Bayesian optimization-based closed-loop deep brain stimulation ADRC control method and system

Abstract

The invention belongs to the technical field of neural regulation and control, and discloses a closed-loop deep brain stimulation ADRC control method and system based on Bayesian optimization, and the system comprises a neural signal processing module, an ADRC controller, a Bayesian optimization module, and a stimulator module. The method comprises the following steps: acquiring a target brain region neuroelectric signal and extracting beta-band oscillation power as feedback; an ADRC controller is adopted, an extended state observer of the ADRC controller is utilized to estimate and compensate total disturbance of the system on line, real-time tracking and suppression of the oscillation power are achieved, and a stimulation frequency control signal is output; a Bayesian optimization algorithm is further introduced, comprehensive performance indexes with control errors and stimulation energy as targets are constructed, automatic global optimization is carried out on key parameters of an ADRC controller through Gaussian process modeling and collection function guidance, the defects of unsmooth control, poor interference immunity and the like in an existing closed-loop stimulation strategy are overcome, and the method has the advantages of being high in robustness, high in reliability and the like. And an effective solution is provided for individualized closed-loop treatment of Parkinson's disease.

Description

Technical Field

[0001] This invention relates to the field of neuromodulation technology, specifically to a closed-loop deep brain stimulation (ADRC) control method and system based on Bayesian optimization. Background Technology

[0002] Deep brain stimulation (DBS) is an effective treatment for movement disorders such as Parkinson's disease. A DBS treatment system consists of three core components: implantable stimulation electrodes, extension leads, and a subcutaneous programmable pulse generator.

[0003] Traditional open-loop deep brain stimulation (DBS) continuously applies stimulation with fixed parameters, which can alleviate symptoms, but it cannot adapt to the diurnal fluctuations of patients' symptoms, changes in drug effects, and individual differences in condition, potentially leading to unstable treatment effects, increased side effects, and increased energy consumption. Therefore, adaptive closed-loop DBS (aDBS), which can adjust stimulation parameters in real time based on physiological signal feedback, has become an important development direction. Specifically, closed-loop DBS quantitatively assesses the patient's disease state by real-time detection of physiological signals highly correlated with Parkinson's disease symptoms, and then designs stimulation parameters based on control theory for real-time optimization and adjustment. Among these, the abnormal enhancement of pathological β-band oscillations in local field potentials (LFP) originating from targets such as the subthalamic nucleus (STN) has been widely confirmed to be closely related to the severity of motor symptoms in Parkinson's disease and is a recognized key biomarker.

[0004] Based on this biomarker, various aDBS control methods have been proposed. One typical approach, such as Chinese patent CN120053882B, analyzes the STN-LFP signal to calculate the ratio of low-frequency β energy to high-frequency β energy (LHR). Based on a preset threshold, it classifies the patient's state into discrete categories such as "drug effective," "drug ineffective - resting," or "drug ineffective - movement," thereby triggering corresponding high, medium, and low fixed stimulus intensities. This method achieves stimulus switching based on state classification and belongs to a rule-driven control strategy. However, this type of strategy has significant limitations: its control output is discrete and step-like, unable to achieve continuous and fine-tuning of β oscillation power; furthermore, its fixed threshold-based rules lack the ability to estimate and compensate for unknown dynamics within the nervous system and external disturbances, resulting in limited control accuracy and adaptability when facing individual differences and complex physiological environmental changes.

[0005] A broader analysis of existing aDBS control strategies reveals that they can be categorized into two main approaches: data-driven and model-driven, both of which face inherent challenges.

[0006] Data-driven approaches, such as reinforcement learning (RL), deep reinforcement learning (DDPG), and various deep learning models, can optimize stimulus parameters by interacting with the patient's brain's physiological environment in real time, especially in feature extraction and dimensionality compression in high-dimensional state spaces. Bayesian control, based on Gaussian process regression, efficiently explores stimulus parameters by quantifying uncertainty, avoiding the blindness of traditional optimization methods. Although these methods successfully circumvent the challenge of relying on precise neural system models, they face problems in clinical neuromodulation scenarios, including data scarcity, high training costs, potential safety risks during the exploration process, poor model interpretability, and difficulty in adequately addressing significant individual differences.

[0007] Model-driven methods, such as those based on the Rubin and Terman neural computational models, fractional-order PID controllers, or fuzzy control, rely on known physiological mechanisms and mathematical theories of neural circuits. These methods offer clear control logic and facilitate theoretical analysis; however, due to the nonlinearity and complexity of the nervous system, the simplified models they construct inevitably mismatch with the real physiological system, affecting the accuracy and stability of adaptive regulation.

[0008] In summary, although existing technologies have made progress in the field of closed-loop DBS, both specific biomarker-based threshold switching schemes and broader model-driven or data-driven control paradigms have their own bottlenecks. To truly achieve efficient, stable, safe, and personalized neuromodulation of Parkinson's disease, the key lies in overcoming the limitations of existing control strategies and better addressing changes in pathological states and individual differences through more flexible feedback mechanisms and real-time optimization of control parameters. Summary of the Invention

[0009] In view of this, the present invention aims to overcome the above-mentioned defects and proposes a closed-loop active disturbance rejection control (ADRC) control method and system for deep brain stimulation based on Bayesian optimization. The architecture uses the β oscillation power of neural nuclei as the feedback signal, adopts an active disturbance rejection controller to realize real-time adjustment of the stimulation frequency, and designs Bayesian optimization to system tune the ADRC controller parameters.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, a closed-loop deep brain stimulation self-disturbance rejection control system based on Bayesian optimization is provided, comprising: (1) A neural signal processing module is used to collect neural electrical signals of the target brain region in real time and extract the β-band oscillation power reflecting the pathological state of Parkinson's disease as a feedback signal. (2) Active disturbance rejection controller (ADRC) receives feedback signals and preset target values, plans a smooth reference trajectory through tracking differentiator (TD), estimates the total system disturbance in real time using extended state observer (ESO), and generates stimulation frequency control signals based on nonlinear state error feedback (NLSEF) and disturbance compensation. (3) Bayesian optimization module, used to automatically optimize the key parameters of ADRC controller; this module establishes a proxy model of control performance based on Gaussian process, explores and utilizes the acquisition function balance, iteratively searches for the optimal parameter combination that minimizes the comprehensive objective function in the simulation environment, and configures it to ADRC controller; (4) Stimulator module: Based on the control signal output by ADRC controller, it generates electrical stimulation of the corresponding frequency and applies it to the target brain region.

[0011] Furthermore, the ADRC controller includes: (1) Tracking Differentiator (TD) is used to dynamically filter the preset target power to generate a smooth tracking signal and its derivative; (2) Extended State Observer (ESO) is used to estimate the system output, output rate of change and total disturbance estimate in real time based on system input and control output; (3) Nonlinear State Error Feedback Unit (NLSEF) is used to generate a preliminary control quantity based on the tracking signal output by the tracking differentiator, the derivative of the tracking signal, the error between the system output estimated by the extended state observer and the rate of change of output, and in combination with a preset nonlinear function. (4) Disturbance compensation unit, used to subtract the total disturbance estimate from the preliminary control quantity and perform gain normalization to obtain the final stimulation frequency control signal.

[0012] Furthermore, the Bayesian optimization module is configured to perform the following steps: S1: Initialization, obtaining the initial parameter sample set; S2: Gaussian process modeling, which sets the prior of the objective function as a Gaussian process and calculates the mean and variance of the posterior prediction distribution based on the sample set; S3: Optimize the acquisition function. Use the expected improvement function as the acquisition function. Based on the posterior prediction distribution, determine the next parameter to be evaluated by maximizing the acquisition function. S4: Iterative update: Evaluate the performance indicators corresponding to the parameter points in the simulation environment, update the historical evaluation data and Gaussian process model, and repeat the acquisition function optimization and iterative evaluation steps until the stopping condition is met. S5: Output optimal parameters, selecting the parameters that optimize performance metrics from all historical evaluation data.

[0013] Secondly, a closed-loop deep brain stimulation (ADRC) control method based on Bayesian optimization is provided, comprising the following steps: Step 1, Signal Acquisition and Processing: Real-time acquisition of neural electrical signals from the target brain region, and extraction of β-band oscillation power as a feedback signal; Step 2, Active Disturbance Rejection Control: Input the feedback signal and the preset target power value into the active disturbance rejection controller; the active disturbance rejection controller performs the following operations: a. Calculate the smoothed tracking signal and its derivative using a tracking differentiator; b. Estimate the system state and its estimated values ​​using an extended state observer; c. Over-nonlinear state error feedback law, based on error calculation of preliminary control quantity; d. Perform disturbance compensation and gain normalization on the preliminary control quantity to obtain the stimulation frequency control signal, and apply amplitude and rate of change constraints to it; Step 3, Bayesian optimization: In the simulation environment, the Bayesian optimization algorithm is used to optimize the parameter vector of the active disturbance rejection controller, find the optimal parameters that minimize the objective function, and configure them to the active disturbance rejection controller; the Bayesian optimization algorithm uses a Gaussian process as a surrogate model and uses the expected improvement function as the acquisition function to guide the parameter search; Step 4, Stimulation Execution: Electrical stimulation is generated according to the stimulation frequency control signal and applied to the target brain region.

[0014] Furthermore, the parameters in the parameter vector of the ADRC controller are selected from a parameter set, which includes the sampling period, TD bandwidth, ESO bandwidth, system nominal gain, error channel gain, nonlinear function power parameters, and stimulus frequency safety constraints; the Bayesian optimization module searches for key subsets of these parameters.

[0015] Thirdly, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the method.

[0016] Fourthly, a computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method when executing the program.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: The closed-loop control method proposed in this invention achieves precise and smooth suppression of pathological nerve signals through an ADRC controller. Its extended state observer can estimate and compensate for internal and external disturbances in real time, thereby significantly improving the accuracy and robustness of control and overcoming the shortcomings of traditional strategies such as unsmooth response and weak anti-interference ability.

[0018] Meanwhile, the introduction of Bayesian optimization (BO) algorithm enables personalized tuning of key parameters of ADRC, which can intelligently optimize based on individual physiological characteristics, ensuring efficacy while optimizing stimulation energy, effectively balancing control effect and treatment safety, and providing an efficient and stable solution for individualized closed-loop treatment of Parkinson's disease.

[0019] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0020] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a structural diagram of the design scheme of the present invention; Figure 2 This is a block diagram of the ADRC control algorithms for TD, NLSEF, and ESO; Figure 3 This is a flowchart of the Bayesian optimized ADRC control algorithm; Figure 4 This is a graph showing the relationship between stimulation frequency and power (mean ± standard deviation). Figure 5 This is a comparison of three closed-loop control methods (PI control, fuzzy control, and ADRC control) with open-loop 130Hz stimulation frequencies (a1, b1, c1, d1) and tracking effects (a2, b2, c2, d2) for tracking a target power of 110. Figure 6 Is the ADRC controller tracking constant before and after parameter optimization? Power performance indicators; Figure 7 The tracking performance of different control systems under time-varying Parkinson's disease (PD) conditions is as follows: (a) β-band power variation (no stimulus conditions); (b) proportional-integral (PI) control; (c) fuzzy control; (d) the tracking performance of optimized ADRC on β-band power variation; (e) the tracking performance of optimized ADRC on β-band power variation. Detailed Implementation

[0021] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0022] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0023] This invention provides a closed-loop deep brain stimulation (ADRC) control method and system based on Bayesian optimization. The invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] Example 1: Overall Architecture and Workflow of a Bayesian-Optimized Closed-Loop Deep Brain Stimulation (ADRC) Control System like Figure 1 As shown, the core of this invention lies in constructing a two-layer intelligent control closed loop: the inner layer is a real-time dynamic control loop composed of ADRC controllers, responsible for rapidly and smoothly suppressing pathological nerve signals; the outer layer is a parameter self-tuning loop composed of Bayesian optimization (BO) algorithm, responsible for offline or periodic global optimization of key parameters of the inner layer controller. The specific workflow of the system is as follows: 1.1 Signal Acquisition and Feature Extraction: The system implants electrodes into the patient's target brain region (preferably the dorsolateral subthalamic nucleus, STN) via stereotactic surgery to ensure precise positioning of the stimulating electrodes within the target area. Neural activity from the medial globus pallidus is acquired in real-time as neuronal cluster firing and transmitted to the neural signal processing module. The signal processing module performs preprocessing such as filtering and noise reduction on the raw signal, and then extracts key biomarkers through power spectrum analysis. In this embodiment, the biomarker is preferably the β-band (13-30 Hz) oscillation power, which is strongly correlated with the motor symptoms of Parkinson's disease. y(k) , where k represents a discrete time series.

[0025] 1.2 Target Setting and Error Generation: Based on clinical assessment, a target β-power value representing symptom relief is set. The real-time extracted β power y(k) With target value The comparison generates a real-time tracking error signal.

[0026] 1.3 ADRC Real-time Control: The tracking error signal is fed into the ADRC as the main input. Internally, the ADRC plans a smooth tracking trajectory using a tracking differentiator (TD), estimates the sum of unmodeled system dynamics and external disturbances in real time using an extended state observer (ESO), and calculates the optimal stimulus frequency control value at the current moment by combining a nonlinear state error feedback (NLSEF) law with disturbance compensation. u(k) .

[0027] 1.4 Stimulus execution: Control volume u(k) The signals are sent to an implantable pulse generator (IPG), which drives the stimulating electrodes to apply electrical stimulation at the appropriate frequency to the STN nuclei, thereby modulating neural circuit activity and inhibiting abnormal β oscillations.

[0028] 1.5 Bayesian Optimization Parameter Tuning: During system initialization or periodic maintenance, the outer Bayesian optimization module is activated. This module runs in a simulation environment based on the cortex-basal ganglia-thalamus computational model. It uses ADRC control parameters (such as ESO bandwidth, feedback gain, etc.) as optimization variables and a comprehensive evaluation function of control performance indicators (such as tracking error and stimulus energy) as the objective. Guided by Gaussian process modeling and acquisition functions, it automatically performs multiple rounds of simulation-evaluation-update to ultimately find a set of ADRC parameters that optimize the overall system performance. This optimized set of parameters is then configured into the real-time running ADRC controller.

[0029] Example 2: Detailed Design and Implementation of ADRC Controller The overall structure of the ADRC controller is as follows: Figure 2 As shown, its core design concept originates from canonical form and state observer theory. Its key innovation lies in classifying the unknown model characteristics of the system into the special state category of "extended state" and using the extended state observer (ESO) to estimate this state in real time, thereby achieving effective handling of unknown disturbances and model uncertainties. The ADRC control algorithm consists of three main components, including tracking differential (TD), nonlinear state feedback (NLSEF), and extended state observer (ESO).

[0030] Step S1: System Model Abstraction and Disturbance Definition Considering the complexity of neural circuits, v For target tracking power, at stimulation frequency For the controller's control output, GPi β band power The neural circuit representing the controlled output signal of the system can be abstracted into a class of uncertain discrete dynamic systems, whose expression is: ; in, This represents an unknown nonlinear function within the system. For the system nominal gain, External disturbances The sampling period.

[0031] To simplify the processing, the unknown dynamics and disturbances are uniformly defined as the "total disturbance" z(k): ; Therefore, the system model simplifies to: ; The above formula shows the controlled output Control signal With total disturbance Joint decision-making. The core idea of ​​ADRC is to estimate through observers. And compensate for it in the control law, thereby achieving model-independent dynamic control.

[0032] Step S2: Tracking Differentiator (TD) Design To ensure the smoothness of the control signal, a second-order discrete tracking differentiator is used to analyze the target power signal. Perform dynamic filtering: ; in, For the smoothed reference trajectory, for rate of change, This refers to the TD bandwidth parameter.

[0033] This structure is equivalent to a second-order critical damping filter, which can generate a continuous differentiable trajectory when the desired signal undergoes a step or disturbance, thus avoiding drastic fluctuations in the stimulation frequency caused by sudden changes in the target value.

[0034] Step S3: Design of Extended State Observer (ESO) and Perturbation Estimation ESO is key to ADRC's ability to resist disturbances. It extends the total disturbance z(k) into a new state of the system for observation, defining the extended state vector to abstract the controlled object without explicit modeling: ;; in, This indicates the output estimate. Indicates the rate of change of output. To the total disturbance The estimate.

[0035] The update law for a third-order discrete ESO is: ; Among them, observer gain From a single ESO bandwidth parameter Decide: ; By setting up appropriately By taking a certain value, ESO can quickly and accurately estimate the total disturbance, including model uncertainties and external disturbances, and approximate the true state of the system within a finite time. ≈ .

[0036] Step S4: Generation of Nonlinear State Error Feedback (NLSEF) Law First, calculate the error between the expected trajectory given by TD and the system state estimated by ESO: ; And use nonlinear functions To construct the control law: ; ; in, For power-order parameters, The threshold for the linear region. , These are the error channel gains, This is a nonlinear feedback. The above equation accelerates the response speed through nonlinear power amplification in the large error region, while maintaining linearity in the small error region to obtain a stable output, thereby achieving a balance between dynamic and steady-state performance.

[0037] Step S5: Disturbance Compensation and Final Control Output ESO estimation results for disturbances By introducing a control law for feedforward compensation, the final control quantity, i.e., the stimulus frequency u(k) to be output, is obtained: ; This compensation term can effectively offset unknown disturbances and uncertainties in the system, so that the output dynamics are mainly determined by... Decide.

[0038] In practical control, the stimulation frequency must meet the requirements of physical constraints and neural safety; therefore, the control signal must be limited to a feasible range. ; in, and These are the safe upper and lower limits for stimulation frequency. This represents the maximum allowable frequency change step size within adjacent cycles. This constraint ensures smooth frequency changes and avoids generating non-physiological stimuli to neural tissue.

[0039] Combining all the formulas, the parameter set of discrete ADRC is defined as follows: ; Thus, in each sampling period k, ADRC sequentially executes steps S2 to S5 to generate a smooth reference trajectory from TD. ESO estimates system state and disturbances NLSEF calculation of nonlinear feedback The final output is formed by combining disturbance compensation. This forms a closed-loop control.

[0040] Example 3: Automatic Tuning of Controller Parameters Based on Bayesian Optimization ADRC parameters are numerous and highly coupled, making manual tuning to the optimal value extremely difficult. This invention introduces a Bayesian optimization (BO) framework to perform global search and adaptive optimization of the control parameters. The process is as follows: Figure 3 As shown.

[0041] 3.1 Define the optimization problem The parameter tuning problem of ADRC is formalized as a black-box optimization problem. When the objective function is not analytical or computationally expensive, the optimal solution is approximated stepwise through a small number of simulations. The core idea is to use a Gaussian process (GP) to establish a posterior surrogate model of the objective function, and to balance exploration and utilization through an acquisition function, thereby finding the optimal combination of control parameters within a finite number of simulations.

[0042] Let the objective function to be optimized be... for: ;in, This is the ADRC control parameter vector.

[0043] Assume that performance metrics can be observed for this parameter in each simulation: ; That is, the true cost function Determined but unknown; observations contain Gaussian noise. .

[0044] 3.2 Establishing a Gaussian Process (GP) surrogate model Because each evaluation is performed on the simulation model Both have computational costs. BO uses a probabilistic surrogate model with a Gaussian process prior as the objective function: ;in, It is a mean function. ) is the covariance function, used to measure the correlation between parameter points.

[0045] Given the currently obtained sample set The posterior prediction distribution of a Gaussian process is defined as: ; in, , Let be the covariance matrix, and its elements are: , and They represent the points respectively. The predicted mean and uncertainty of the objective function.

[0046] 3.3 Select the acquisition function and determine the next evaluation point BO uses the Expected Improvement-Plus (EI) function for data acquisition. + The search process aims to achieve a balance between exploration and exploitation. (EI) + The function introduces a positive "margin" on top of the traditional Expectation Improvement (EI) method. >0, to avoid prematurely falling into local optima. For the cost function minimization problem, EI... + Defined as: ;in, This represents the minimum objective function value that has been observed so far.

[0047] According to the posterior distribution of the Gaussian process, EI + In parsing form: ; in, and These are the cumulative distribution function and probability density function of the standard normal distribution, respectively. When When very close to 0, = 0 to avoid numerical instability.

[0048] 3.4 Iterative Optimization and Convergence The optimization process is carried out in cycles until the maximum number of iterations is reached or the value of the acquisition function is lower than the set threshold.

[0049] First, by solving Determine the next parameter to be evaluated. Based on this, simulations are performed to update the dataset. .

[0050] Secondly, when When the number of iterations reaches a preset limit, the Bayesian optimization is considered to have converged, and the ADRC parameters are stopped from being updated. This limits the optimization time and avoids ineffective searches. Finally, the optimal parameters are output. .

[0051] In terms of performance evaluation, a normalized error index is defined to quantify the tracking accuracy of the control system; and an average stimulus intensity index is introduced to further evaluate the energy consumption of the control system. ; ; in, Indicates the first The output power obtained from the simulation For the target power, This represents the total number of simulation steps. For the first The stimulation frequency output by the controller in this simulation.

[0052] Example 4: Simulation Verification of Experimental Results 4.1 Selecting a Test Platform To verify the effectiveness of the method of the present invention, a computational model of the cortical-basal ganglia-thalamus (CBGT) closed-loop neural circuit was established in the MATLAB R2023b environment as a simulation test platform.

[0053] The Bayesian optimization process is executed offline in the same software environment, automatically updating the parameters of the ADRC controller based on the closed-loop simulation results to minimize the comprehensive objective function. In Parkinson's closed-loop DBS control, setting an appropriate tracking target is crucial for Parkinson's state conditioning.

[0054] 4.2 Control Performance Benchmark Test To verify the basic control performance of this invention, the relationship between the DBS stimulation frequency and the β-segment power of the neural nucleus in the CBGT calculation model was analyzed, and the results are as follows: Figure 4 As shown in the figure. Initially, at time 0, the patient is in a Parkinsonian state without DBS stimulation. The power gradually decreases as the stimulation frequency increases above 50 Hz.

[0055] The initial parameter settings for the ADRC controller are shown in Table 1. The STN was selected as the stimulation target in the simulation, and closed-loop DBS was applied starting from 2 seconds with an initial stimulation frequency of 5 Hz. The controller tracked the β-segment power of 110 dB calculated under open-loop stimulation at 130 Hz.

[0056] Table 1 ADRC Basic Parameters

[0057] The ADRC control of this invention is compared with open-loop 130Hz stimulation, traditional proportional-integral (PI) control, and fuzzy control. The stimulation frequency tracking and β power suppression effects are as follows: Figure 5 As shown in Table 2, the key performance indicators are compared.

[0058] Table 2. Closed-loop and open-loop DBS in tracking constant... Power performance indicators

[0059] The results show that, with similar stimulus energy (average frequency), ADRC significantly reduced the tracking error by about 20%, demonstrating superior control accuracy.

[0060] 4.3 Verification of Bayesian Optimization Effect Subsequently, the aforementioned Bayesian optimization framework was applied to a subset of key parameters of ADRC. A comparison of the average stimulus frequency and relative error before and after optimization is shown below. Figure 6 As shown in the figure, after Bayesian optimization (BO), the average stimulation frequency of the ADRC controller is 123.86 Hz, and the average relative error has decreased significantly to 14.41%. While maintaining the average stimulation frequency essentially unchanged or even slightly reduced, the control error was further reduced by approximately 16% compared to before optimization, and the error fluctuation range was significantly narrowed. This demonstrates that Bayesian optimization pushes the parameters into the stable feasible region of "low error – compliant frequency," not by increasing frequency to gain accuracy, but by reducing steady-state deviation and fluctuations through more effective parameter coordination; therefore, it achieves simultaneous improvement in control accuracy and stability without significantly increasing the stimulation burden.

[0061] 4.4 System Generalization Performance Test To evaluate the generalization performance of the optimized ADRC system, the system's performance under time-varying pathological conditions was tested to simulate real fluctuations in Parkinson's disease.

[0062] A parameter pd (0-1) is set to represent the severity of the disease. This parameter also represents the corresponding change in β oscillation power. Here, pd = 0 represents the normal state and pd = 1 represents the most severe PD state.

[0063] In the simulation above, pd changed over different time periods: pd=1 from 0 to 10 seconds; pd=0.8 from 10 to 20 seconds; and pd=0.1 from 20 to 30 seconds. When the stimulus was turned off, the β power changed with the pd value as follows: Figure 7 As shown in (a).

[0064] In this dynamic scenario, the tracking capabilities of ADRC, PI control, and fuzzy control before and after optimization are compared. The results are as follows: Figure 7As shown in (ae): the ADRC controller after Bayesian optimization tuning performs best. It exhibits "brief transition-rapid alignment" at the two PD state switching points, with almost no spikes or long tails, minimal fluctuations after switching, and the most stable steady state. In contrast, the unoptimized ADRC can suppress the ripple at the switching point, but the ripple is more obvious and there are low-frequency fluctuations in the later stage. The PI has strong initial suppression, but is prone to high-amplitude bursts and drifts after the second switching. The fuzzy control is generally stable, but deviates from the target for a long time and has a slow transition.

[0065] This test fully verifies that the ADRC control based on Bayesian optimization has advantages in settling time, transient ripple and steady-state variance, and has stable operation capability under dynamic changes in PD state.

[0066] Example 5: Possible Forms of System Implementation Based on the above basic structure, the present invention also includes the following alternative or additional technical features, all of which are compatible with the core technical solution: 5.1 Feedback Signal Replacement or Fusion: In addition to STN / GPi LFP signals, the system can select or fuse cortical ECoG, scalp EEG, electromyography (EMG), or inertial sensor motion signals as supplementary biomarkers. The control and optimization unit only needs to add the corresponding feature extraction channel to the signal preprocessing module to achieve multimodal feedback within a unified framework.

[0067] 5.2 Expansion of control quantity and control target: The control target can be not only the reduction of β power, but also a proxy indicator of tremor amplitude, mean square value of acceleration, or clinical score (such as UPDRS subscale); the control quantity can be expanded into a multi-channel, multi-target stimulation parameter vector to support the synergistic regulation of bilateral STN or STN and GPi.

[0068] 5.3 Variations and Combinations of Control Algorithms: The ADRC module can be replaced with linear ADRC, nonlinear ADRC, or a hybrid structure of ADRC with fuzzy control and PID neural networks, depending on the implementation. In addition to Gaussian processes, the surrogate model of the Bayesian optimization module can also be a t-process, random forest, or deep kernel Gaussian process. The acquisition function can be replaced with upper confidence bound (UCB), probability improvement (PI), etc. The sampling strategy remains unchanged in its basic principle of "iteratively selecting the next optimal test point in the parameter space".

[0069] 5.4 Implementation and Deployment: The control and optimization unit can be integrated as low-power hardware / firmware within the implantable IPG, or it can communicate with the implanted device via wired / wireless means as an external wearable or desktop application. For the Bayesian optimization part that requires greater computing power, edge devices or cloud computing can be used, and the optimization results are written back to the local controller through a secure communication channel.

[0070] These alternatives and extensions maintain the same structural connections as the basic implementation: they operate through a closed-loop chain of feedback signal → preprocessing → control and optimization → stimulus generation → neural response, differing only in the signal source, the form of the control law, or the location of computational deployment.

[0071] In summary, this invention develops a novel closed-loop control algorithm within the ADRC framework, which uses its extended state observer (ESO) to estimate the dependence of traditional control strategies on precise mathematical models in real time; it employs a Bayesian optimization algorithm to achieve automated tuning of key parameters of the ADRC controller, and solves the problem of difficulty in manually adjusting multiple parameters by constructing a probabilistic model of the objective function and designing a data acquisition function to guide parameter search.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A closed-loop deep brain stimulation self-disturbance rejection control system based on Bayesian optimization, characterized in that, include: (1) A neural signal processing module is used to collect neural electrical signals of the target brain region in real time and extract the β-band oscillation power reflecting the pathological state of Parkinson's disease as a feedback signal y(k); (2) An active disturbance rejection controller (ADRC) receives the feedback signal y(k) and the preset target power value r at its input terminal and outputs the stimulation frequency control signal u(k) at its output terminal; the ADRC is configured to generate the stimulation frequency control signal u(k) based on the tracking differentiator TD, the extended state observer ESO and the nonlinear state error feedback law NLSEF. (3) Bayesian optimization module, used for optimizing the parameter vector of the active disturbance rejection controller. Optimize the parameter vector The parameter set included in the active disturbance rejection controller The subset to be optimized; the Bayesian optimization module is configured to: establish the objective function through a Gaussian process. The proxy model, and based on the acquisition function iteratively search for... The minimum optimal parameters are then configured into the active disturbance rejection controller. (4) Stimulator module, used to apply electrical stimulation of a corresponding frequency to the electrodes implanted in the target brain region according to the stimulation frequency control signal u(k).

2. The closed-loop deep brain stimulation self-disturbance rejection control system based on Bayesian optimization according to claim 1, characterized in that, The active disturbance rejection (ADRC) controller includes: (1) Tracking differentiator TD, used to track the preset target power Dynamic filtering is performed to generate a smooth tracking signal. and its differential ; (2) Extended State Observer (ESO) is used to estimate the system output in real time based on the system input y(k) and the control output u(k). (k) Output rate of change (k) and total disturbance estimate (k); (3) Nonlinear state error feedback law unit NLSEF, used to calculate the output of the tracking differentiator. , The estimated by the extended state observer (k) Error between (k) , And combined with a preset nonlinear function Generate preliminary control quantities ; (4) Disturbance compensation unit, used to convert the preliminary control quantity Subtract the total disturbance estimate (k) and perform gain normalization to obtain the final stimulus frequency control signal. .

3. The closed-loop deep brain stimulation self-disturbance rejection control system based on Bayesian optimization according to claim 2, characterized in that, The discrete update law of the extended state observer ESO is: ； Where Ts is the sampling period; , , For observer gain, the bandwidth parameter of a single extended state observer is used. Decision, satisfaction: .

4. A closed-loop deep brain stimulation self-disturbance rejection control system based on Bayesian optimization according to claim 1 or 3, characterized in that, The Bayesian optimization module is configured to perform the following steps: S1: Initialization, obtaining the initial parameter sample set ,in, , For ADRC control parameter vectors, Let be the objective function to be optimized. It is Gaussian noise; S2: Gaussian process modeling, the objective function The prior is set as a Gaussian process. ,in, It is a mean function. ) is the covariance function; Based on sample set Calculate the mean of the posterior prediction distribution. and variance : ； in, , Let be the covariance matrix, and its elements are: , , and They represent points respectively. The predicted mean and uncertainty of the objective function; S3: Optimize the acquisition function by adopting the desired improvement function. As a data acquisition function: ； in, The minimum objective function value currently observed, margin > 0, and These are the cumulative distribution function and probability density function of the standard normal distribution, respectively. Based on the aforementioned posterior prediction distribution, the solution is obtained by maximizing the acquisition function. Determine the next parameter to be evaluated. ; S4: Iterative update, parameter evaluation get Update the sample set Then return to step S2 until the stopping condition is met; S5: Output optimal parameters, selecting the parameters that optimize performance metrics from all historical evaluation data. Output.

5. A closed-loop deep brain stimulation self-disturbance rejection control system based on Bayesian optimization according to claim 4, characterized in that, Define the normalized error index To improve the tracking accuracy of the quantification control system: ； in, Indicates the first The output power obtained from the simulation For the target power, This represents the total number of simulation steps.

6. The closed-loop deep brain stimulation self-disturbance rejection control system based on Bayesian optimization according to claim 4, characterized in that, Introducing the average stimulus intensity index Further evaluation of the energy consumption of the control system: ； in, This represents the total number of simulation steps. For the first The stimulation frequency output by the controller in this simulation.

7. A closed-loop deep brain stimulation (ADRC) control method based on Bayesian optimization, characterized in that, Includes the following steps: Step 1, Signal Acquisition and Processing: Real-time acquisition of neural electrical signals from the target brain region, and extraction of β-band oscillation power. y(k) As a feedback signal; Step 2, Active Disturbance Rejection Control: Input the feedback signal y(k) and the preset target power value r into the active disturbance rejection controller; the active disturbance rejection controller performs the following operations: a. Calculate the smoothed tracking signal using a tracking differentiator. and its differential ; b. Estimating the system state using an extended state observer. (k) (k) and estimated value (k); c. Over-nonlinear state error feedback law, based on error Calculate the initial control quantity ; d. The initial control quantity Perturbation compensation and gain normalization are performed to obtain the stimulation frequency control signal. and to Apply amplitude and rate of change constraints; Step 3, Bayesian optimization: In the simulation environment, the Bayesian optimization algorithm is used to optimize the parameter vector of the active disturbance rejection controller. Optimize and find a way to make the objective function Minimum optimal parameters and will Configured to the active disturbance rejection controller; the Bayesian optimization algorithm uses a Gaussian process as a surrogate model and employs the desired improvement function. Used as a guide for parameter search in the acquisition function; Step 4, Stimulation Execution: Based on the stimulation frequency control signal... u(k) It generates electrical stimulation and acts on the target brain region.

8. The method according to claim 7, characterized in that, The parameter vector of the active disturbance rejection controller The parameters are selected from the parameter set. ； in, The sampling period is For TD bandwidth parameters, For ESO bandwidth parameters, The nominal gain of the system; For the error channel gain, For power-order parameters, The threshold for the linear region. This represents the maximum allowable frequency variation step size within adjacent periods. These are the upper and lower limits for the safe frequency of stimulation.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in claim 7 or 8.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in claim 7 or 8.

Images