Closed-loop neuromodulation system and method based on heart-brain interaction

Abstract

The application relates to a closed-loop nerve regulation system and method based on heart-brain interaction, which comprises four modules of multi-modal physiological signal acquisition, cross-modal feature fusion evaluation, Bayesian parameter optimization and median nerve electric stimulation output, can synchronously collect signs, fuse features to give scores, optimize stimulation parameters and apply electric stimulation, and forms a closed loop. Through the whole-link design of multi-modal signal synchronous acquisition, cross-modal feature intelligent fusion, stimulation parameter dynamic optimization and closed-loop electric stimulation intervention, the closed-loop nerve regulation system based on heart-brain interaction according to the embodiment of the application can effectively solve the core pain points of traditional consciousness disorder patient wake-up treatment, such as lack of real-time feedback, one-sided single-index evaluation, insufficient use of autonomic nervous signals and non-intelligent control of non-invasive stimulation, realize precise evaluation and individualized intervention, improve the wake-up efficiency, reduce the risk, and promote the data-driven change of nerve regulation.

Description

Technical Field

[0001] This application relates to the fields of neural modulation, biological signal processing and intelligent medical devices, and in particular to a closed-loop neural modulation system and method based on heart-brain interaction. Background Technology

[0002] Disorders of consciousness are a group of clinical syndromes that seriously affect patients' quality of life and prognostic assessment, commonly seen in cases of severe brain injury, stroke, and resuscitation after cardiac arrest. Typical states of consciousness include vegetative state (VS), minimally conscious state (MCS), and coma. Their pathogenesis involves multiple factors, including functional decoupling between the cerebral cortex and thalamus, dysfunction of central autonomic regulation, and impaired neural network connections. Currently, clinical treatments for promoting awakening often employ methods such as pharmacological stimulation (e.g., zolpidem, amantadine), physical stimulation (e.g., hyperbaric oxygen therapy), and neuromodulation (e.g., transcranial electrical stimulation, spinal cord stimulation, vagus nerve stimulation).

[0003] However, these intervention methods generally suffer from the following problems: a lack of real-time feedback mechanisms, making individualized and precise control difficult; most current neurostimulation or arousal therapies adopt an open model, i.e., continuous stimulation under pre-set parameters by the doctor, ignoring changes in the patient's real-time physiological state, which may lead to insufficient or excessive stimulation intensity, or even adverse reactions; traditional single-channel indicators are insufficient to comprehensively reflect the patient's state of consciousness; while commonly used electroencephalography (EEG) monitoring can reflect central nervous system function to some extent, in patients with impaired consciousness, the signal is often affected by artifacts and has complex characteristics, making it difficult to accurately judge the dynamic changes in consciousness level with a single EEG parameter; autonomic nervous system indicators are not fully utilized as a basis for regulation; heart rate variability (HRV), as an important indicator reflecting the activity of the autonomic nervous system, has been proven to be closely related to the state of consciousness and arousal level. For example, patients with restored consciousness often have improved sympathetic-parasympathetic balance and an increased HRV index. However, HRV has not yet been incorporated into the closed-loop control system in current clinical arousal interventions; non-invasive stimulation methods lack intelligent regulation mechanisms; median nerve electrical stimulation, as an emerging non-invasive neuromodulation technology, has shown potential in activating the ascending reticular activating system and improving arousal levels. It can act directly on the median nerve through a skin patch on the wrist, offering high safety and good patient compliance. However, there is currently no closed-loop control system for the median nerve based on real-time physiological feedback.

[0004] Therefore, there is an urgent need to develop a closed-loop neuromodulation system that integrates multimodal physiological signals such as HRV and EEG to dynamically assess a patient's state of consciousness and adjust stimulation parameters in real time accordingly. This system will help improve the efficiency of arousal in patients with impaired consciousness, reduce stimulation risks, and promote the transformation of neuromodulation from "experience-driven" to "data-driven." Summary of the Invention

[0005] In view of this, this application proposes a closed-loop neural modulation system based on heart-brain interaction, comprising: The multimodal physiological signal acquisition module is used to simultaneously acquire electrocardiogram (ECG) and electroencephalogram (EEG) signals from patients with impaired consciousness, and extract heart rate variability features from ECG signals and consciousness-related EEG features from EEG signals. The cross-modal feature fusion assessment module receives heart rate variability features and consciousness-related EEG features output by the multimodal physiological signal acquisition module. It performs cross-modal fusion processing on the two types of features through a Transformer-based multi-head self-attention convolutional fusion network and outputs a quantitative score of the patient's neurophysiological state. The Bayesian parameter optimization module is used to search and determine the individualized optimal combination of stimulation parameters within the preset median nerve electrical stimulation parameter space, based on the Gaussian process regression algorithm of the ARDMatern5 / 2 kernel, with the state score output by the cross-modal feature fusion evaluation module as the optimization target. The median nerve electrical stimulation output module receives the optimal combination of stimulation parameters output by the Bayesian parameter optimization module, applies electrical stimulation with corresponding parameters to the patient's median nerve through non-invasive skin electrodes, and transmits the stimulation feedback signal to the multimodal physiological signal acquisition module to form a closed-loop control.

[0006] In one possible implementation, the processing of ECG and EEG signals by the multimodal physiological signal acquisition module includes: The signal preprocessing unit filters, removes artifacts, and segments the acquired raw electrocardiogram and electroencephalogram signals to obtain signals with interference noise removed. The feature extraction unit extracts time-domain features, frequency-domain features, and nonlinear features from the preprocessed electrocardiogram signal as heart rate variability features; Multi-band power spectral density, bandwidth ratio, brain inter-brain phase synchronization index, and complexity class features were extracted from the preprocessed EEG signals as consciousness-related EEG features.

[0007] In one possible implementation, the Transformer multi-head self-attention-based convolutional fusion network in the cross-modal feature fusion evaluation module includes: The feature encoding network unit inputs heart rate variability features into a one-dimensional convolutional channel, and processes them through one-dimensional convolution, ReLU activation, max pooling, and fully connected layers to obtain a fixed-length representation vector of heart rate variability features; After converting consciousness-related EEG features into pseudo-image form, they are input into a two-dimensional convolutional channel. Through two-dimensional convolution, ReLU activation, max pooling, and fully connected layers, a fixed-length representation vector of the EEG features is obtained. The cross-modal attention fusion network unit generates a query using EEG feature representation vectors and generates a key and value using heart rate variability feature representation vectors. Attention weights are calculated using ScaledDot-ProductAttention, and then dynamic alignment and weighted fusion of the two types of features are achieved through multi-head attention extension to obtain context-enhanced fusion features. The fused features are input into a classifier, and after processing through average pooling, multilayer perceptron, and activation function, a quantified score of the patient's neurophysiological state is output.

[0008] In one possible implementation, the parameter optimization process of the Bayesian parameter optimization module includes: The parameter space definition unit uses the stimulation frequency, current magnitude, and pulse width of the median nerve electrical stimulation as optimization parameters to obtain a discretized stimulation parameter space. The Gaussian process regression optimization unit uses a Gaussian process regression algorithm based on the ARDMatern5 / 2 kernel to fit the response surface between the stimulus parameter combination and the patient's neurophysiological state score. First, the optimal current parameter is determined through single-parameter optimization, and then the parameter combination corresponding to the lowest point of the state score is searched in the multi-parameter joint optimization space as the individualized optimal stimulus parameter combination.

[0009] In one possible implementation, the non-invasive skin electrode in the median nerve electrical stimulation output module is placed on the patient's inner forearm at the wrist, and the electrode covers the main trunk of the median nerve.

[0010] In one possible implementation, the patient's neurophysiological state score output by the cross-modal feature fusion assessment module uses the clinical consciousness assessment scale score and diagnosis results as supervision labels. After discretization and normalization, the score range is set to 0-100, where a lower score indicates a better neurophysiological state and a higher score indicates a worse neurophysiological state.

[0011] This application also provides a closed-loop neural modulation method based on heart-brain interaction, comprising the following steps: The electrocardiogram (ECG) and electroencephalogram (EEG) signals of patients with impaired consciousness were simultaneously acquired using a multimodal physiological signal acquisition module, and heart rate variability features and consciousness-related EEG features were extracted from the ECG and EEG signals, respectively. The cross-modal feature fusion assessment module receives heart rate variability features and consciousness-related EEG features, which are then processed by a Transformer-based multi-head self-attention convolutional fusion network to output a quantitative score of the patient's neurophysiological state. Using the Bayesian parameter optimization module with state score as the optimization objective, and based on the Gaussian process regression algorithm of the ARDMatern5 / 2 kernel, the individualized optimal combination of stimulation parameters is searched and determined within the preset median nerve electrical stimulation parameter space. The optimal combination of stimulation parameters is received by the median nerve electrical stimulation output module, and electrical stimulation with corresponding parameters is applied to the patient's median nerve via non-invasive skin electrodes. The stimulation feedback signal is then transmitted to the multimodal physiological signal acquisition module to form a closed-loop control.

[0012] In one possible implementation, the electrocardiogram (ECG) and electroencephalogram (EEG) signals of a patient with impaired consciousness are simultaneously acquired via a multimodal physiological signal acquisition module, and heart rate variability features and consciousness-related EEG features are extracted from the ECG and EEG signals, respectively, including the following steps: The signal preprocessing unit of the multimodal physiological signal acquisition module filters, removes artifacts, and segments the acquired raw electrocardiogram (ECG) and electroencephalogram (EEG) signals to obtain pure ECG and EEG signals with interference noise removed. The feature extraction unit of the multimodal physiological signal acquisition module extracts time-domain features, frequency-domain features, and nonlinear features from the pure electrocardiogram signal to form heart rate variability features; The feature extraction unit of the multimodal physiological signal acquisition module extracts multi-band power spectral density, frequency band ratio, brain inter-brain phase synchronization index and complexity features from pure EEG signals to form consciousness-related EEG features.

[0013] In one possible implementation, the heart rate variability features and consciousness-related EEG features are received by a cross-modal feature fusion assessment module, processed by a Transformer-based multi-head self-attention convolutional fusion network, and a quantified patient neurophysiological state score is output, including the following steps: The feature encoding network unit of the cross-modal feature fusion evaluation module inputs the heart rate variability features into a one-dimensional convolutional channel, and processes them sequentially through one-dimensional convolution, ReLU activation, max pooling and fully connected layers to obtain a fixed-length representation vector of the heart rate variability features. The feature encoding network unit of the cross-modal feature fusion evaluation module converts consciousness-related EEG features into pseudo-image form and inputs them into a two-dimensional convolutional channel. The features are then processed sequentially through two-dimensional convolution, ReLU activation, max pooling, and fully connected layers to obtain a fixed-length representation vector of the EEG features. The cross-modal attention fusion network unit of the cross-modal feature fusion evaluation module generates a query using EEG feature representation vectors and a key and value using heart rate variability feature representation vectors. After calculating the attention weights through ScaledDot-ProductAttention, the two types of features are dynamically aligned and weighted fused through multi-head attention extension to obtain context-enhanced fusion features. The classifier of the cross-modal feature fusion assessment module sequentially processes the fused features through average pooling, multilayer perceptron, and activation function to output a quantitative score of the patient's neurophysiological state.

[0014] In one possible implementation, using a Bayesian parameter optimization module with state scores as the optimization objective, and based on the Gaussian process regression algorithm of the ARDMatern5 / 2 kernel, the individualized optimal combination of stimulation parameters is searched and determined within a preset median nerve electrical stimulation parameter space. This includes the following steps: By using the parameter space definition unit of the Bayesian parameter optimization module, the stimulation frequency, current magnitude, and pulse width of the median nerve electrical stimulation are set as optimization parameters, and a discretized stimulation parameter space is constructed and output. The Gaussian process regression optimization unit of the Bayesian parameter optimization module uses the Gaussian process regression algorithm based on the ARDMatern5 / 2 kernel to fit the response surface between the stimulus parameter combination and the patient's neurophysiological state score. The optimal current parameter is determined by first analyzing the correspondence between the stimulus current and the state score through the Gaussian process regression optimization unit of the Bayesian parameter optimization module. Through the Gaussian process regression optimization unit of the Bayesian parameter optimization module, in the multi-parameter joint optimization space composed of stimulation frequency, stimulation pulse width and state score, the parameter combination corresponding to the lowest point of state score and the optimal current parameter are searched to form an individualized optimal stimulation parameter combination.

[0015] The beneficial effects of this invention are: Through a full-link technology design encompassing multimodal signal synchronous acquisition, intelligent cross-modal feature fusion, dynamic optimization of stimulation parameters, and closed-loop electrical stimulation intervention, this application addresses the technical challenges of traditional awakening treatments for patients with impaired consciousness, such as lack of real-time feedback, one-sided single-indicator assessment, underutilization of autonomic nerve signals, and lack of intelligent regulation of non-invasive stimulation. The multimodal physiological signal acquisition module simultaneously acquires ECG and EEG signals, covering both autonomic and central nervous system physiological mechanisms, providing a comprehensive and accurate data foundation for state assessment and avoiding the limitations of single-channel signal assessment. The cross-modal feature fusion assessment module, based on a Transformer multi-head self-attention convolutional fusion network, achieves dynamic matching of HRV and heterogeneous EEG features. The system integrates quantification and weighted analysis to output a 0-100 point score of neurophysiological state, accurately matching the patient's actual level of consciousness and improving the accuracy of state assessment. The Bayesian parameter optimization module, relying on the Gaussian process regression algorithm of the ARDMatern5 / 2 kernel, quickly searches for the individualized optimal parameter combination within the discrete stimulation parameter space, avoiding the drawbacks of insufficient or excessive stimulation by traditional fixed parameters, and achieving precise control for each individual. The median nerve electrical stimulation output module applies optimized electrical stimulation through non-invasive electrodes and sends feedback signals back to the acquisition module to form a closed loop. This automated cycle of acquisition, analysis, optimization, and intervention can be completed without manual intervention, significantly improving the arousal efficiency of patients with impaired consciousness, reducing stimulation risks, and promoting the transformation of neuromodulation from experience-driven to data-driven.

[0016] Other features and aspects of this application will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0017] The accompanying drawings, which are included in and form part of this specification, illustrate exemplary embodiments, features, and aspects of this application together with the specification and serve to explain the principles of this application.

[0018] Figure 1 A block diagram of a closed-loop neuromodulation system based on heart-brain interaction according to an embodiment of this application is shown. Figure 2 A flowchart illustrating a closed-loop neural modulation method based on heart-brain interaction according to an embodiment of this application is shown; Figure 3 A flowchart illustrating a closed-loop neural modulation method based on heart-brain interaction according to an embodiment of this application is shown. Figure 4 A line graph showing the closed-loop neural modulation system and method based on heart-brain interaction according to an embodiment of this application; Figure 5 The image shows a time-frequency diagram of the electroencephalogram (EEG) signal according to an embodiment of this application; Figure 6A grouped bar chart showing the closed-loop neural modulation system and method based on heart-brain interaction according to embodiments of this application; Figure 7 An architecture diagram of a deep learning model according to an embodiment of this application is shown; Figure 8 The response surface plot and single-factor trend plot of the embodiments of this application are shown; Figure 9 A scatter plot showing the closed-loop neuromodulation system and method based on heart-brain interaction according to an embodiment of this application; Figure 10 A line graph showing the closed-loop neural modulation system and method based on heart-brain interaction according to an embodiment of this application; Figure 11 The response surface plot and single-factor trend plot of the embodiments of this application are shown; Figure 12 The graph shows the trend of multiple variables changing with the number of optimizations in an embodiment of this application; Detailed Implementation Various exemplary embodiments, features, and aspects of this application will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0019] It should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this application or to simplify the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0020] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0021] 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.

[0022] Furthermore, to better illustrate this application, numerous specific details are provided in the following detailed embodiments. Those skilled in the art should understand that this application can be implemented without certain specific details. In some instances, methods, means, components, and circuits well-known to those skilled in the art have not been described in detail in order to highlight the main points of this application.

[0023] This application presents a closed-loop neuromodulation system and method based on heart-brain interaction. This system integrates multimodal physiological signal analysis and intelligent parameter optimization, enabling automated signal acquisition, feature fusion assessment, stimulation parameter optimization, and closed-loop electrical stimulation intervention throughout the entire process. Applied to the fields of neuromodulation, biological signal processing, and intelligent medical devices, it accurately assesses the neurophysiological state of patients with consciousness disorders, dynamically generates individualized median nerve electrical stimulation programs, and addresses the pain points of traditional arousal therapy, such as lack of real-time feedback, biased assessment, and unintelligent regulation. This improves the arousal efficiency of patients with consciousness disorders, reduces stimulation risks, and promotes the transformation of neuromodulation from experience-driven to data-driven approaches.

[0024] Specific references Figures 1-12 As a specific embodiment of a closed-loop neural modulation system and method based on heart-brain interaction according to this application, the closed-loop neural modulation system based on heart-brain interaction includes: The multimodal physiological signal acquisition module 110 is used to simultaneously acquire electrocardiogram (ECG) and electroencephalogram (EEG) signals of patients with impaired consciousness, and extract heart rate variability features from ECG signals and consciousness-related EEG features from EEG signals.

[0025] Specifically, the multimodal physiological signal acquisition module 110 serves as the data entry point for the system's closed-loop control. Its core function is to solve the problem of synchronous acquisition and accurate feature extraction of ECG and EEG signals, providing high-quality physiological data support for subsequent state assessment. In the signal acquisition stage, it synchronously acquires ECG and multi-channel EEG signals from patients with impaired consciousness using dedicated monitoring equipment, avoiding analytical biases caused by signal asynchrony. In the feature extraction stage, for ECG signals, it focuses on extracting heart rate variability (HRV) features, covering time-domain, frequency-domain, and nonlinear indicators. For EEG signals, it focuses on extracting features directly related to the level of consciousness, including time-frequency domain features, functional connectivity features, and complexity features. All extracted features undergo preprocessing such as filtering, artifact removal, and segmentation to ensure data validity and reliability.

[0026] The cross-modal feature fusion assessment module 120 is used to receive heart rate variability features and consciousness-related EEG features output by the multimodal physiological signal acquisition module 110, and to perform cross-modal fusion processing on the two types of features through a Transformer-based multi-head self-attention convolutional fusion network to output a quantitative score of the patient's neurophysiological state.

[0027] Specifically, the cross-modal feature fusion assessment module 120 is the core of the system's intelligent analysis. By deeply fusing the heterogeneous features of ECG and EEG, it achieves a precise quantitative assessment of the patient's neurophysiological state. It first receives HRV features and consciousness-related EEG features output from the multimodal physiological signal acquisition module 110, and then processes them using a Transformer-based multi-head self-attention convolutional fusion network: For a one-dimensional HRV feature vector, a one-dimensional convolutional channel of "Conv1D+ReLU+MaxPooling+fully connected layer" is used to extract local rhythmic patterns in the time dimension and compress them into a fixed-length HRV feature representation. For multi-channel EEG signals, a two-dimensional convolutional channel of "Conv2D+ReLU+MaxPooling+fully connected layer" is used to capture spatial activity differences and multi-band spectral structures in different brain regions, generating an EEG feature representation. The key cross-modal fusion step is achieved through a multi-head self-attention mechanism: EEG features are used to generate a query, HRV features are used to generate a key and value, the attention weight matrix of the two in the time dimension is calculated, and then the multi-head extension is used to enhance the model's representation ability of different feature subspaces. Finally, the enhanced cross-modal features are concatenated or weighted with the original features and input into a classifier of "average pooling + MLP + Softmax / regression", and the output is a quantitative neurophysiological state score of 0-100 points, that is, 0 points represent the best state and 100 points represent the worst state. This score is directly used as the basis for subsequent stimulation parameter optimization.

[0028] The Bayesian parameter optimization module 130 is used to search and determine the individualized optimal combination of stimulation parameters within the preset median nerve electrical stimulation parameter space, based on the state score output by the cross-modal feature fusion evaluation module 120 as the optimization target and the Gaussian process regression algorithm of the ARDMatern5 / 2 kernel.

[0029] Specifically, the Bayesian parameter optimization module 130 is key to the system's individualized control. Using the state score output by the cross-modal feature fusion evaluation module 120 as the optimization target, it searches for the optimal median nerve electrical stimulation parameters within a preset parameter space using an efficient algorithm. First, it defines a discrete stimulation parameter space: the core stimulation parameters are divided into discrete values ​​with fixed step sizes according to clinical needs and equipment capabilities. This includes 18 logarithmic intervals for frequencies of 2-512Hz, 20 linear intervals for currents of 1-40mA, and 5 linear intervals for pulse widths of 100-500µs, forming 1800 potential parameter combinations. Subsequently, a Gaussian process regression algorithm based on the ARDMatern5 / 2 kernel is used for Bayesian optimization. Unlike traditional linear regression, this algorithm can achieve localized surface fitting through the kernel function (ARDMatern5 / 2 kernel), where the length scale of each parameter axis determines the surface flexibility. During the optimization process, Gaussian process regression estimates the average value and confidence of the state score for each point in the parameter space. By exploring and combining strategies, the optimal current is first determined in the single-parameter optimization stage, and then a three-dimensional space of frequency, pulse width and state score is constructed in the multi-parameter joint optimization stage. The parameter combination with the lowest score is searched, and finally the individualized optimal stimulus parameters are output. The entire optimization process takes an average of about 12.42 hours, balancing optimization accuracy and efficiency.

[0030] The median nerve electrical stimulation output module 140 is used to receive the optimal combination of stimulation parameters output by the Bayesian parameter optimization module 130, apply electrical stimulation with corresponding parameters to the patient's median nerve through non-invasive skin electrodes, and transmit the stimulation feedback signal to the multimodal physiological signal acquisition module to form a closed-loop control.

[0031] Specifically, the median nerve electrical stimulation output module 140 is the execution end of the system for intervention feedback. It is responsible for converting the optimized stimulation parameters into actual electrical stimulation operations and transmitting feedback signals to complete the closed-loop link. In terms of hardware implementation, non-invasive skin electrodes are used to avoid the risks of invasive operations and improve patient compliance. The stimulation parameters strictly follow the optimal combination output by the Bayesian optimization module 130, covering stimulation frequency, current intensity, and pulse width, and supporting single, periodic, or adaptive rhythm stimulation modes to meet the neural response characteristics of different patients. At the same time, the module is connected to the system's central processing unit in real time through a microcontroller: on the one hand, it receives the optimal parameter instructions to execute electrical stimulation; on the other hand, it transmits feedback signals such as the device's operating status and actual output parameters during the stimulation process to the multimodal physiological signal acquisition module 110. This allows the multimodal physiological signal acquisition module 110 to simultaneously monitor changes in ECG and EEG after stimulation, forming a complete closed-loop control link of acquisition, analysis, optimization, intervention, and re-acquisition, ensuring that the stimulation effect can be evaluated in real time and the parameters can be dynamically adjusted.

[0032] Furthermore, such as Figure 1 As shown, the closed-loop neuromodulation system 100 based on heart-brain interaction includes a multimodal physiological signal acquisition module 110, which consists of a signal preprocessing unit 111 and a feature extraction unit 112, used to simultaneously acquire ECG and EEG signals of patients with impaired consciousness, and extract heart rate variability features and consciousness-related EEG features; a cross-modal feature fusion assessment module 120, which consists of a feature encoding network unit 121 and a cross-modal attention fusion network unit 122, which performs cross-modal fusion of two types of features through a Transformer-based multi-head self-attention convolutional fusion network, and outputs the patient's neurophysiological state score; a Bayesian parameter optimization module 130, which consists of a parameter space definition unit 131 and a Gaussian process regression optimization unit 132, which uses the state score as the optimization target and determines the individualized optimal median nerve electrical stimulation parameter combination based on the ARDMatern5 / 2 kernel Gaussian process regression algorithm; and a median nerve electrical stimulation output module 140, which receives the optimal parameter combination, applies electrical stimulation, and transmits the feedback signal to the multimodal physiological signal acquisition module to form a closed-loop modulation.

[0033] Specifically, the closed-loop neuromodulation method based on heart-brain interaction includes: collecting and simultaneously acquiring multimodal physiological signals (including electrocardiogram and electroencephalogram signals), performing feature extraction and fusion analysis on the signals, and calculating and outputting assessment indicators or scores reflecting the subject's current neurophysiological state based on the extracted multimodal features. The closed-loop neuromodulation system based on heart-brain interaction includes a stimulation parameter optimization module and a stimulation output module. The stimulation parameter optimization module automatically selects and determines the optimal combination of stimulation parameters based on the state assessment results and predetermined goals using a Bayesian optimization algorithm; the stimulation output module executes corresponding neural stimulation operations based on the optimization results. By constructing a closed-loop modulation link of "identification-optimization-intervention," this invention can achieve individualized and intelligent neurofunctional intervention, significantly improving the accuracy and effectiveness of neuromodulation.

[0034] Furthermore, such as Figure 3As shown, the closed-loop neuromodulation system 100 based on heart-brain interaction of this application mainly includes three core parts: ECG and EEG signal acquisition, feature fusion and state assessment, and stimulation optimization and output. First, the system acquires the subject's ECG and EEG signals and extracts the heart rate variability (HRV) features of the ECG signals and the time-frequency domain features of the EEG signals. Then, multimodal feature fusion is achieved through convolution operations, and the fused features are analyzed using a deep learning model to output a comprehensive score reflecting the patient's neurophysiological state. Based on this state score, the system uses a Bayesian optimization algorithm to adaptively adjust the neural stimulation parameters in real time, optimizing key parameters such as stimulation frequency, current intensity, and stimulation pulse width. Finally, the stimulation output module executes corresponding neural stimulation operations according to the optimization results, thereby constructing a closed-loop regulation link of "acquisition—analysis—optimization—intervention" to achieve individualized and intelligent regulation of neural activity.

[0035] In one possible implementation, the processing of ECG and EEG signals by the multimodal physiological signal acquisition module includes: The signal preprocessing unit 111 performs filtering, artifact removal, and segmentation on the acquired raw electrocardiogram and electroencephalogram signals to obtain signals with interference noise removed.

[0036] The feature extraction unit 112 extracts time-domain features, frequency-domain features, and nonlinear features from the preprocessed electrocardiogram signal as heart rate variability features.

[0037] Multi-band power spectral density, bandwidth ratio, brain inter-brain phase synchronization index, and complexity class features were extracted from the preprocessed EEG signals as consciousness-related EEG features.

[0038] Specifically, this application proposes a closed-loop neural modulation system based on heart-brain interaction, comprising the following key modules: Combined ECG and EEG monitoring: This project employs a combined cardiac and brain monitoring method, which can simultaneously collect electrocardiogram (ECG) and electroencephalogram (EEG) signals to solve the problem of asynchronous ECG and EEG signals.

[0039] In ECG acquisition, raw ECG signals are acquired, RR interval sequences are extracted, and HRV features are obtained by analyzing time-domain, frequency-domain, and nonlinear indicators (such as SDNN, RMSSD, LF / HF ratio, Poincaré plot indicators, etc.).

[0040] Multichannel EEG signals were collected on EEG, with a focus on extracting features reflecting the level of consciousness, such as the α / δ ratio, slow wave power, entropy indices, functional connectivity parameters, and event-related potentials (ERPs).

[0041] All signals are processed through preprocessing (filtering, artifact removal, segmentation) and feature extraction modules for subsequent state assessment.

[0042] ECG and EEG feature fusion module: This module is the core algorithm of the system, designed to deeply fuse and comprehensively analyze the acquired multimodal physiological signals (ECG and EEG) to accurately assess the patient's current state of consciousness and provide precise basis for the generation of subsequent stimulus parameters. This invention employs a convolutional neural network structure based on attention mechanisms (CNN with attention fusion), specifically designed for joint modeling of the heterogeneous features of ECG HRV and EEG.

[0043] Feature extraction stage: HRV Feature Extraction: Extract standard time-domain features (such as SDNN, RMSSD), frequency-domain features (such as LF, HF, LF / HF), and nonlinear features from ECG signals to form a one-dimensional feature vector; EEG Feature Extraction: Wavelet transform or short-time Fourier transform is performed on the raw EEG signal to extract multi-band power spectral density (e.g., ...). , , , Two-dimensional time-frequency feature maps or multi-channel time-series feature matrices are constructed using features such as wave power, frequency band ratio, brain inter-brain phase synchronization index (e.g., PLI), and complexity class features.

[0044] Furthermore, such as Figure 4 and Figure 5 As shown, before performing multimodal feature fusion, the system first extracts features from the electrocardiogram (ECG) and electroencephalogram (EEG) signals respectively. Specifically, heart rate variability (HRV) features are extracted from the ECG signals. HRV can quantify subtle fluctuations in ECG signals and reflect the balance between the sympathetic and parasympathetic nervous system activities of the subject. Frequency-division time-spectral features are extracted from the EEG signals. These features retain both frequency and time dimensions of EEG signal information while effectively characterizing the dynamic response of brain activity to changes in external stimuli. Through the above feature extraction process, a high time-frequency resolution physiological feature foundation is provided for subsequent multimodal feature fusion and state assessment.

[0045] Furthermore, such as Figure 6 As shown, the accuracy of the model trained using fused ECG and EEG features is significantly higher than that trained using a single model feature.

[0046] In one possible implementation, the Transformer multi-head self-attention-based convolutional fusion network in the cross-modal feature fusion evaluation module includes: The feature encoding network unit 121 inputs the heart rate variability features into a one-dimensional convolutional channel, and processes them through one-dimensional convolution, ReLU activation, max pooling and fully connected layers to obtain a fixed-length representation vector of the heart rate variability features.

[0047] After converting consciousness-related EEG features into pseudo-image form, they are input into a two-dimensional convolutional channel. Through two-dimensional convolution, ReLU activation, max pooling, and fully connected layers, a fixed-length representation vector of the EEG features is obtained.

[0048] The cross-modal attention fusion network unit 122 generates a query using EEG feature representation vectors and generates a key and value using heart rate variability feature representation vectors. Attention weights are calculated using ScaledDot-ProductAttention, and then dynamic alignment and weighted fusion of the two types of features are achieved through multi-head attention extension to obtain context-enhanced fusion features.

[0049] The fused features are input into a classifier, and after processing through average pooling, multilayer perceptron, and activation function, a quantified score of the patient's neurophysiological state is output.

[0050] Specifically, the feature encoding network structure: The HRV feature vector is input into a one-dimensional convolutional channel and processed through multiple Conv1D+ReLU+MaxPool combinations to extract a temporally localized deep representation.

[0051] Structural components: Conv1D→ReLU→MaxPooling→FCLayer.

[0052] Function: Conv1D is used to extract local temporal patterns (such as abrupt changes and trends) from HRV. After stacking multiple layers, it can learn rhythmic changes at different scales. Finally, a fully connected layer is used to compress it into a fixed-length representation vector. .

[0053] EEG signals are input into a two-dimensional convolutional channel in a multi-channel temporal form, and Conv2D layers are used to extract the spatial dependence and spectral structure between different brain regions and different frequency bands.

[0054] Input format: Where C is the number of leads and T is the time point, it is converted into a pseudo-image form and input into Conv2D.

[0055] Structural components: Conv2D→ReLU→MaxPooling→FCLayer.

[0056] Functions: Conv2D extracts spatial patterns (such as activity differences between the frontal and occipital regions) and spectral distribution patterns (such as alpha-delta wave power differences) between brain regions; it is particularly suitable for capturing local structures in "multi-channel + multi-band" EEG; the FC layer embeds them as vectors. .

[0057] Cross-modalAttentionFusion: In this stage, a multi-head attention mechanism (MHA) is used to align and weight the high-order features of the EEG and HRV modalities.

[0058] We use the classic ScaledDot-ProductAttention from Transformer as the basis to build a cross-modal attention module: 1. Query-Key-Value generation: Generate a query from the EEG modality and a key and value from the HRV modality.

[0059] in, , , These are learnable parameters.

[0060] 2. Attention weight calculation: Calculate the attention of the EEG modality to the HRV modality. in, It is the attention weight matrix between EEG and HRV in the time dimension, representing the attention of each EEG time point to all HRV time points.

[0061] 3. Multi-Head Attention: To enhance the model's ability to represent different subspaces, a multi-head mechanism is employed. Where H is the number of heads. An independent projection matrix for each head.

[0062] Output and Status Evaluation: Finally, through cross-modal attention, the EEG modality obtains a context-enhanced representation from the HRV modality: This enhanced feature can be concatenated or weighted and fused with the original EEG feature: Similarly, a symmetric attention mechanism for HRV←EEG can be designed to construct a bidirectional cross-attention model.

[0063] The fused representation Input classifier: Furthermore, such as Figure 7 As shown, this application employs a deep learning model based on the Transformer framework to analyze multimodal fusion features and output a patient status score. The model's score ranges from 0 to 100, where 0 represents the patient's best neurophysiological state and 100 represents the worst state. During model training, clinical scale scores and diagnostic results are used as supervision labels, and they are discretized and normalized to ensure that the score results reflect the patient's overall neurofunctional state within a unified numerical range, thus achieving a quantitative assessment of individual physiological status.

[0064] In one possible implementation, the parameter optimization process of the Bayesian parameter optimization module includes: The parameter space definition unit 131 uses the stimulation frequency, current magnitude, and stimulation pulse width of the median nerve electrical stimulation as optimization parameters to obtain a discretized stimulation parameter space.

[0065] The Gaussian process regression optimization unit 132 uses a Gaussian process regression algorithm based on the ARDMatern5 / 2 kernel to fit the response surface between the stimulus parameter combination and the patient's neurophysiological state score. First, the optimal current parameter is determined through single-parameter optimization, and then the parameter combination corresponding to the lowest point of the state score is searched in the multi-parameter joint optimization space as the individualized optimal stimulus parameter combination.

[0066] Specifically, the Bayesian-based closed-loop electrical stimulation parameter control module: (1) Define the discrete parameter space: In the optimization of stimulation parameters, the optimized parameters include: stimulation frequency (2-512H divided into 18 logarithmic intervals), current magnitude (1-40mA divided into 20 linear intervals), and stimulation pulse width (100-500µs divided into 5 linear intervals). Therefore, the discretized parameter space contains 1800 combinations of stimulation parameters. The output state of the evaluation model, obtained through the feature fusion described above, is used as the parameter (output) to evaluate the effect of the stimulation combination.

[0067] (2) Bayesian optimization based on Gaussian regression: Bayesian optimization on Gaussian process regression is a sequential approach to spatial exploration to efficiently find the optimal value. Similar to the more common linear regression, Gaussian process regression fits a surface to the data points (see appendix). Figure 5 However, unlike linear regression, Gaussian process regression can bend locally to better fit the data; the flexibility of the response surface is determined by the kernel function. We used the ARDMatern5 / 2 kernel, where the surface flexibility comes from the length scale of each parameter axis. ARDMatern5 / 2 kernel: in The length scale (ARD) for each dimension. Let V be the signal variance. The smaller the value, the more "flexible" the model is in that dimension (more sensitive to local changes and less affected by neighborhood influences). The larger the value, the smoother the surface and the stronger its generalization ability. Pulse width dimension (5 steps). Stimulation frequency dimension (18 steps) Stimulation current dimension (20 steps) In the fitted plane, at each point in the plane, Gaussian process regression estimates the mean and confidence level of the optimization variables.

[0068] Furthermore, such as Figure 8 As shown, this application uses the state score output by the aforementioned deep learning model as the basis for optimizing stimulation parameters and employs a Bayesian optimization algorithm to search for the optimal combination of stimulation parameters. Specifically, the system intervenes in the patient's physiological state under different combinations of stimulation parameters and evaluates the intervention effect corresponding to each parameter combination based on the changes in the state score output by the model. First, in the single-parameter optimization stage, the current intensity at the lowest point of the corresponding score is determined by analyzing the binary relationship curve between the stimulation current intensity and the state score, and is taken as the optimal current parameter. Subsequently, in the multi-parameter joint optimization stage, the stimulation frequency, stimulation pulse width, and state score are constructed into a three-dimensional optimization space, and the Bayesian optimization algorithm is used to search for the lowest point of the state score in this space; the corresponding stimulation frequency and pulse width are the optimal parameter combination.

[0069] Furthermore, such as Figure 9 As shown, the average time for finding the optimal parameter combination through parameter optimization is 12.42 hours.

[0070] In one possible implementation, the non-invasive skin electrode in the median nerve electrical stimulation output module is placed on the patient's inner forearm at the wrist, and the electrode covers the main trunk of the median nerve.

[0071] In one possible implementation, the patient's neurophysiological state score output by the cross-modal feature fusion assessment module uses the clinical consciousness assessment scale score and diagnosis results as supervision labels. After discretization and normalization, the score range is set to 0-100, where a lower score indicates a better neurophysiological state and a higher score indicates a worse neurophysiological state.

[0072] Furthermore, such as Figure 10 , Figure 11 and Figure 12 As shown, the patient was in a state of impaired consciousness upon admission, clinically assessed as minimally conscious, with a CRS-R score of 8. The effect of 14 days of median nerve intervention with fixed parameters was as follows: Figure 10 As shown, the CRS-R scale score showed no significant change. Subsequently, the patient underwent closed-loop median nerve electrical stimulation intervention. After 8.3 hours of optimal parameter search, the optimal combination of stimulation parameters was obtained. The optimal parameter plane is shown below. Figure 11 As shown, the optimization path for each parameter is as follows: Figure 12 As shown, after intervention using optimal parameters, their CRS-R score significantly improved over 14 consecutive days, eventually reaching a level of 16.

[0073] This application also provides a closed-loop neural modulation method based on heart-brain interaction, comprising the following steps: 100. The electrocardiogram (ECG) and electroencephalogram (EEG) signals of patients with impaired consciousness are simultaneously acquired through the multimodal physiological signal acquisition module, and heart rate variability features and consciousness-related EEG features are extracted from the ECG and EEG signals, respectively.

[0074] Specifically, step 100 is the data foundation construction stage of this method. Its core objective is to simultaneously acquire high-quality ECG and EEG signals and extract features strongly correlated with the state of consciousness, providing reliable input for subsequent evaluation. In the signal acquisition stage, dedicated monitoring equipment is used to simultaneously acquire ECG and EEG signals. ECG signal acquisition focuses on the raw ECG waveform, particularly capturing the RR intervals within the heartbeat cycle to form an RR interval sequence. EEG signals are acquired using multi-electrode acquisition, covering key brain regions related to consciousness to ensure the signals reflect the overall state of cerebral cortex activity. In the feature extraction stage, for ECG signals, three types of heart rate variability (HRV) features are extracted from the RR interval sequence: time-domain features, frequency-domain features, and nonlinear features. For EEG signals, consciousness-related features are extracted, including time-frequency domain features, functional connectivity features, and complexity features. All features undergo preprocessing before extraction to ensure their effectiveness and accuracy.

[0075] 200. The heart rate variability feature and consciousness-related EEG feature are received through the cross-modal feature fusion assessment module, processed by a Transformer-based multi-head self-attention convolutional fusion network, and output a quantitative score of the patient's neurophysiological state.

[0076] Specifically, step 200 is the core intelligent analysis step of this method. By deeply fusing heterogeneous ECG and EEG features, it achieves a precise quantitative assessment of the patient's neurophysiological state. First, the cross-modal feature fusion assessment module receives the HRV features and consciousness-related EEG features output from the previous step. Then, it processes these features in stages through a Transformer-based multi-head self-attention convolutional fusion network: The first stage is feature encoding. HRV features are input into a one-dimensional convolutional channel to extract local rhythmic patterns in the time dimension and compress them into fixed-length HRV feature vectors. EEG features are converted into pseudo-images and input into a two-dimensional convolutional channel to capture spatial activity differences and multi-band spectral structures in different brain regions, generating EEG feature vectors. The second stage is cross-modal fusion. Using the Transformer multi-head self-attention mechanism, the EEG feature vector generates a query, and the HRV feature vector generates a key and value. The attention weight matrix of the two in the time dimension is calculated. Multi-head extension enhances the model's representation ability for different feature subspaces. Finally, the enhanced cross-modal features are concatenated or weighted with the original features. The third stage is the output of the state score. The fused features are processed by average pooling and multilayer perceptron (MLP) and output as a quantitative neurophysiological state score of 0-100 through regression or softmax function. This score directly reflects the patient's current neurophysiological state related to consciousness and provides the core basis for subsequent optimization of stimulation parameters.

[0077] 300. Using the Bayesian parameter optimization module with state score as the optimization objective, and based on the Gaussian process regression algorithm of the ARDMatern5 / 2 kernel, the individualized optimal combination of stimulation parameters is searched and determined within the preset median nerve electrical stimulation parameter space.

[0078] Specifically, step 300 is the parameter optimization decision-making stage of this method. Using the neurophysiological state score as the optimization objective, an efficient algorithm determines individualized median nerve electrical stimulation parameters within a preset parameter space. First, a discrete stimulation parameter space is defined: core stimulation parameters are divided according to clinical needs and equipment capabilities. Stimulation frequency (2-512Hz) is divided into 18 logarithmic interval steps, current magnitude (1-40mA) into 20 linear interval steps, and stimulation pulse width (100-500µs) into 5 linear interval steps, ultimately forming 1800 potential parameter combinations. Subsequently, a Gaussian process regression algorithm based on the ARDMatern5 / 2 kernel is used to perform Bayesian optimization. This algorithm achieves localized surface fitting through a kernel function. The length scale parameter of the ARDMatern5 / 2 kernel determines the surface flexibility: pulse width dimension length scale l_w=120µs, frequency dimension l_f=3.4dB, and current dimension l_A=2mA. The smaller the length scale, the more sensitive the model is to local changes in that parameter dimension. During the optimization process, Gaussian process regression estimates the average value and confidence level of the state score for each point in the parameter space. It balances search efficiency and accuracy through exploration and utilization strategies: first, the optimal current is determined in the single-parameter optimization stage, and then a three-dimensional space of frequency, pulse width and state score is constructed in the multi-parameter joint optimization stage to search for the parameter combination with the lowest score. Finally, it outputs the individualized optimal stimulation parameters that are suitable for the patient's current state. The average optimization time is about 12.42 hours, which balances accuracy and clinical applicability.

[0079] 400. The optimal combination of stimulation parameters is received through the median nerve electrical stimulation output module, and electrical stimulation with corresponding parameters is applied to the patient's median nerve through non-invasive skin electrodes. The stimulation feedback signal is then transmitted to the multimodal physiological signal acquisition module to form a closed-loop control.

[0080] Specifically, step 400 is the intervention and feedback closed-loop link of this method, responsible for translating the optimized parameters into actual stimulation operations and maintaining closed-loop regulation through feedback signals. First, the median nerve electrical stimulation output module receives the optimal combination of stimulation parameters output by the Bayesian parameter optimization module. Then, stimulation is performed through non-invasive skin electrodes placed on the patient's inner forearm at the wrist, precisely covering the main trunk of the median nerve, avoiding the risks of invasive procedures and improving patient compliance. The stimulation parameters strictly adhere to the optimal combination, covering stimulation frequency, current intensity, and pulse width, supporting single, periodic, or adaptive rhythm modes to suit the different neural response characteristics of patients. Simultaneously, the module is connected in real-time to the system's central processing unit via a microcontroller, transmitting key feedback signals during the stimulation process to the multimodal physiological signal acquisition module. After receiving the feedback signal, the multimodal physiological signal acquisition module will simultaneously acquire new ECG and EEG signals after stimulation, and re-execute the feature extraction, fusion assessment, and parameter optimization process to form a complete closed-loop control link of acquisition, analysis, optimization, intervention, and re-acquisition. This ensures that the stimulation parameters can be adjusted in real time according to the dynamic changes in the patient's neurophysiological state, thereby achieving individualized and dynamic neuromodulation.

[0081] Furthermore, such as Figure 2 As shown, the closed-loop neuromodulation method based on heart-brain interaction first executes step 100, simultaneously acquiring electrocardiogram (ECG) and electroencephalogram (EEG) signals of patients with impaired consciousness through a multimodal physiological signal acquisition module, and extracting heart rate variability features and consciousness-related EEG features from them, respectively. Next, step 200 executes, receiving the above two types of features through a cross-modal feature fusion evaluation module, processing them through a Transformer-based multi-head self-attention convolutional fusion network, and outputting a quantified neurophysiological state score for the patient. Then, step 300 executes, using the state score as the optimization target, a Bayesian parameter optimization module searches and determines the individualized optimal combination of stimulation parameters within a preset median nerve electrical stimulation parameter space based on an ARDMatern5 / 2 kernel Gaussian process regression algorithm using a Bayesian parameter optimization module. Finally, step 400 executes, receiving the optimal combination of stimulation parameters through a median nerve electrical stimulation output module, applying electrical stimulation with corresponding parameters to the patient's median nerve via non-invasive skin electrodes, and transmitting the stimulation feedback signal to the multimodal physiological signal acquisition module, forming a closed-loop modulation.

[0082] In one possible implementation, the electrocardiogram (ECG) and electroencephalogram (EEG) signals of a patient with impaired consciousness are simultaneously acquired via a multimodal physiological signal acquisition module, and heart rate variability features and consciousness-related EEG features are extracted from the ECG and EEG signals, respectively, including the following steps: The signal preprocessing unit of the multimodal physiological signal acquisition module filters, removes artifacts, and segments the acquired raw electrocardiogram (ECG) and electroencephalogram (EEG) signals to obtain pure ECG and EEG signals with interference noise removed.

[0083] The feature extraction unit of the multimodal physiological signal acquisition module extracts time-domain features, frequency-domain features, and nonlinear features from the pure electrocardiogram signal to form heart rate variability features.

[0084] The feature extraction unit of the multimodal physiological signal acquisition module extracts multi-band power spectral density, frequency band ratio, brain inter-brain phase synchronization index and complexity features from pure EEG signals to form consciousness-related EEG features.

[0085] In one possible implementation, the heart rate variability features and consciousness-related EEG features are received by a cross-modal feature fusion assessment module, processed by a Transformer-based multi-head self-attention convolutional fusion network, and a quantified patient neurophysiological state score is output, including the following steps: The feature encoding network unit of the cross-modal feature fusion evaluation module inputs the heart rate variability features into a one-dimensional convolutional channel, and processes them sequentially through one-dimensional convolution, ReLU activation, max pooling and fully connected layers to obtain a fixed-length representation vector of the heart rate variability features.

[0086] The feature encoding network unit of the cross-modal feature fusion evaluation module converts consciousness-related EEG features into pseudo-image form and inputs them into a two-dimensional convolutional channel. The features are then processed sequentially through two-dimensional convolution, ReLU activation, max pooling, and fully connected layers to obtain a fixed-length representation vector of the EEG features.

[0087] The cross-modal attention fusion network unit of the cross-modal feature fusion evaluation module generates a query using EEG feature representation vectors and a key and value using heart rate variability feature representation vectors. After calculating the attention weights through ScaledDot-ProductAttention, the two types of features are dynamically aligned and weighted fused through multi-head attention extension to obtain context-enhanced fused features.

[0088] The classifier of the cross-modal feature fusion assessment module sequentially processes the fused features through average pooling, multilayer perceptron, and activation function to output a quantitative score of the patient's neurophysiological state.

[0089] In one possible implementation, using a Bayesian parameter optimization module with state scores as the optimization objective, and based on the Gaussian process regression algorithm of the ARDMatern5 / 2 kernel, the individualized optimal combination of stimulation parameters is searched and determined within a preset median nerve electrical stimulation parameter space. This includes the following steps: By using the parameter space definition unit of the Bayesian parameter optimization module, the stimulation frequency, current magnitude, and pulse width of the median nerve electrical stimulation are set as optimization parameters, and a discretized stimulation parameter space is constructed and output.

[0090] The Gaussian process regression optimization unit of the Bayesian parameter optimization module uses the Gaussian process regression algorithm based on the ARDMatern5 / 2 kernel to fit the response surface between the stimulus parameter combination and the patient's neurophysiological state score.

[0091] The optimal current parameters are determined by first analyzing the correspondence between the stimulus current and the state score through the Gaussian process regression optimization unit of the Bayesian parameter optimization module.

[0092] Through the Gaussian process regression optimization unit of the Bayesian parameter optimization module, in the multi-parameter joint optimization space composed of stimulation frequency, stimulation pulse width and state score, the parameter combination corresponding to the lowest point of state score and the optimal current parameter are searched to form an individualized optimal stimulation parameter combination.

[0093] This application presents a closed-loop neuromodulation system and method based on heart-brain interaction. This system simultaneously acquires ECG and EEG signals from patients with impaired consciousness via a multimodal physiological signal acquisition module, preprocessing and extracting heart rate variability and consciousness-related EEG features. A cross-modal feature fusion assessment module, relying on a Transformer-based multi-head self-attention convolutional fusion network, encodes and dynamically aligns the two types of features, outputting a 0-100 score quantifying the neurophysiological state. A Bayesian parameter optimization module, based on an ARDMatern5 / 2 kernel Gaussian process regression algorithm, first optimizes single parameters to determine the optimal current within a discretized stimulation parameter space, then searches for the optimal combination of multiple parameters. A median nerve electrical stimulation output module applies optimized parameter electrical stimulation via a non-invasive electrode at the wrist on the inner forearm and transmits the feedback signal back to the acquisition module, forming a closed loop. This addresses the pain points of traditional arousal therapy, such as lack of real-time feedback, one-sided assessment, and unintelligent regulation. It achieves precise assessment of the patient's neurophysiological state and individualized electrical stimulation intervention, significantly improving the arousal efficiency of patients with impaired consciousness, reducing stimulation risks, and promoting the transformation of neuromodulation from experience-driven to data-driven approaches.

[0094] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A closed-loop neural modulation system based on heart-brain interaction, characterized in that, include: The multimodal physiological signal acquisition module is used to simultaneously acquire electrocardiogram (ECG) and electroencephalogram (EEG) signals from patients with impaired consciousness, and extract heart rate variability features from ECG signals and consciousness-related EEG features from EEG signals. The cross-modal feature fusion assessment module is used to receive heart rate variability features and consciousness-related EEG features output by the multimodal physiological signal acquisition module, and to perform cross-modal fusion processing on the two types of features through a Transformer-based multi-head self-attention convolutional fusion network to output a quantitative score of the patient's neurophysiological state. The Bayesian parameter optimization module is used to search and determine the individualized optimal combination of stimulation parameters within the preset median nerve electrical stimulation parameter space, based on the Gaussian process regression algorithm of the ARDMatern5 / 2 kernel, with the state score output by the cross-modal feature fusion evaluation module as the optimization target. The median nerve electrical stimulation output module is used to receive the optimal combination of stimulation parameters output by the Bayesian parameter optimization module, apply electrical stimulation with corresponding parameters to the patient's median nerve through non-invasive skin electrodes, and transmit the stimulation feedback signal to the multimodal physiological signal acquisition module to form a closed-loop control.

2. The closed-loop neural modulation system based on heart-brain interaction according to claim 1, characterized in that, The multimodal physiological signal acquisition module processes electrocardiogram (ECG) and electroencephalogram (EEG) signals, including: The signal preprocessing unit filters, removes artifacts, and segments the acquired raw electrocardiogram and electroencephalogram signals to obtain signals with interference noise removed. The feature extraction unit extracts time-domain features, frequency-domain features, and nonlinear features from the preprocessed electrocardiogram signal as heart rate variability features; Multi-band power spectral density, bandwidth ratio, brain inter-brain phase synchronization index, and complexity class features were extracted from the preprocessed EEG signals as consciousness-related EEG features.

3. The closed-loop neural modulation system based on heart-brain interaction according to claim 1, characterized in that, The cross-modal feature fusion evaluation module includes a Transformer-based multi-head self-attention convolutional fusion network, which comprises: The feature encoding network unit inputs heart rate variability features into a one-dimensional convolutional channel, and processes them through one-dimensional convolution, ReLU activation, max pooling, and fully connected layers to obtain a fixed-length representation vector of heart rate variability features; After converting consciousness-related EEG features into pseudo-image form, they are input into a two-dimensional convolutional channel. Through two-dimensional convolution, ReLU activation, max pooling, and fully connected layers, a fixed-length representation vector of the EEG features is obtained. The cross-modal attention fusion network unit generates a query using EEG feature representation vectors and generates a key and value using heart rate variability feature representation vectors. Attention weights are calculated using ScaledDot-ProductAttention, and then dynamic alignment and weighted fusion of the two types of features are achieved through multi-head attention extension to obtain context-enhanced fusion features. The fused features are input into a classifier, and after processing through average pooling, multilayer perceptron, and activation function, a quantified score of the patient's neurophysiological state is output.

4. The closed-loop neural modulation system based on heart-brain interaction according to claim 1, characterized in that, The parameter optimization process of the Bayesian parameter optimization module includes: The parameter space definition unit uses the stimulation frequency, current magnitude, and pulse width of the median nerve electrical stimulation as optimization parameters to obtain a discretized stimulation parameter space. The Gaussian process regression optimization unit uses a Gaussian process regression algorithm based on the ARDMatern5 / 2 kernel to fit the response surface between the stimulus parameter combination and the patient's neurophysiological state score. First, the optimal current parameter is determined through single-parameter optimization, and then the parameter combination corresponding to the lowest point of the state score is searched in the multi-parameter joint optimization space as the individualized optimal stimulus parameter combination.

5. The closed-loop neural modulation system based on heart-brain interaction according to claim 1, characterized in that, The non-invasive skin electrode in the median nerve electrical stimulation output module is placed on the inner side of the patient's forearm at the wrist, and the electrode covers the main trunk of the median nerve.

6. The closed-loop neural modulation system based on heart-brain interaction according to claim 1, characterized in that, The patient neurophysiological status score output by the cross-modal feature fusion assessment module uses the clinical consciousness assessment scale score and diagnosis results as supervision labels. After discretization and normalization, the score range is set to 0-100, where a lower score indicates a better neurophysiological status and a higher score indicates a worse neurophysiological status.

7. A closed-loop neural modulation method based on heart-brain interaction, used to implement the closed-loop neural modulation system based on heart-brain interaction as described in any one of claims 1-6, characterized in that, Includes the following steps: The electrocardiogram (ECG) and electroencephalogram (EEG) signals of patients with impaired consciousness were simultaneously acquired using a multimodal physiological signal acquisition module, and heart rate variability features and consciousness-related EEG features were extracted from the ECG and EEG signals, respectively. The cross-modal feature fusion assessment module receives heart rate variability features and consciousness-related EEG features, which are then processed by a Transformer-based multi-head self-attention convolutional fusion network to output a quantitative score of the patient's neurophysiological state. Using the state score as the optimization target, the Bayesian parameter optimization module searches and determines the individualized optimal combination of stimulation parameters within the preset median nerve electrical stimulation parameter space based on the Gaussian process regression algorithm of the ARDMatern5 / 2 kernel. The optimal combination of stimulation parameters is received by the median nerve electrical stimulation output module, and electrical stimulation with corresponding parameters is applied to the patient's median nerve via non-invasive skin electrodes. The stimulation feedback signal is then transmitted to the multimodal physiological signal acquisition module to form a closed-loop control.

8. The closed-loop neural modulation method based on heart-brain interaction according to claim 7, characterized in that, The process of simultaneously acquiring electrocardiogram (ECG) and electroencephalogram (EEG) signals from patients with impaired consciousness using a multimodal physiological signal acquisition module, and extracting heart rate variability features and consciousness-related EEG features from the ECG and EEG signals respectively, includes the following steps: The signal preprocessing unit of the multimodal physiological signal acquisition module filters, removes artifacts, and segments the acquired raw electrocardiogram (ECG) and electroencephalogram (EEG) signals to obtain pure ECG and EEG signals with interference noise removed. The feature extraction unit of the multimodal physiological signal acquisition module extracts time-domain features, frequency-domain features, and nonlinear features from the pure electrocardiogram signal to form heart rate variability features; The feature extraction unit of the multimodal physiological signal acquisition module extracts multi-band power spectral density, frequency band ratio, brain inter-brain phase synchronization index and complexity features from pure EEG signals to form consciousness-related EEG features.

9. The closed-loop neural modulation method based on heart-brain interaction according to claim 7, characterized in that, The process of receiving heart rate variability features and consciousness-related EEG features through a cross-modal feature fusion assessment module, processing them through a Transformer-based multi-head self-attention convolutional fusion network, and outputting a quantified patient neurophysiological state score includes the following steps: The feature encoding network unit of the cross-modal feature fusion evaluation module inputs the heart rate variability features into a one-dimensional convolutional channel, and processes them sequentially through one-dimensional convolution, ReLU activation, max pooling and fully connected layers to obtain a fixed-length representation vector of the heart rate variability features. The feature encoding network unit of the cross-modal feature fusion evaluation module converts consciousness-related EEG features into pseudo-image form and inputs them into a two-dimensional convolutional channel. The features are then processed sequentially through two-dimensional convolution, ReLU activation, max pooling, and fully connected layers to obtain a fixed-length representation vector of the EEG features. The cross-modal attention fusion network unit of the cross-modal feature fusion evaluation module generates a query using EEG feature representation vectors and a key and value using heart rate variability feature representation vectors. After calculating the attention weights through ScaledDot-ProductAttention, the two types of features are dynamically aligned and weighted fused through multi-head attention extension to obtain context-enhanced fusion features. The classifier of the cross-modal feature fusion assessment module sequentially processes the fused features through average pooling, multilayer perceptron, and activation function to output a quantitative score of the patient's neurophysiological state.

10. The closed-loop neural modulation method based on heart-brain interaction according to claim 7, characterized in that, The step of using the Bayesian parameter optimization module with the state score as the optimization objective, and based on the Gaussian process regression algorithm of the ARDMatern5 / 2 kernel, to search for and determine the individualized optimal combination of stimulation parameters within the preset median nerve electrical stimulation parameter space includes the following steps: By using the parameter space definition unit of the Bayesian parameter optimization module, the stimulation frequency, current magnitude, and pulse width of the median nerve electrical stimulation are set as optimization parameters, and a discretized stimulation parameter space is constructed and output. The Gaussian process regression optimization unit of the Bayesian parameter optimization module uses the Gaussian process regression algorithm based on the ARDMatern5 / 2 kernel to fit the response surface between the stimulus parameter combination and the patient's neurophysiological state score. The optimal current parameter is determined by first analyzing the correspondence between the stimulus current and the state score through the Gaussian process regression optimization unit of the Bayesian parameter optimization module. Through the Gaussian process regression optimization unit of the Bayesian parameter optimization module, in the multi-parameter joint optimization space composed of stimulation frequency, stimulation pulse width and state score, the parameter combination corresponding to the lowest point of state score and the optimal current parameter are searched to form an individualized optimal stimulation parameter combination.

Images