A disease attack early warning system based on brain and muscle electrical signals

Abstract

The application provides a disease attack early warning system based on brain and myoelectric signals, comprising: a spiky wave-muscle spasm space-time domain joint detection and fusion module; a pre-attack precursor feature dynamic extraction module based on deep learning, used for receiving the joint feature vector, extracting deep time sequence features identifying pre-attack precursor states through a pre-trained neural network model; a multi-dimensional risk quantification and probability evolution evaluation module, used for calculating a real-time attack risk probability value at the current time according to the deep time sequence features, and predicting the evolution trend of the probability in a short time domain; and a personalized threshold adaptive learning and calibration module, used for dynamically adjusting a risk probability threshold triggering a warning according to user historical attack data and warning records. In the above technical solution, through synchronous fusion analysis of electroencephalogram and myoelectric signals, earlier and more accurate warning of disease attacks such as epilepsy is realized.

Description

Technical Field

[0001] This application relates to the field of digital medical technology, and in particular to a disease onset early warning system based on brain electromyography signals. Background Technology

[0002] Epilepsy and other paroxysmal neurological disorders severely impact the quality of life and safety of tens of millions of patients worldwide. Unpredictable, sudden seizures can lead to injury, accidents, and even sudden death. While medication is effective for some patients, approximately 30% suffer from drug-resistant epilepsy, where seizures cannot be fully controlled. Therefore, early warning systems capable of predicting seizures are crucial for patients to take preventative measures (such as finding a safe location and taking emergency medication), reduce risks, and improve their ability to live independently.

[0003] The existing technology has the following main drawbacks:

[0004] Short lead time and low accuracy: Single EEG signal analysis is difficult to provide a highly reliable early warning in the early stages of an attack (such as the prodromal period).

[0005] High false alarm rate: Artificial signs generated by daily activities (movement, blinking, etc.) are easily misjudged as abnormal signals, affecting user experience and system credibility.

[0006] Lack of personalization: Early warning thresholds are usually fixed or based on population statistics, which cannot adapt to the huge differences between individual patients and the changes in their condition over time.

[0007] System rigidity and inability to evolve: Most systems have fixed parameters after deployment, making it impossible to learn from actual use and optimize performance.

[0008] Disconnect between early warning and response: After an early warning is issued, there is a lack of a mechanism for efficient linkage with the user's environment and emergency contacts. Summary of the Invention

[0009] This application provides a disease onset early warning system based on brain electromyography signals, which can achieve earlier and more accurate early warning of the onset of diseases such as epilepsy by synchronously fusing and analyzing brain and muscle signals.

[0010] Firstly, a disease onset early warning system based on brain electromyography signals is provided, including:

[0011] The spatiotemporal joint detection and fusion module for spike waves and electromyography (EMG) is used to detect EEG spike wave features and EMG features from the preprocessed signal, and perform spatiotemporal alignment and feature-level fusion to generate a joint feature vector.

[0012] The deep learning-based dynamic extraction module for prodromal features is used to receive the joint feature vector and extract deep temporal features that identify the prodromal state of an attack through a pre-trained neural network model.

[0013] The multi-dimensional risk quantification and probability evolution assessment module is used to calculate the real-time occurrence risk probability value at the current moment based on the deep temporal characteristics, and predict the evolution trend of the probability in the short time domain.

[0014] The personalized threshold adaptive learning and calibration module is used to dynamically adjust the risk probability threshold for triggering warnings based on the user's historical attack data and warning records.

[0015] The graded early warning and multimodal emergency notification module is used to generate different levels of early warning information based on the magnitude and evolution trend of the probability value when the real-time risk probability value exceeds the adjusted threshold, and send it to the user and at least one emergency contact via a mobile smart terminal.

[0016] In the above technical solution, a spatiotemporal joint detection and fusion module for spike-myoclonus is set up to detect EEG spike features and EMG clonus features from the preprocessed signal, and perform spatiotemporal alignment and feature-level fusion to generate a joint feature vector; a deep learning-based dynamic extraction module for prodromal features is used to receive the joint feature vector and extract deep temporal features that identify the prodromal state of the seizure through a pre-trained neural network model; a multi-dimensional risk quantification and probability evolution assessment module is used to calculate the real-time seizure risk probability value at the current moment based on the deep temporal features and predict the evolution trend of the probability in the short time domain; a personalized threshold adaptive learning and calibration module is used to dynamically adjust the risk probability threshold for triggering the warning based on the user's historical seizure data and warning records; a graded warning and multimodal emergency notification module is used to generate different levels of warning information based on the probability value and evolution trend when the real-time risk probability value exceeds the adjusted threshold, and send it to the user and at least one emergency contact via a mobile smart terminal; by synchronously fusing and analyzing EEG and EMG signals, earlier and more accurate warnings for seizures of diseases such as epilepsy are achieved.

[0017] In one specific implementation scheme, it also includes:

[0018] The multimodal signal synchronous acquisition and adaptive preprocessing module is used to denoise, remove artifacts, and normalize the amplitude of synchronously acquired EEG and EMG signals.

[0019] In one specific implementation scheme, it also includes:

[0020] The closed-loop feedback and system self-evolution optimization module is used to collect user feedback information on the warning and verification information on whether the warning has occurred, in order to optimize the neural network model and the warning threshold.

[0021] In one specific implementation scheme, the spatiotemporal joint detection and fusion module for spike wave-electromyoclonus includes:

[0022] The EEG spike detection unit is used to locate spikes, spikes, or spike-and-slow-wave discharges in EEG signals using wavelet transform and morphological analysis methods.

[0023] The electromyography (EMG) clonus detection unit uses a method based on the Teager-Kaiser energy operator and active segmentation to detect transient, burst-like muscle discharge clusters in EMG signals.

[0024] The spatiotemporal alignment unit is used to establish a physiological conduction delay model between EEG leads and EMG channels, and to perform temporal correlation analysis on detected spike events and electromyoclonus events.

[0025] The feature fusion unit is used to concatenate the features of spike-array pairs with strong spatiotemporal correlation and calculate the synergy index to form the joint feature vector.

[0026] In one specific implementation scheme, the deep learning-based pre-onset feature dynamic extraction module employs a hybrid neural network structure, including:

[0027] A one-dimensional convolutional neural network layer used to capture local temporal dependencies;

[0028] A bidirectional long short-term memory network layer used for modeling long-term temporal contexts;

[0029] An attention mechanism layer used to focus on key time points and feature channels.

[0030] In one specific implementation scheme, the execution process of the multi-dimensional risk quantification and probability evolution assessment module includes:

[0031] The deep temporal features are input into a fully connected layer, and a real-time risk probability P_t between 0 and 1 is output.

[0032] Input the feature sequences of the current and past time windows into a recurrent neural network or a temporal convolutional network to predict the change curve of risk probability P_(t+1:t+T) within the next T minutes;

[0033] The overall risk level is calculated by combining the rising slope and peak value predictions of P_t and P_(t+1:t+T).

[0034] In one specific implementation scheme, the personalized threshold adaptive learning and calibration module includes:

[0035] In the initial phase, a default threshold based on population statistics was used;

[0036] Record each warning event and mark it as a true positive, false positive, or false negative;

[0037] Using a contextual gambling machine algorithm, the system dynamically balances sensitivity and specificity based on individual user feedback on different alert sensitivities and actual event outcomes, adjusting the optimal alert threshold.

[0038] In one specific implementation scheme, the warning levels of the graded early warning and multimodal emergency notification module include at least:

[0039] Low-risk alert is used to display a prompt on the device screen or mobile application interface when the risk probability value exceeds a first threshold.

[0040] Medium-risk warning is used to issue a warning to users via mobile terminals when the risk probability value exceeds the second threshold;

[0041] High-risk alerts are used to issue a strong warning to users when the risk probability value exceeds the third threshold or the predicted probability will rise sharply, and automatically send an alert message containing location information to at least one preset emergency contact.

[0042] In one specific implementation scheme, the closed-loop feedback and system self-evolution optimization module includes:

[0043] The mobile application interface provides a simple feedback button for users to confirm whether an attack occurred, it was a false alarm, or there was no attack.

[0044] When the system connects to the cloud server, it uploads anonymous feature data, early warning results, user feedback and verification information to the cloud database.

[0045] The deep learning model is regularly updated via federated learning in the cloud.

[0046] Secondly, a disease onset early warning method based on brain electromyography signals is provided, including the following steps:

[0047] The spatiotemporal joint detection and fusion module of spike wave and electromyography (EMG) is used to detect EEG spike wave features and EMG features from the preprocessed signal, and spatiotemporal alignment and feature-level fusion are performed to generate a joint feature vector.

[0048] The joint feature vector is received by a deep learning-based pre-seizure feature dynamic extraction module, and deep temporal features identifying the pre-seizure state are extracted through a pre-trained neural network model.

[0049] The multi-dimensional risk quantification and probability evolution assessment module calculates the real-time occurrence risk probability value at the current moment based on the deep temporal characteristics, and predicts the evolution trend of this probability in the short time domain.

[0050] The personalized threshold adaptive learning and calibration module dynamically adjusts the risk probability threshold for triggering warnings based on the user's historical attack data and warning records.

[0051] When the real-time risk probability value exceeds the adjusted threshold, the tiered early warning and multimodal emergency notification module generates different levels of early warning information based on the probability value and its evolution trend, and sends it to the user and at least one emergency contact via a mobile smart terminal.

[0052] In the above technical solution, a spatiotemporal joint detection and fusion module for spike-myoclonus is set up to detect EEG spike features and EMG clonus features from the preprocessed signal, and perform spatiotemporal alignment and feature-level fusion to generate a joint feature vector; a deep learning-based dynamic extraction module for prodromal features is used to receive the joint feature vector and extract deep temporal features that identify the prodromal state of the seizure through a pre-trained neural network model; a multi-dimensional risk quantification and probability evolution assessment module is used to calculate the real-time seizure risk probability value at the current moment based on the deep temporal features and predict the evolution trend of the probability in the short time domain; a personalized threshold adaptive learning and calibration module is used to dynamically adjust the risk probability threshold for triggering the warning based on the user's historical seizure data and warning records; a graded warning and multimodal emergency notification module is used to generate different levels of warning information based on the probability value and evolution trend when the real-time risk probability value exceeds the adjusted threshold, and send it to the user and at least one emergency contact via a mobile smart terminal; by synchronously fusing and analyzing EEG and EMG signals, earlier and more accurate warnings for seizures of diseases such as epilepsy are achieved. Attached Figure Description

[0053] Figure 1 A structural block diagram of a disease onset early warning system based on electromyography signals provided in an embodiment of this application;

[0054] Figure 2 A flowchart illustrating a disease onset early warning method based on brain electromyography signals provided in this application embodiment. Detailed Implementation

[0055] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. Through these descriptions, the features and advantages of the present application will become clearer and more apparent.

[0056] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments. Although various aspects of embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless specifically indicated otherwise.

[0057] Furthermore, the technical features involved in the different embodiments of this application described below can be combined with each other as long as they do not conflict with each other.

[0058] To facilitate understanding of the disease seizure early warning system based on electroencephalography (EEG) signals provided in this application embodiment, its application scenario is first explained. The disease seizure early warning system based on EEG signals provided in this application embodiment is used to achieve earlier and more accurate early warnings for diseases such as epilepsy by synchronously fusing and analyzing EEG and EMG signals. Existing technologies have the following main drawbacks: Short warning lead time and low accuracy: Single EEG signal analysis is difficult to provide highly reliable early warnings in the early stages of a seizure (such as the prodromal phase). High false alarm rate: Artifacts generated by daily activities (movement, blinking, etc.) are easily misjudged as abnormal signals, affecting user experience and system reliability. Lack of personalization: Warning thresholds are usually fixed or based on group statistics, unable to adapt to the significant differences between individual patients and changes in their condition over time. System rigidity and inability to evolve: Most systems have fixed parameters after deployment, making it impossible to learn and optimize performance from actual use. Disconnect between warning and response: After a warning is issued, there is a lack of a mechanism for efficient linkage with the user environment and emergency contacts. Therefore, this application provides a disease attack early warning system based on electroencephalography (EEG) signals, which achieves earlier and more accurate early warning of diseases such as epilepsy by synchronously fusing and analyzing EEG and EMG signals. The following detailed description, in conjunction with specific accompanying drawings, illustrates the embodiments.

[0059] refer to Figure 1 and Figure 2 , Figure 1 A structural block diagram of a disease onset early warning system based on electromyography signals provided in an embodiment of this application; Figure 2 A flowchart illustrating a disease onset early warning method based on brain electromyography signals provided in this application embodiment.

[0060] exist Figure 1 In this application, an embodiment provides a disease onset early warning system based on electromyography (EMG) signals, comprising:

[0061] The spatiotemporal joint detection and fusion module for spike waves and electromyography (EMG) is used to detect EEG spike wave features and EMG features from the preprocessed signal, and perform spatiotemporal alignment and feature-level fusion to generate a joint feature vector.

[0062] The deep learning-based dynamic extraction module for prodromal features is used to receive the joint feature vector and extract deep temporal features that identify the prodromal state of an attack through a pre-trained neural network model.

[0063] The multi-dimensional risk quantification and probability evolution assessment module is used to calculate the real-time occurrence risk probability value at the current moment based on the deep temporal characteristics, and predict the evolution trend of the probability in the short time domain.

[0064] The personalized threshold adaptive learning and calibration module is used to dynamically adjust the risk probability threshold for triggering warnings based on the user's historical attack data and warning records.

[0065] The graded early warning and multimodal emergency notification module is used to generate different levels of early warning information based on the magnitude and evolution trend of the probability value when the real-time risk probability value exceeds the adjusted threshold, and send it to the user and at least one emergency contact via a mobile smart terminal.

[0066] In the above technical solution, a spatiotemporal joint detection and fusion module for spike-myoclonus is set up to detect EEG spike features and EMG clonus features from the preprocessed signal, and perform spatiotemporal alignment and feature-level fusion to generate a joint feature vector; a deep learning-based dynamic extraction module for prodromal features is used to receive the joint feature vector and extract deep temporal features that identify the prodromal state of the seizure through a pre-trained neural network model; a multi-dimensional risk quantification and probability evolution assessment module is used to calculate the real-time seizure risk probability value at the current moment based on the deep temporal features and predict the evolution trend of the probability in the short time domain; a personalized threshold adaptive learning and calibration module is used to dynamically adjust the risk probability threshold for triggering the warning based on the user's historical seizure data and warning records; a graded warning and multimodal emergency notification module is used to generate different levels of warning information based on the probability value and evolution trend when the real-time risk probability value exceeds the adjusted threshold, and send it to the user and at least one emergency contact via a mobile smart terminal; by synchronously fusing and analyzing EEG and EMG signals, earlier and more accurate warnings for seizures of diseases such as epilepsy are achieved.

[0067] Specifically, the beneficial effects include:

[0068] Beneficial effects

[0069] I. Improve the early detection and accuracy of disease onset warnings

[0070] Advantages of Multimodal Signal Fusion: This system innovatively combines and fuses spike features from EEG signals with clonic features from EMG signals in the spatiotemporal domain. Traditional single EEG signal monitoring is susceptible to noise interference, and some pre-seizure EEG changes are not significant; while EMG signals can reflect abnormal muscle activity and complement EEG signals. By generating a joint feature vector through spatiotemporal alignment and feature-level fusion, the information from both signals can be comprehensively utilized to capture weaker and earlier prodromal signs of seizures, significantly improving the early warning of disease onset and buying valuable intervention time for patients.

[0071] Deep Learning-Based Precise Feature Extraction: A deep learning-based module for dynamically extracting prodromal features utilizes a pre-trained neural network model to perform in-depth analysis of joint feature vectors. Deep learning models possess powerful nonlinear fitting and feature learning capabilities, automatically uncovering deep temporal features hidden within complex signals. These features more accurately identify prodromal states. Compared to traditional methods based on manual experience for feature extraction, deep learning models can handle massive amounts of more complex data, extracting more representative and discriminative features, thus significantly improving the accuracy of early warning.

[0072] II. Reduce the false alarm rate of the system and improve the specificity of the early warning.

[0073] Personalized Threshold Adaptive Learning: The personalized threshold adaptive learning and calibration module is a major highlight of this system. Different users have different disease attack patterns and physiological characteristics, making it difficult for a uniform warning threshold to meet personalized needs, easily leading to false alarms or missed alarms. This module dynamically adjusts the risk probability threshold for triggering warnings based on the user's historical attack data and warning records. Through continuous learning and optimization, the threshold becomes more closely aligned with the user's actual situation, effectively reducing false alarms caused by individual differences, improving the specificity of warnings, and increasing the trust of patients and their families in the system's warning information.

[0074] Multi-dimensional Risk Quantification and Trend Prediction: The multi-dimensional risk quantification and probability evolution assessment module not only calculates the real-time probability of an attack at the current moment, but also predicts the evolution trend of this probability over a short time domain. This multi-dimensional risk assessment method avoids the one-sidedness of issuing warnings based solely on the risk value at a single point in time. By analyzing the risk evolution trend, the system can more accurately determine the likelihood of an attack, further reducing the false alarm rate. For example, even if the current risk probability value is slightly higher than the threshold, if it is predicted to show a downward trend, the system may not trigger an alert, thus avoiding unnecessary panic and interference.

[0075] III. Enhancing System Usability and User Compliance

[0076] A complete technical solution forms a closed-loop feedback: This system establishes a complete and evolvable technical solution from signal perception, risk quantification, intelligent early warning to closed-loop feedback. In the signal perception stage, it accurately acquires and preprocesses electromyographic signals; in the risk quantification stage, it precisely calculates the probability of seizures and predicts trends; in the intelligent early warning stage, it generates tiered early warning information based on personalized thresholds; simultaneously, the system can collect user feedback, such as actual seizure occurrences after an early warning, to further optimize the model and adjust parameters, forming a closed-loop feedback mechanism. This complete technical solution enables the system to continuously improve and evolve, maintaining high early warning performance and enhancing its practicality.

[0077] Tiered early warning and multimodal notifications enhance user experience: The tiered early warning and multimodal emergency notification module generates different levels of early warning information based on real-time risk probability values ​​and evolution trends, and sends them to the user and at least one emergency contact via mobile smart terminal. Different levels of early warning information allow users and their families to clearly understand the urgency of an attack and take appropriate measures. Multimodal notification methods (such as sound, vibration, and SMS) ensure that early warning information is delivered to relevant personnel in a timely and effective manner, attracting attention even when the user does not check their phone immediately. This user-friendly design increases user acceptance and compliance with the system, making patients more willing to use the system for long-term disease monitoring and early warning.

[0078] In one specific implementation scheme, it also includes:

[0079] The multimodal signal synchronous acquisition and adaptive preprocessing module is used to denoise, remove artifacts, and normalize the amplitude of synchronously acquired EEG and EMG signals.

[0080] In one specific implementation scheme, it also includes:

[0081] The closed-loop feedback and system self-evolution optimization module is used to collect user feedback information on the warning and verification information on whether the warning has occurred, in order to optimize the neural network model and the warning threshold.

[0082] In one specific implementation scheme, the spatiotemporal joint detection and fusion module for spike wave-electromyoclonus includes:

[0083] The EEG spike detection unit is used to locate spikes, spikes, or spike-and-slow-wave discharges in EEG signals using wavelet transform and morphological analysis methods.

[0084] The electromyography (EMG) clonus detection unit uses a method based on the Teager-Kaiser energy operator and active segmentation to detect transient, burst-like muscle discharge clusters in EMG signals.

[0085] The spatiotemporal alignment unit is used to establish a physiological conduction delay model between EEG leads and EMG channels, and to perform temporal correlation analysis on detected spike events and electromyoclonus events.

[0086] The feature fusion unit is used to concatenate the features of spike-array pairs with strong spatiotemporal correlation and calculate the synergy index to form the joint feature vector.

[0087] In the above technical solutions, the following features are achieved: Precise detection of abnormal signals: The EEG spike detection unit utilizes wavelet transform and morphological analysis. Wavelet transform allows for multi-scale analysis of EEG signals, while morphological analysis enhances signal characteristics. The combination of these two methods accurately locates abnormal discharges such as spikes and sharp waves, providing reliable EEG data for subsequent analysis. The electromyography (EMG) clonic spasm detection unit employs the Teager-Kaiser energy operator and activity segmentation. The former is sensitive to the instantaneous energy of the signal, while the latter effectively divides activity segments, accurately detecting brief bursts of muscle discharges and ensuring the accuracy of EMG signal abnormality detection. Accurate spatiotemporal alignment analysis: The physiological conduction delay model established by the spatiotemporal alignment unit fully considers the physiological relationship between EEG leads and EMG channels. Based on this model, temporal correlation analysis of spike events and EMG clonic spasm events accurately identifies their temporal and spatial correspondence, avoiding analysis errors caused by signal conduction delays and providing accurate spatiotemporal information for feature fusion. Efficient feature fusion generation: The feature fusion unit concatenates spike-clonic spasm pairs with strong spatiotemporal correlation and calculates the co-exponentiation to form a joint feature vector. This fusion approach not only integrates key features of EEG and EMG signals, but also reflects the interaction between the two through a synergy index, making the joint feature vector contain richer and more comprehensive information. This provides more representative and discriminative feature basis for subsequent disease onset warnings, helping to improve the accuracy and reliability of warnings.

[0088] In one specific implementation scheme, the deep learning-based pre-onset feature dynamic extraction module employs a hybrid neural network structure, including:

[0089] A one-dimensional convolutional neural network layer used to capture local temporal dependencies;

[0090] A bidirectional long short-term memory network layer used for modeling long-term temporal contexts;

[0091] An attention mechanism layer used to focus on key time points and feature channels.

[0092] In the above technical solutions, precise capture of local temporal features is achieved: the one-dimensional convolutional neural network layer possesses powerful local feature extraction capabilities. When processing temporal data such as electroencephalogram (EEG) signals, it can automatically learn and capture local temporal dependencies in the signal by sliding the convolutional kernel along the time axis. For example, it can accurately identify transient spike discharge patterns in EEG signals or local fluctuations of specific frequencies in EMG signals. These local features are often important markers of the prodromal state of disease onset, providing crucial basic information for subsequent analysis. Effective modeling of long-term temporal context is also achieved: the bidirectional long short-term memory (LSTM) network layer overcomes the gradient vanishing problem in traditional neural networks when processing long-sequence data. It can simultaneously consider past and future temporal information to model the long-term temporal context of EEG signals. Through this bidirectional information flow, the evolution of the signal in different time periods can be better understood, capturing potential prodromal features that span longer periods, thus more comprehensively reflecting the dynamic changes before disease onset. Focusing on key information improves performance: the introduction of the attention mechanism layer enables the model to automatically focus on key time points and feature channels in EEG signals. In complex signals, not all information is equally important for disease onset warning. Attention mechanisms can dynamically assign weights to different parts, highlighting features that are more helpful for early warning. This helps reduce interference from irrelevant information, improves the model's sensitivity to important features, and thus enhances the accuracy and reliability of the entire system in extracting precursor features and issuing early warnings for disease onset.

[0093] In one specific implementation scheme, the execution process of the multi-dimensional risk quantification and probability evolution assessment module includes:

[0094] The deep temporal features are input into a fully connected layer, and a real-time risk probability P_t between 0 and 1 is output.

[0095] Input the feature sequences of the current and past time windows into a recurrent neural network or a temporal convolutional network to predict the change curve of risk probability P_(t+1:t+T) within the next T minutes;

[0096] The overall risk level is calculated by combining the rising slope and peak value predictions of P_t and P_(t+1:t+T).

[0097] In one specific implementation scheme, the personalized threshold adaptive learning and calibration module includes:

[0098] In the initial phase, a default threshold based on population statistics was used;

[0099] Record each warning event and mark it as a true positive, false positive, or false negative;

[0100] Using a contextual gambling machine algorithm, the system dynamically balances sensitivity and specificity based on individual user feedback on different alert sensitivities and actual event outcomes, adjusting the optimal alert threshold.

[0101] In the above technical solution, both universality and initial effectiveness are considered: In the initial stage, a default threshold based on population statistics is adopted. This design fully considers the lack of individual data for new users. Population statistics are general patterns derived from the analysis of a large number of samples, providing new users with a relatively reasonable and universal starting threshold. This ensures that the system has a certain warning capability from the initial use, avoiding the inability to carry out warning work due to the lack of initial values, and laying the foundation for subsequent personalized adjustments. Accurate recording aids data accumulation: Every warning event is recorded in detail and accurately marked as a true positive, false positive, or false negative. This process accumulates rich and crucial individual data for the system. This data truly reflects the system's warning performance under different circumstances and the user's actual health status, serving as an important basis for subsequent optimization of the warning threshold. Through the analysis of this data, a deeper understanding of the user's disease onset patterns and the accuracy of the system's warnings can be obtained, providing strong support for personalized adjustments. Intelligent algorithms achieve dynamic optimization: Using a contextual gambling machine algorithm, based on the user's individual feedback on different warning sensitivities and actual onset outcomes, sensitivity and specificity are dynamically balanced to adjust the optimal warning threshold. This algorithm intelligently learns the characteristics and needs of users, ensuring that no real seizures are missed (high sensitivity) while minimizing false alarms (high specificity). As usage time increases, the system continuously adapts to individual differences, making the warning thresholds increasingly aligned with the user's actual situation, significantly improving the accuracy and reliability of warnings, and enhancing user trust and compliance with the system.

[0102] In one specific implementation scheme, the warning levels of the graded early warning and multimodal emergency notification module include at least:

[0103] Low-risk alert is used to display a prompt on the device screen or mobile application interface when the risk probability value exceeds a first threshold.

[0104] Medium-risk warning is used to issue a warning to users via mobile terminals when the risk probability value exceeds the second threshold;

[0105] High-risk alerts are used to issue a strong warning to users when the risk probability value exceeds the third threshold or the predicted probability will rise sharply, and automatically send an alert message containing location information to at least one preset emergency contact.

[0106] In the above technical solution, risk differentiation is detailed and precise: the module sets at least three warning levels—low, medium, and high—and classifies different risk scenarios based on risk probability values ​​and trends. This detailed risk classification can accurately match risk situations of different severity levels, avoiding a "one-size-fits-all" approach to warnings, allowing users to clearly understand their own risk level and providing a clear basis for formulating subsequent response measures. Notification methods are flexible and diverse: various notification methods are used for different warning levels. For low-risk situations, a prompt is displayed only on the device screen or mobile application interface, which is both non-intrusive and serves as a reminder; for medium-risk situations, a warning is issued via mobile terminal, enhancing the reminder and ensuring users are informed promptly; for high-risk situations, a strong alarm is issued and an alarm message containing location information is automatically sent to emergency contacts, comprehensively protecting user safety and meeting the notification needs under different risk levels. Emergency response is efficient and timely: the special design of the high-risk alarm, when the risk probability value exceeds the third threshold or the prediction rises sharply, not only reminds the user but also quickly notifies emergency contacts and shares location information. This helps emergency contacts quickly understand the situation and take action, such as contacting medical facilities or going to the scene for rescue, greatly shortening emergency response time, improving the efficiency of dealing with sudden risks, and effectively reducing the harm that risks may cause.

[0107] In one specific implementation scheme, the closed-loop feedback and system self-evolution optimization module includes:

[0108] The mobile application interface provides a simple feedback button for users to confirm whether an attack occurred, it was a false alarm, or there was no attack.

[0109] When the system connects to the cloud server, it uploads anonymous feature data, early warning results, user feedback and verification information to the cloud database.

[0110] The deep learning model is regularly updated via federated learning in the cloud.

[0111] In one specific implementation scheme, the disease seizure early warning system based on electroencephalogram (EEG) and electromyography (EMG) signals achieves earlier and more accurate early warning of seizures for diseases such as epilepsy by simultaneously fusing and analyzing EEG and EMG signals. Through deep learning and personalized adaptive technology, the system's false alarm rate is reduced, and the specificity of the warning is improved. By establishing a complete and evolvable technical solution from signal perception, risk quantification, intelligent early warning to closed-loop feedback, the system's practicality and user compliance are enhanced. This includes:

[0112] Wearable signal acquisition device: Designed as a headband or headband, integrating dry electrodes or microneedle electrodes for acquiring multi-lead (at least 2 leads) scalp EEG signals, while simultaneously integrating flexible dry electrodes at key muscle groups such as the forehead, neck, or limbs to acquire at least one channel of surface EMG signals. The device has a built-in high-precision synchronization clock to ensure strict time synchronization between EEG and EMG signals. The device communicates with mobile smart terminals via Bluetooth Low Energy.

[0113] Mobile intelligent terminal: Serving as the system's local processing core and human-computer interaction interface. Dedicated applications run on it, responsible for receiving raw signals, running lightweight local early warning models, performing real-time risk calculations, triggering early warnings, and managing the user interface. Simultaneously, it acts as a gateway for communication with the cloud.

[0114] Cloud servers are responsible for storing users' long-term anonymized data and performing complex model training, federated learning, and system-level optimization analysis. When users permit and network conditions allow, mobile devices can upload data to the cloud and download updated model parameters.

[0115] Specifically, including:

[0116] Module 1: Multimodal Signal Synchronous Acquisition and Adaptive Preprocessing Module

[0117] This module serves as the system's data entry point and quality assurance mechanism. It is used for:

[0118] Hardware synchronization and software coordination: At the hardware level, the EEG and EMG acquisition circuits use the same master clock to ensure microscopic alignment of sampling timestamps. At the software level, the data stream carries a unified timestamp sequence.

[0119] Adaptive preprocessing pipeline:

[0120] For EEG signals, an eye-tracking and ECG artifact removal method based on a combination of independent component analysis and deep learning is employed. First, ICA is used to separate suspected artifact components. Then, a trained lightweight CNN network is used to determine whether the component is an artifact, thereby removing artifacts more accurately and preserving possible EEG abnormalities.

[0121] For electromyographic signals, an adaptive bandwidth filter is used to dynamically adjust the filtering bandwidth according to the power spectrum characteristics of the signal itself, so as to preserve the clonic burst characteristics in the best way and suppress baseline fluctuations and motion artifacts.

[0122] Amplitude normalization: Dynamic normalization based on a sliding time window is adopted, using the signal statistical characteristics of the user over a period of time as a benchmark to eliminate the amplitude influence caused by changes in electrode contact and individual differences.

[0123] Module 2: Spatiotemporal Joint Detection and Fusion Module of Spike Wave-Electromyoclonus

[0124] This module is crucial for improving detection specificity. Traditional independent detection of EEG spikes or EMG discharges is susceptible to interference. This module performs a correlation analysis between the two in the spatiotemporal domain. It includes:

[0125] EEG spike detection unit: Employs a multi-scale wavelet transform combined with morphological filters. Wavelet transform captures transient spikes at multiple scales, while morphological filters (such as top-hat transform) further extract events with spike-like morphologies (steep rise, slow fall) from background activity, calculating their amplitude, slope, duration, and other characteristics.

[0126] Electromyography (EMG) clonic spasm detection unit: Employs the Teager-Kaiser energy operator (TKEO). TKEO significantly enhances the transient burst energy in EMG signals and is highly sensitive to brief clonic discharges. Subsequently, active segment detection is performed to extract features such as the duration, average amplitude, and number of pulses of the clonic clusters.

[0127] Spatiotemporal Alignment and Fusion Unit (Core Innovation): This is the core of this module. Considering the physiological delay (usually tens to hundreds of milliseconds) in the transmission of abnormal cortical discharges to muscles and the generation of electromyographic responses via neural networks, this unit establishes a learnable delay-matching model. For each detected EEG spike event, an electromyographic clonic event is searched within a physiologically reasonable delay window (e.g., 50-300 ms). If found, it is considered a "spike-clonic pair," possessing a high physiological correlation and likely originating from the spread of abnormal cortical discharges. Subsequently, the features of this pair of events (such as spike amplitude and clonic energy) are concatenated, and a "brain-muscle synergy index" (e.g., correlation coefficient, mutual information) is calculated, forming a joint feature vector. This spatiotemporal joint detection can effectively distinguish between genuine pathological discharges and random, irrelevant artifacts or physiological activities, greatly reducing false alarms from single signal detection.

[0128] Module 3: Dynamic Extraction Module of Pre-Symptom Features Based on Deep Learning

[0129] The goal of this module is to automatically learn and extract deep patterns that are most relevant to the prodromal state of an attack and may be difficult for the human eye to recognize from continuous and complex joint feature sequences.

[0130] Network architecture design: This invention designs a 1D-CNN-BiLSTM-Attention hybrid neural network. It includes:

[0131] 1D Convolutional Layer: First, multiple one-dimensional convolutional kernels are slid across the time series to extract pattern combinations within local time periods, such as the co-occurrence patterns of spikes of a specific frequency and clonic patterns of a specific morphology within a short period of time.

[0132] Bidirectional LSTM layer: The local feature sequences extracted by CNN are input into BiLSTM. BiLSTM can learn the dependencies of long-term sequences in both forward and backward directions, capturing the evolution of slow-accumulating, long-term brain-muscle synergistic activities that may exist in the prodromal period.

[0133] Attention Mechanism Layer: An attention mechanism is introduced into the output of BiLSTM. This mechanism can automatically learn and assign different weights to features at different time points, so that the model pays more attention to those time points that are more critical to judging whether an attack is approaching (such as the period when spike-clonic pairs occur densely), while ignoring irrelevant or interfering time periods, thus enhancing the interpretability and robustness of the model.

[0134] Output: The output of this module is a fixed-dimensional, highly abstract deep temporal feature vector that comprehensively represents the abnormal patterns of brain-muscle coordination activities in the current and past periods.

[0135] Module Four: Multi-dimensional Risk Quantification and Probability Evolution Assessment

[0136] This module transforms abstract, deep features into intuitive, actionable risk indicators. These include:

[0137] Real-time risk probability calculation: The deep temporal feature vector is input into a fully connected neural network, followed by a sigmoid activation function. The output is a scalar value P_t between 0 and 1, representing the real-time attack risk probability at the current moment. P_t > 0.5 indicates that the risk is higher than the normal baseline.

[0138] Risk Evolution Trend Prediction: To further improve the foresight of early warning, this module introduces a time-series prediction submodule. Deep feature sequences (or historical sequences of P_t) from a past period (e.g., 5 minutes) are input into a temporal convolutional network or GRU network to predict the change curve P_(t+1:t+T) of the risk probability within a future time window (e.g., the next 3 minutes). TCN performs excellently in this task due to its parallel computing advantages and controllable receptive field.

[0139] Multi-dimensional comprehensive assessment: The final risk assessment is based not only on the current P_t, but also on the characteristics of the prediction curve, such as the future peak P_max, the average upward slope S, and the predicted time T_toAlert for reaching the high-risk threshold (e.g., 0.8). The system comprehensively calculates an overall risk level (e.g., low, medium, high) based on these dimensions, providing richer evidence for tiered early warning.

[0140] Module 5: Personalized Threshold Adaptive Learning and Calibration Module

[0141] Fixed warning thresholds are unsuitable for all patients. This module enables the system to "learn" each user's unique patterns. It includes:

[0142] Initial threshold setting: When initializing the system, the threshold suggested by a general model trained on large sample data can be used, or it can be set by the clinician based on the user's initial assessment.

[0143] Online adaptive learning: The system runs continuously, recording every warning event (time, risk characteristics, trigger threshold level) as well as the results of user feedback on "true seizures" and "false alarms" or verification by automatic analysis in the background (such as the subsequent occurrence of clear electromyographic / motor signals of tonic-clonic seizures).

[0144] Contextual Bandit Algorithm: Employs the Contextual Bandit algorithm to dynamically adjust thresholds. The system uses the user's physiological state (e.g., sleep / awake, determined by accelerometer), historical alert performance, and the user's sensitivity preference to false alarms (configurable within the application) as the "context." Each time a decision is needed to issue an alert, the algorithm selects from several candidate threshold strategies (e.g., "high sensitivity - low threshold," "high specificity - high threshold") based on the current context. According to the "reward" after an alert (positive reward for correct alerts, negative reward for false alarms, and a large negative reward for missed alerts), the algorithm continuously updates the expected returns of each strategy under different contexts, gradually converging to the optimal alert threshold strategy best suited to the user's current state. This allows the system to minimize user interference while ensuring security.

[0145] Module Six: Tiered Early Warning and Multimodal Emergency Notification Module

[0146] This module is responsible for translating risk assessment results into timely, appropriate, and effective user interventions. This includes:

[0147] Three-level early warning system:

[0148] Level 1 (Monitoring Prompt): When P_t continuously exceeds a low threshold (e.g., 0.4) and the trend is flat, only a mild prompt icon or text (e.g., "Please take a break") is displayed on the device screen or application interface, without disturbing the user.

[0149] Level 2 (Early Warning): When P_t exceeds a moderate threshold (e.g., 0.65), or if a prediction indicates a rapid increase in risk in the near future, the mobile terminal will alert the user through a specific vibration pattern and a customized warning ringtone. The user can choose to acknowledge receipt, initiate mitigation measures (such as guided deep breathing exercises), or mark it as a false alarm.

[0150] Level 3 (Emergency Alert): When P_t exceeds a high threshold (e.g., 0.85), or a high-probability event is predicted to occur, the system triggers an emergency alert. The user terminal emits a strong, continuous audible, visual, and vibration alarm. Simultaneously, the system automatically sends an alert message to pre-set emergency contacts (family members, caregivers). The message may include: "[Epilepsy Warning Alert] User [Anonymous ID] is currently at extremely high risk of seizure. Last known location: [geographic coordinates or text address], time: [timestamp]". The alert can be sent via SMS, app push notifications, or even automatic phone calls.

[0151] Environmental linkage: If the user's home has a smart home ecosystem (such as through Tencent Lianlian and other platforms), the three-level alarm can simultaneously trigger preset safety scenarios, such as automatically turning on the lights in the room, lowering the TV volume, unlocking the door (for easy rescue), and even laying a safety air mattress next to the bed (if the device supports it), creating a safer environment for possible attacks.

[0152] Module 7: Closed-Loop Feedback and System Self-Evolution Optimization Module

[0153] This module ensures the system can be continuously improved, becoming more accurate with use. It includes:

[0154] User feedback loop: After an alert is triggered, the application interface provides a minimalist feedback button ("Is it an attack?", "Is it not an attack?"). User feedback is a valuable monitoring signal.

[0155] Data validation closed loop: The system automatically saves the original signals and all intermediate features for a period of time before and after the warning (e.g., 10 minutes before and after). If a clear seizure pattern is subsequently detected (e.g., a typical electromyographic burst of tonic-clonic seizures), even if no warning is triggered (missed report), it will be recorded as an "unwarned seizure" event. This data, together with user feedback, constitutes labeled training / optimization samples.

[0156] Federated learning-driven cloud-edge co-evolution: While protecting user privacy, mobile terminals calculate updated gradients for model parameters based on newly added labeled data locally (instead of uploading the original data), and then upload the encrypted gradients to the cloud server. The cloud server aggregates gradients from thousands of devices and updates the global model. The updated global model parameters are then encrypted and distributed to each terminal. Through this federated learning paradigm, all users' devices contribute knowledge, enabling the central model and each local model to continuously evolve, improving the ability to identify various pre-existing patterns and new user patterns, while ensuring that user data does not leave their local device, thus protecting privacy.

[0157] In this embodiment, the beneficial effects include:

[0158] Earlier and more accurate early warning: Through multimodal spatiotemporal joint detection of "brain-muscle", the complete physiological chain of abnormal neural electrical activity from the cortex to the triggering of muscle response is fully utilized. It can capture earlier and more specific prodromal signals than single EEG or EMG, which significantly increases the advance time of early warning (up to the minute level) and reduces the false alarm rate.

[0159] Highly personalized and adaptive: Through online learning algorithms and federated learning frameworks, the system can continuously adapt to each user's unique physiological patterns and disease changes, dynamically optimize warning thresholds and model parameters, achieve "one model per person", and improve the individual applicability and long-term effectiveness of warnings.

[0160] Intelligent risk evolution assessment: It not only assesses current risks but also predicts short-term risk trends, achieving a leap from static threshold judgment to dynamic trend prediction, making early warning decisions more forward-looking and scientific.

[0161] Tiered Response and Multimodal Collaboration: A three-tiered early warning system was established, enabling a refined response from mild alerts to emergency intervention. Integration with emergency contacts and smart home systems created a closed loop from individual early warning to social support and environmental safety assurance, significantly improving patient safety.

[0162] The system has the ability to continuously evolve: the innovative closed-loop feedback and federated learning mechanism enables the system to continuously learn from massive amounts of real-world usage data, constantly improve overall performance, and solve the problem of performance stagnation after the deployment of traditional medical devices.

[0163] Improving user compliance and quality of life: More accurate warnings reduce unnecessary panic (reducing false positives), earlier alerts give patients valuable preparation time, and the closed-loop learning mechanism makes users feel that the system is "tailor-made" for them. All of these greatly improve users' willingness to wear and use the system and their long-term compliance, thereby truly improving patients' ability to live independently and their sense of psychological security.

[0164] Privacy and Scalability: Employing federated learning, raw data does not need to be uploaded to the cloud, protecting the privacy of users' sensitive medical data. The system's modular design facilitates the integration of new biosignals (such as ECG and blood oxygen) or new algorithm models, and can be expanded for early warning of other paroxysmal diseases (such as migraine aura and panic attacks).

[0165] In one specific feasible implementation, the system's hardware configuration and data acquisition include:

[0166] The wearable device is designed as a lightweight headband with three dry electrode EEG acquisition points (FP1, FP2, and a reference point) on the forehead and a flexible surface electromyography (EMG) electrode integrated on each side of the temporalis muscle. The device has a built-in 24-bit ADC, an EEG sampling rate of 256Hz, an EMG sampling rate of 512Hz, and a synchronization error of less than 1ms. The device connects to the user's smartphone (with a dedicated app installed) via Bluetooth 5.2. The app receives data, processes it locally in real time, and displays the risk status.

[0167] In one feasible implementation, user A wears the device while resting at home. The system's continuously running alert process includes:

[0168] Module 1: Real-time acquisition and preprocessing of EEG and EMG signals.

[0169] Module 2: At time t0, a significant spike wave (feature F1) is detected in the central region. At t0+120ms, a brief clonic burst of the right temporalis muscle is detected (feature F2). The spatiotemporal alignment unit determines that both are within the physiological delay range and have strong morphological correlation, generating a joint feature vector J.

[0170] Module 3: J, along with the feature sequences from the previous few minutes, is input into a locally deployed lightweight CNN-BiLSTM-Attention hybrid model. The model extracts the implicit, progressively increasing anomalous synchronization patterns hidden in the current feature sequence and outputs a deep feature vector D.

[0171] Module 4: The risk quantification module receives D. The fully connected layer outputs the current risk probability P_t = 0.52. The time series prediction subnetwork analyzes the characteristics of the past 5 minutes and predicts that P_t will continue to rise rapidly in the next 3 minutes, and may reach 0.88 after 2 minutes.

[0172] Module 5: Personalized Threshold Module. Based on user A's historical data (A is more sensitive to false alarms and has a higher threshold preference), and considering that the user is currently awake, the module provides the following three threshold levels: alert threshold 0.5, warning threshold 0.7, and alarm threshold 0.8.

[0173] Decision and Early Warning: Since the current P_t = 0.52 > the alert threshold of 0.5, but less than the warning threshold, the system enters a Level 1 warning state. The user sees a softly flashing yellow indicator light on the mobile app. One minute later, P_t rises to 0.73, exceeding the warning threshold, and the predicted trend is strong. Module Six initiates a Level 2 warning: The user's phone begins to vibrate in a specific pattern and plays a soothing but alerting piece of music. The screen displays "Increased risk of seizure detected. It is recommended to sit down or lie down immediately." The user receives the reminder and sits down next to the sofa. After another 90 seconds, P_t reaches 0.9, triggering a Level 3 emergency alarm. The phone emits a sharp alarm sound and vibrates strongly, and the screen displays "High risk! Please maintain a safe posture!" in full red. At the same time, the system automatically sends an alarm message to the user's wife via WeChat and SMS: "[Epilepsy Warning Alert] User A you are following has an extremely high risk of seizure! Location: Living room at home (last updated at XX:XX)". The smart lights in the home receive the alarm through the Internet of Things protocol and automatically adjust the living room lights to maximum brightness.

[0174] Module Seven: Subsequently, the user experienced a brief loss of consciousness and arm twitching (seizure). After recovering, the user clicked the "Seizure" feedback button on the app. The system packaged the data from the 15 minutes before and after this event, all intermediate features, warning records, and user feedback. After connecting to Wi-Fi, it encrypted the model parameter update gradient and uploaded it to the cloud for federated learning. After the cloud model is updated, the models distributed to User A and other users in the future will be able to better identify such seizure precursor patterns.

[0175] In a specific feasible implementation, the personalized threshold self-learning process includes:

[0176] When User B first used the system, the threshold was relatively conservative. In the first week, the system triggered several Level 2 alerts due to B's electromyographic artifacts (such as rapid head shaking), even though no actual episodes occurred (false positives). User B marked these as "false alarms" on the app. The Contextual Bandit algorithm in the personalized threshold module (Module 5) received this negative feedback. When the algorithm detected a similar pattern of electromyographic activity without significant abnormalities in the EEG, considering the context of "User B's low tolerance for false alarms," ​​it tended to choose a higher alert threshold strategy, thus suppressing the alert. After a period of learning, the system found a higher threshold for B in a "awake and active" state, significantly reducing false alarms caused by daily activities, while still effectively alerting users when genuine abnormalities in brain-muscle coordination occurred.

[0177] exist Figure 2 In this application, an embodiment provides a method for early warning of disease onset based on electromyography (EMG) signals, comprising the following steps:

[0178] The spatiotemporal joint detection and fusion module of spike wave and electromyography (EMG) is used to detect EEG spike wave features and EMG features from the preprocessed signal, and spatiotemporal alignment and feature-level fusion are performed to generate a joint feature vector.

[0179] The joint feature vector is received by a deep learning-based pre-seizure feature dynamic extraction module, and deep temporal features identifying the pre-seizure state are extracted through a pre-trained neural network model.

[0180] The multi-dimensional risk quantification and probability evolution assessment module calculates the real-time occurrence risk probability value at the current moment based on the deep temporal characteristics, and predicts the evolution trend of this probability in the short time domain.

[0181] The personalized threshold adaptive learning and calibration module dynamically adjusts the risk probability threshold for triggering warnings based on the user's historical attack data and warning records.

[0182] When the real-time risk probability value exceeds the adjusted threshold, the tiered early warning and multimodal emergency notification module generates different levels of early warning information based on the probability value and its evolution trend, and sends it to the user and at least one emergency contact via a mobile smart terminal.

[0183] In the above technical solution, a spatiotemporal joint detection and fusion module for spike-myoclonus is set up to detect EEG spike features and EMG clonus features from the preprocessed signal, and perform spatiotemporal alignment and feature-level fusion to generate a joint feature vector; a deep learning-based dynamic extraction module for prodromal features is used to receive the joint feature vector and extract deep temporal features that identify the prodromal state of the seizure through a pre-trained neural network model; a multi-dimensional risk quantification and probability evolution assessment module is used to calculate the real-time seizure risk probability value at the current moment based on the deep temporal features and predict the evolution trend of the probability in the short time domain; a personalized threshold adaptive learning and calibration module is used to dynamically adjust the risk probability threshold for triggering the warning based on the user's historical seizure data and warning records; a graded warning and multimodal emergency notification module is used to generate different levels of warning information based on the probability value and evolution trend when the real-time risk probability value exceeds the adjusted threshold, and send it to the user and at least one emergency contact via a mobile smart terminal; by synchronously fusing and analyzing EEG and EMG signals, earlier and more accurate warnings for seizures of diseases such as epilepsy are achieved.

[0184] Those skilled in the art will know that this application can be implemented as a system, method, or computer program product.

[0185] Therefore, this disclosure can be implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this application can also be implemented as a computer program product in one or more computer-readable media, the computer-readable media containing computer-readable program code.

[0186] Any combination of one or more computer-readable media may be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0187] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application. Based on this, various substitutions and improvements can be made to this application, all of which fall within the protection scope of this application.

Claims

1. A disease onset early warning system based on brain electromyography signals, characterized in that, include: The spatiotemporal joint detection and fusion module for spike waves and electromyography (EMG) is used to detect EEG spike wave features and EMG features from the preprocessed signal, and perform spatiotemporal alignment and feature-level fusion to generate a joint feature vector. The deep learning-based dynamic extraction module for prodromal features is used to receive the joint feature vector and extract deep temporal features that identify the prodromal state of an attack through a pre-trained neural network model. The multi-dimensional risk quantification and probability evolution assessment module is used to calculate the real-time occurrence risk probability value at the current moment based on the deep temporal characteristics, and predict the evolution trend of the probability in the short time domain. The personalized threshold adaptive learning and calibration module is used to dynamically adjust the risk probability threshold for triggering warnings based on the user's historical attack data and warning records. The graded early warning and multimodal emergency notification module is used to generate different levels of early warning information based on the probability value and evolution trend when the real-time risk probability value exceeds the adjusted threshold, and send it to the user and at least one emergency contact via a mobile smart terminal.

2. The disease attack early warning system based on brain electromyography signals according to claim 1, characterized in that, Also includes: The multimodal signal synchronous acquisition and adaptive preprocessing module is used to denoise, remove artifacts, and normalize the amplitude of synchronously acquired EEG and EMG signals.

3. The disease attack early warning system based on brain electromyography signals according to claim 2, characterized in that, Also includes: The closed-loop feedback and system self-evolution optimization module is used to collect user feedback information on the warning and verification information on whether the warning has occurred, in order to optimize the neural network model and the warning threshold.

4. The disease attack early warning system based on brain electromyography signals according to claim 3, characterized in that, The spatiotemporal joint detection and fusion module for spike wave-electromyoclonus includes: The EEG spike detection unit is used to locate spikes, spikes, or spike-and-slow-wave discharges in EEG signals using wavelet transform and morphological analysis methods. The electromyography (EMG) clonus detection unit uses a method based on the Teager-Kaiser energy operator and active segmentation to detect transient, burst-like muscle discharge clusters in EMG signals. The spatiotemporal alignment unit is used to establish a physiological conduction delay model between EEG leads and EMG channels, and to perform temporal correlation analysis on detected spike events and electromyoclonus events. The feature fusion unit is used to concatenate the features of spike-array pairs with strong spatiotemporal correlation and calculate the synergy index to form the joint feature vector.

5. The disease attack early warning system based on brain electromyography signals according to claim 4, characterized in that, The deep learning-based pre-seizure feature dynamic extraction module uses a hybrid neural network structure, including: A one-dimensional convolutional neural network layer used to capture local temporal dependencies; A bidirectional long short-term memory network layer used for modeling long-term temporal contexts; An attention mechanism layer used to focus on key time points and feature channels.

6. The disease attack early warning system based on brain electromyography signals according to claim 5, characterized in that, The execution process of the multi-dimensional risk quantification and probability evolution assessment module includes: The deep temporal features are input into a fully connected layer, and a real-time risk probability P_t between 0 and 1 is output. Input the feature sequences of the current and past time windows into a recurrent neural network or a temporal convolutional network to predict the change curve of risk probability P_(t+1:t+T) within the next T minutes; The overall risk level is calculated by combining the rising slope and peak value predictions of P_t and P_(t+1:t+T).

7. The disease attack early warning system based on brain electromyography signals according to claim 6, characterized in that, The personalized threshold adaptive learning and calibration module includes: In the initial phase, a default threshold based on population statistics was used; Record each warning event and mark it as a true positive, false positive, or false negative; Using a contextual gambling machine algorithm, the system dynamically balances sensitivity and specificity based on individual user feedback on different alert sensitivities and actual event outcomes, adjusting the optimal alert threshold.

8. The disease attack early warning system based on electromyography signals according to claim 7, characterized in that, The warning levels of the tiered early warning and multimodal emergency notification module include at least: Low-risk alert is used to display a prompt on the device screen or mobile application interface when the risk probability value exceeds a first threshold. Medium-risk warning is used to issue a warning to users via mobile terminals when the risk probability value exceeds the second threshold; High-risk alerts are used to issue a strong warning to users when the risk probability value exceeds the third threshold or the predicted probability will rise sharply, and automatically send an alert message containing location information to at least one preset emergency contact.

9. The disease attack early warning system based on brain electromyography signals according to claim 8, characterized in that, The closed-loop feedback and system self-evolution optimization module includes: The mobile application interface provides a simple feedback button for users to confirm whether an attack occurred, it was a false alarm, or there was no attack. When the system connects to the cloud server, it uploads anonymous feature data, early warning results, user feedback and verification information to the cloud database. The deep learning model is regularly updated via federated learning in the cloud.

10. A method for early warning of disease onset based on electromyography (EMG) signals, characterized in that, Includes the following steps: The spatiotemporal joint detection and fusion module of spike wave and electromyography (EMG) is used to detect EEG spike wave features and EMG features from the preprocessed signal, and spatiotemporal alignment and feature-level fusion are performed to generate a joint feature vector. The joint feature vector is received by a deep learning-based pre-seizure feature dynamic extraction module, and deep temporal features identifying the pre-seizure state are extracted through a pre-trained neural network model. The multi-dimensional risk quantification and probability evolution assessment module calculates the real-time occurrence risk probability value at the current moment based on the deep temporal characteristics, and predicts the evolution trend of this probability in the short time domain. The personalized threshold adaptive learning and calibration module dynamically adjusts the risk probability threshold for triggering warnings based on the user's historical attack data and warning records. When the real-time risk probability value exceeds the adjusted threshold, the tiered early warning and multimodal emergency notification module generates different levels of early warning information based on the probability value and its evolution trend, and sends it to the user and at least one emergency contact via a mobile smart terminal.

Images