Wearing nerve regulation and control method and system based on cloud AI control

Abstract

The invention belongs to the field of big data, and provides a wearable nerve regulation and control method and system based on cloud AI control, and the method comprises the steps: collecting user data through a wearable nerve regulation and control device worn on the limb part of a user, the user data being physiological data and / or motion data; uploading the user data to a cloud server, wherein the cloud server is provided with an artificial intelligence model; the artificial intelligence model receives the user data, and generates personalized nerve electrical stimulation parameters in combination with pre-stored user historical treatment data and / or medical history data; the cloud server issues the personalized nerve electrical stimulation parameters to the wearable nerve regulation and control device, and the wearable nerve regulation and control device executes transcutaneous nerve electrical stimulation operation based on the personalized nerve electrical stimulation parameters.

Description

Technical Field

[0001] This invention belongs to the field of big data, and specifically relates to a wearable neural modulation method and system based on cloud AI control. Background Technology

[0002] In recent years, neuromodulation technology has received widespread attention for the treatment of neurological disorders such as essential tremor, Parkinson's disease, dystonia, and neuropathic pain. With the development of wearable physiological sensors and neurostimulators, neuromodulation methods are gradually evolving from traditional invasive deep brain stimulation to non-invasive, personalized, and mobile approaches. Against this backdrop, utilizing physiological sensors to collect user status data and generating electrical stimulation control commands has become an important pathway to expand the application of this technology.

[0003] In existing technologies, most common neurostimulation devices employ local control, meaning they integrate fixed algorithm modules or preset control parameters, generating stimulation outputs directly on the terminal device based on the user's current state. The control algorithms are often simplified rule-based models, such as proportional-integral-derivative (PID) control, timed pulse output, fixed stimulation sequences, or manually set parameters, combined with simple threshold triggering strategies for stimulation decisions. This localized processing mode offers advantages such as fast response speed and less reliance on the network environment, but it has significant limitations in handling the high dynamism and individual differences in user states.

[0004] Some studies have attempted to introduce local AI modules to achieve adaptive control, such as running lightweight machine learning algorithms or neural networks on the device. However, limited by terminal computing power, storage resources, and energy consumption constraints, such models often cannot handle highly complex data fusion and state modeling tasks. Therefore, the parameter update frequency is low, the algorithm's generalization ability is weak, and it is difficult to achieve real-time, accurate, and personalized treatment strategy delivery.

[0005] Furthermore, existing methods generally lack closed-loop feedback mechanisms, and the adjustment of stimulus parameters often relies on subjective user feedback or manual intervention, failing to achieve real-time iterative updates based on actual physiological responses. In addition, physiological signals exhibit strong modal heterogeneity and time-varying characteristics, making it difficult for traditional single-modal inputs and static model structures to accurately capture the coupling evolution process between multimodal states, resulting in low efficiency and poor stability in the construction of individualized stimulus paths. Summary of the Invention

[0006] To address the problems in the prior art, this invention provides a wearable neural modulation method based on cloud-based AI control, comprising the following steps:

[0007] User data is collected by wearable neuromodulation devices worn on the user's limbs. The user data includes physiological data and / or motion data.

[0008] The user data is uploaded to a cloud server, where an artificial intelligence model is deployed.

[0009] The artificial intelligence model receives the user data and combines it with pre-stored user history treatment data and / or medical history data to generate personalized neurostimulation parameters.

[0010] The cloud server sends the personalized neural electrical stimulation parameters to the wearable neuromodulation device, and the wearable neuromodulation device performs transcutaneous neural electrical stimulation based on the personalized neural electrical stimulation parameters.

[0011] Furthermore, the artificial intelligence model is a cross-modal feedback graph network structure, the network structure including:

[0012] The multimodal temporal coding module is used to perform preliminary feature extraction and unified representation of different modal sensing data and historical data from wearable devices;

[0013] The event-driven node generation module is used to automatically generate new nodes in the graph when the system detects a significant physiological event.

[0014] The modal coupling edge construction module is responsible for constructing edges between nodes to express the mutual influence and synergy mechanism between physiological modalities; the feedback response regulation module is used to collect the user's physiological response after each round of neural electrical stimulation and convert it into a feedback tensor, which is then injected back into the graph structure for adjustment.

[0015] The graph reasoning and embedding update module is used to perform high-dimensional neural state modeling tasks on the current graph.

[0016] The closed-loop path evaluation and stimulus parameter generation module is the terminal output module of this network structure.

[0017] Furthermore, the feedback response control module includes:

[0018] Feedback collection unit, used to monitor target indicators;

[0019] The response evaluation function is used to combine multiple objective functions, set the direction of target improvement and penalty mechanism, and calculate the response score; the graph structure adjustment function is used to adjust the weight values ​​of nodes on the currently active path according to the score, suppress low response paths, strengthen the connection density of high response paths, or retopologically adjust the connection relationships of edges.

[0020] Furthermore, the graph reasoning and embedding update module specifically includes:

[0021] The state partitioning module partitions the graph structure into response flow clusters based on the temporal similarity between nodes and the distribution of dominant modal components.

[0022] The intrastream propagation submodule executes a time-gated graph convolution mechanism within each response stream cluster;

[0023] The inter-flow competition control submodule controls the flow between different response flow clusters through the response performance factor.

[0024] The feedback-guided dynamic edge reconstruction unit dynamically reconstructs the connection strength and directionality of edges in the graph structure based on the feedback response information after each round of reasoning, forming an instantaneous causal response graph.

[0025] The time-aware node embedding update mechanism updates all node embeddings through a long-short-term state fusion mechanism after graph convolution.

[0026] In another aspect, the present invention provides a wearable neuromodulation system based on cloud-based AI control, comprising the following modules:

[0027] Wearable neuromodulation devices are used to be worn on a user's limbs and collect user data, including physiological data and / or motion data.

[0028] The communication module is used to upload the user data to the cloud server;

[0029] A cloud server, in which an artificial intelligence model is deployed, is used to receive the user data and, in combination with pre-stored user historical treatment data and / or medical history data, generate personalized neuro-electrical stimulation parameters;

[0030] The sending module is used to send the personalized neuro-electrical stimulation parameters to the wearable neuromodulation device;

[0031] An execution module, located in the wearable neuromodulation device, is used to perform transcutaneous electrical nerve stimulation based on the personalized electrical nerve stimulation parameters.

[0032] Furthermore, the artificial intelligence model is a cross-modal feedback graph network structure, the network structure including:

[0033] The multimodal temporal coding module is used to extract features from different modal sensing data and historical data from the wearable neuromodulation device and generate a unified state representation.

[0034] The event-driven node generation module is used to construct corresponding state nodes when a sudden change in the user's physiological state is detected.

[0035] The modal coupling edge construction module is used to establish graph edge relationships that characterize the intermodal cooperation mechanism;

[0036] The feedback response control module is used to update the graph structure based on the user's physiological feedback data collected after stimulation.

[0037] The graph reasoning and embedding update module is used for node state modeling and graph structure information propagation.

[0038] The closed-loop path evaluation and stimulus parameter generation module is used to output personalized stimulus parameters based on the graph structure.

[0039] Furthermore, the feedback response control module includes:

[0040] Feedback collection unit is used to collect user feedback data in real time after electrical stimulation;

[0041] The response evaluation function unit is used to calculate the feedback response score based on a multi-objective function.

[0042] The graph structure adjustment unit is used to adjust the weight values ​​of nodes on the currently active path according to the score, suppress low-response paths, strengthen the connection density of high-response paths, or retopologically adjust the connection relationships of edges.

[0043] Furthermore, the graph reasoning and embedding update module includes:

[0044] The state partitioning module is used to divide the graph structure into multiple response stream clusters based on time series similarity and modality dominance labels;

[0045] The in-stream propagation submodule is used to perform time-gated graph convolution operations in each response stream cluster;

[0046] The inter-flow competition control submodule is used to adjust the propagation priority of each flow cluster based on the response performance factor;

[0047] Feedback-guided dynamic edge reconstruction unit is used to dynamically optimize edge connection strength and directionality by combining historical feedback;

[0048] A time-aware node embedding update mechanism is used to fuse historical states with the current state to generate the final node embedding representation.

[0049] Furthermore, the response efficacy factor is calculated from the score change of each response stream cluster within a predetermined feedback period, and is used to characterize the actual contribution of each cluster in the neural modulation process; when the response efficacy factor of a certain cluster decreases beyond a preset threshold, all its propagation paths will be suppressed.

[0050] The cross-modal feedback graph network structure proposed in this invention significantly overcomes the modeling bottleneck of traditional graph neural networks in neural modulation scenarios. This network structure uses multimodal temporal physiological data as its core input, physiological events as its driving mechanism, and feedback responses as its regulatory basis to dynamically construct a sustainably evolving graph structure. This enables high-precision modeling and regulatory path inference of the individual's neural state evolution process. By introducing a modal attention mechanism and a feedback closed-loop structure, the system can distinguish dominant modalities and identify high-value paths during state evolution, and reconstruct the graph after stimulus execution, thereby achieving real-time optimization of stimulus parameters.

[0051] The dynamic graph structure constructed in this invention no longer relies on static topology settings, but rather dynamically generates and evolves by combining event detection and feedback data. Its information propagation is no longer homogeneous diffusion, but rather weighted adjustment based on intermodal causal coupling and historical response value. Therefore, the system can accurately respond to individual differences and state fluctuations. Ultimately, this structure provides a closed-loop mechanism for personalized neuromodulation, encompassing the entire process of perception—mapping—reasoning—feedback—re-mapping, significantly improving the intelligence, adaptability, and therapeutic safety of neural electrical stimulation systems. It is particularly suitable for applications involving multimodal data fusion, time-varying neural state evolution, and real-time modulation tasks. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a flowchart of the method of the present invention;

[0054] Figure 2 This is a system structure diagram of the present invention;

[0055] Figure 3 This is a schematic diagram of a cross-modal feedback graph network structure. Detailed Implementation

[0056] The invention will now be described in preferred form with reference to the accompanying drawings and specific embodiments.

[0057] like Figure 1As shown, in one embodiment, this invention discloses a wearable neuromodulation method based on cloud AI control. This method constructs a remote collaborative closed-loop system with a cloud computing platform at its core and an intelligent wearable neurostimulation device as the execution terminal. It can achieve dynamic perception of the user's individual physiological state and optimization of neurostimulation parameters based on artificial intelligence algorithms, thereby executing personalized and precise percutaneous neurostimulation treatment operations. Specifically, it includes the following steps:

[0058] User data is collected by wearable neuromodulation devices worn on the user's limbs. The user data includes physiological data and / or motion data.

[0059] The user data is uploaded to a cloud server, where an artificial intelligence model is deployed.

[0060] The artificial intelligence model generates personalized neurostimulation parameters based on the user data and in combination with pre-stored user historical treatment data and / or medical history data.

[0061] The cloud server sends the personalized neural electrical stimulation parameters to the wearable neuromodulation device, and the wearable neuromodulation device performs transcutaneous neural electrical stimulation based on the personalized neural electrical stimulation parameters.

[0062] In implementation, cloud-based AI refers to a technology that uses artificial intelligence models deployed on remote cloud servers to analyze user data uploaded from terminal devices in real time or periodically, and generates control commands or feedback suggestions based on the model's inference results. Wearable neuromodulation refers to the precise regulation of nerve function by applying transcutaneous electrical stimulation to specific peripheral nerves (such as the radial, median, and ulnar nerves) in the user's wrist or forearm, using electrical signals to intervene in abnormal nerve discharges, reconstruct neural circuit connections, and activate sensory-motor feedback loops. This method is applicable to various neurological diseases. Firstly, it can be used to treat essential tremor by intervening in the thalamus-cortex-cerebellum tremor circuit to alleviate postural and action-related hand tremors. Secondly, it is suitable for patients with tremor-type Parkinson's disease, helping to alleviate resting and postural tremors and improve bradykinesia and rigidity symptoms. In addition, this method can also be used for patients with dystonia, reducing the incidence of spasms by regulating the activity of overexcited motor units; it is suitable for neuropathic pain, achieving pain relief by activating the endogenous analgesic system and blocking abnormal pain transmission; and it has a certain regulatory effect on autonomic dysfunction (such as anxiety-related heart rate abnormalities and vegetative nervous system disorders).

[0063] In this implementation, firstly, a wearable neuromodulation device worn on the user's limbs is used to monitor the user in real time and collect user data. This user data includes at least one physiological and / or motion data. The physiological data are parameters reflecting the user's autonomic nervous system state, muscle activity state, or nervous system functional state, including but not limited to: heart rate, heart rate variability, skin conductance response, skin temperature, electromyography (EMG) signals, local tremor amplitude, and neural reflex latency. The motion data are motion characteristic data acquired through inertial sensors, including but not limited to: triaxial acceleration, angular velocity, limb trajectory, posture angle changes, tremor frequency, movement stability, and motor coordination. The data is collected by an integrated sensing module in the device, which includes, but is not limited to: an inertial measurement unit (IMU), an electromyography (EMG) sensor, a capacitive electrophysiological sensor, a pressure sensor, and a temperature sensor.

[0064] Secondly, the collected user data is transmitted to a mobile terminal via the wireless communication module in the device, or directly uploaded to a cloud server via Wi-Fi, Bluetooth, or cellular network through the communication chip embedded in the device. The cloud server is equipped with an artificial intelligence model, which is constructed based on deep learning, graph neural networks, reinforcement learning, Bayesian inference, or a combination thereof, and is used to perform data analysis and stimulus parameter generation tasks.

[0065] Next, after receiving the user data, the artificial intelligence model further retrieves the user's pre-established user profile on the cloud server. This profile includes the user's historical treatment data and / or medical history data. The historical treatment data includes, but is not limited to: previously used neural electrical stimulation parameters (such as frequency, pulse width, amplitude), stimulation duration, stimulation site, treatment frequency, efficacy evaluation results, and user subjective feedback information. The medical history data includes, but is not limited to: diagnostic information for neurological diseases such as essential tremor, Parkinson's disease, and dystonia, including onset time, disease stage, neurological function score, previous interventions, medication records, and comorbidities. The artificial intelligence model constructs an individual feature vector based on the current user state information and historical data, and generates a personalized set of neural electrical stimulation parameters that best matches the current neurological state based on the trained model.

[0066] The personalized neural electrical stimulation parameters are a set of control parameters that constitute a complete stimulation instruction set. These parameters include, but are not limited to: stimulation frequency (in Hz), stimulation pulse width (in μs), pulse current amplitude (in mA), waveform morphology (such as monophasic square wave, biphasic symmetrical wave, triphasic multi-segment wave, etc.), stimulation mode (continuous, intermittent, variable frequency), stimulation path (the types of nerves involved and the stimulation sequence, such as radial nerve, median nerve, ulnar nerve, etc.), multi-channel on / off control strategy, parameter execution duration and dynamic change rules, etc.

[0067] Subsequently, the cloud server sends the personalized neuro-electrical stimulation parameters to the wearable neuromodulation device via the wireless communication module. After receiving the stimulation parameters, the device calls its internal control module to parse the parameters and transmit them to the electrical stimulation output unit. The electrical stimulation output unit, by setting the electrode arrangement to fit closely to the skin, performs percutaneous electrical stimulation on specific nerve areas of the user. The electrical stimulation operation uses a low-frequency pulse mode to act on superficial peripheral nerves. The target nerves stimulated include, but are not limited to, the median nerve, radial nerve, and ulnar nerve. The stimulation process control accuracy is not less than ±0.1mA, the waveform remains stable, and the electrode impedance is automatically corrected.

[0068] In an optional implementation, while performing the neural electrical stimulation operation, the device continues to collect the user's physiological response data and uploads the response data to the cloud server in real time. The artificial intelligence model performs neural state update assessment based on the feedback data and dynamically adjusts subsequent stimulation parameters according to the assessment results, thereby forming a personalized, closed-loop controlled remote neuromodulation system.

[0069] In another alternative implementation, the system sets a daily maximum safe stimulation threshold, which is determined by the doctor during the initial fitting and dynamically updated based on the user's self-evaluation, to ensure that the stimulation intensity is within the physiological tolerance range and to avoid discomfort or side effects caused by excessive stimulation.

[0070] In order to achieve highly accurate modeling of the user's neural state under the constraints of different modal inputs and multiple feedback information, in a further improved approach, the artificial intelligence model is a cross-modal feedback graph network.

[0071] While traditional Graph Neural Networks (GNNs) demonstrate strong capabilities in modeling graph-structured data, they suffer from several structural limitations when handling tasks involving multimodal physiological data fusion, time-varying neural state modeling, and real-time electrical stimulation feedback modulation. First, traditional GNNs are typically based on static or weakly dynamic graphs with predefined graph structures and nodes and edges determined during network initialization. This lacks the ability to build graphs in real-time to accommodate sudden changes in user physiological states or feedback responses, failing to accurately capture nonlinear abrupt changes and multi-factor driven characteristics of neural states. Second, traditional GNNs often assume homogeneous or equal-weighted information across nodes during information transmission, making it difficult to distinguish the primary and secondary coupling relationships between different modalities in neural response modulation and hindering the modeling of complex causal relationships between modalities. Third, traditional GNNs lack an intrinsic feedback mechanism, preventing them from adjusting their graph structure or node states based on the physiological response results after electrical stimulation. This prevents the neural modulation system from forming a truly closed-loop adaptive optimization path. Therefore, existing GNN methods cannot meet the demands for multidimensional collaborative control based on accurate perception of individual neural states, path selection, and parameter optimization.

[0072] To address the aforementioned issues, this embodiment constructs a cross-modal feedback graph network structure. Its core principle is to fuse and represent multi-source physiological data from users across time, modality, and feedback dimensions, and then dynamically construct a high-order graph structure with perception-driven, modal coupling, and feedback loop characteristics. This network no longer relies on predefined graph structures; instead, it uses each significant physiological state change (such as EMG mutations or increased tremor) as the graph-driving event, generating nodes and edges in real time based on the interaction patterns between multimodal data at the time of the event. Each node not only represents the physiological state at a given moment but also carries a state vector that integrates multiple modal codes, while edges represent the interaction paths and feedback response records between modalities.

[0073] The modality-aware mechanism employed in this network can identify the dominance of different physiological modalities in the neural modulation process and adjust the modality influence weights accordingly, thereby prioritizing the retention of paths that have the greatest impact on the neural state. Simultaneously, the network design incorporates a feedback response tensor. After each round of electrical stimulation, user feedback signals (such as heart rate changes and tremor relief amplitude) are collected, and the feedback results are mapped to the original graph structure. The graph morphology is dynamically adjusted through node weight correction, edge weight deformation, and path access control, enabling the network to reflect on and reconstruct after stimulation, ultimately achieving true closed-loop neural state management throughout the entire process of perception-modulation-feedback-reperception.

[0074] This cross-modal feedback graph network is no longer a continuation of the static information transmission of traditional GNNs, but fundamentally reshapes the graph structure into a physiological state-driven regulatory engine. It not only models the spatial distribution of the current state, but also continuously updates its internal structure and infers the optimal stimulation path in the future based on feedback. Therefore, it is suitable for scenarios with high response accuracy, high physiological adaptability, and long-term individual evolutionary regulation, significantly improving the intelligence and personalization of neural electrical stimulation.

[0075] like Figure 3 As shown, in one specific implementation, the cross-modal feedback graph network structure includes the following six main substructures: a multimodal temporal encoding module, an event-driven node generation module, a modal coupling edge construction module, a feedback response regulation module, a graph reasoning and embedding update module, and a closed-loop path evaluation and stimulus parameter generation module. These substructures each undertake different information processing and control responsibilities, collaboratively constructing a continuously self-updating graph structure regulation system, aiming to address the limitations of traditional graph neural networks in processing time-varying multimodal physiological data and electrical stimulation feedback problems.

[0076] The multimodal temporal coding module is used for preliminary feature extraction and unified representation of different modal sensing data and historical data from wearable devices. This module contains multiple modality-specific coding networks, each meticulously designed according to modal characteristics. For example, the substructure for electromyography (EMG) signals uses a one-dimensional convolutional network combined with a gated recurrent unit (GRU) to extract muscle discharge intensity and rhythm variations; the substructure for electrical conductance analysis (EDA) uses threshold segmentation and change-point detection algorithms to encode the bursts of sympathetic activity; the substructure for heart rate variability (HRV) uses wavelet transform to extract the low-frequency / high-frequency power ratio (LF / HF); and the acceleration data for the inertial measurement unit (IMU) uses a dual-branch LSTM to process angular velocity and linear acceleration separately to obtain posture changes and tremor features. The feature vectors output by each modal subnetwork are aligned and weighted through a modal attention layer, and finally fused into a unified multimodal state representation vector. This module ensures that the network can understand and compare state changes under different modalities at a uniform scale, providing high-quality state embeddings for subsequent graph structure construction.

[0077] The event-driven node generation module automatically generates new nodes in the graph when the system detects significant physiological events. Events include not only the application of external electrical stimulation but also key state transitions detected by the system, such as a sudden increase in EMG amplitude, a sharp decrease in HRV, or a tremor frequency exceeding a threshold. This module employs an event detector combined with a threshold judgment model. Once triggered, it calls the current fused state vector to construct a graph node, attaching a timestamp, event source, and modal mutation index as node attributes. For example, if a sudden increase in EMG accompanied by a decrease in HRV is detected within a certain time period, the module will generate a state node whose attributes will include event type: co-occurrence outbreak, outbreak modality: EMG+HRV, and signal level: intermediate risk. The design purpose of this module is to allow the graph structure to realistically reflect the dynamic evolution trajectory in the physiological state space and to dynamically expand the state space according to the suddenness of events, thus avoiding the limitations of static graph structures.

[0078] The modal coupling edge construction module is responsible for building edges between nodes to represent the mutual influence and synergistic mechanisms between physiological modalities. Edge construction is divided into two categories: modal structural edges and historical feedback edges. Modal structural edges are based on physiological models; for example, causal edges are often built between EDA and HRV marginal resonances, and action output path edges are established between EMG and acceleration. Historical feedback edges originate from response pathways formed by similar nodes in the past after executing stimuli. After edge construction, the system assigns attributes to each edge, such as modal pathway type, conduction delay, neural response efficiency, stimulus accessibility, and edge reliability. In subsequent graph reasoning, these edge attributes will be used to adjust information flow weights, allowing different physiological mechanism pathways to play different roles in the network.

[0079] The feedback response modulation module collects the user's physiological response after each round of neural electrical stimulation and converts it into a feedback tensor, which is then injected back into the graph structure for adjustment. This module consists of three parts: a feedback collection unit, a response evaluation function, and a graph structure modulation function. The feedback collection unit monitors target indicators, such as whether EMG amplitude decreases within 3 seconds after stimulation, whether tremor amplitude decreases, and whether EDA waveform stabilizes. The evaluation function uses a combination of multi-objective functions, such as setting a target improvement direction (tremor reduction) and a penalty mechanism (heart rate fluctuation), and calculates a response score. Finally, the graph structure modulation function adjusts the weight values ​​of nodes on the current activation path based on the score, suppressing low-response paths, strengthening the connection density of high-response paths, and even re-topologically reconstructing edge connections. This module constructs a closed-loop modulation feedback loop, enabling the network to not only generate stimulation parameters via feedforward but also achieve self-optimization in each feedback iteration.

[0080] The graph inference and embedding update module performs high-dimensional neural state modeling tasks on the current graph. Its core is a graph convolutional layer (such as a Graph Attention Network) combined with a state propagation path controller, which propagates state information in a time-dependent order while adjusting the propagation strength based on edge attributes. For example, for a path dominated by HRV, the inference process assigns higher weights to HRV embeddings and weakens the influence of EDA. The updated node carries predictive value about its position in the entire graph, including future response potential, activation efficiency, and load risk. This module ultimately generates a new state vector for each node in the graph, which is the core basis for subsequent selection of activation paths and parameter calculations.

[0081] The closed-loop path evaluation and stimulus parameter generation module is the terminal output module of this network structure. Its working mechanism involves searching for the optimal response path in the graph structure. The starting point of the path is the current state node, and the ending point is a preset target state (such as the disappearance of tremor). The path evaluation function integrates multiple dimensions such as path length, edge weight, response potential, and energy budget, and uses a heuristic search algorithm (such as A*) to select the optimal path. After selecting a path, the system synthesizes stimulus parameters based on the node state and historical edge responses, including stimulus frequency, voltage, pulse width, and action channel sequence, and packages and sends them to the wearable device for execution. The entire process achieves real-time personalized parameter reconstruction driven by state, supporting high-response and precise control under dynamic changes.

[0082] The six substructures described above form a highly coupled, collaborative system. The multimodal encoding module ensures the diversity and accuracy of data input, the event-driven module enables the graph structure to dynamically expand, the modal coupling edges introduce physiological mechanism constraints and empirical feedback paths, the feedback response module achieves closed-loop adaptive regulation, the graph reasoning module performs deep state understanding and potential estimation, and the path evaluation module completes high-dimensional reasoning and parameter refinement. The biggest difference between the overall network and traditional graph neural networks lies in the following: its graph structure is dynamically generated driven by physiological states, rather than a statically defined topology; its information propagation is controlled by both modality and feedback, rather than a unified diffusion; and its outcome orientation is not classification or regression, but path synthesis and physiological intervention control. Therefore, this network has significant advantages in intelligence, adaptability, and individual accuracy in handling real-time, multi-source, feedback-driven neural modulation tasks, which is difficult for traditional graph neural networks to achieve.

[0083] In one specific example, a patient with tremor-dominant Parkinson's disease, while receiving home-based wearable neurostimulation therapy, had the system collect the following multimodal physiological data from a device worn on their wrist: electromyography (EMG), triaxial acceleration (for tremor frequency extraction), electrical skin response (EDA), heart rate variability (HRV), and skin temperature (TEMP). This data was uploaded to the cloud in real time via Bluetooth.

[0084] EMG signals revealed a 6Hz tremor rhythm in the thumb muscles, with high-amplitude features extracted using a one-dimensional CNN. Accelerometer data, after Fourier transform, showed a continuous tremor in the wrist between 5.8Hz and 6.2Hz. EDA signals showed elevated skin conductance, indicating sympathetic activation. HRV calculations showed an LF / HF ratio of 3.2, indicating abnormal autonomic nervous system function with sympathetic activation. Temperature showed no significant fluctuations. These modal data, extracted from their respective subnetworks, were then fused using modal attention to form a unified neural state vector for the current moment.

[0085] The system detected that the EMG amplitude exceeded the baseline by 50% for 10 consecutive seconds, and the tremor frequency remained above 6Hz, while the EDA fluctuated significantly. Therefore, a tremor enhancement event was triggered, and a new state node A was created in the graph structure. The node attributes include tremor frequency = 6Hz, EMG amplitude = high, HRV = sympathetic activation, and EDA = medium-high. Node A is marked as the current major physiological abnormality.

[0086] The system identifies multiple node paths from historical treatment maps that previously yielded significant effects after electrical stimulation at similar state nodes, and constructs modal coupling edges connecting these paths to the current node A, such as A→B and A→C. Path A→B represents a state of low HRV + high tremor where low-frequency stimulation of the right ulnar nerve leads to tremor relief; path A→C represents a moderate relief effect achieved through median nerve pulse change pathway modulation. Based on these historical edges, the system constructs edge weight parameters, including average response rate, channel coupling reliability, and feedback score.

[0087] The system generates a stimulation protocol for path A→B with the following parameters: frequency 20Hz, pulse width 200μs, current amplitude 1.2mA, targeting the right ulnar nerve of the wrist, and duration 20 seconds. After stimulation, the system re-acquires EMG, acceleration, and HRV. The EMG signal decreased to 50% of its original value within 10 seconds, the tremor frequency decreased to 5Hz, HRV improved, and EDA returned to stability. The feedback response tensor assessment result indicates a high-response path with a feedback score of 87 / 100. The system increases the edge weights of the A→B path in the graph structure and prioritizes storing this path for subsequent recommendations.

[0088] Based on the updated graph structure, the system identifies node B as the current high-value state, increases its weight, and predicts the state path the user might enter within the next hour. The system updates the inference path A→B→D, where D is a node with no symptoms of tremor. It identifies stability plus decreased sympathetic awareness as the optimal transition state and recommends node B as the next target state.

[0089] The system re-evaluates the stimulation path based on path A→B→D. After comparing the current HRV, it adjusts the next stimulation to an intermittent mode, reduces the current amplitude to 1.0mA, maintains the stimulation pulse width, and fine-tunes the frequency to 22Hz. Simultaneously, skin cooling pretreatment is added to improve comfort. The stimulation command is sent to the device for execution, completing a full closed loop of individual state-driven—map inference—feedback evaluation—path adjustment—parameter issuance.

[0090] In a further implementation, the structure of the graph reasoning and embedding update module is further transformed into an enhanced graph state modeling structure, which aims to address the performance bottlenecks of traditional graph neural networks in handling non-stationary time-varying states, heterogeneous propagation between modalities, and feedback loop failure.

[0091] In existing graph neural inference, graph structures are often simplified to static node-edge relationships in a convolutional propagation form, which is insufficient to handle the dynamic changes in feedback flow during real-world user neural modulation, such as abrupt physiological changes, rapid decay of short-term feedback responses after electrical stimulation, and complex temporal scenarios where modal pathway weights alternately dominate. Neural modulation is a task extremely sensitive to the temporal correlation of the response process; therefore, it is necessary to introduce temporal-dimensional reinforcement modeling mechanisms and streaming pathway scheduling mechanisms into the graph inference process.

[0092] The principle of the improved module proposed in this embodiment is as follows: taking the graph node state as the center, combining historical feedback, modal signal timing patterns and path activity indicators, dynamically dividing the graph into response flow clusters with consistent physiological response tendencies, and on this basis, executing intra-flow enhanced propagation mechanism and inter-flow competitive selection mechanism, ultimately obtaining a node embedding representation that is more time-sensitive and has consistent feedback loop.

[0093] In one specific implementation, the enhanced graph state modeling structure includes the following sub-modules:

[0094] The state partitioning module divides the graph structure into response flow clusters based on the temporal series similarity between nodes and the distribution of dominant modal components. Dynamic Time Warped Distance (DTW) is used to measure the nonlinear similarity between node temporal feature vectors. Combined with dominant modal labels (e.g., EMG-dominant, HRV-dominant), a spectral clustering algorithm is employed to generate multiple dynamic response clusters. Each cluster represents a synchronous evolutionary flow of a physiological pathway, ensuring that the propagation of graph convolutions within the flow follows temporal consistency.

[0095] Within each response stream cluster, the intra-stream propagation submodule executes a time-gated graph convolution mechanism. This mechanism adjusts the influence intensity of adjacent nodes using a gating function based on the time-sensitive weights of each node. For example, if a node's historical feedback time curve shows a high-frequency burst response, it will exert a stronger positive stimulus on surrounding nodes during propagation. The convolution kernel parameters are updated in real time through a feedback tuning mechanism, enabling the propagation process to more accurately reflect physiological dynamics.

[0096] The inter-flow competition regulation submodule regulates the flow clusters through a response performance factor. This factor is calculated based on the score improvement of each cluster in recent feedback rounds, representing the contribution of that pathway to actual regulation. The system introduces a suppression mechanism: when the response performance of a cluster decreases beyond a threshold, all its propagation paths are suppressed, thereby preventing ineffective pathways from consuming graph computational resources or forming erroneous regulation paths.

[0097] The feedback-guided dynamic edge reconstruction unit dynamically reconstructs the connection strength and directionality of edges in the graph structure based on feedback response information after each round of inference, forming an instantaneous causal response graph. Using a variational edge updater (VEBU), the conductance parameters of each edge are Bayesian-optimized based on the actual physiological score of the path containing that edge, reflecting the latest neuromodulation results. This ensures that the inference structure maintains an optimal graph configuration in each round.

[0098] After graph convolution, the time-aware node embedding update mechanism updates all node embeddings through a long-short-term state fusion mechanism. Each node retains its historical state vector sequence, analyzes the state evolution trend through a bidirectional gated recurrent unit (Bi-GRU), and performs weighted fusion with the current graph convolution output to form a final embedding representation that combines short-term response and long-term trend awareness.

[0099] This structure not only retains the ability of graph structures to model complex neural state spaces, but also introduces a flow partitioning and propagation mechanism based on feedback and temporal logic, thereby improving the sensitivity and response speed of the control strategy to changing states and phased feedback. Especially in scenarios with drastic fluctuations in tremor frequency, delayed neural reflexes, or alternating dominance of multimodal signals, it effectively avoids the instability or distortion problems of traditional graph inference results, achieving a leap from graph structure modeling to graph state flow control.

[0100] In one specific example, the patient wore a smart neuromodulation device incorporating EMG, HRV, EDA, and IMU sensors. After the system initiated real-time data acquisition, it detected increased hand tremors (tremor frequency increased to 6.5 Hz, wrist angular acceleration fluctuated dramatically, and EMG amplitude significantly increased). At this time, the device simultaneously recorded the following data:

[0101] EMG data: Increased sudden electromyographic discharge activity;

[0102] HRV: Low-frequency power is reduced, and the LF / HF value is high, reflecting increased sympathetic nerve activity;

[0103] IMU: Three-axis accelerometer displays periodic jitter waveforms;

[0104] EDA: Enhanced skin conductance indicates an increased stress state.

[0105] The system constructs a time-series state vector for each time window (2-second interval) sampled over the past 3 minutes. Using the Dynamic Time Warping (DTW) algorithm, the nonlinear time delay of state changes between nodes is registered, assigning multiple high-tremor state nodes to EMG-dominant cluster A; nodes with highly coordinated changes in EDA and HRV are assigned to autonomic neural cluster B; and the remaining stable states are assigned to low-response cluster C.

[0106] The node convolution employs a temporal gating mechanism. If a node experiences a sudden increase in its tremor frequency from 5Hz to 6.5Hz within the past 5 seconds, the system assigns a high propagation weight to this node using a gating function, guiding it to spread the intensified tremor state to neighboring nodes. This propagation behavior prompts the system to form a high-intensity anomaly detection path within the cluster.

[0107] The system analyzed the recent increase in response scores of each cluster after stimulation. The results showed that the EMG-dominant cluster A achieved a tremor relief rate of 20% in the past two rounds of feedback, while the EDA-HRV co-operation cluster B only achieved 5%. Therefore, the system increased the priority of cluster A, giving it more inference resources, while the side propagation probability of cluster B was suppressed, avoiding over-reliance on low feedback paths.

[0108] After electrical stimulation, the user reported a reduction in tremor. The feedback tensor showed that the current channel combination (median nerve + radial nerve dual-channel pulses) produced a significant tremor-suppressing effect. The system invoked the VEBU module to redistribute path edge weights based on the feedback: strengthening edges leading to high-response nodes and weakening the conductivity of previously ineffective edges. The directionality of some edges was also reversed due to the feedback effect, used to describe the delayed positive modulation trend.

[0109] After graph convolution, the node state vectors are fed into the Bi-GRU module. Historical tremor frequency curves, EMG patterns, and stimulus response results are used together for short-term and long-term state trend analysis. The final node representation reflects a decreasing trend in tremor, but short-term fluctuations remain severe. Based on this, the system generates new stimulus parameters: slightly reducing the pulse frequency, extending the duration of a single stimulus, and maintaining the dual-channel output strategy.

[0110] Compared to traditional graph neural networks, this structure truly transforms graphs from static representations into dynamic control controllers, possessing key breakthrough value in precise neural modulation applications.

[0111] like Figure 2 As shown, in another embodiment, the present invention also provides a wearable neuromodulation system based on cloud AI control, comprising:

[0112] Wearable neuromodulation devices are used to be worn on a user's limbs and collect user data, including physiological data and / or motion data.

[0113] The communication module is used to upload the user data to the cloud server;

[0114] A cloud server, in which an artificial intelligence model is deployed, is used to receive the user data and, in combination with pre-stored user historical treatment data and / or medical history data, generate personalized neuro-electrical stimulation parameters;

[0115] The sending module is used to send the personalized neuro-electrical stimulation parameters to the wearable neuromodulation device;

[0116] An execution module, located in the wearable neuromodulation device, is used to perform transcutaneous electrical nerve stimulation based on the personalized electrical nerve stimulation parameters.

[0117] It should be noted that the explanations and descriptions of the aforementioned wearable neuromodulation method based on cloud AI control also apply to the devices in the embodiments of this application, and will not be repeated here.

[0118] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0119] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0120] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0121] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. For some module structures not specifically defined in this invention, the content described in the prior art shall prevail. The prior art mentioned in the foregoing background and specific embodiments sections can be considered as part of this invention and used to understand the meaning of some technical features or parameters.

Claims

1. A wearable neural modulation method based on cloud-based AI control, characterized in that, The method includes the following steps: User data is collected by wearable neuromodulation devices worn on the user's limbs. The user data includes physiological data and / or motion data. The user data is uploaded to a cloud server, where an artificial intelligence model is deployed. The artificial intelligence model receives the user data and combines it with pre-stored user history treatment data and / or medical history data to generate personalized neurostimulation parameters. The cloud server sends the personalized neural electrical stimulation parameters to the wearable neuromodulation device, and the wearable neuromodulation device performs transcutaneous neural electrical stimulation based on the personalized neural electrical stimulation parameters.

2. The cloud-based AI-controlled wearable neural modulation method according to claim 1, characterized in that, The artificial intelligence model is a cross-modal feedback graph network structure, which includes: The multimodal temporal coding module is used to perform preliminary feature extraction and unified representation of different modal sensing data and historical data from wearable devices; The event-driven node generation module is used to automatically generate new nodes in the graph when the system detects a significant physiological event. The modal coupling edge construction module is responsible for constructing edges between nodes to express the mutual influence and synergy mechanism between physiological modalities; the feedback response regulation module is used to collect the user's physiological response after each round of neural electrical stimulation and convert it into a feedback tensor, which is then injected back into the graph structure for adjustment. The graph reasoning and embedding update module is used to perform high-dimensional neural state modeling tasks on the current graph. The closed-loop path evaluation and stimulus parameter generation module is the terminal output module of this network structure.

3. The wearable neural modulation method based on cloud-based AI control according to claim 2, characterized in that, The feedback response control module includes: Feedback collection unit, used to monitor target indicators; The response evaluation function is used to combine multiple objective functions, set the direction of target improvement and penalty mechanism, and calculate the response score; the graph structure adjustment function is used to adjust the weight values ​​of nodes on the currently active path according to the score, suppress low response paths, strengthen the connection density of high response paths, or retopologically adjust the connection relationships of edges.

4. The wearable neural modulation method based on cloud AI control according to claim 2, characterized in that, The graph reasoning and embedding update module specifically includes: The state partitioning module partitions the graph structure into response flow clusters based on the temporal similarity between nodes and the distribution of dominant modal components. The intrastream propagation submodule executes a time-gated graph convolution mechanism within each response stream cluster; The inter-flow competition control submodule controls the flow between different response flow clusters through the response performance factor. The feedback-guided dynamic edge reconstruction unit dynamically reconstructs the connection strength and directionality of edges in the graph structure based on the feedback response information after each round of reasoning, forming an instantaneous causal response graph. The time-aware node embedding update mechanism updates all node embeddings through a long-short-term state fusion mechanism after graph convolution is completed.

5. The wearable neural modulation method based on cloud AI control according to claim 4, characterized in that, The response efficacy factor is calculated from the score improvement of each cluster in the most recent rounds of feedback, representing the contribution of the pathway in actual regulation; when the response efficacy factor of a cluster decreases beyond a preset threshold, all its propagation paths will be suppressed.

6. A wearable neural modulation system based on cloud-based AI control, characterized in that, The system includes the following modules: Wearable neuromodulation devices are used to be worn on a user's limbs and collect user data, including physiological data and / or motion data. The communication module is used to upload the user data to the cloud server; A cloud server, in which an artificial intelligence model is deployed, is used to receive the user data and, in combination with pre-stored user historical treatment data and / or medical history data, generate personalized neuro-electrical stimulation parameters; The sending module is used to send the personalized neuro-electrical stimulation parameters to the wearable neuromodulation device; An execution module, located in the wearable neuromodulation device, is used to perform transcutaneous electrical nerve stimulation based on the personalized electrical nerve stimulation parameters.

7. The wearable neuromodulation system based on cloud-based AI control according to claim 6, characterized in that, The artificial intelligence model is a cross-modal feedback graph network structure, which includes: The multimodal temporal coding module is used to extract features from different modal sensing data and historical data from the wearable neuromodulation device and generate a unified state representation. The event-driven node generation module is used to construct corresponding state nodes when a sudden change in the user's physiological state is detected. The modal coupling edge construction module is used to establish graph edge relationships that characterize the intermodal cooperation mechanism; The feedback response control module is used to update the graph structure based on the user's physiological feedback data collected after stimulation. The graph reasoning and embedding update module is used for node state modeling and graph structure information propagation. The closed-loop path evaluation and stimulus parameter generation module is used to output personalized stimulus parameters based on the graph structure.

8. The wearable neuromodulation system based on cloud AI control according to claim 7, characterized in that, The feedback response control module includes: Feedback collection unit is used to collect user feedback data in real time after electrical stimulation; The response evaluation function unit is used to calculate the feedback response score based on a multi-objective function. The graph structure adjustment unit is used to adjust the weight values ​​of nodes on the currently active path according to the score, suppress low-response paths, strengthen the connection density of high-response paths, or retopologically adjust the connection relationships of edges.

9. The wearable neuromodulation system based on cloud-based AI control according to claim 7, characterized in that, The graph reasoning and embedding update module includes: The state partitioning module is used to divide the graph structure into multiple response stream clusters based on time series similarity and modality dominance labels; The in-stream propagation submodule is used to perform time-gated graph convolution operations in each response stream cluster; The inter-flow competition control submodule is used to adjust the propagation priority of each flow cluster based on the response performance factor; Feedback-guided dynamic edge reconstruction unit is used to dynamically optimize edge connection strength and directionality by combining historical feedback; A time-aware node embedding update mechanism is used to fuse historical states with the current state to generate the final node embedding representation.

10. The wearable neuromodulation system based on cloud-based AI control according to claim 9, characterized in that, The response efficacy factor is calculated from the score change of each response stream cluster within a predetermined feedback period, and is used to characterize the actual contribution of each cluster in the neural modulation process; when the response efficacy factor of a certain cluster decreases beyond a preset threshold, all its propagation paths will be suppressed.

Images