation that does not affect the mutation point, and the mutation point after this test is the final signal mutation point;

[0032] Furthermore, the specific implementation of step 4 is as follows: select a mode from each group of modes, Different combinations of

[0033] Furthermore, the present invention also provides an epileptic seizure pre-prediction system based on electrocardiogram and deterministic learning, which specifically includes the following modules:

[0034] Preprocessing module: preprocess the ECG data before epileptic seizure and divide them into three normal groups and one seizure sign group;

[0035] Pattern library construction module: for different groups of original ECG signals, a discrete high-gain observer is used to obtain the ECG state, and a radial basis function neural network is constructed based on deterministic learning theory to extract the ECG dynamics pattern, which is finally saved uniformly to form a standardized initial pattern library;

[0036] Residual recognition module: Based on the dynamic estimator, specific judgment rules and mutation prompt signals are designed to compare the ECG state of the original ECG signal with the difference between different ECG state patterns in the pattern library, and output a point-to-point signal change graph, and then establish a verification link to ensure the accuracy of the mutation prompt signal;

[0037] Combination analysis module: Randomly select any one ECG state pattern from each of the four groups of patterns for combination experiments, and analyze the differences in the selection of different ECG state patterns to visualize the changes in ECG signals; select different episodes of the same case to build a pattern library for repeated experiments, analyze the commonality and specificity of the experimental results, and thus develop an exclusive pattern library for each case data.

[0038] The beneficial effects of the present invention are as follows:

[0039] 1. The ECG changes before an epileptic attack were analyzed, proving that there is a sudden change in the ECG signal before an epileptic attack.

[0040] 2. Analyze from the perspective of electrocardiodynamics, not limited to the commonly used HRV features. It is proved that electrocardiodynamics can well describe the changes of ECG signals before epileptic seizures.

[0041] 3. The method based on discrete high-gain observer and deterministic learning theory can obtain the signal indicating ECG mutation within 10 minutes before the onset (mostly within 5 minutes before the onset), and can obtain the precise time node of the mutation.

[0042] 4. A pattern library dedicated to each case of data can be developed for clinical ECG diagnosis, assisting doctors in timely obtaining the patient's ECG mutation status. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Flow chart of an example of the present invention.

[0044] Figure 2 Schematic diagram of radial basis function neural network identification of cardiac electrodynamics.

[0045] Figure 3 Schematic diagram of residual results and mutation prompt signals of normal test data.

[0046] Figure 4 Schematic diagram of residual results and mutation prompt signals for special test data. DETAILED DESCRIPTION

[0047] Here, the specific implementation modes of the present invention are described in detail with reference to the accompanying drawings.

[0048] Figure 1 These are the relevant steps of the method for studying the pre-epilepsy attack based on electrocardiogram and determination learning, which have been introduced in detail in the content of the invention, that is, the technical solution of the present invention mainly includes the following steps:

[0049] Step 1: pre-processing the ECG data in the pre-epilepsy period and dividing them into three normal groups and one symptom group; the pre-epilepsy period is the period from 20 minutes before the onset of the epilepsy to the onset of the epilepsy;

[0050] Step 2: For different groups of original ECG signals, a discrete high-gain observer is used to obtain the ECG state, and a neural network is constructed based on deterministic learning theory to extract the ECG dynamics pattern, which is finally saved uniformly to form a standardized initial pattern library;

[0051] Step 3: Based on the dynamic estimator, a specific judgment rule and mutation prompt signal are designed to compare the ECG state of the original ECG signal with the difference between different ECG state patterns in the pattern library, and output a point-to-point signal change graph, and then establish a verification link to ensure the accuracy of the mutation prompt signal;

[0052] Step 4: Randomly select any one ECG state pattern from each of the four groups of patterns for combined experiments, and analyze the differences in the changes in visualized ECG signals caused by different ECG state pattern selections; select different episodes of the same case to build a pattern library for repeated experiments, analyze the commonality and specificity of the experimental results, and thus develop an exclusive pattern library for each case data.

[0053] The specific implementation of step 1 is as follows:

[0054] 1-1. Prepare original ECG data: For patients with at least 2 attacks and an interval between attacks greater than 1 hour, obtain ECG data from 20 minutes before the attack to the end of the attack;

[0055] 1-2. Preprocess the original ECG signal using multiple filters: first, remove the power frequency interference through a 50Hz notch filter, then remove the baseline drift through a median filter, and then use a 1-70Hz bandpass filter to remove the interference of the electromyographic signal and noise; finally, perform normalization processing;

[0056] 1-3. The pre-processed data were divided into three normal groups and one seizure symptom group; the normal groups corresponded to the time periods of 15-20 minutes, 10-15 minutes, and 5-10 minutes before the seizure, respectively, while the seizure symptom group corresponded to the time period of 0-5 minutes before the seizure;

[0057] 1-4. Each group is divided into five 1-minute segments as ECG analysis units.

[0058] The specific operations of step 2 are as follows:

[0059] 2-1. For the ECG signal with a sampling frequency of 1000 Hz, for each group of ECG units y r (k), use a discrete high-gain observer to estimate the ECG state. The observer is defined as: Where x(k) is the original single-lead ECG data, is the ECG state estimated by the observer, and O(k) is the state vector of the observer. Among the methods of discretizing high-gain observers, the observation error of the bilinear transformation method is small and can effectively guarantee the stability of mapping the continuous-time transfer function to the discrete-time transfer function. Therefore, according to the bilinear transformation method,

[0060] A d =(I+(α / 2)A o )(I-(α / 2)Ao ) -1 , B d =α(I-(α / 2)A o ) -1 H o ,

[0061] C d =D -1 (I-(α / 2)A o ) -1 , D d =(α / 2)C d H, A o =A-HC, α=T / ε.

[0062] Here, T=1000HZ is the sampling frequency of the original ECG. Settings:

[0063]

[0064] 2-2. Normalize the ECG state estimated by the observer and combine it with the original ECG signal to form a two-dimensional signal.

[0065] 2-3. According to the deterministic learning theory, a radial basis function neural network identifier is constructed for the two-dimensional signal composed of the original ECG and the ECG state to extract the electrocardiographic dynamics. The radial basis function neural network is a neural network that uses the radial basis function as the activation function. Its output is a linear combination of the network input and the radial basis function. The network model structure is as follows:

[0066]

[0067] in, is the regression vector of the radial basis function, is the ECG state of the input network, W=[w1,w2,…,w N ] T is the weight vector of the network, and N is the number of neurons. In deterministic learning, the radial basis function is the Gaussian function, and its expression is as follows:

[0068]

[0069] Among them, σ i (i=1,…,N) is the position of the neuron identifier in the network, and η is the effective interval of the identifier. As long as the appropriate identifier position and the spacing between adjacent identifiers are deployed, for any neural network, the signal trajectory in the local range can be accurately identified when N is large enough. Specifically, the dimension of the network is set to two dimensions, and the network boundary, the spacing between adjacent identifiers, the effective interval of the identifier and other parameters are set according to the boundary of the two-dimensional ECG signal to ensure that the deployed network identifier can fully cover the two-dimensional ECG signal, such as Figure 2 As shown. Based on deterministic learning theory, each identifier can accurately identify nearby sampled signals under the condition of continuous excitation (PE). The identifier formula is as follows:

[0070]

[0071] Among them, χ m (k) represents the state of the mth identifier at the kth moment, β is the design gain of the identifier, T is the sampling frequency, is the constructed radial basis function neural network, where is the weight of the network at the kth moment, is the regression vector of the network. The weight update law is:

[0072]

[0073] Where δ is the update gain. The electrocardiographic dynamics obtained by the network are expressed in the form of constant network weights. Saved in the initial pattern library, three groups of normal patterns and one group of seizure symptom patterns are obtained, and each group contains 5 pattern units.

[0074] The specific operations of step 3 are as follows:

[0075] 3-1. Prepare a new set of ECG data as test data, which is from before the other attack of the same case to the end of the attack. Also use a high-gain observer to obtain the ECG state of the test data.

[0076] 3-2. Select one pattern from each of the four groups of patterns and build a discrete dynamic estimator to compare the difference between the test data and the four patterns. Each pattern can build a corresponding dynamic estimator. The formula is as follows:

[0077]

[0078] in, represents the state of the mth dynamic estimator at the kth moment, represents the state of the test data, γ is the design gain of the estimator, is the constant network weight previously saved in the pattern library, is the calculated network regression vector. The difference between the test data and the four models can be expressed as residual Expression, finally expressed as the average residual norm, the formula is as follows:

[0079]

[0080] Among them, T eThe design parameters are preferably the period or multiple of the ECG state. If conditions permit, T e The larger the value, the smoother the residual graph trajectory generated. Since the average period of the ECG signal is 600ms, it is set to 12000 here.

[0081] 3-3. The smaller the average residual norm generated by the dynamic estimator, the higher the similarity between the test data and the corresponding pattern. Therefore, we focus on the change of the residual between the test data and the seizure symptom pattern. The signal mutation point is the moment when the residual between the test data and the seizure symptom pattern is lower than the residual between the test data and all normal patterns.

[0082] 3-4. Formulate a mutation signal. When the residual corresponding to any normal mode is lower than the seizure symptom mode, the mutation signal is 0; if and only if the residual corresponding to the seizure symptom mode is lower than the residual of all normal modes, the mutation signal becomes 1. Input the test data from 5 minutes before the seizure to the end of the seizure into the dynamic estimator, and visualize the average residual norm of the test data and the four modes, as shown in Figure 3 As shown. It can be clearly seen that before 197763ms, the residual of the test data with the normal mode is lower than the residual of the test data with the attack symptom mode. After 197763ms, the residual of the test data with the attack symptom mode is lower than the residual of the test data with the normal mode, indicating that from this moment on, the similarity between the test data and the attack symptom mode is the highest. Therefore, 197763ms (3.296min, i.e. 1.704min before the attack) is the mutation point of the ECG signal.

[0083] 3-5. Establish a mutation signal inspection mechanism. Since the ECG signal is a time series signal with periodic fluctuations and will be disturbed by external factors such as body movement and emotional fluctuations, the calculated residual will fluctuate accordingly. In order to ensure the accuracy of the mutation prompt signal and prevent the signal mutation caused by the residual change caused by short-term abnormal ECG fluctuations, special judgment rules are added to determine whether the previous mutation prompt signal is a false alarm. The moment when the mutation prompt signal changes to 1 is used as the initial mutation point. As time goes on, determine whether the signal changes to 0 (the residual of the symptom mode is higher than the residual of at least one normal mode). If not, the initial mutation point is maintained. If so, start calculating the ratio of the time it takes for the signal to change to 0 to the time from the current moment to the mutation point. Assuming the ratio is greater than or equal to 0.5, it means that the mutation prompt signal is a false alarm, and then look for a new mutation point. Assuming the ratio is less than 0.5, it means that it is only a short-term residual fluctuation that does not affect the mutation point. The mutation point after layer-by-layer inspection is the final signal mutation point. Figure 4As shown, the initial warning signal was issued at 195134ms, but at 202208-204077ms, perhaps due to daily actions, the residual of the seizure symptom mode was temporarily higher than the residual of normal mode 1. However, compared with the time when the mutation prompt signal was issued, the proportion of this recovery time was less than 0.5, so the initial mutation point continued to remain.

[0084] 3-6. Experiments were conducted one by one on the data of 7 patients with a total of 29 episodes, of which 7 episodes were used to build a dedicated pattern library for each case of data, and the remaining 22 episodes were used for testing. The results are shown in Table 1.

[0085] Table 1 Display of signal mutation points for each data case

[0086] serial number Number of attacks Signal mutation point (minutes) 1 3 1.57 / 1.08 2 4 4.29 / 3.95 / 1.19 3 7 3.13 / 3.3 / 4.68 / 2.61 / 2.08 / 4.46 4 4 1.47 / 3.75 / 1.12 5 2 1.03 6 2 6.69 7 7 2.61 / 1.70 / 6.04 / 0.8 / 1.82 / 4.27

[0087] As shown in the table above, each test of each case of data has a signal mutation before the attack, and the signal mutation time is within 10 minutes before the attack, and most of them are within 5 minutes before the attack. This proves that the present invention can effectively capture the mutation signal of the electrocardiogram before the epileptic attack and contain the signal prompt of the precise time.

[0088] The specific operations of step 4 are as follows:

[0089] 4-1. The key point of the present invention lies in the selection of patterns. Whether the selection of different patterns will affect the capture of mutation signals or the capture time of mutation signals is a question worth exploring. In the pattern library constructed in step 2, each of the four groups of patterns contains 5 patterns. Selecting a pattern from each group of patterns can have There are different combinations. The patterns are numbered 1-20 according to the distance from the onset time, that is, 15-20 minutes before the onset corresponds to 1-5, and 5 minutes before the onset corresponds to 15-20. Here, the results of case 7 are presented and analyzed in detail. Different groups of patterns are randomly selected to experiment on the same test data, and the results are shown in Table 2.

[0090] Table 2 Results of different mode combinations

[0091]

[0092]

[0093] As can be seen from Table 2, the signal mutation point of each test is within 5 minutes before the onset. The signal mutation point capture results of different mode combinations for the same test data are slightly different, and the difference is within 1 minute.

[0094] 4-2. Then, what are the differences and commonalities when constructing pattern libraries for different episodes of data and testing mutation signals with other episodes of data? This is an extended question. The 7 episodes of the 7th case of data are used to construct pattern libraries for testing the other 6 episodes. Record as many combinations as possible with signal mutations, and find the combinations with the most occurrences and list them in Table 3.

[0095] Table 3 Test results of different mode combinations at different times

[0096]

[0097]

[0098] Depend on Figure 3 It can be seen that for the same pattern library, different attack test data have different signal mutation times. For pattern libraries constructed with different attack times, the mutation signal times captured using the same test data are all within milliseconds. This shows that in terms of mutation signal capture, no matter which attack data is used to build the pattern library, and using different combinations, the results are not much different.

[0099] 4-3. Build a unique pattern library for each data. There are 625 different combinations, but testing all pattern combinations will take a lot of time, so it is necessary to find a specific combination for each data. As can be seen from Table 3, there are 3 combinations that appear more often, namely 1; 10; 15; 17, 5; 10; 14; 20, 5; 7; 14; 19. As long as the mutation signal can be accurately captured before the attack, it is a meaningful result. Therefore, for this data, no matter which attack is selected as training, selecting these 3 pattern combinations in the test can obtain the signal mutation point within 10 minutes (except for the second attack). Looking at the original ECG signal of this case, the second attack occurred at about 1 a.m., and the ECG amplitude of this attack was smaller and more stable than other attacks. This may be the reason why the pattern of the second attack is not easy to select and there are differences, but it is a minority case.

Claims

1. A method for predicting the pre-epilepsy attack based on electrocardiogram and deterministic learning, characterized by: Step 1: pre-processing the ECG data in the pre-epilepsy period and dividing them into three normal groups and one symptom group; the pre-epilepsy period is the period from 20 minutes before the onset of the epilepsy to the onset of the epilepsy; Step 2: For different groups of original ECG signals, a discrete high-gain observer is used to obtain the ECG state, and a neural network is constructed based on deterministic learning theory to extract the ECG dynamics pattern, which is finally saved uniformly to form a standardized initial pattern library; Step 3: Based on the dynamic estimator, a specific judgment rule and mutation prompt signal are designed to compare the ECG state of the original ECG signal with the difference between different ECG state patterns in the pattern library, and output a point-to-point signal change graph, and then establish a verification link to ensure the accuracy of the mutation prompt signal; Step 4: Randomly select any one ECG state pattern from each of the four groups of patterns for combined experiments, and analyze the differences in the changes in visualized ECG signals caused by different ECG state pattern selections; select different episodes of the same case to build a pattern library for repeated experiments, analyze the commonality and specificity of the experimental results, and thus develop an exclusive pattern library for each case data.

2. According to the method for predicting the pre-epilepsy attack based on electrocardiogram and deterministic learning according to claim 1, the specific implementation of step 1 is as follows: 1-1. Prepare original ECG data: For patients with at least 2 attacks and an interval between attacks greater than 1 hour, obtain ECG data from 20 minutes before the attack to the end of the attack; 1-2. Preprocess the original ECG signal using multiple filters: first, remove the power frequency interference through a 50Hz notch filter, then remove the baseline drift through a median filter, and then use a 1-70Hz bandpass filter to remove the interference of the electromyographic signal and noise; finally, perform normalization processing; 1-3. The pre-processed data were divided into three normal groups and one seizure symptom group; the normal groups corresponded to the time periods of 15-20 minutes, 10-15 minutes, and 5-10 minutes before the seizure, respectively, while the seizure symptom group corresponded to the time period of 0-5 minutes before the seizure; 1-4. Each group is divided into five 1-minute segments as ECG analysis units.

3. According to the method for predicting the pre-epilepsy attack based on electrocardiogram and deterministic learning according to claim 2, the specific implementation of step 2 is as follows: 2-1. For each group of ECG analysis units, a discrete high-gain observer is used to estimate the ECG state; the observer is defined as: in, x(k) is the original single-lead ECG data, is the ECG state estimated by the observer, O(k) is the state vector of the observer; A d ,B d ,C d ,D d is the designed observer parameter; k represents the kth moment; 2-2. Normalize the ECG state estimated by the observer and combine it with the original ECG signal to form a two-dimensional signal; 2-3. According to the deterministic learning theory, a radial basis function neural network is constructed for two-dimensional signals to identify the intrinsic dynamics of the ECG. The ECG dynamics obtained by the network are stored in the initial pattern library in the form of constants, thus obtaining three groups of normal patterns and one group of seizure symptom patterns, and each group contains 5 patterns, so there are 25 patterns in total in the four groups.

4. According to the method for predicting the pre-epilepsy attack based on electrocardiogram and deterministic learning according to claim 1 or 3, the specific implementation of step 3 is as follows: 3-1. Prepare a new set of ECG data as test data. The data comes from the period before and after the attack of the same case. Also use a high-gain observer to obtain the ECG state of the test data and normalize it. 3-2. Select one pattern from each of the four groups, and compare the ECG state of the test data with the four selected patterns through the dynamic estimator. The difference is expressed as the residual Expression, finally expressed as the mean residual norm; Represents the ECG states of the four modes at time k; 3-3. Input the ECG state of the test data into the dynamic estimator, compare the average residual norm of the test data with that of the four modes, and record the moment when the residual of the test data with the seizure symptom mode is lower than the residual of all normal modes with the test data. This moment is the signal mutation moment; The normal mode refers to the ECG state of a mode randomly selected from the normal group; The seizure symptom pattern refers to the electrocardiographic state of a pattern randomly selected from the seizure symptom group; 3-4. Develop mutation prompt signals to show changes in ECG signals before an epileptic seizure; 3-5. Establish a mutation warning signal inspection mechanism; for the mutation warning signal caused by the residual change caused by the short-term abnormal ECG fluctuation, the inspection mechanism shall be used to determine whether the previous mutation warning signal is a false alarm; 3-6. Experiment with other episodes of the same case in turn, and output the residual change graph and mutation signal prompt graph for the test data.

5. The method for predicting the pre-epilepsy attack based on electrocardiogram and deterministic learning according to claim 4, characterized in that: The mutation prompt signal in step 3-4 is as follows: When the residual corresponding to any normal mode is lower than that of the seizure symptom mode, the mutation prompt signal is 0; if and only if the residual corresponding to the seizure symptom mode is lower than the residuals of all normal modes, the mutation prompt signal becomes 1.

6. The method for predicting the pre-epilepsy attack based on electrocardiogram and deterministic learning according to claim 4 or 5, characterized in that: The specific inspection mechanism of steps 3-5 is as follows: The moment when the mutation prompt signal changes to 1 for the first time is taken as the initial mutation point; then, as time goes on, it is determined whether there is a situation where the mutation prompt signal changes to 0, that is, the residual of the seizure symptom mode is higher than the residual of at least one normal mode; if not, the initial mutation point is maintained; if so, the ratio of the duration when the mutation prompt signal changes to 0 to the duration from the current moment to the initial mutation point is calculated; if the ratio is greater than or equal to 0.5, it indicates that the mutation prompt signal is a false alarm and a new mutation point needs to be found; if the ratio is less than 0.5, it indicates that it is only a short-term residual fluctuation that does not affect the mutation point, and the mutation point after this test is the final signal mutation point.

7. According to the method for predicting the pre-epilepsy attack based on electrocardiogram and deterministic learning in claim 1, the specific implementation of step 4 is as follows: a pattern is selected from each group of patterns, Different combinations.

8. An epileptic seizure pre-prediction system based on electrocardiogram and deterministic learning, characterized in that: Includes the following modules: Preprocessing module: preprocess the ECG data before epileptic seizure and divide them into three normal groups and one seizure sign group; Pattern library construction module: for different groups of original ECG signals, a discrete high-gain observer is used to obtain the ECG state, and a radial basis function neural network is constructed based on deterministic learning theory to extract the ECG dynamics pattern, which is finally saved uniformly to form a standardized initial pattern library; Residual recognition module: Based on the dynamic estimator, specific judgment rules and mutation prompt signals are designed to compare the ECG state of the original ECG signal with the difference between different ECG state patterns in the pattern library, and output a point-to-point signal change diagram, and then establish a verification link to ensure the accuracy of the mutation prompt signal; Combination analysis module: Randomly select any one ECG state pattern from each of the four groups of patterns for combination experiments, and analyze the differences in the selection of different ECG state patterns to visualize the changes in ECG signals; select different episodes of the same case to build a pattern library for repeated experiments, analyze the commonality and specificity of the experimental results, and thus develop an exclusive pattern library for each case data.

Images