A Low-Power Wearable Epileptic Seizure Detection System Based on Multi-Level Classification

Abstract

The present invention discloses a low-power wearable epilepsy seizure detection system based on multi-level classification. First-level pre-classification is performed according to the characteristics of physiological signal data, and the obtained positive sample data is input into a detection model for fine second-level classification. Then, the final detection result is obtained by combining the first-level pre-classification and the second-level classification results, which can improve the detection accuracy; multiple modalities of physiological signals are synchronously collected to construct a stable epilepsy seizure detection model for the fusion of multiple modalities; data of healthy people's daily activities is added as negative samples to the data set used for constructing the classification model, making the constructed classification model more suitable for the actual life scenario and meeting the real needs of epilepsy patients; for the problem of data imbalance, a data imbalance processing solution based on the prior knowledge of human activities is adopted, and data imbalance processing is carried out by extracting the standard deviation of the resultant acceleration, the main frequency of the resultant acceleration in the frequency domain, the peak-to-peak value of the resultant acceleration, and setting empirical thresholds.

Description

Technical Field

[0001] The present invention belongs to the technical field of biomedical signal processing, and particularly relates to a low-power wearable epilepsy seizure detection system based on multi-level classification. Background Art

[0002] Epilepsy is a common brain dysfunction disease caused by various etiologies, characterized by persistent and spontaneous epileptic seizures. 70%-80% of epilepsy patients can suppress seizures through appropriate drug assistance, while the remaining 20%-30% of chronic or refractory epilepsy cannot be controlled by drugs and is called refractory epilepsy. Refractory epilepsy patients often face greater safety risks and psychological pressures. Generalized tonic-clonic seizure (GTCS) is one of the most dangerous types of epilepsy, often referred to as "grand mal epilepsy". Epileptic seizures will bring a series of physical and psychological impacts to patients and their families at the same time. When epileptic seizures occur, patients often show symptoms such as absence, uncontrolled body, and respiratory arrest, which often directly or indirectly lead to accidental injuries or even death of patients. The main cause of death related to epilepsy is sudden unexpected death in epilepsy (SUDEP).

[0003] The sudden onset of epilepsy poses a great safety risk to epilepsy patients. Especially when patients are alone, epileptic seizures are extremely likely to cause accidental injuries or even death of patients. If patients can receive timely treatment during epileptic seizures, the life safety of patients can be greatly guaranteed. According to the clinical guidelines of the International League Against Epilepsy (ILEA), wearing an epilepsy seizure monitoring device by patients, which can timely alarm and notify relevant personnel when seizures occur, and reduce the occurrence of respiratory dysfunction and hypoxemia caused by epileptic seizures, is currently the only definitely effective important method to reduce the occurrence of sudden unexpected death in epilepsy. In addition, currently, doctors mostly diagnose patients by the oral description of the patient's condition by the patient or their caregiver to evaluate the patient's condition and subsequent treatment. However, relying on the recollection of patients and their families to describe the condition often has many subjective factors, resulting in deviations in the diagnosis results. Detecting epileptic seizures based on wearable devices can not only ensure the safety of patients, but also record various information of patients during seizures in detail, providing more dimensional information for doctors to diagnose and treat, and comprehensively guaranteeing the health and safety of patients.

[0004] Since epileptic patients are no different from healthy people when not having seizures and the seizure duration of most epileptic patients only accounts for a very small part of their lives, patients and their families hope that seizure detection devices can accurately detect seizure events while minimally affecting their normal lives. Therefore, wearable seizure detection devices better meet the needs of patients. In recent years, there have been more and more lightweight wearable devices. Detecting seizures based on wearable devices can reduce the risk of patients getting injured and ensure their safety. At the same time, it can also protect the privacy of patients, reduce the stigma of the disease, and enable patients to live like healthy people.

[0005] Existing seizure detection methods are mainly divided into two categories. One is to monitor patients based on audio signals, and detect seizures by analyzing video and sound signals. The other is to collect and analyze the physiological parameters of patients through various human physiological signal acquisition devices, and then detect seizures in epileptic patients. Common physiological parameters collected include electroencephalogram (EEG) signals, electrocardiogram (ECG) signals, three-axis acceleration, three-axis gyroscope, skin conductance signals, electromyogram (EMG) signals, body temperature, photoplethysmogram, etc.

[0006] In existing implementation schemes, the method based on audio analysis restricts patients to be in an audio monitoring environment, interfering with their normal lives. At the same time, there are also problems such as large computational load and patient privacy leakage when detecting seizures in an audio monitoring environment. Although EEG signals are regarded as the gold standard for judging epileptic seizures clinically, the device is not portable and concealed enough, and the data collection process affects the normal lives of patients. The collection of the remaining physiological parameters is relatively convenient and can all be collected through wearable devices on the wrist of the human body. However, relying solely on motion signals is prone to false alarms due to interference from daily behaviors. ECG signals are easily interfered by other strenuous activities and have poor stability. There are many types of epilepsy, and there are significant differences in seizure manifestations among different patients. The physical changes of the same patient also vary during different seizure periods. Collecting a single physiological modality signal cannot comprehensively reflect the physiological state differences between epileptic patients during seizures and non-seizures, resulting in low accuracy and many false alarms of the detection model.

[0007] Most existing seizure detection methods are in the video electroencephalogram (vEEG) monitoring environment in the hospital. Although it can ensure accurate calibration of the dataset, the range of patient activities is restricted during data collection, that is, the collection scenario is too ideal. The data collected in the hospital cannot truly reflect the daily lives of patients, and the detection models designed based on in-hospital data often perform worse when applied to actual life scenarios. Patients hope more that seizure detection devices can ensure high accuracy and low false alarm rates in daily life.

[0008] Although some existing epilepsy seizure detection models for multi - physiological modality signals integrate multiple physiological signals, signal acquisition requires wearing multiple signal acquisition devices at multiple positions on the patient's body, which brings great inconvenience to the patient. In addition, increasing the signal modality will also increase the computational complexity of the model and the power consumption of the device. The existing multi - level classification methods mainly construct multiple classification models according to the hierarchical relationship. Although it can reduce the device power consumption to a certain extent by reducing the amount of calculation, it is necessary to train multiple classification models, increasing the complexity of model construction and adjustment.

[0009] Most epilepsy patients have a low frequency of epilepsy seizures, and the cumulative duration of epilepsy seizures only accounts for a very small part of daily life, that is, there is a serious imbalance between positive and negative samples in the data. Sample imbalance will seriously affect the performance of the detection model in practical applications. There are a large number of obvious non - epilepsy seizure behaviors in the daily behaviors of the human body, and some normal human behaviors similar to epilepsy seizures are the main reasons for more false alarms. Traditional sample imbalance processing methods such as random oversampling of a small number of samples and random undersampling of a large number of samples do not utilize the prior information of the daily behaviors of the human body, and the distribution of the original sample set is not changed after the imbalance processing. The detection model does not focus more on distinguishing normal behaviors that are very similar to epilepsy seizures, resulting in more false alarms of the model.

[0010] The existing technologies do not effectively unify the solution of data imbalance and the multi - level classification method. Conventional imbalance processing methods can only be used in the classification model construction stage and cannot be used in the actual application stage of the model. The multi - level classification method is mainly used in the actual application stage of the model and cannot help alleviate the data imbalance problem in the model construction stage. At the same time, multi - level classification will increase the complexity of building the model. Summary of the Invention

[0011] In view of this, the purpose of the present invention is to provide a low - power wearable epilepsy seizure detection system based on multi - level classification to improve the detection accuracy.

[0012] A low - power wearable epilepsy seizure detection system based on multi - level classification includes a wearable device and a mobile terminal;

[0013] The wearable device includes a real - time data acquisition module, an online raw data pre - processing module, an epilepsy seizure pre - classification module, and a data real - time wireless transmission module;

[0014] The real - time data acquisition module is used to collect physiological signal parameter data of the user's wrist, including: three - axis acceleration, three - axis gyroscope, surface electromyogram, skin conductance, and body temperature;

[0015] The online raw data pre - processing module is used to pre - process each physiological signal parameter data;

[0016] The epilepsy seizure pre-classification module makes a preliminary judgment on the pre-processed physiological signal parameter data, discriminates epileptic seizure sample data and non-seizure sample data, and obtains the primary detection result of epileptic seizures;

[0017] The data real-time wireless transmission module sends the seizure sample data detected by the epilepsy seizure pre-classification module and the preliminary judgment result to the mobile terminal;

[0018] The mobile terminal includes a secondary epilepsy seizure detection module, an epilepsy seizure detection result decision module, and an epilepsy seizure status warning module;

[0019] The secondary epilepsy seizure detection module includes an epilepsy detection model trained using training samples composed of epileptic seizure data and physiological signal parameter data during non-seizure; this epilepsy detection model is used to identify the seizure sample data sent by the data real-time wireless transmission module to determine whether it is an epileptic seizure or non-seizure, and obtain the secondary detection result of epileptic seizures;

[0020] The epilepsy seizure detection result decision module is used to make a joint judgment based on the primary judgment result sent by the data real-time wireless transmission module and the secondary detection result obtained by the secondary epilepsy seizure detection module to obtain the final result of whether it is an epileptic seizure.

[0021] Preferably, the method for the epilepsy seizure pre-classification module to make a preliminary judgment on the pre-processed physiological signal parameter data is as follows:

[0022] Obtain the standard deviation of the resultant acceleration, the main frequency in the frequency domain of the resultant acceleration, and the peak-to-peak value of the resultant acceleration from the three-axis acceleration data, and determine whether all three are greater than the set threshold. If all are greater, it is determined as epileptic seizure sample data; otherwise, it is determined as non-seizure sample data.

[0023] Preferably, the online raw data pre-processing module obtains each piece of data by performing a sliding window cut on the physiological signal parameter data, with a window length of 10 s and a step length of 2.5 s.

[0024] Furthermore, the online raw data pre-processing module also performs filtering on the data after the sliding window cut.

[0025] Preferably, the construction method of the epilepsy detection model in the secondary epilepsy seizure detection module is as follows:

[0026] Collect the seizure data of epileptic patients as the positive sample data of the training set, and the data of healthy people as the negative sample data of the training set to obtain the training sample data;

[0027] Feature extraction and feature dimensionality reduction are sequentially performed on the training sample data, and then the training data after feature extraction is used to construct and optimize a random forest classification model to obtain an epilepsy detection model.

[0028] Preferably, after performing data imbalance processing on the training sample data, it is used to construct and optimize a random forest classification model. The data imbalance processing method is as follows: obtain the standard deviation of the resultant acceleration, the main frequency of the resultant acceleration in the frequency domain, and the peak-to-peak value of the resultant acceleration based on the triaxial acceleration data, and determine whether all three are greater than the set threshold. If all are greater, retain them as the training sample data; otherwise, eliminate them.

[0029] Preferably, before performing data imbalance processing on the training sample data, the original training data is first obtained, specifically:

[0030] The original physiological signal parameter data is subjected to sliding windowing to obtain the original training data, where the step length of the positive samples is less than that of the negative samples, and the window overlap rate of the positive samples is greater than that of the negative samples.

[0031] Further, before performing data imbalance processing on the training sample data, it also includes filtering processing of the original training data.

[0032] Preferably, for the data real-time wireless transmission module, for the data detected as non-seizure data, the classification result identifier is sent to the mobile terminal; when multiple consecutive samples are pre-classified as seizure data, the complete data of the first sample is sent to the mobile terminal, and the subsequent samples send the partial data intercepted from the back of this data.

[0033] Preferably, the method for the epilepsy seizure detection result decision module to perform joint judgment is as follows:

[0034] When the detection result of a piece of data detected by the pre-classification module is non-seizure, the final detection result of this piece of data is non-seizure, and the detection result value is marked as 0; when the detection result of a piece of data detected by the pre-classification module is seizure, it is sent to the epilepsy seizure detection result decision module of the mobile terminal for secondary detection, and the detection result of this module is used as the final detection result of this piece of data;

[0035] Accumulate the sum value Sn of the final detection results of the current data and the previous n - 1 pieces of data, and at the same time accumulate the sum value Sm of the final detection results of the current data and the previous m - 1 pieces of data, and make a judgment:

[0036] When Sn is greater than the first set threshold Th n or Sm is greater than the second set threshold Th m then it is considered that a seizure has been detected;

[0037] where n < m, n is at least 1, and m is at least 5.

[0038] The present invention has the following beneficial effects:

[0039] First, the present invention performs a first-level pre-classification based on the characteristics of physiological signal data. The obtained positive sample data is input into a detection model for fine second-level classification. Then, by combining the results of the first-level pre-classification and the second-level classification, the final detection result can be obtained, improving the detection accuracy.

[0040] Multiple modalities of physiological signals on the wrist are synchronously collected, and a stable epilepsy seizure detection model can be constructed through the fusion of multiple modalities. The present invention adds the data of healthy people's daily activities as negative samples to the data set for constructing the classification model, making the constructed classification model more suitable for the actual life scenario and meeting the real needs of epilepsy patients.

[0041] Regarding the data imbalance problem, the present invention proposes a solution for data imbalance processing based on the prior knowledge of human activities. The data imbalance is processed by extracting the standard deviation of the resultant acceleration, the main frequency of the resultant acceleration in the frequency domain, the peak-to-peak value of the resultant acceleration, and setting empirical thresholds.

[0042] The present invention sets a result decision strategy based on the cumulative posterior probability, sets two different ranges of continuous sample windows and calculates the corresponding cumulative posterior probabilities. The shorter sample window judges the state of epilepsy patients from a short-time perspective, and the longer sample window judges the patient's state from a longer-time perspective. The decision strategy of two different cumulative posterior probabilities is used to avoid the interference of sudden abnormal behaviors.

[0043] The first-level pre-classification is deployed on the wearable watch, and the second-level epilepsy detection model is deployed on the mobile phone. The wearable watch performs wireless transmission of data based on the results of the first-level pre-classification, thereby reducing the amount of wireless data transmission to achieve the purpose of reducing power consumption. Description of the Drawings

[0044] Figure 1 Is the block diagram of the wearable device;

[0045] Figure 2 Is the flowchart of the threshold judgment for data imbalance processing;

[0046] Figure 3 Is the flowchart of the online data sliding window;

[0047] Figure 4 Is the block diagram of the mobile terminal;

[0048] Figure 5 Is the construction process of the second-level epilepsy seizure detection model;

[0049] Figure 6 Result decision flowchart. Detailed implementation manners

[0050] The present invention will be described in detail below in conjunction with the accompanying drawings and by way of examples.

[0051] The present invention synchronously collects various physiological modality signals of a subject's wrist based on a single wearable wristwatch for analysis and constructs a multimodal seizure detection model. The physiological modality signals synchronously collected on the single wristwatch include three-axis acceleration signal (ACC), three-axis gyroscope signal (GYR), surface electromyogram signal (sEMG), skin conductance value (EDA), and body temperature signal (Temp). In addition, attitude calculation is performed based on the collected three-axis acceleration signal and three-axis gyroscope signal, and the pitch angle and roll angle of the wristwatch are synchronously calculated. The signals in each dimension are sequentially subjected to sliding windowing, filtering and denoising, feature extraction, and feature fusion, and then a seizure detection model is constructed based on the training set.

[0052] The present invention constructs a seizure detection model respectively based on seizure data of epileptic patients and data in the daily life of healthy people. Among them, the seizure data of epileptic patients is used as positive samples, and since the daily behaviors of healthy people definitely do not include seizure behaviors, they are used as negative samples. The seizure data of epileptic patients is calibrated by doctors to ensure the accuracy of the positive sample data. The data of healthy people is collected during the daily life of the subjects, and no constraints are imposed on the behaviors of the subjects to ensure that the data can more comprehensively reflect real daily activities. Using the multimodal data of healthy people instead of the non-seizure data of epileptic patients adds a large number of complex behaviors in daily life to the negative samples, making the seizure detection model closer to the actual application scenario.

[0053] Aiming at the imbalance of epileptic data, the present invention combines the statistical prior knowledge of human daily behaviors to extract feature indicators and set judgment thresholds to achieve the imbalance processing of the training data set. A small number of feature indicators are extracted from the training set and empirical thresholds are set for threshold judgment, so as to eliminate a large number of obvious non-seizure behaviors while retaining the seizure data, thereby achieving the purpose of balancing the training set data.

[0054] In view of the problem of large computational complexity of multi-modal signals, the present invention extracts the same features as those in data imbalance processing to construct a multi-level classification model. The first-level pre-classification model is deployed on the wearable wristwatch. The same feature indicators as those in feature imbalance processing are extracted and thresholds are set. If the data samples do not meet the feature threshold indicators, they will not be transmitted to the mobile phone. Only when the data meets the feature threshold indicators will the data be wirelessly transmitted to the mobile phone. The second-level seizure detection model is deployed on the mobile phone to make a more refined discrimination on the data transmitted from the wearable wristwatch. Through the pre-classification process with a small amount of calculation in the first level, a large number of non-seizure behaviors are excluded first, reducing the computational amount of the model and the power consumption of device wireless transmission. Through the second-level detection model, accurate detection of the patient's seizure state is achieved.

[0055] In the data imbalance processing stage and the first level of multi-level classification, the present invention adopts the same feature indicators and feature thresholds, making the construction of the seizure detection model highly consistent with the deployment of the model, which not only alleviates data imbalance but also reduces the computational complexity of the overall model.

[0056] (1) Wearable wristwatch side

[0057] As Figure 1 shown, the wearable device mainly includes a real-time data acquisition module for user data, an online preprocessing module for raw data, a seizure pre-classification module, and a real-time wireless transmission module for user data.

[0058] (1.1) Real-time data acquisition module

[0059] The data acquisition module synchronously acquires various physiological signal parameters of the user's wrist, including: three-axis acceleration (ACC, 50Hz), three-axis gyroscope (GYR, 50Hz), surface electromyogram (sEMG, 200Hz), skin conductance (EDA, 4Hz), body temperature (Temp, 0.5Hz). The raw data acquired in real time will be sent to the subsequent data preprocessing module for real-time processing.

[0060] (1.2) Online raw data preprocessing module

[0061] The data preprocessing on the wearable wristwatch side includes two parts: online data sliding windowing and raw data filtering.

[0062] (a) Perform sliding windowing on the continuously acquired data to segment the continuous data and obtain each sample data;

[0063] As Figure 3 shown is the online data sliding windowing process. The wearable wristwatch forms a 10s sample window by combining the current 2.5s data with the previously acquired 7.5s data for subsequent processing every 2.5s of data acquisition.

[0064] (b) Motion signal filtering

[0065] The originally collected motion signals include triaxial acceleration signals and triaxial gyroscope signals. Attitude calculation is performed on the collected motion signals to obtain synchronized roll angle and pitch angle information. The motion frequency of the human wrist is relatively low, and the actually collected motion signals are interfered by high-frequency noise, so the motion signals need to be denoised. A sliding mean filter is used to remove the interference of high-frequency noise, and the selected window length of the sliding mean filter is 0.1 s.

[0066] (c) Skin conductance signal filtering

[0067] The change of skin conductance value is slow, and the main energy is distributed in the frequency band below 0.2 Hz. However, the actually sampled skin conductance signal contains high-frequency jitter, so it also needs to be denoised. A sliding mean filter is used to remove the high-frequency noise of skin conductance, and the selected sliding window length of the sliding mean filter is 1 s.

[0068] (d) Epidermal electromyogram signal filtering

[0069] The epidermal electromyogram signal is easily interfered by movement artifact noise and power frequency noise. High-pass filtering (filter cut-off frequency is 10 Hz) is performed on the collected surface electromyogram (sEMG) data to remove movement artifact noise, and a 50 Hz power frequency notch filter is used to remove the interference of power frequency noise.

[0070] (1.3) Epileptic seizure pre-classification module

[0071] The epileptic seizure pre-classification module on the wearable wristwatch extracts the standard deviation of the combined acceleration, the main frequency in the frequency domain of the combined acceleration, and the peak-to-peak value of the combined acceleration and performs threshold judgment to pre-classify the data. The threshold judgment process is as Figure 2 shown.

[0072] (a) Standard deviation of combined acceleration

[0073] Based on the triaxial acceleration signals after filtering and denoising, the combined acceleration signal is calculated. The combined acceleration can reflect the amplitude of motion. The standard deviation of the combined acceleration of each sample window is extracted as a characteristic index, and the unit is gravitational acceleration g (9.8 m / s2). Based on the prior knowledge of human behavior, this characteristic is: when epileptic patients have seizures, it is manifested as violent shaking of the wrist, while in some daily human behaviors, the wrist is close to static. Therefore, a threshold of the standard deviation of the combined acceleration can be set to eliminate behaviors with small activity amplitudes. The set threshold of the standard deviation of the combined acceleration is 0.2 g, and when it is greater than 0.2 g, it is considered to meet the threshold condition.

[0074] (b) Main frequency in the frequency domain of combined acceleration

[0075] The detrended filter is applied to the calculated resultant acceleration, and then the discrete Fourier transform (DFT) is performed to obtain the frequency-domain information of the signal. The detrended filter is mainly used to remove the baseline and thus obtain the main change information. The frequency value corresponding to the maximum amplitude in the frequency domain is calculated as the pre-judgment index, with the unit of Hz. The prior human knowledge on which this characteristic index is based is that during an epileptic seizure, the wrist of the patient twitches, showing higher-frequency jitters, while the frequencies of some daily movements are relatively low. The set threshold for the main frequency in the frequency domain of the resultant acceleration is 2 Hz. When the main frequency in the frequency domain is greater than 2 Hz, it is considered to meet the threshold condition.

[0076] (c) Peak-to-peak value of the resultant acceleration

[0077] The peak-to-peak value of the calculated resultant acceleration is extracted as the characteristic index. The peak-to-peak value can reflect the maximum range of the resultant acceleration fluctuation within a sample window, with the unit of gravitational acceleration g (9.8 m / s²). The prior human knowledge on which this characteristic is based is that during a generalized tonic-clonic seizure, a large range of wrist movements occur, while the wrist movement range of some daily human behaviors is relatively small. A threshold for the peak-to-peak value of the resultant acceleration is set to exclude behaviors with a small movement range. The set threshold for the peak-to-peak value of the resultant acceleration is 0.8 g. When it is greater than 0.8 g, it is considered to meet the threshold condition.

[0078] When the data collected in real time does not meet any of the characteristic indexes, the data of this record is determined as a non-seizure sample. When all three indexes in a piece of data meet the characteristic index threshold conditions, the data is sent to the subsequent epileptic seizure detection model for classification and judgment.

[0079] (1.4) Data real-time wireless transmission module

[0080] The data real-time wireless transmission module transmits data according to the classification results of the pre-classification module. The pre-classification module pre-classifies the multi-modal physiological data collected in real time into seizure and non-seizure. Among the data classified as seizure, there is still a large amount of non-seizure data similar to epileptic seizures, which still needs to be more finely discriminated by the second-level classification model on the mobile phone side. The data classified as non-seizure is behavior that is very different from epileptic seizures and does not require subsequent discrimination. When the data is detected as seizure by the pre-classification module, the data is real-time wirelessly transmitted to the mobile phone side for subsequent judgment.

[0081] When the data sample window is detected as non-seizure by the pre-classification module, the original data is no longer transmitted, and only the pre-classification result identifier is transmitted. When multiple consecutive pieces of data are pre-classified as seizure, to avoid duplicate data transmission, the first piece of data transmits the complete 10 s of data, and the subsequent data only transmits the last 2.5 s of data. As Figure 3As shown in the figure, when the (n - 1)-th data and the n-th data are continuously pre-classified as seizures (the (n - 2)-th data is pre-classified as non-seizure and not transmitted), the (n - 1)-th data window transmits 10s of complete data, and the n-th data only transmits its last 2.5s of data. Data transmission is carried out according to the results of the pre-classification module, avoiding the transmission of a large amount of data of obvious non-epileptic seizure behaviors, reducing the overall data transmission volume of the wristwatch, and thus reducing the power consumption of the device.

[0082] (2) Mobile terminal

[0083] The mobile phone receives the preprocessed multi-modal physiological parameter data or the pre-classification result identification transmitted by the wristwatch. As Figure 4 shown, the mobile phone terminal includes a second-level epileptic seizure detection module (used to perform a second-level detection on the data retained by the wristwatch pre-classification), an epileptic seizure detection result judgment module, and an epileptic seizure status warning module. The specific functions of each module are as follows.

[0084] (2.1) Second-level epileptic seizure detection module

[0085] The mobile phone performs a second-level epileptic seizure detection on the user using the epileptic seizure detection model to obtain the second-level detection result. The second-level epileptic seizure detection module makes a more refined discrimination of the user's behavior, ensuring the accuracy rate while reducing the false alarm rate.

[0086] Among them, as Figure 5 shown, the process of constructing the epileptic seizure detection model is as follows:

[0087] (2.1.1) Original data preprocessing:

[0088] Use the seizure data of epileptic patients as the positive samples of the training set, and the data of healthy people as the negative samples of the training set. Since there are no restrictions on the data collection scenarios and activity types when collecting data from healthy people, a large number of daily behaviors are included in the collected negative samples. By adding a large number of daily behaviors in real-life scenarios, the epileptic seizure detection model is more applicable in daily life. The original data preprocessing module mainly performs offline preprocessing on the collected multi-modal data, mainly including data sliding windowing and original signal filtering. Among them, the original signal filtering includes motion signal filtering, skin conductance signal filtering, and surface electromyogram signal filtering. A set of denoised multi-modal data samples is obtained through the original data preprocessing module.

[0089] The original data preprocessing includes:

[0090] The positive and negative samples in the original training set are respectively subjected to sliding windowing to obtain positive and negative sample sets. Among them, the positive samples are intercepted with a window length of 10s and a step length of 1s (the sliding window overlap rate is 90%), and the negative samples are intercepted with a window length of 10s and a step length of 10s (the sliding window overlap rate is 0%).

[0091] Then, filter the data samples, and perform moving average filtering on the motion signal, moving average filtering on the skin conductance signal, and filtering and denoising on the surface electromyogram signal in sequence. The methods and parameters used for filtering each physiological modality data are the same as those used in the filtering of the online raw data preprocessing module on the wristwatch side.

[0092] (2.1.2) Further process data imbalance:

[0093] The duration of seizure data is much shorter than that of non-seizure behaviors. Although the data imbalance degree has been effectively reduced during the sliding window stage of data preprocessing, the proportion of data in the training set is still extremely disparate. Therefore, it is necessary to further solve the data imbalance problem. The method adopted is the same as that used in the seizure pre-classification module.

[0094] The present invention extracts three characteristic indexes of triaxial acceleration data based on human prior knowledge, and eliminates a large number of obvious non-seizure negative samples through threshold judgment to achieve further data imbalance processing. The data retained by the preprocessing module includes seizure data and non-seizure data similar to seizure behaviors, and these data constitute the training set for constructing the subsequent seizure detection model.

[0095] (2.1.3) Construct and optimize the detection model

[0096] By processing data imbalance, a large number of obvious non-seizure behavior data in the training set are eliminated, and the remaining data are seizure data and daily behavior data similar to seizures. Use the training set retained after data imbalance processing to construct a seizure detection model. As shown in the block diagram of the seizure detection model construction module in Figure 5 , perform feature extraction, feature dimensionality reduction, construction and optimization of a random forest classification model on each data sample window respectively. Extract the time domain, frequency domain, time-frequency domain, and non-linear features of each modality data of each sample window respectively, and use the maximum relevance minimum redundancy (mRMR) algorithm to select features to obtain the feature set of the training set data. Construct a seizure detection model based on the random forest (RF) model and perform grid search on the parameters to optimize the detection model. A more refined seizure detection model is obtained through the seizure detection model construction module to distinguish some normal behaviors of the user similar to seizure behaviors.

[0097] (2.2) Seizure detection result decision module

[0098] The pre-classification and seizure detection model classification steps continuously output the posterior probability values of each piece of data (the posterior probability of samples directly excluded by pre-classification is regarded as 0). By comparing with the threshold, the detection result of a single piece of data can be obtained. However, the seizure of an epileptic patient is an event, and the duration of the seizure event is not fixed. The detection result of a single sample cannot fully reflect the human body's activity state. It is necessary to design a decision-making strategy to integrate multiple samples to obtain the final judgment result of the epileptic seizure event. The epileptic seizure detection result decision module performs threshold judgment based on the double cumulative posterior probability and outputs the state judgment result of the patient. Two consecutive numbers of cumulative posterior probabilities are set for judgment. Figure 6 This is the result decision-making process. When the detection result of a piece of data in the pre-classification module is non-seizure, this piece of data is not sent to the mobile terminal, and the final detection result of this piece of data is non-seizure, and the detection result value is marked as 0; when the detection result of a piece of data in the pre-classification module is seizure, it is sent to the epileptic seizure detection result decision module of the mobile terminal for secondary detection, and the detection result of this module (a probability value output by the detection model) is used as the final detection result of this piece of data; accumulate the sum value Sn of the detection results of the previous n - 1 pieces of data of the current data. When Sn is greater than the set threshold Th n then it is considered that a seizure has been detected. To detect seizures on a longer time scale, a value m larger than n is set, and the sum value Sm of the detection results of the previous m - 1 pieces of data of the current data is accumulated. When Sm is greater than the set threshold Th m then it is considered that a seizure has been detected. If Sn is greater than the set threshold Th n and Sm is greater than the set threshold Th m at the same time, it is considered to be the same seizure. In the present invention, n is taken as 3 at least, m is taken as 12 at least, n < m, n = 3, m = 12. The consecutive n sample windows are designed to judge the state of the patient from a short time segment, and the consecutive m sample windows are designed to judge the state of the patient from a long time segment.

[0099] Considering that there may be phenomena such as "severe seizure - seizure stop - severe seizure" during the same seizure process of an epileptic patient and some status epilepticus (seizure duration is relatively long), when the distance between two detected epileptic seizure events is less than 5 minutes, the two detected seizure events are regarded as the same seizure.

[0100] (2.3) Epileptic seizure status alarm module

[0101] When the epileptic seizure detection result decision module outputs a result indicating that a seizure event has been detected, the epileptic seizure status alarm module will be triggered. The epileptic seizure status alarm module uses sound and light alarms to remind people around to give timely help to the patient and ensure the safety of the patient.

[0102] In summary, the above are only the preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A low-power wearable epilepsy seizure detection system based on multi-level classification, characterized in that, It includes a wearable device and a mobile terminal; The wearable device includes a real-time data acquisition module, an online raw data preprocessing module, a seizure pre-classification module, and a real-time data wireless transmission module; The real-time data acquisition module is used to collect physiological signal parameter data of the user's wrist, including: three-axis acceleration, three-axis gyroscope, surface electromyogram, skin conductance, and body temperature; The online raw data preprocessing module is used to preprocess each physiological signal parameter data; The seizure pre-classification module makes a preliminary judgment on each preprocessed physiological signal parameter data, discriminates seizure sample data and non-seizure sample data, and obtains a primary detection result of seizure; The real-time data wireless transmission module sends the seizure sample data detected by the seizure pre-classification module and the preliminary judgment result to the mobile terminal; The mobile terminal includes a secondary seizure detection module, a seizure detection result judgment module, and a seizure status warning module; The secondary seizure detection module includes a seizure detection model trained by using training samples composed of seizure data and physiological signal parameter data during non-seizure; this seizure detection model is used to identify the seizure sample data sent by the real-time data wireless transmission module, and determine whether it is a seizure or non-seizure, and obtain a secondary detection result of seizure; The seizure detection result judgment module is used to make a joint judgment based on the primary judgment result sent by the real-time data wireless transmission module and the secondary detection result obtained by the secondary seizure detection module, and obtain the final result of whether it is a seizure; The construction process of the seizure detection model in the secondary seizure detection module is as follows: Collect seizure data of epilepsy patients as positive sample data of the training set, and data of healthy people as negative sample data of the training set to obtain training sample data; Successively perform feature extraction and feature dimensionality reduction on the training sample data, and then use the training data after feature extraction to construct and optimize a random forest classification model to obtain a seizure detection model.

2. The low-power wearable epilepsy seizure detection system based on multi-level classification according to claim 1, wherein, The process of the seizure pre-classification module making a preliminary judgment on each preprocessed physiological signal parameter data includes: Obtain the standard deviation of the resultant acceleration, the main frequency of the resultant acceleration in the frequency domain, and the peak-to-peak value of the resultant acceleration according to the three-axis acceleration data, and judge whether all three are greater than the set threshold. If all are greater, it is determined as seizure sample data, otherwise it is determined as non-seizure sample data.

3. The low-power wearable epilepsy seizure detection system based on multi-level classification according to claim 2, characterized in that, The online raw data preprocessing module obtains each data by performing a sliding window on the physiological signal parameter data, with a window length of 10 s and a step length of 2.5 s.

4. The low-power wearable epilepsy seizure detection system based on multi-level classification according to claim 3, wherein The online raw data preprocessing module also performs filtering on the data after the sliding window processing.

5. The low-power wearable epilepsy seizure detection system based on multi-level classification according to claim 1, characterized in that After performing data imbalance processing on the training sample data, then use it to construct and optimize a random forest classification model. Among them, the data imbalance processing method is: obtain the standard deviation of the resultant acceleration, the main frequency of the resultant acceleration in the frequency domain, and the peak-to-peak value of the resultant acceleration according to the three-axis acceleration data, and judge whether all three are greater than the set threshold. If all are greater, it is retained as training sample data, otherwise it is eliminated.

6. A low-power wearable epilepsy seizure detection system based on multi-level classification according to claim 1 or 5, characterized in that, Before performing data imbalance processing on the training sample data, the original training data is first obtained, specifically: The original physiological signal parameter data is subjected to sliding windowing to obtain the original training data, where the step length of the positive samples is less than that of the negative samples, and the window overlap rate of the positive samples is greater than that of the negative samples.

7. The low-power wearable epilepsy seizure detection system based on multi-level classification according to claim 6, characterized in that Before performing data imbalance processing on the training sample data, it also includes filtering the original training data.

8. The low-power wearable epilepsy seizure detection system based on multi-level classification according to claim 1, wherein For the data real-time wireless transmission module, for the data detected as non-seizure data, the classification result identifier is sent to the mobile terminal; when multiple consecutive samples are pre-classified as seizure data, the complete data of the first sample is sent to the mobile terminal, and the subsequent samples send the part of the data intercepted from the back.

9. The low-power wearable epilepsy seizure detection system based on multi-level classification according to claim 1, wherein, The method for the epilepsy seizure detection result decision module to make a joint judgment is as follows: When the detection result of a piece of data detected by the pre-classification module is non-seizure, the final detection result of this piece of data is non-seizure, and the detection result value is marked as 0; when the detection result of a piece of data detected by the pre-classification module is seizure, it is sent to the epilepsy seizure detection result decision module of the mobile terminal for secondary detection, and the detection result of this module is used as the final detection result of this piece of data; The sum value Sn of the final detection results of the current data and the previous n - 1 pieces of data is accumulated, and at the same time, the sum value Sm of the final detection results of the current data and the previous m - 1 pieces of data is accumulated, and a judgment is made: When Sn is greater than the first set threshold Th n or Sm is greater than the second set threshold Th m then it is considered that an attack has been detected; Where n < m, n is at least 1, and m is at least 5.

Images