Artificial intelligence spinal nerve regulation and control system and method based on multi-modal physiology-nerve-experience closed loop

Abstract

The invention provides an artificial intelligence spinal nerve regulation and control system and method based on a multi-modal physiology-nerve-experience closed loop, and relates to the technical field of medical instruments, the system comprises an implantable stimulation and sensing subsystem, a body surface and wearable sensing subsystem, a communication and time synchronization subsystem and a control processing unit, the implantable stimulation and sensing subsystem, the body surface / wearable sensing subsystem and the control processing unit are in communication connection in pairs through the communication and time synchronization subsystem. According to the method, the fundamental limitation that in the prior art, only dosage is maintained, and brain and subjective experience is ignored is broken through, and more stable, more efficient and more personalized accurate pain management is realized through multi-modal fusion and hierarchical cooperative control.

Description

Technical Field

[0001] This invention relates to the field of medical device technology, and in particular to an artificial intelligence spinal cord nerve modulation system and method based on a multimodal physiological-neurological-experience closed loop. Background Technology

[0002] Spinal cord stimulation (SCS) is a well-established therapy for treating chronic intractable pain. Traditional open-loop SCS systems use preset, fixed parameters for stimulation. Their core drawback is that they cannot adapt to changes in the relative distance between the electrodes and the spinal cord during patients' daily activities (such as bending over or coughing), leading to unstable stimulation doses. This can result in discomfort from overstimulation or decreased efficacy from understimulation.

[0003] To address this problem, closed-loop SCS technology emerged. Currently, the most advanced closed-loop SCS technology is exemplified by Saluda Medical. Systems and Medtronic's Inceptiv TM The system, represented by this approach, uses evoked compound action potentials (ECAPs) as its core biomarker. ECAPs are the direct electrophysiological response of the spinal cord to electrical stimulation. By monitoring ECAPs in real time and adjusting the stimulation intensity, the system can maintain a constant level of neural activation (i.e., "dose"), thus solving the problem of stimulation consistency. This technical approach has become a widely adopted technique.

[0004] However, existing ECAP-based closed-loop systems have the following fundamental limitations:

[0005] 1. Steady-state maintenance rather than intelligent optimization: The system is essentially an automatic adjustment mechanism whose function is to maintain the target ECAP value manually set by clinicians, but it does not have the ability to actively explore and discover better combinations of treatment parameters.

[0006] 2. Limited physiological information: The system relies solely on the ECAP signal and cannot perceive other physiological states closely related to pain, such as muscle tension, autonomic nervous system arousal level, and the patient's overall activity status, thus lacking contextual awareness.

[0007] 3. Ignoring the brain and subjective experience: The ultimate sensation of pain originates in the brain. Current SCS technology operates entirely at the spinal cord level, failing to perceive brain activity related to pain emotion and cognitive components, and further failing to incorporate the patient's true subjective experience into the regulatory loop, resulting in a disconnect between the "spinal cord-brain-mind".

[0008] Therefore, it is essential to design an artificial intelligence spinal cord nerve modulation system and method based on a multimodal physiological-neurological-experience closed loop. SUMMARY

[0009] In order to overcome the deficiencies of the prior art, the purpose of the present application is to provide a multi-modal physiological-neural-experience closed-loop artificial intelligence spinal cord nerve regulation system and method.

[0010] To achieve the above-mentioned purpose, the present application provides the following scheme:

[0011] The present application provides a multi-modal physiological-neural-experience closed-loop artificial intelligence spinal cord nerve regulation system, comprising:

[0012] An implantable stimulation and sensing subsystem, a body surface and wearable sensing subsystem, a communication and time synchronization subsystem, a control processing unit, the implantable stimulation and sensing subsystem, the body surface / wearable sensing subsystem and the control processing unit are communicatively connected to each other through the communication and time synchronization subsystem;

[0013] The implantable stimulation and sensing subsystem is used to output stimulation pulses to the target nerve tissue according to the input stimulation control parameters, and collect evoked compound action potentials within the set sensing time window after stimulation;

[0014] The body surface and wearable sensing subsystem is used to acquire electroencephalogram, electromyogram, electrocardiogram and multiple physiological and behavioral signals;

[0015] The communication and time synchronization subsystem is used for bidirectional communication between the implantable stimulation and sensing subsystem, the body surface and wearable sensing subsystem and the control processing unit, and realizes time alignment of data flow;

[0016] The control processing unit comprises a processor and a memory, the memory stores instructions executable on the processor, and the processor is used to execute the instructions.

[0017] Preferably, the implantable stimulation and sensing subsystem comprises an implantable pulse generator and an epidural electrode electrically connected thereto, the epidural electrode is implanted in the epidural space of the patient's spinal cord, and the implantable pulse generator cooperates with the epidural electrode to output stimulation pulses to the target nerve tissue according to the input stimulation control parameters, and collect ECAP within the set sensing time window after stimulation.

[0018] Preferably, the body surface and wearable sensing subsystem comprises a motion sensor, an electromyography sensor, a galvanic skin sensor, a heart rate variability sensor and an electroencephalogram sensor;

[0019] The motion sensor is used to monitor the posture and activity level of the patient;

[0020] The electromyography sensor is used to monitor the tension of the target muscle group related to pain;

[0021] The skin conductance sensor and heart rate variability sensor are used to monitor autonomic nervous system activity related to pain and stress.

[0022] The EEG sensor is used to monitor specific frequency bands of brain electrical activity in the cerebral cortex associated with pain, emotion, and cognitive state.

[0023] Preferably, the control processing unit is equipped with a three-layer closed-loop control and a finite state machine arbitration module, wherein:

[0024] The first closed-loop control loop is a millisecond-level ECAP tracking loop, used to modulate the stimulus amplitude at the single-pulse level to track the target ECAP value. Target The first closed-loop control loop adopts proportional-integral control and has an integral anti-saturation mechanism. The update delay of the control loop is no more than 5ms, and safety constraints on charge density, pulse width, frequency and maximum amplitude are applied.

[0025] The second closed-loop control loop is a model predictive control loop based on EEG supervision. It is used to predict state variables estimated from EEG biomarkers, utilizing an individualized state-space model and cost function, within the treatment window and ECAP. Target Calculate ΔECAP under the maximum rate of change constraint. Target And send it to the first layer closed-loop control loop;

[0026] The third-layer closed-loop control loop is a Bayesian optimization loop based on experience and energy consumption, used to maximize the health score within the safety parameter domain. The health score integrates at least patient-reported outcomes, energy consumption, and sensory abnormality penalty terms, and outputs the optimization results used to adjust the cost function weights and / or parameter vector θ of the second closed-loop control loop.

[0027] The finite state machine arbitration module is used to arbitrate the activation, freezing, and rollback of the first-layer closed-loop control loop, the second-layer closed-loop control loop, and the third-layer closed-loop control loop based on motion sensor data, EEG sensor data, sleep and rest states, and safety events.

[0028] This invention also provides an artificial intelligence spinal cord nerve modulation method based on a multimodal physiological-neurological-experience closed loop, applied to the aforementioned artificial intelligence spinal cord nerve modulation system based on a multimodal physiological-neurological-experience closed loop, comprising:

[0029] Step 1: Apply electrical stimulation pulses to the patient via epidural electrodes and sense the ECAP signal;

[0030] Step 2: Execute the first-level closed-loop control loop, adjusting the parameters of subsequent electrical stimulation pulses in real time based on the ECAP signal to maintain a preset target ECAP value.Target ;

[0031] Step 3: Perform the second level closed-loop control loop to calculate the value of ΔECAP Target , based on which to dynamically adjust ECAP Target ;

[0032] Step 4: Perform the third level closed-loop control loop to receive patient reported outcomes and based on which to generate optimized stimulation parameters, based on which to update the first level closed-loop control loop and the second level closed-loop control loop;

[0033] Step 5: Arbitrate by the finite state machine arbitration module based on motion sensor data, electroencephalography sensor data, sleep and resting state, and safety events to activate, freeze, and fallback the first level closed-loop control loop, the second level closed-loop control loop, and the third level closed-loop control loop.

[0034] Preferably, in step 2, the first level closed-loop control loop is performed to adjust the parameters of the subsequent electrical stimulation pulses in real-time based on the ECAP signal to maintain a preset target value of ECAP ECAP Target , specifically:

[0035] The first level closed-loop control loop is performed to track the target value of ECAP ECAP Target to maintain a preset target value of ECAP ECAP Target , the first level closed-loop control loop has an integral anti-windup mechanism, and the update delay of the control loop is no more than 5 ms, and safety constraints of charge density, pulse width, frequency, and maximum amplitude are imposed.

[0036] Preferably, in step 3, the second level closed-loop control loop is performed to calculate the value of ΔECAP Target , based on which to dynamically adjust ECAP Target , specifically:

[0037] The second level closed-loop control loop is performed to adopt a model predictive control loop based on electroencephalography supervision, to calculate the value of ΔECAP Target based on state variables estimated by electroencephalography biomarkers, using an individualized state space model and a cost function, under the constraints of a therapeutic window and a maximum rate of change of ECAP Target , based on which to dynamically adjust ECAP Target . Target .

[0038] Preferably, in step 4, the third level closed-loop control loop is performed to receive patient reported outcomes and based on which to generate optimized stimulation parameters, based on which to update the first level closed-loop control loop and the second level closed-loop control loop, specifically:

[0039] performing a third level closed loop control loop, adopting a Bayesian optimization loop based on experience and energy consumption to maximize a health score in a safety parameter domain, wherein the health score at least integrates patient reported outcomes, energy consumption and a paresthesia penalty term, generating optimized stimulation parameters, updating the first level closed loop control loop and the second level closed loop control loop based on the optimized stimulation parameters.

[0040] According to the specific embodiments provided by the application, the following technical effects are disclosed:

[0041] The application provides an artificial intelligence spinal cord nerve modulation system and method based on a multi-modal physiological-neural-experience closed loop, which comprises an implantable stimulation and sensing subsystem, a body surface and wearable sensing subsystem, a communication and time synchronization subsystem, and a control processing unit, wherein the implantable stimulation and sensing subsystem, the body surface / wearable sensing subsystem and the control processing unit are communicatively connected in pairs through the communication and time synchronization subsystem, and the method comprises applying an electrical stimulation pulse to a patient through an epidural electrode and sensing an ECAP signal, performing a first level closed loop control loop, adjusting the parameters of a subsequent electrical stimulation pulse in real time based on the ECAP signal to maintain a preset target value ECAP Target of the ECAP, performing a second level closed loop control loop, calculating a value of ΔECAP Target , and dynamically adjusting the ECAP based on the value of ΔECAP Target , performing a third level closed loop control loop, receiving patient reported outcomes, and generating optimized stimulation parameters based on the patient reported outcomes, updating the first level closed loop control loop and the second level closed loop control loop based on the optimized stimulation parameters, and arbitrating the first level closed loop control loop, the second level closed loop control loop and the third level closed loop control loop based on motion sensor data, electroencephalogram sensor data, sleep and resting state and safety events by a finite state machine arbitration module. The application has the following advantages:

[0042] 1. The stimulation stability and efficiency are synergistically improved: in the scenario of simulating daily posture changes (such as from a sitting posture to a standing posture), the fluctuation (measured by standard deviation) of the ECAP amplitude output by the system of the application is reduced by more than 35%, which indicates that the application can provide more stable nerve modulation "dose" than the prior art, and at the same time, this improvement in stability is achieved with a reduction of at least 20% in total energy delivery (TED), in the prior art, pursuing higher stability usually means more frequent adjustment and higher energy consumption, while the application can improve both of the two mutually contradictory technical indicators through hierarchical prediction and optimization, thereby prolonging the service life of the implanted pulse generator (IPG) battery;

[0043] 2. Convert the improvement of objective engineering indicators into quantifiable patient experience improvement: The improvement of the above-mentioned objective engineering indicators can be converted into the improvement of the subjective experience of the patient. The simulation results show that the daily fluctuation of the daily NRS pain score of the user of the system is reduced by 35%, which indicates that the high stability of the system can bring the patient a more predictable and reliable pain relief experience, and reduce the discomfort caused by the fluctuation of the stimulation. This conversion effect from the stability of the objective physiological signal to the stability of the subjective feeling cannot be directly predicted from simply adding ECAP, EEG and AI technology, which embodies the synergistic effect of the architecture of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0045] Figure 1 The system architecture schematic diagram provided for the embodiments of the present application;

[0046] Figure 2 The implantable pulse generator (IPG) and electrode connection block diagram provided for the embodiments of the present application;

[0047] Figure 3 The signal chain and artifact suppression flowchart provided for the embodiments of the present application;

[0048] Figure 4 The three-level closed-loop control block diagram provided for the embodiments of the present application;

[0049] Figure 5 The model predictive control (MPC) constraint and cost function schematic diagram provided for the embodiments of the present application;

[0050] Figure 6 The system finite state machine (FSM) state and transition condition schematic diagram provided for the embodiments of the present application;

[0051] Figure 7 The multi-data stream time alignment and timestamp schematic diagram provided for the embodiments of the present application;

[0052] Figure 8 The health score composition and Bayesian optimization flowchart provided for the embodiments of the present application;

[0053] Figure 9 The ECAP measurement window and peak determination schematic diagram provided for the embodiments of the present application;

[0054] Figure 10The security coverage and watchdog rollback flowchart provided for the embodiments of the present application. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0056] The purpose of the present application is to provide an artificial intelligence spinal cord nerve regulation system and method based on a multi-modal physiological-neural-experience closed loop. The present application initiates a three-level architecture of "ECAP steady state-EEG supervision-AI experience optimization", and realizes intelligent arbitration through a finite state machine. First, the millisecond-level ECAP-PI loop ensures the basic stability of the stimulation dose; second, the EEG-based MPC controller converts the brain's pain emotional state into dynamic adjustment instructions for the ECAP target, realizing feedback across from the spinal cord to the brain; finally, the Bayesian optimization engine that fuses multi-modal physiological data and patient subjective reports aims to achieve long-term health scores, and autonomously explores personalized optimal treatment parameters. The finite state machine dynamically schedules the working modes of each level according to the situation (such as movement, sleep, signal quality), ensuring the robustness and safety of the system. The present application breaks through the fundamental limitation of the prior art that only maintains the dose and ignores the brain and subjective experience, and through multi-modal fusion and hierarchical collaborative control, realizes more stable, more efficient and more personalized precision pain management.

[0057] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0058] The present application provides an artificial intelligence spinal cord nerve regulation system based on a multi-modal physiological-neural-experience closed loop, comprising: an implantable stimulation and sensing subsystem, a body surface and wearable sensing subsystem, a communication and time synchronization subsystem, a control processing unit, the implantable stimulation and sensing subsystem, the body surface / wearable sensing subsystem and the control processing unit are communicatively connected to each other through the communication and time synchronization subsystem;

[0059] The implantable stimulation and sensing subsystem is used to output a stimulation pulse to the target nerve tissue according to the input stimulation control parameter, and collect the evoked compound action potential within the set sensing time window after stimulation;

[0060] The body surface and wearable sensing subsystem is used to acquire electroencephalogram, electromyogram, electrocardiogram and multiple physiological and behavioral signals;

[0061] The communication and time synchronization subsystem is used for bidirectional communication between the implantable stimulation and sensing subsystem, the body surface and wearable sensing subsystem and the control processing unit, and realizes time alignment of data flow.

[0062] The control processing unit comprises a processor and a memory, the memory stores instructions executable on the processor, and the processor is used to execute the instructions.

[0063] The implantable stimulation and sensing subsystem comprises an implantable pulse generator (IPG) and an epidural electrode electrically connected thereto, the epidural electrode is implanted in the epidural space of the patient's spinal cord, and the implantable pulse generator cooperates with the epidural electrode to output a stimulation pulse to the target nerve tissue according to the input stimulation control parameter, and collect ECAP in the set sensing time window after stimulation.

[0064] The body surface and wearable sensing subsystem comprises a motion sensor, an electromyography (EMG) sensor, a skin conductance (SC) sensor, a heart rate variability (HRV) sensor and an electroencephalogram (EEG) sensor.

[0065] The motion sensor (such as a three-axis accelerometer) is used to monitor the posture and activity level of the patient.

[0066] The electromyography (EMG) sensor is used to monitor the tension of the target muscle group related to pain.

[0067] The skin conductance (SC) sensor and the heart rate variability (HRV) sensor are used to monitor the autonomic nervous system activity related to pain emotion and stress.

[0068] The electroencephalogram (EEG) sensor is used to monitor the electroencephalogram activity of the specific frequency band of the cerebral cortex related to the pain emotion and the cognitive state, and can adopt a wearable (such as a head ring, a patch) or a minimally invasive implantable device.

[0069] According to specific needs, the application can further provide a patient terminal device and a cloud artificial intelligence (AI) engine, wherein the patient terminal device, such as a smart phone application, is used for input of patient reported outcomes (PROs) and communication with the IPG and the cloud server; the cloud artificial intelligence (AI) engine is used for storage and processing of massive data and running of complex machine learning optimization algorithms, wherein the patient terminal device and the cloud artificial intelligence (AI) engine can respectively represent the processor and the memory, or can be connected with the processor and the memory, which is not limited herein.

[0070] The system architecture of the application balances the delay, power consumption and computing capability by distributing different hardware components to specific control levels through a distributed design, and specifically comprises:

[0071] 1. Data Uplink Pathway: Implantable spinal cord electrodes sense ECAP signals, and a multimodal physiological sensor array (EEG, motion, EMG, SC, etc.) collects in vitro physiological data. All data streams are converged wirelessly to the patient's terminal device for timestamp synchronization (e.g., ...). Figure 7 (As shown), and uploaded to the cloud AI engine for storage and advanced processing;

[0072] 2. Controlling the Downlink Path: The cloud-based AI engine runs the third-level optimization algorithm to generate long-term policy parameters θ. These parameters are sent to the IPG as macro-level guidance for the second-level MPC controller. The second-level MPC controller (which can run on the IPG) calculates the target value ECAP required by the first level based on real-time EEG biomarkers. Target The first-level fast feedback loop is implemented entirely within the IPG, according to ECAP. Target The stimulation pulses applied by the spinal cord electrodes can be adjusted in real time.

[0073] This distributed computing architecture, which physically distributes the three layers of control logic across the IPG, patient terminal, and cloud, is an innovative trade-off and optimization for specific application scenarios of neuromodulation. It ensures that the system can achieve high-order intelligent optimization while meeting extremely low latency requirements.

[0074] In addition, such as Figure 4 As shown, the control processing unit is equipped with a three-layer closed-loop control and a finite state machine arbitration module, wherein:

[0075] 1. First-level closed-loop control circuit: Spinal cord autoregulator (rapid reflex layer)

[0076] Function: Maintains immediate stability and consistency of stimuli on a millisecond timescale.

[0077] Working mechanism: This level is a fast feedback loop based on ECAP (such as...) Figure 9 As shown, Figure 9 A typical ECAP waveform is shown, including stimulus artifacts, measurement time windows, and N1 and P1 peaks used to calculate peak-to-peak values. This loop is implemented in engineering as a proportional-integral (PI) controller, entirely on dedicated circuitry within the IPG. Its operating mechanism is as follows:

[0078] (1) Error calculation: In each control cycle k, the IPG will measure the ECAP value in real time. measured (k) and the target value ECAP issued by the second level Target (k) is compared, and the current error e(k) = ECAP is calculated. Target (k)-ECAP measured(k);

[0079] (2) PI algorithm: The processor calculates the amplitude of the next stimulation pulse u(k) according to the current error and the accumulation of historical errors. Its mathematical form is:

[0080]

[0081] where Kp is the proportional gain and Ki is the integral gain, which together ensure that the controller can quickly respond and accurately track the target value without steady-state error;

[0082] Integral Anti-Windup: To prevent dangerous stimulation overshoot due to continuous accumulation of the integral term when the hardware output reaches the upper limit, the system includes an integral anti-windup mechanism. When the calculated stimulation amplitude u(k) reaches the physical upper limit, the algorithm will suspend the accumulation of the integral term, ensuring that the system can quickly and smoothly reduce the output after the target value falls back, ensuring patient safety.

[0083] The system applies stimulation pulses through the spinal cord electrode and immediately senses the returned ECAP signal. When the patient's body movement causes the ECAP amplitude to deviate from the preset ECAP target value (ECAP Target ), a first control loop with a response delay of less than 5 milliseconds immediately adjusts the amplitude of the next stimulation pulse, making the ECAP amplitude quickly return to the vicinity of the target value.

[0084] 2. Second closed-loop control loop: Brain-spinal cord supervisor (slow emotional layer)

[0085] Function: Supervise and adjust the treatment strategy of the underlying layer according to the brain's pain-related state.

[0086] Working mechanism: This layer is a slow feedback loop based on EEG, and its feedback signal is a low-dimensional neurobiological marker highly related to the pain emotional state after processing, such as the power P ACC,γ of the anterior cingulate cortex (ACC) specific frequency band (such as Gamma band). The MPC controller achieves supervision by prospectively adjusting the ECAP Target value of the first layer. This layer uses a Model Predictive Control (MPC) framework. The core architecture of the MPC framework includes:

[0087] (1) State space prediction model: A personalized mathematical model learned from patient calibration data that can predict the evolution trend of future neurobiological markers. Its typical linear time-invariant (LTI) state space form is:

[0088] x(k+1) = A · x(k) + B · u(k) (2)

[0089] where the state vector x(k) contains biomarkers P ACC,γ (k), and the control input u(k) is ΔECAP Target (k). The system matrices A and B together constitute a personalized model of the patient’s “brain-spinal cord” response characteristics.

[0090] (2) Optimization problem and cost function: At each control period, the MPC finds the optimal control sequence that minimizes the cumulative cost by solving an optimization problem that looks N time steps into the future.

[0091] (3) Constraints: The optimization problem is solved under strict clinical safety constraints (as shown in Figure 5 ), ensuring that the MPC’s output never exceeds a predefined safety range.

[0092] This MPC controller operates on a slower time scale (e.g., every 200-500 milliseconds) and its core task is to dynamically adjust the ECAP Target values of the first tier. The MPC controller makes decisions based on a data-driven patient neurophysiological state prediction model that predicts the future evolution of key biomarkers (e.g., gamma power in the anterior cingulate cortex ACC). At each control period, the MPC solves an optimization problem to minimize a predefined cost function that aims to penalize predicted, pain-related neural states while penalizing excessive stimulation target changes to ensure patient comfort. A specific cost function can be formalized as:

[0093]

[0094] The output of this controller is subject to a series of strict, clinically-safety-specification-derived hard constraints, such as preset limits on the maximum rate of change of ECAP Target to ensure stimulation modulation smoothness and patient comfort, and to ensure it always stays within the clinician-set treatment window.

[0095] 3. Third tier closed-loop control loop: Mind-System Regulator (slow cognitive and optimization tier)

[0096] Function: To achieve therapy personalization and self-optimization with the patient’s long-term, comprehensive subjective experience as the ultimate optimization target.

[0097] Working mechanism: This tier is a slow feedback and learning loop based on AI and multi-modal data, using a Bayesian Optimization (BO) framework based on a multi-objective cost function (as shown in Figure 8The BO framework enables feedback and learning through a continuous "belief update-decision-evaluation" loop. Its framework components are as follows:

[0098] 1. Surrogate Model: This is the core of BO, usually implemented as a Gaussian Process (GP). GP uses the existing observations to construct a probabilistic model of the unknown objective function (i.e., "Patient Wellness Score") W(θ). This model can provide a prediction (mean μ * (θ) and variance σ GP (θ)) with uncertainty for any untested parameter point θ. *

[0099] 2. Acquisition Function: This is a strategy function that guides the next sampling, balancing "exploitation" and "exploration". A commonly used acquisition function is Upper Confidence Bound (UCB), which is defined as:

[0100]

[0101] where μ GP (θ) is the exploitation term, encouraging sampling where the model performs well; σ GP (θ) is the exploration term, encouraging sampling where the model is uncertain; and β is the trade-off parameter.

[0102] BO algorithm aims to maximize a composite "Patient Wellness Score" (W), which is a function of normalized PROs (e.g., NRS pain score, sleep quality, functional score) and penalty terms (e.g., energy consumption, paresthesia events). A specific objective function can be formalized as:

[0103]

[0104] where θ represents the multi-dimensional parameter combination (frequency, pulse width, electrode combination, etc.) that needs to be optimized. BO algorithm uses surrogate models such as Gaussian Process to efficiently explore and exploit within the "safe parameter domain" defined by clinical safety guidelines, to find the parameter combination that maximizes the patient's long-term subjective experience in fewer iterations, and to update the first and second level control strategies.

[0105] ​​The three control levels of the present application are not independent, but constitute a tightly coupled hierarchical control cascade operating on different time scales, with the following characteristics:

[0106] Top-down instruction flow: the BO algorithm of the third level (day / week level) provides strategic guidance for the second level by optimizing the core parameter set θ and adjusting the cost function weights of the second level MPC. The MPC controller of the second level (hundreds of milliseconds level) provides a dynamic tactical target ECAP for the first level according to the strategic guidance and real-time EEG Target . The PI controller of the first level (millisecond level) as the final executor accurately tracks the dynamic target.

[0107] Bottom-up information flow: ECAP data is used by the first level to calculate the error; EEG data (its signal chain processing flow is shown in Figure 3 ) is used by the second level for state prediction; and multi-modal data such as PROs, movements, EMG, etc. are used by the third level to calculate the objective function, driving the entire learning process.

[0108] This hierarchical goal-setting cascade structure is the core idea of the present application, which cleverly decomposes complex control problems on different time scales and abstraction levels, making the entire system modular, controllable and safe.

[0109] 4. Finite state machine arbitration module: finite state machine (FSM), top-level arbitration and robustness guarantee

[0110] The finite state machine (FSM) is not any one of the three control loops, but a top-level arbitration logic (as shown in Figure 6 ) that overrides them, where Figure 6 There are five different system states: resting normal, dynamic activity, sleep, high artifact, safety override). Its role is to dynamically manage and schedule the operation of the three controllers according to the macroscopic state of the patient (such as resting, active, sleep, etc.) judged from the sensor fusion data, and is the key to ensuring the robustness and safety of the system in real-world variable environments.

[0111] The correspondence between FSM and the three control processes is that FSM selectively activates, disables or modifies the behavior of each control level by executing a set of pre-defined, context-based rules. It is a "rule engine" that ensures that the system always adopts the most suitable control strategy for the current context. When the system detects a safety event or watchdog timeout, it will trigger the safety override and fallback process (the specific process is shown in Figure 10 , Figure 10The ultimate safety mechanism is demonstrated, the starting point of the flow is "system normal operation", the branch condition can be "communication interruption> preset time length" or "controller output touches the hard constraint", no matter which condition triggers, the flow should point to the two actions of "parameter rollback to the preset safety set" and "send alarm". The following table details how the FSM interacts with the three-layer control logic in different states.

[0112] Table 1: FSM interaction with three-layer control logic in different states

[0113]

[0114]

[0115]

[0116] The application also provides an embodiment of a system overall architecture based on a patient terminal device and a cloud AI engine, the architecture schematic diagram is as shown in Figure 1 The core components include:

[0117] Implanted part: in the middle or left side, including an implanted pulse generator (IPG) (containing a processor, a power supply, an ECAP sensing / stimulation circuit, and a motion sensor) and a spinal cord electrode;

[0118] Extracorporeal part: right side or periphery, including wearable multi-modal sensors (containing EEG, EMG, and SC sensors) and a patient terminal device (such as a smart phone, containing an App);

[0119] Remote part: top or cloud, representing a cloud AI engine.

[0120] The application also provides an embodiment of an implanted pulse generator (IPG), the implanted pulse generator (IPG) and electrode connection block diagram is as shown in Figure 2 The diagram should demonstrate the internal structure of the IPG, including the shell, the ceramic radio frequency wave window, and the connection with the electrode lead wire, and the internal stimulation front end, sensing front end, processor, power supply, communication module, and memory should be schematically drawn.

[0121] The application also provides a signal chain and artifact suppression flow, specifically as shown in Figure 3 The diagram demonstrates the processing process of the original multi-modal signal (especially EEG / IMU) in the form of a flowchart, including: original signal input -> adaptive filtering (using IMU data) -> independent component analysis (ICA) artifact removal -> feature extraction -> output to the second level controller.

[0122] The application also provides a complete and specific embodiment which can be directly implemented, wherein the system hardware structure comprises an implantable IPG, an ECAP sensing and stimulation circuit, a three-axis accelerometer and a Bluetooth communication module integrated in the IPG, two eight-contact or twelve-contact spinal cord electrodes, a wearable flexible neck ring or chest patch 17 integrated with dry EEG sensors, EMG sensors and SC sensors, and a smart phone preloaded with a special App as a patient terminal.

[0123] The three-level closed-loop method is implemented as follows:

[0124] After the patient is implanted with the system, in the initial programming stage:

[0125] The first layer calibration: the clinician determines the initial stimulation parameters effective for the patient and the ECAP therapeutic window (i.e. the minimum ECAP amplitude that produces an effect and the maximum ECAP amplitude that causes discomfort), and the ECAP Target The initial value is set in the middle of the window;

[0126] The second layer MPC controller implementation: the patient wears an EEG device, and the system starts monitoring the EEG rhythm of the patient in the resting state and during pain attacks, and the system algorithm (such as fast Fourier transform) calculates the gamma band power of the ACC and PFC regions in real time. When the MPC controller predicts that the gamma power will continue to rise and exceed the comfort zone based on its internal model, it calculates an optimal DeltaECAP Target adjustment (e.g. 5-15% up), which aims to pull the predicted gamma power back to the safety baseline under the premise of meeting the maximum change rate and therapeutic window constraints;

[0127] The third layer BO engine activation: the patient records NRS pain scores, sleep quality and functional status PROs through the mobile phone App every day, and the cloud BO engine collects at least a week of continuous data (ECAP, EEG, motion, EMG, SC and PROs). The BO algorithm starts to learn the correlation between these data. For example, the model may find that when the combination of "high EMG signal + low motion + enhanced gamma waves" appears after 1 hour, the patient is likely to report a high NRS score. Based on this, the BO engine will use its acquisition function to decide a new exploratory parameter combination (e.g. fine-tune the stimulation frequency from 40Hz to 50Hz), and observe the changes in PROs in the next 24 hours. Through the continuous "exploration-exploitation" cycle, the system autonomously "customizes" the optimal therapy for the patient;

[0128] System engineering implementation: synchronization, arbitration and robustness, specifically:

[0129] Synchronization and data fusion: all data streams from IPG, wearable devices and smart phone are time-stamped with high precision by a unified time protocol, ensuring the maximum allowed time alignment error between data points used for fusion is less than 20ms, smart phone App as data sink node, packaging and uploading time-aligned data, a hybrid model combining early and late fusion (such as CNN-LSTM) is used to extract a unified feature vector that can fully represent the current physiological-neural state of the patient;

[0130] Arbitration logic: as described in the above scheme, the system uses a finite state machine (FSM) as the top-level arbitrator, which determines the macro state of the patient according to the fused sensor data, and switches between different states to execute different control priorities and strategies;

[0131] Artifact rejection and safety: the system integrates a multi-stage artifact processing pipeline, including adaptive filtering using IMU data and blind source separation algorithm based on independent component analysis (ICA), to remove motion and electromyographic artifacts, ensuring the purity of the EEG signal input to the second level controller, in addition, the system solidifies the non-crossable hard-coded safety rules, including absolute parameter limits, charge density limits (such as below 30μC / cm 2 ), watchdog timer and automatic shutdown function in strong electromagnetic environment.

[0132] Based on the above embodiment, a comparative research scheme and simulation verification are carried out, and the following A / B comparative research scheme is designed:

[0133] Table 2 A / B comparative research scheme

[0134]

[0135]

[0136] The application also provides an artificial intelligence spinal cord nerve regulation method based on a multi-modal physiological-neural-experience closed loop, which is applied to the artificial intelligence spinal cord nerve regulation system based on the multi-modal physiological-neural-experience closed loop and comprises the following steps:

[0137] Step 1: applying an electrical stimulation pulse to the patient through an epidural electrode and sensing an ECAP signal;

[0138] Step 2: executing a first level closed loop control loop, adjusting the parameters of the subsequent electrical stimulation pulse in real time based on the ECAP signal, so as to maintain a preset ECAP target value ECAP Target ;

[0139] Step 3: executing a second level closed loop control loop, calculating a ΔECAP Target value, and dynamically adjusting the ECAPTarget ;

[0140] Step 4: performing a third level closed-loop control loop, receiving patient reported outcomes, and generating optimized stimulation parameters based on which updating the first and second level closed-loop control loops;

[0141] Step 5: by the finite state machine arbitration module, based on motion sensor data, electroencephalogram sensor data, sleep and resting state, and safety events, activating, freezing, and reverting the first, second, and third level closed-loop control loops.

[0142] In Step 2, a first level closed-loop control loop is performed, and parameters of subsequent electrical stimulation pulses are adjusted in real time based on ECAP signals to maintain a preset ECAP target value ECAP Target , specifically:

[0143] The first level closed-loop control loop is performed, and the ECAP target value ECAP Target is tracked based on a proportional-integral control to maintain a preset ECAP target value ECAP Target , the first level closed-loop control loop has an integral anti-windup mechanism, and the update delay of the control loop is not greater than 5 ms, and safety constraints of charge density, pulse width, frequency, and maximum amplitude are applied.

[0144] In Step 3, a second level closed-loop control loop is performed, and a ΔECAP Target value is calculated, based on which ECAP Target is dynamically adjusted, specifically:

[0145] The second level closed-loop control loop is performed, and a model predictive control loop based on electroencephalogram supervision is adopted, according to state variables estimated by electroencephalogram biomarkers, using an individualized state space model and a cost function, under the constraints of a therapeutic window and a maximum ECAP Target change rate, a ΔECAP Target is calculated, and based on the ΔECAP Target , ECAP Target is dynamically adjusted.

[0146] In Step 4, a third level closed-loop control loop is performed, patient reported outcomes are received, and optimized stimulation parameters are generated based on which the first and second level closed-loop control loops are updated, specifically:

[0147] A third level closed-loop control loop is executed to maximize a health score within a safety parameter domain using a Bayesian optimization loop based on experience and energy consumption, wherein the health score at least integrates patient reported outcomes, energy consumption, and a paresthesia penalty term to generate optimized stimulation parameters, and the first and second level closed-loop control loops are updated based on the optimized stimulation parameters.

[0148] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be mutually referred to.

[0149] The principles and implementation manners of the present application are described by using specific examples in the specification. The above description of the embodiments is only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, the specific implementation manners and application ranges can be changed according to the idea of the present application. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. An artificial intelligence spinal cord nerve modulation system based on a multimodal physiological-neurological-experience closed loop, characterized in that, include: The system includes an implantable stimulation and sensing subsystem, a body surface / wearable sensing subsystem, a communication and time synchronization subsystem, and a control and processing unit. The implantable stimulation and sensing subsystem, the body surface / wearable sensing subsystem, and the control and processing unit are connected in pairs via the communication and time synchronization subsystem. The implantable stimulation and sensing subsystem is used to output stimulation pulses to the target nerve tissue according to the input stimulation control parameters, and to collect evoked compound action potentials within a set sensing time window after stimulation. The body surface and wearable sensing subsystem is used to acquire electroencephalogram (EEG), electromyogram (EMG), electrocardiogram (ECG), and multiple physiological and behavioral signals. The communication and time synchronization subsystem is used to enable bidirectional communication between the implantable stimulation and sensing subsystem, the body surface and wearable sensing subsystem and the control processing unit, and to achieve time alignment of the data stream. The control processing unit includes a processor and a memory. The memory stores instructions that can be executed on the processor, and the processor is used to execute the instructions.

2. The system according to claim 1, characterized in that, The implantable stimulation and sensing subsystem includes an implantable pulse generator and an epidural electrode electrically connected thereto. The epidural electrode is implanted in the epidural space of the patient's spinal cord. The implantable pulse generator, in conjunction with the epidural electrode, is used to output stimulation pulses to the target nerve tissue according to the input stimulation control parameters and to collect ECAP within a set sensing time window after stimulation.

3. The system according to claim 2, characterized in that, The body surface and wearable sensing subsystem includes motion sensors, electromyography (EMG) sensors, electrodermal (ED) sensors, heart rate variability sensors, and electroencephalogram (EEG) sensors. The motion sensor is used to monitor the patient's posture and activity level; The electromyography sensor is used to monitor the tension of target muscle groups associated with pain; The skin conductance sensor and heart rate variability sensor are used to monitor autonomic nervous system activity related to pain and stress. The EEG sensor is used to monitor specific frequency bands of brain electrical activity in the cerebral cortex associated with pain, emotion, and cognitive state.

4. The system according to claim 3, characterized in that, The control processing unit is equipped with a three-layer closed-loop control and a finite state machine arbitration module, wherein: The first closed-loop control loop is a millisecond-level ECAP tracking loop, used to modulate the stimulus amplitude at the single-pulse level to track the target ECAP value. Target The first closed-loop control loop adopts proportional-integral control and has an integral anti-saturation mechanism. The update delay of the control loop is no more than 5ms, and safety constraints on charge density, pulse width, frequency and maximum amplitude are applied. The second closed-loop control loop is a model predictive control loop based on EEG supervision. It is used to predict state variables estimated from EEG biomarkers, utilizing an individualized state-space model and cost function, within the treatment window and ECAP. Target Calculate ΔECAP under the maximum rate of change constraint. Target And send it to the first layer closed-loop control loop; The third-layer closed-loop control loop is a Bayesian optimization loop based on experience and energy consumption, used to maximize the health score within the safety parameter domain. The health score integrates at least the patient-reported outcome, energy consumption, and sensory abnormality penalty, and outputs the optimization results used to adjust the cost function weights and / or parameter vector θ of the second closed-loop control loop. The finite state machine arbitration module is used to arbitrate the activation, freezing, and rollback of the first-layer closed-loop control loop, the second-layer closed-loop control loop, and the third-layer closed-loop control loop based on motion sensor data, EEG sensor data, sleep and rest states, and safety events.

5. An artificial intelligence spinal cord nerve modulation method based on a multimodal physiological-neurological-experience closed loop, applied to the artificial intelligence spinal cord nerve modulation system based on a multimodal physiological-neurological-experience closed loop as described in any one of claims 1-4, characterized in that, include: Step 1: Apply electrical stimulation pulses to the patient via epidural electrodes and sense the ECAP signal; Step 2: Execute the first-level closed-loop control loop, adjusting the parameters of subsequent electrical stimulation pulses in real time based on the ECAP signal to maintain a preset target ECAP value. Target ; Step 3: Execute the second-level closed-loop control loop and calculate ΔECAP. Target Value, based on which ECAP is dynamically adjusted Target ; Step 4: Execute the third-level closed-loop control loop, receive the patient-reported outcome, generate optimized stimulation parameters based on it, and update the first-level and second-level closed-loop control loops based on the optimized stimulation parameters. Step 5: The finite state machine arbitration module performs activation, freezing, and rollback arbitration on the first-layer closed-loop control loop, the second-layer closed-loop control loop, and the third-layer closed-loop control loop based on motion sensor data, EEG sensor data, sleep and rest states, and safety events.

6. The method according to claim 5, characterized in that, In step 2, the first-level closed-loop control loop is executed, and the parameters of subsequent electrical stimulation pulses are adjusted in real time based on the ECAP signal to maintain a preset target value of ECAP. Target Specifically: The first-level closed-loop control loop is executed, and the target value of ECAP is tracked based on proportional-integral control. Target To maintain a preset target value for ECAP. Target The first-level closed-loop control loop has an integral anti-saturation mechanism, and the update delay of the control loop is no more than 5ms. Safety constraints are applied to charge density, pulse width, frequency and maximum amplitude.

7. The method according to claim 6, characterized in that, In step 3, the second-level closed-loop control loop is executed, and ΔECAP is calculated. Target Value, based on which ECAP is dynamically adjusted Target Specifically: The second-level closed-loop control loop is executed using a model-predictive control loop based on EEG supervision. Based on state variables estimated from EEG biomarkers, an individualized state-space model and cost function are used within the treatment window and ECAP. Target Calculate ΔECAP under the maximum rate of change constraint. Target Based on ΔECAP Target Dynamically adjust ECAP Target .

8. The method according to claim 7, characterized in that, In step 4, the third-level closed-loop control loop is executed, the patient-reported outcome is received, and optimized stimulation parameters are generated based on it. The first-level and second-level closed-loop control loops are then updated based on the optimized stimulation parameters. Specifically: The third-level closed-loop control loop is executed, employing a Bayesian optimization loop based on experience and energy consumption to maximize the health score within the safety parameter domain. The health score integrates at least patient-reported outcomes, energy consumption, and sensory abnormality penalty items to generate optimized stimulus parameters. The first-level and second-level closed-loop control loops are updated based on the optimized stimulus parameters.

Images