Central autonomic nervous system stability monitoring method and system

Abstract

The application discloses a central autonomic nervous system stability monitoring method and system. The method comprises the following steps: synchronously collecting electroencephalogram and electrocardiogram signals and aligning; taking the R-wave peak time as an anchor point to intercept a historical background window and a target response window, and constructing an electroencephalogram background signal sequence, a heart beat evoked potential observation true value sequence and a heart rhythm feature vector; inputting the electroencephalogram background signal sequence and the heart rhythm vector into a multi-modal collaborative attention reconstruction network, outputting coupled latent variables by using a mixed gate formula of a multiplication interaction term for calculating brain-heart matching degree and a peripheral independent significant term representing cardiac autonomic nervous system, and predicting a heart beat evoked potential waveform by decoding and restoring; calculating residual energy of the true value sequence and the predicted waveform to generate a brain-heart stability index; and triggering an early warning when the index continuously exceeds a standard. The application can realize real-time quantification of brain-heart decoupling process and early warning of non-specific instability state without relying on pathological samples.

Description

Technical Field

[0001] This invention belongs to the field of brain-computer interface (BCI) and biomedical signal processing technology, specifically relating to a method and system for monitoring the stability of the central autonomic nervous system. Background Technology

[0002] There is a close and complex bidirectional interactive regulatory mechanism between the central nervous system and the autonomic nervous system. This dynamic coupling between the brain and heart is key to maintaining the body's physiological homeostasis. In a healthy state, the heart's pulsating signals are transmitted to the brain via the vagus nerve, inducing specific cortical neural responses (i.e., cardiac evoked potentials). The brain, in turn, controls the heart rhythm by regulating autonomic nerve tone.

[0003] However, existing physiological monitoring technologies have the following limitations:

[0004] First, existing monitoring methods often rely on single-modal features, resulting in a lack of high-dimensional coupling information. Clinically, EEG or ECG is often analyzed independently, for example, focusing only on abnormal discharge waveforms or arrhythmia segments. This fragmented monitoring approach ignores the millisecond-level dynamic information interaction between the brain and heart, making it difficult to identify many systemic instability states that are not obvious in a single signal but have already shown signs of decoupling at the "brain-heart coupling" level.

[0005] Secondly, traditional methods heavily rely on supervised learning and lack the ability to broadly monitor unknown unstable states. Existing technologies are mostly based on classification training using specific pathological features, which not only requires a large number of scarce, diagnosed pathological samples, but also limits the model's generalization ability. For unknown, non-specific, or sudden central-autonomic nervous system dysfunctions, the model often cannot make effective judgments because it has never seen such samples before, easily leading to missed detections.

[0006] Finally, the current lack of objective indicators that can quantify the dynamic coupling strength in real time leads to delayed early warnings. Clinically, devices often only alarm when physiological indicators show significant abnormalities (such as severe convulsions or cardiac arrest). There is a lack of a quantitative tool that can model based on normal physiological patterns and reflect the potential instability risk of the system by calculating real-time deviations, thus missing the optimal window for early intervention. Summary of the Invention

[0007] The technical objective of this invention is to address the shortcomings of current physiological monitoring technologies, which rely on single-modal features and supervised learning paradigms, resulting in an inability to capture deep brain-heart coupling mechanisms and identify unknown abnormal states. This invention provides a method and system for monitoring the stability of the central autonomic nervous system. By constructing a multimodal signal reconstruction mechanism based on normal physiological laws, it achieves real-time quantification and early warning of functional decoupling of the system.

[0008] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution.

[0009] In a first aspect, embodiments of the present invention provide a method for monitoring the stability of the central autonomic nervous system, including:

[0010] Acquire scalp EEG and ECG signals that are strictly aligned on the same timeline for the user;

[0011] The peak time of the R wave in the electrocardiogram signal is detected; with each peak time of the R wave as the intercept time, an EEG background signal sequence is constructed based on the historical background window before the intercept time, and a true value sequence of cardiac evoked potential observations is constructed based on the target response window after the intercept time; and a cardiac rhythm feature vector is extracted based on the implicit cardiac rhythm calculation window.

[0012] The real-time acquired and processed EEG background signal sequence and the heart rhythm feature vector are input into a trained multimodal collaborative attention reconstruction network model, which outputs a predicted cardiac evoked potential waveform. The multimodal collaborative attention reconstruction network model includes a feature encoder, a collaborative attention gating module, and a signal reconstruction decoder. It is obtained through self-supervised training using multimodal data under normal physiological homeostasis.

[0013] The residual energy between the real-time received cardiac evoked potential observation true sequence and the predicted cardiac evoked potential waveform is calculated to obtain the instantaneous reconstruction error, and the brain-heart stability index is obtained by moving average processing; when the brain-heart stability index continuously exceeds the set threshold and continues for a preset time, a functional decoupling warning signal is automatically triggered.

[0014] Secondly, embodiments of the present invention also provide a central autonomic nervous system stability monitoring system for implementing the central autonomic nervous system stability monitoring method provided in any possible implementation of the first aspect, the system comprising:

[0015] Signal acquisition and alignment module: used to synchronously acquire scalp EEG and ECG signals that are strictly aligned on the same time axis;

[0016] The temporal window feature construction module is used to detect the peak time of the R wave in the electrocardiogram signal; with each peak time of the R wave as the intercept time, it constructs an EEG background signal sequence based on the historical background window before the intercept time, and constructs a true value sequence of cardiac evoked potential observations based on the target response window after the intercept time; and extracts the heart rhythm feature vector based on the implicit heart rhythm calculation window.

[0017] The deep generative reconstruction network module includes a built-in multimodal collaborative attention reconstruction network model trained under self-supervised conditions based on healthy physiological homeostasis data. This model takes the real-time acquired and processed EEG background signal sequence and the heart rhythm feature vector as input to the trained multimodal collaborative attention reconstruction network model and outputs a predicted cardiac evoked potential waveform. The multimodal collaborative attention reconstruction network model comprises an ECG encoder, an EEG background encoder, a collaborative attention gating module, and a signal reconstruction decoder. It is obtained through self-supervised training using multimodal data from normal physiological homeostasis.

[0018] Quantitative assessment and early warning module: used to calculate the residual energy between the real-time received cardiac evoked potential observation true value sequence and the predicted cardiac evoked potential waveform, obtain the instantaneous reconstruction error, and obtain the brain-heart stability index through moving average processing; when the brain-heart stability index continuously exceeds the set threshold and continues for a preset time, a functional decoupling early warning signal is automatically triggered.

[0019] Thirdly, embodiments of the present invention also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the central autonomic nervous system stability monitoring method provided in any possible implementation of the first aspect.

[0020] Compared with existing technologies, the central autonomic nervous system stability monitoring method and system provided in this invention have the following beneficial technical effects: Eliminating dependence on pathological samples: The multimodal collaborative attention reconstruction network model adopts a self-supervised learning paradigm, requiring only the learning of a "normal mode" without the need to collect scarce pathological data, thus enabling the monitoring of various unknown non-specific unstable states. Achieving early warning: The Brain-Heart Stability Index (CSI) can keenly capture "reconstruction deviation" at the microscopic level, detecting instability signs before clinical symptoms appear (e.g., 45 seconds in advance). Strong physiological interpretability: The model structure and loss function highly fit the real neural gating and sparse conduction mechanisms; an increase in CSI directly corresponds to the "brain-heart decoupling" process, providing intuitive interpretation. Strong anti-interference capability: R-wave time-locked analysis naturally filters out non-time-locked noise, and combined with a sliding window smoothing algorithm, effectively reduces the false alarm rate.

[0021] It should be understood that the summary section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0022] The accompanying drawings described herein are for illustrative purposes only and are not intended to limit the scope of this application in any way. In the drawings:

[0023] Figure 1A schematic diagram of the process for monitoring the stability of the central autonomic nervous system provided in this embodiment;

[0024] Figure 2 This is a schematic diagram of the multimodal collaborative attention reconstruction network model structure in the embodiment;

[0025] Figure 3 This is a schematic diagram illustrating the physiological signal windowing and feature extraction principle based on R-wave anchoring in the embodiment.

[0026] Figure 4 This is a schematic diagram illustrating the brain-heart coupling remodeling effect under normal physiological conditions in the embodiment.

[0027] Figure 5 This is a schematic diagram illustrating the brain-heart coupling remodeling effect under pathological instability conditions in the embodiment.

[0028] Figure 6 This is a time-series graph showing the CSI index changes during continuous monitoring of a user with epilepsy (chb01) in the example.

[0029] Figure 7 This is the ROC curve of anomaly detection performance on the test dataset in the embodiment;

[0030] Figure 8 This is a complete curve of the online monitoring stability index (CSI) of real epilepsy cases over time in the examples;

[0031] Figure 9 This is a comparison chart of the brain-heart coupling waveform reconstruction effects corresponding to the normal steady-state time node (node ​​A) in the embodiment;

[0032] Figure 10 This is a comparison chart of the brain-heart coupling waveform reconstruction effect corresponding to the pre-seizure drift time node (node ​​B) in the embodiment;

[0033] Figure 11 This is a comparison diagram of the brain-heart coupling waveform reconstruction effect corresponding to the epileptic seizure outbreak time node (node ​​C) in the embodiment. Detailed Implementation

[0034] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0035] It should be fully understood that the user's EEG and ECG signals involved in this application are information and data authorized by the user or fully authorized by all parties. The use of user information should comply with industry privacy policies and practices that are generally considered to meet or exceed the standards for protecting user privacy. The collection, use and processing of related data should comply with relevant laws, regulations and standards, and provide corresponding operation access points for users to choose to authorize or refuse.

[0036] This invention aims to address the technical problem of how to effectively monitor and provide early warning of non-specific central-autonomic nervous system instability under the existing technical conditions of lacking specific pathological samples and insufficient multimodal coupling feature mining. It achieves real-time quantification and early warning of system functional decoupling by constructing a multimodal signal reconstruction mechanism based on normal physiological laws.

[0037] The present invention provides a method and system for monitoring the stability of the central autonomic nervous system, the method comprising:

[0038] Multimodal signal synchronous acquisition: Acquire scalp electroencephalogram (EEG) and electrocardiogram (ECG) signals from the user (monitoring subject) with strict time alignment.

[0039] Feature construction based on R-wave anchoring: Using the peak time of the ECG R-wave as the time anchor, the historical EEG background before the peak time of the R-wave and the target segments of the heartbeat evoked potential (HEP) after the peak time of the R-wave are extracted respectively, and the heart rhythm variability features are extracted.

[0040] Normal pattern modeling: A multimodal collaborative attention reconstruction network model (MCAR-Net, hereinafter referred to as the "model") is constructed, which includes a dual-channel EEG / ECG encoder, a collaborative attention gating module, and a signal reconstruction decoder. The model is self-supervised and trained using only healthy steady-state data, combined with a loss function including L1 regularization, so that the model learns the normal spatiotemporal mapping pattern of brain-heart coupling.

[0041] Online monitoring and early warning: Real-time EEG background signal sequences and ECG rhythm feature vectors are input into MCAR-Net to calculate the brain-heart stability index (CSI) between the actual and predicted cardiac evoked potential waveforms. When the CSI continuously exceeds a threshold, system instability is determined and an early warning is triggered.

[0042] This invention uses a deep generative model to learn the brain-heart coupling pattern under normal physiological conditions, and infers system stability by monitoring the reconstruction error of real-time signals.

[0043] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0044] Methods for monitoring the stability of the central autonomic nervous system, such as Figure 1 As shown, it includes:

[0045] Collect synchronized EEG and ECG signals from users and strictly align them on the same timeline;

[0046] The peak time of the R wave in the electrocardiogram signal is detected; each peak time of the R wave is used as the intercept time, and an EEG background signal sequence is constructed based on the historical background window before the intercept time. At the same time, a true sequence of cardiac evoked potential observations is constructed based on the target response window after the intercept time; and a cardiac rhythm feature vector is extracted based on the implicit cardiac rhythm calculation window.

[0047] The real-time acquired and processed EEG background signal sequence and heart rhythm feature vector are input into the trained multimodal collaborative attention reconstruction network model, which outputs the predicted cardiac evoked potential waveform. The multimodal collaborative attention reconstruction network model includes an ECG encoder, an EEG background encoder, a collaborative attention gating module, and a signal reconstruction decoder. It is obtained by self-supervised training using multimodal data under normal physiological homeostasis.

[0048] The residual energy between the real-time received cardiac evoked potential observation true sequence and the predicted cardiac evoked potential waveform is calculated to obtain the instantaneous reconstruction error, which is then processed by moving average to obtain the brain-heart stability index (CSI). When the brain-heart stability index (CSI) continuously exceeds the set threshold and continues for a preset duration, a functional decoupling warning signal is automatically triggered.

[0049] In this embodiment, a portable wireless EEG acquisition device or a scalp EEG measurement system can be used to acquire the user's EEG signals. A dynamic electrocardiogram recorder or a multi-lead electrocardiogram monitor can be used to acquire the user's ECG signals.

[0050] In some embodiments, signal preprocessing of the acquired electroencephalogram (EEG) or electrocardiogram (ECG) signals is included. Examples specifically include:

[0051] ECG processing: The raw ECG signal underwent full-band preprocessing (e.g., Butterworth bandpass filtering from 0.5-40Hz) to remove baseline drift and high-frequency electromyography interference. Subsequently, the Pan-Tompkins algorithm was used to detect the R-wave peak value and obtain the R-wave peak time sequence.

[0052] EEG processing: The original EEG signal is preprocessed with full-band artifact suppression (e.g., bandpass filtering from 0.5 to 45 Hz), and independent component analysis (ICA) algorithm is used to remove electrooculography (EOG) and electromyography (EMG) artifacts, followed by Z-score normalization.

[0053] To capture brain-heart interaction features within a single cardiac cycle, this invention uses each R-wave peak moment... To capture specific moments (i.e., time anchors), pairs of "input-target" samples are constructed (see appendix). Figure 3 ), Figure 3 It intuitively demonstrates how to use the peak value of the ECG R wave as a benchmark to extract the historical background window (input) and the target response window (true value) respectively, and extract the correspondence of heart rhythm features.

[0054] In this embodiment, an EEG background signal sequence is constructed based on a historical background window forward of the intercepted time. Can be set to capture time Previous historical time series window span (like The EEG signal segment within the window. This is the history background window. Include A sampling point is used to characterize the current background state of the brain (such as cortical excitability levels). This historical background window is designed to capture the baseline cortical excitability state of the brain prior to a heartbeat. Neurophysiological studies have shown that the background state of the brain before receiving afferent signals from the heart directly affects the efficiency and manner in which it processes those signals. The 2-second window length is sufficient to cover multiple complete slow-wave EEG cycles (such as Delta / Theta rhythms), providing the model with ample contextual information to predict upcoming neural responses.

[0055] In the embodiment, it is possible to extract After a moment to The EEG signal segment (as a target response window). This target response window contains... Each sampling point physically corresponds to the interval where the Hemorrhage Episode (HEP) occurs. The central nervous system's specific processing of the heartbeat is mainly manifested in the N250 and P400 potential components within 200-600 ms after the peak of the R wave. Extracting this specific interval not only covers the main energy peak of brain-heart interaction but also naturally filters out ECG artifacts at the moment of R wave occurrence, ensuring that the model learns a pure neural response rather than conduction noise. The EEG signal segments within this time window directly constitute the observed true value of a single HEP in the current cardiac cycle.

[0056] While traditional HEP analysis typically relies on multiple stacking averages to eliminate background noise, this invention employs a self-supervised generative learning paradigm designed to capture transient changes at the level of a single heartbeat. Therefore, cross-cycle stacking averages are unnecessary. The original EEG waveform sequence within the target response window is defined as the target signal for model training, and its corresponding physical object is the transient HEP response including the noisy background. Through training, the model learns how to predict this specific neural response component from historical context.

[0057] In the embodiments, and The value is determined by the length of the corresponding preset historical background window and target response window, and the ECG signal sampling rate. Decision. Specifically, the number of sampling points for the historical background window. The historical time series window span It can be set from 2.0 seconds to 5.0 seconds to cover sufficiently long background EEG rhythms (such as Delta wave cycles); the number of target response window sampling points. ,in This is the initial latency period of HEP. The end time is defined by this window, which covers the main energy range of the HEP.

[0058] In this embodiment, heart rhythm feature vectors are extracted based on an implicit heart rhythm calculation window. For the detected first The peak time of the R wave Construct a low-dimensional vector containing the current instantaneous heart rate features and local variability features. , The instantaneous RR interval is the time difference between the current R-wave peak time and the previous R-wave peak time. This indicator reflects the current instantaneous heart rate level.

[0059] In the embodiment, in order to capture the transient tension of the autonomic nervous system (especially the vagus nerve), calculations are performed using... The closest one before the current deadline The root mean square difference of each heart cycle is used as a local heart rate variability. :

[0060] .

[0061] Implicit heart rate calculation window coverage from The most recent time to go back in time One heart cycle (e.g.) ).

[0062] In the embodiment, the heart rhythm feature vector is finally processed. Z-score standardization is performed to eliminate numerical differences caused by different units (seconds and milliseconds), making its distribution suitable for the input requirements of the neural network. By combining instantaneous values ​​with local window statistics, this feature vector not only describes how fast the heart beats, but also how steady it is, providing a key basis for the subsequent collaborative attention mechanism to determine whether the heart is in an "independent oscillation state".

[0063] In this embodiment, the construction of the multimodal collaborative attention reconstruction network model is as follows: Figure 2As shown. This embodiment of the invention constructs a deep neural network based on an Encoder-Decoder architecture, which can be seen in [reference needed]. Figure 2 This demonstrates the connection methods and data flow of the EEG encoder (which can use CNN+LSTM), the ECG encoder (which can use MLP), the collaborative attention gating module, and the signal reconstruction decoder. The core of the model lies in introducing a "collaborative attention gating mechanism" to simulate the central nervous system's gating processing of cardiac information. MCAR-Net specifically includes a feature encoding module, a collaborative attention gating module, and a signal reconstruction decoding module.

[0064] Feature encoders include ECG encoders and EEG background encoders.

[0065] In this embodiment, the ECG encoder uses a multilayer perceptron to process low-dimensional heart rhythm feature vectors. Mapped to a high-dimensional heart state embedding vector : ;in, and For learnable weights and biases, This is the ReLU activation function.

[0066] In this embodiment, the EEG background encoder employs a one-dimensional convolutional neural network (BNN). Extracting historical background window The local temporal features are then fed into a Long Short-Term Memory (LSTM) network layer to capture long-range dependencies, ultimately outputting an EEG background embedding vector. :

[0067] ;

[0068] The collaborative attention gating module is used to calculate the “neural attention” that cardiac signals should receive in the current brain context.

[0069] In this embodiment, the collaborative attention gating module embeds the heart state into a vector. As a query item Perform linear projection to obtain the linear projection matrix of the query. Embedding EEG background vectors Each serves as a key item (Key, ) and value item (Value, Perform linear projections to obtain the linear projection matrices of the bonds. The linear projection matrix of the value The brain-heart coupling latent variables were calculated using a hybrid gating formula that combines multiplicative interaction terms and additive bias terms.

[0070] To preserve the heart's independent nature as a peripheral actuator (an organ located outside the central nervous system that receives nerve impulses and produces specific physiological effects) (referring to the inherent dynamic properties of a peripheral actuator that are not entirely dependent on central commands; for the heart, this independence is manifested in the automatic rhythm dominated by the sinoatrial node pacemaker), some embodiments involve using a hybrid gating formula incorporating "multiplicative interaction" and "additive bias" to calculate the brain-heart coupling latent variables. :

[0071] ;

[0072] in Represents the heart state embedding vector. Represents the EEG background embedding vector; : Linear projection matrices for query, key, and value, respectively, used to map features to the same latent space. Multiplicative interaction term. This term characterizes the "matching degree" or "synergy" between the brain background and the heart state. Under physiological homeostasis (such as during deep sleep), brain-heart synchronicity is strong, and this value is relatively high; in pathological decoupling, this value decreases significantly. Additive bias term. This metric represents the "independent significance" of cardiac signals. When the heart experiences dramatic rhythmic fluctuations (such as premature ventricular contractions and bigeminy) but the background noise on the electroencephalogram (EEG) is high, this metric ensures that the model can still capture cardiac abnormalities. This is the adaptive weight for the additive bias term; This is a scaling factor used to prevent gradient vanishing due to excessively large dot product results. As a non-linear activation function, this invention uses the Sigmoid function to map the interaction score to... The interval is used as a gating coefficient, which physically represents "the degree to which the brain accepts the cardiac signal characteristics" (0 represents inhibition, 1 represents activation). The output brain-heart coupling latent variable implies "the standard HEP response pattern that the brain should produce under the current physiological state".

[0073] As an example, the method also includes: calculating in real time the absolute deviation between the instantaneous RR interval of the current heartbeat cycle and the average instantaneous RR interval of the past K heartbeat cycles in the heart rhythm feature vector; when the absolute deviation is greater than the preset fluctuation threshold, it indicates that the heart has generated a cardiac abnormal rhythm, and the weight value of the additive bias term is increased to amplify the weight of the additive bias term in the hybrid gating formula, so as to ensure that the multimodal collaborative attention reconstruction network model can still capture cardiac abnormal features in a directional manner under the interference of whole-brain background noise.

[0074] In the embodiments, the signal reconstruction decoder aims to abstract latent variables. The waveform is restored to the time domain. In this embodiment, the signal reconstruction decoder uses a fully connected cascaded deconvolutional layer (...). ), coupling latent variables of brain and heart The mapping is restored to the same dimension as the true sequence of cardiac evoked potential observations (dimension is...). The predicted cardiac evoked potential waveform is represented as follows:

[0075] ;

[0076] in, That is, the predicted cardiac evoked potential waveform that the model predicts and that should theoretically exist. and These represent the learnable weight matrix and bias vector of the signal reconstruction decoder network layer, respectively. During model training, these two parameter matrices... and It will continuously update through the backpropagation algorithm, eventually learning how to abstract latent variables. Linear mapping and restoration to high-dimensional time series waveform data.

[0077] In this embodiment, the model is pre-built and trained. Multimodal data (health baseline) under normal physiological homeostasis can be selected from the MIT-BIH Polysomnography Database (slpdb). Segments labeled "NREM sleep stage (stage 2 or 3)" or "awake and calm stage" without apnea events are selected. This type of data represents the standard coupling pattern of the central nervous system and autonomic nervous system under physiological homeostasis.

[0078] Test data (pathological instability) can be selected from the CHB-MIT Scalp EEG Database (containing synchronized EEG / ECG signals). Pre- and intra-seizure segments are extracted as instability samples to validate the model's ability to detect brain-heart decoupling.

[0079] As an example, a complete self-supervised training strategy includes: inputting only the extracted health baseline data (MIT-BIH slpdb) into the model, without using any epilepsy or malignant arrhythmia data for training.

[0080] The loss function uses reconstruction error loss as the optimization objective, forcing the model to learn the distribution manifold of normal samples, as shown in the following formula:

[0081] ;

[0082] The physical meaning of this formula is: to calculate all... The true waveform of each sample With predicted waveform The average Euclidean distance between them. Where the superscript... Indicates the first in the current training batch There are 1 sample index, and N represents the size of the training batch. For the first Predicted cardiac evoked potential waveforms for each sample. This represents the true value of the corresponding cardiac evoked potential observation. For the brain-mind coupling latent variables output by the collaborative attention gating module, The sparse penalty coefficient is used to constrain the sparse transmission of latent spatial information by minimizing the joint loss function.

[0083] The first term is the mean squared error (MSE), which measures the similarity between the predicted waveform and the actual healthy waveform. The second term is the L1 regularization term, which constrains the sparsity of latent variables. From a neurophysiological perspective, normal brain-heart information interaction should be efficiently transmitted through specific neural pathways, rather than diffuse whole-brain activation. Through L1 regularization, the model is forced to filter out redundant background noise, retaining only the most significant brain-heart coupling feature dimensions, thereby improving the model's sensitivity to abnormal decoupling states.

[0084] Optimization process: Using the Adam optimizer with a learning rate of 0.001, network parameters are updated via backpropagation until the loss converges. The trained model is considered a nonlinear observer in a "normal brain-heart coupling mode".

[0085] To strictly ensure the objectivity of anomaly detection and prevent data leakage, the following strict data isolation strategy is adopted in this embodiment:

[0086] Training Phase: The model is input only with healthy baseline data. The model learns the brain-heart coupling manifold in a healthy state by continuously minimizing the reconstruction loss. During this phase, the model has never encountered any pathological samples of epilepsy or arrhythmia.

[0087] Convergence monitoring: Reserve 10% of the healthy data as a validation set to monitor the loss curve during training. When the loss on the validation set no longer decreases, stop training and save the model parameters to prevent the model from overfitting to specific healthy samples.

[0088] Blind Testing Strategy: The CHB-MIT epilepsy dataset mentioned above is strictly defined as a test set. This dataset is not involved in any model training, gradient updates, or parameter adjustments. It is only input into the system as "unknown data" after the model has been trained and its parameters have been fixed, to evaluate the sensitivity and specificity of the CSI index when abnormal events occur (see subsequent experimental results for validity verification based on real data). This strategy ensures that the model's identification of abnormal states is entirely based on its judgment of deviations from "normal patterns," rather than its memory of pathological features.

[0089] In this embodiment, the trained model is used for online monitoring of the stability of the central autonomic nervous system and early warning of anomalies.

[0090] In the inference phase of the implementation, the model is trained by inputting the acquired user's EEG background signal sequence and heart rhythm feature vector, and outputs a predicted cardiac evoked potential waveform. Simultaneously acquire the corresponding true sequence of cardiac evoked potentials. .

[0091] The residual energy between the real-time received observation sequence of cardiac evoked potentials and the predicted cardiac evoked potential waveforms is calculated and defined as the instantaneous reconstruction error. Then, a sliding average smoothing was applied (the sliding window length was...). For example, 10 heartbeat cycles), to obtain the Central-Autonomic Stability Index (CSI), the formula is as follows:

[0092] ;

[0093] ;

[0094] in, The summation index variable represents the value relative to the current time. Tracing back to the first Each sample point. The physical meaning of this formula is: for the current time... and its past Instantaneous reconstruction error within a given moment An arithmetic mean is calculated. This moving average process can filter out random noise interference in the calculation of a single heartbeat and reflect the overall stability trend of the brain-heart coupling state over a period of time (such as the most recent 10 heartbeats).

[0095] In this embodiment, when the Brain-Heart Stability Index (CSI) continuously exceeds a set threshold for a preset duration, a functional decoupling warning signal is automatically triggered.

[0096] As an example, the decision logic includes:

[0097] Steady state (normal): When the user is in a normal state, the brain-heart coupling mechanism conforms to the model's learned patterns, and the true sequence of cardiac evoked potentials is observed. With predicted cardiac evoked potential waveforms High overlap, CSI values ​​remain at a low baseline (e.g.) , (For threshold).

[0098] Instability (abnormality): When the monitored object is in a stage of physiological homeostasis imbalance—whether it stems from regulatory dysfunction of the autonomic nervous system (such as precursors to cardiogenic abnormalities) or from abnormal discharge interference of the central neural network (such as precursors to brain-related abnormalities)—the underlying brain-heart information interaction mechanism will change, causing the actual coupling response to deviate from the normal pattern. At this time, the model is still based on normal logical output. And the real The two exhibit either a disordered or flat appearance, leading to significant differences. The index rose sharply.

[0099] Warning Trigger: The system monitors the CSI curve in real time, and when... continuous If the time exceeds the preset threshold (e.g., 5 seconds), the error will be recorded. A warning will be triggered at that time.

[0100] In some embodiments, the threshold It can adaptively adjust based on the user's steady-state CSI sequence. For example, it includes: calculating the arithmetic mean μ (representing the individual's normal baseline level) and standard deviation σ (representing the random noise level of the individual's normal physiological fluctuation amplitude) of the steady-state CSI sequence.

[0101] ;

[0102] in It is a sensitivity adjustment coefficient, which can optionally be 2.0 to 3.0.

[0103] In this embodiment, to verify the effectiveness of the method, offline testing was conducted using epilepsy user data from the CHB-MIT database.

[0104] 1. Explanation of Dataset Differences: Considering the significant differences in domain distribution between the CHB-MIT database and the SLPDB database used for training, particularly in user age, acquisition environment, and baseline physiological characteristics, direct cross-database testing may lead to a high false alarm rate. Therefore, an "individualized baseline calibration" strategy is introduced. Before applying the universally self-supervised multimodal collaborative attention reconstruction network model to the user's central autonomic nervous system stability monitoring, multimodal calibration physiological signals are acquired, indicating the user is in a known physiological homeostasis phase. These signals are then input into the multimodal collaborative attention reconstruction network model. With minimizing the reconstruction error of the multimodal calibration physiological signals as the optimization objective, only the network parameters of the signal reconstruction decoder in the multimodal collaborative attention reconstruction network model are frozen and updated, allowing the model to adaptively adjust to the user's individualized background characteristics.

[0105] 2. Experiment setup and calibration procedure:

[0106] General model loading: First, load the general MCAR-Net model that has been pre-trained using only data from healthy individuals. This model already has basic brain-heart coupling feature extraction capabilities.

[0107] Baseline calibration: A clinically confirmed seizure-free period was selected from the test subject (chb01). During this period, the user was in a relatively stable physiological state. This data was used to perform lightweight transfer learning or parameter freeze fine-tuning on the general model. This process only adapted the model to the user's individualized background EEG characteristics, without inputting any epileptic seizure samples, ensuring that the basic principles of anomaly detection were not violated.

[0108] Offline testing: Keeping the calibrated model parameters unchanged, test by inputting continuous records containing epileptic seizure events.

[0109] 3. Results Analysis:

[0110] Interictal period (steady-state response): such as Figure 5 As shown, during periods away from the onset of symptoms, the CSI index remained low and oscillated (CSI≈0.2). This indicates that the model was able to accurately reconstruct the user's HEP waveform, the brain's neural response to the heartbeat was in a stable "time-locked" state, and the central-autonomic neural circuit function was normal.

[0111] Pre-seizure aura (early warning): Approximately 50 seconds before the onset time indicated by clinical experts, the monitoring system detected a non-random, significant upward trend in the CSI index, continuously exceeding a preset adaptive threshold (e.g., 0.3). This phenomenon corresponds to the "prodromal period" before a seizure, during which the excitability of the brain's neural networks changes, leading to subtle instability in its gating regulation of cardiac signals. Based on this, the system issues an early warning signal, validating the sensitivity of this method in capturing abnormalities in the "subclinical phase."

[0112] Acute phase (severe decoupling): As the clinical flare-up progresses, the CSI index experiences explosive growth and reaches its peak. At this point, the actual HEP waveform and the predicted waveform completely change from a positive correlation to no correlation or even a negative correlation (see appendix). Figure 7 This reflects that the disruption of the whole-brain network caused by epileptic discharge has completely destroyed the normal brain-heart coupling mechanism.

[0113] Conclusion: The experimental results strongly demonstrate that the method of this invention does not require prior learning of a large number of scarce pathological samples (such as rare epileptic seizures or malignant arrhythmia data). It achieves accurate identification and early warning of central-autonomic nervous system instability simply by learning and monitoring an individual's "reconstruction deviation" from normal physiological patterns. Compared to traditional methods that rely on abnormal feature classification, this method has stronger generalization ability and clinical practical value.

[0114] In the embodiments, Figure 4 and Figure 5 These are comparison images of the brain-heart coupling remodeling effect under normal physiological and pathological unstable conditions. Figure 4 The model-predicted heart evoked potential (HEP) waveform (dashed line) and the observed true sequence of heart evoked potentials (solid line) under healthy conditions are highly consistent, with a correlation coefficient Corr=0.87. Figure 5 The results show that under unstable conditions, the model is unable to reconstruct the EEG background based on ECG features, and the predicted waveform deviates significantly from the actual waveform, with the correlation coefficient dropping to Corr=-0.70.

[0115] Figure 6 This is an example of a time-series chart showing the CSI index changes during continuous monitoring of an epileptic user (chb01) using the method of the present invention. The chart illustrates the evolution of the normalized CSI index from the interictal period to the preictal phase and then to the ictal phase.

[0116] Figure 7 This is the ROC curve of the anomaly detection performance of the method of this invention on the test dataset. The figure shows the sensitivity and specificity of the model in distinguishing between normal and epileptic states, with an AUC of 0.892.

[0117] Figure 8 - Figure 11 This is a diagram illustrating the online monitoring effect and mechanism analysis of the present invention in real epilepsy cases.

[0118] Figure 8 This is a complete curve of the online monitoring stability index (CSI) of real epilepsy cases in this embodiment of the invention changing over time. Different colored blocks are used to mark the premonitory period and the ictal period, and three key time points are marked: A (normal), B (premonitory drift), and C (ictal outbreak).

[0119] Figure 9 In the embodiments of the present invention, corresponding to Figure 8 A comparison of the brain-heart coupling waveform reconstruction effect at time point A (normal steady state). At this point, the system is in a brain-heart coupling steady state with an extremely high correlation coefficient.

[0120] Figure 10 In the embodiments of the present invention, corresponding to Figure 8 Comparison of brain-heart coupling waveform reconstruction effects at time point B (pre-seizure drift). At this point, brain-heart decoupling occurs, and the correlation coefficient approaches 0.

[0121] Figure 11 In the embodiments of the present invention, corresponding to Figure 8 A comparison of brain-heart coupling waveform reconstruction results at time point C (seizure outbreak). At this point, the disruption of the whole-brain network leads to brain-heart decoupling, and the actual waveform shows a negative correlation with the predicted waveform.

[0122] Based on the same inventive concept as the central autonomic nervous system stability monitoring method provided in the above embodiments, this invention also provides a central autonomic nervous system stability monitoring system, including:

[0123] Signal acquisition and alignment module: used to synchronously acquire scalp EEG and ECG signals that are strictly aligned on the same time axis;

[0124] The temporal window feature construction module is used to detect the peak time of the R wave in the electrocardiogram signal; taking each peak time of the R wave as the intercept time, it constructs the EEG background signal sequence based on the historical background window before the intercept time, and constructs the true value sequence of cardiac evoked potential observation based on the target response window after the intercept time; and extracts the heart rhythm feature vector based on the implicit heart rhythm calculation window.

[0125] The deep generative reconstruction network module includes a built-in multimodal collaborative attention reconstruction network model trained under self-supervised conditions based on healthy physiological homeostasis data. This model takes real-time acquired and processed EEG background signal sequences and cardiac rhythm feature vectors as input to the trained multimodal collaborative attention reconstruction network model and outputs predicted cardiac evoked potential waveforms. The multimodal collaborative attention reconstruction network model includes an ECG encoder, an EEG background encoder, a collaborative attention gating module, and a signal reconstruction decoder. It is obtained through self-supervised training using multimodal data under normal physiological homeostasis.

[0126] Quantitative assessment and early warning module: used to calculate the residual energy between the real-time received heart rate evoked potential observation true value sequence and the predicted heart rate evoked potential waveform, obtain the instantaneous reconstruction error, and obtain the brain-heart stability index (CSI) through moving average processing; when the brain-heart stability index (CSI) continuously exceeds the set adaptive threshold and continues for a preset duration, the functional decoupling early warning signal is automatically triggered.

[0127] This invention also provides an electronic device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the central autonomic nervous system stability monitoring method provided in the above embodiments.

[0128] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or physical entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer can be, for example, a personal computer, a laptop computer, or a tablet computer, or any combination of these devices.

[0129] The above provides a detailed description of the central autonomic nervous system stability monitoring method and system provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the concept of this application and should not be construed as limiting the scope of protection of this application.

Claims

1. A method for monitoring the stability of the central autonomic nervous system, characterized in that, include: Acquire scalp EEG and ECG signals that are strictly aligned on the same timeline for the user; The peak time of the R wave in the electrocardiogram signal is detected; with each peak time of the R wave as the intercept time, an EEG background signal sequence is constructed based on the historical background window before the intercept time, and a true value sequence of cardiac evoked potential observations is constructed based on the target response window after the intercept time; and a cardiac rhythm feature vector is extracted based on the implicit cardiac rhythm calculation window. The real-time acquired and processed EEG background signal sequence and the heart rhythm feature vector are input into a trained multimodal collaborative attention reconstruction network model, which outputs a predicted cardiac evoked potential waveform. The multimodal collaborative attention reconstruction network model includes a feature encoder, a collaborative attention gating module, and a signal reconstruction decoder. It was obtained through self-supervised training using multimodal data under normal physiological homeostasis. The residual energy between the real-time received cardiac evoked potential observation true sequence and the predicted cardiac evoked potential waveform is calculated to obtain the instantaneous reconstruction error, and the brain-heart stability index is obtained by moving average processing; when the brain-heart stability index continuously exceeds the set threshold and continues for a preset time, a functional decoupling warning signal is automatically triggered.

2. The method for monitoring the stability of the central autonomic nervous system according to claim 1, characterized in that, The heart rhythm feature vector includes the instantaneous RR interval of the current heartbeat cycle calculated from the time difference between the current R wave peak time and the previous R wave peak time, and the root mean square difference of the instantaneous RR intervals of M heartbeat cycles traced back from the current intercept time. The heart rhythm feature vector is output after standardization.

3. The method for monitoring the stability of the central autonomic nervous system according to claim 1, characterized in that, The feature encoder includes an electrocardiogram encoder and an electroencephalogram background encoder; The electrocardiogram encoder uses a multilayer perceptron to map the heart rhythm feature vector into a heart state embedding vector. The EEG background encoder uses a one-dimensional convolutional neural network cascaded with a long short-term memory network to extract local and long-range temporal features of the EEG background signal sequence and outputs an EEG background embedding vector. The collaborative attention gating module uses the heart state embedding vector as a query term for linear projection to obtain the linear projection matrix of the query. The EEG background embedding vector is used as both key and value terms for linear projection, resulting in linear projection matrices for the keys. The linear projection matrix of the value The brain-heart coupling latent variables were calculated using a hybrid gating formula that combines multiplicative interaction terms and additive bias terms. : ； in, Represents the heart state embedding vector. Represents the EEG background embedding vector, with multiplication interaction terms. The additive bias term characterizes the matching or synergy between the brain background and the heart state. Independent significance characterizing cardiac signals; This is a scaling factor used to prevent gradient vanishing due to excessively large dot product results. It is a non-linear activation function. The weights of the additive bias terms; The signal reconstruction decoder employs fully connected cascaded deconvolutional layers to convert the brain-heart coupled latent variables. The mapping restores the predicted cardiac evoked potential waveforms to the same dimension as the observed true sequence of cardiac evoked potentials.

4. The method for monitoring the stability of the central autonomic nervous system according to claim 3, characterized in that, The method further includes: calculating in real time the absolute deviation between the instantaneous RR interval of the current heartbeat cycle and the average instantaneous RR interval of the past K heartbeat cycles in the heart rhythm feature vector; When the absolute deviation value is greater than the preset fluctuation threshold, it indicates that the heart is producing a cardiac abnormal rhythm. The weight value of the additive bias term is increased to ensure that the multimodal collaborative attention reconstruction network model can still capture cardiac abnormal features in a targeted manner under the interference of whole-brain background noise.

5. The method for monitoring the stability of the central autonomic nervous system according to claim 1, characterized in that, The optimization objective of the self-supervised training adopts a joint loss function that includes an L1 regularization term. The mathematical expression of the joint loss function is as follows: ； Among them, superscript This represents the sample index in the current training batch, and N represents the size of the training batch. The predicted cardiac evoked potential waveform for the j-th sample. This represents the true value of the corresponding cardiac evoked potential observation. For the brain-mind coupling latent variables output by the collaborative attention gating module, The sparse penalty coefficient is used to constrain the sparse propagation of latent spatial information by minimizing the joint loss function.

6. The method for monitoring the stability of the central autonomic nervous system according to claim 1, characterized in that, The method further includes an individualized baseline calibration step: before applying the multimodal collaborative attention reconstruction network model that has completed general self-supervised training to the user's central autonomic nervous system stability monitoring, multimodal calibration physiological signals of the user known to be in a physiological homeostasis period are acquired; the multimodal calibration physiological signals are input into the multimodal collaborative attention reconstruction network model, with minimizing the instantaneous reconstruction error of the multimodal calibration physiological signals as the optimization objective, and only the network parameters of the signal reconstruction decoder in the multimodal collaborative attention reconstruction network model are frozen and updated, so that the multimodal collaborative attention reconstruction network model adaptively adjusts to the user's individualized background features.

7. The method for monitoring the stability of the central autonomic nervous system according to claim 1, characterized in that, The formula for calculating the brain-heart stability index is as follows: ; ; in, The brain-heart stability index at time t. The waveform of the predicted cardiac evoked potential at time t. The true value of the cardiac evoked potential observed at time t; For instantaneous reconstruction error, The length of the sliding window. For summation index variables.

8. The method for monitoring the stability of the central autonomic nervous system according to claim 1, characterized in that, The method further includes processing the acquired raw EEG and ECG signals, including: performing full-band preprocessing on the raw ECG signals to remove baseline drift, and using the Pan-Tompkins algorithm to extract the R-wave peak time series. The raw EEG signal was preprocessed with full-band artifact suppression, and EEG and EMG artifacts were removed by independent component analysis algorithm. Then, the denoised EEG signal was standardized.

9. A central autonomic nervous system stability monitoring system, characterized in that, The system for implementing the method as described in any one of claims 1 to 8 comprises: Signal acquisition and alignment module: used to synchronously acquire scalp EEG and ECG signals that are strictly aligned on the same time axis; The temporal window feature construction module is used to detect the peak time of the R wave in the electrocardiogram signal; with each peak time of the R wave as the intercept time, it constructs an EEG background signal sequence based on the historical background window before the intercept time, and constructs a true value sequence of cardiac evoked potential observations based on the target response window after the intercept time; and extracts the heart rhythm feature vector based on the implicit heart rhythm calculation window. The deep generative reconstruction network module includes a built-in multimodal collaborative attention reconstruction network model trained under self-supervised conditions based on healthy physiological homeostasis data. This model takes the real-time acquired and processed EEG background signal sequence and the heart rhythm feature vector as input to the trained multimodal collaborative attention reconstruction network model and outputs a predicted cardiac evoked potential waveform. The multimodal collaborative attention reconstruction network model comprises an ECG encoder, an EEG background encoder, a collaborative attention gating module, and a signal reconstruction decoder. It is obtained through self-supervised training using multimodal data from normal physiological homeostasis. Quantitative assessment and early warning module: used to calculate the residual energy between the real-time received cardiac evoked potential observation true value sequence and the predicted cardiac evoked potential waveform, obtain the instantaneous reconstruction error, and obtain the brain-heart stability index through moving average processing; when the brain-heart stability index continuously exceeds the set threshold and continues for a preset time, a functional decoupling early warning signal is automatically triggered.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 8.

Images