A neural closed-loop regulation method and system based on prediction coding residual and multi-quantization accuracy joint mapping

Abstract

The application discloses a neural closed-loop regulation method and system based on prediction coding residual and multi-modulation accuracy joint mapping. k The method separates layered prediction errors δ(t) from neural electrical signals, extracts at least three modulation substitute accuracies Π(t), and calculates a stimulation quantity vector through a joint mapping function S(t)=∑w The application breaks through the limitation of traditional single variable threshold feedback, and can be widely applied to various neural regulation scenes such as implantable DBS and non-invasive tDCS / TMS / tFUS. Simulation verification shows that the residual MSE is reduced by 27% and the energy consumption is reduced to 6% compared with the open loop.

Description

Technical Field

[0001] This invention relates to the fields of neural engineering and intelligent medical technology, specifically to a neural closed-loop modulation method and system based on predictive coding residuals and multi-quality tuning precision joint mapping, which is applicable to individualized and precise neural modulation of movement disorders, mood disorders and cognitive impairments. Background Technology

[0002] Deep brain stimulation (DBS) has been widely used in the treatment of neurological disorders such as Parkinson's disease, dystonia, obsessive-compulsive disorder, and treatment-resistant depression. Existing DBS systems are mainly divided into two categories: (1) Open-loop stimulation system: It outputs stimulation with fixed frequency, fixed pulse width and fixed amplitude, which cannot be dynamically adjusted according to the patient's real-time neurological state. It has problems such as mismatch between stimulation parameters and pathological state, rapid battery depletion and accumulation of side effects.

[0003] (2) Simple feedback closed-loop systems: such as the Medtronic Percept PC system (US 2018 / 0256282 A1), which uses the power of the β band (13-30Hz) of the local field potential (LFP) as a feedback signal, triggering stimulation or adjusting the amplitude when the β power exceeds a threshold. Bronte-Stewart et al. (US 2020 / 0346005 A1) disclosed a β power triggering control algorithm based on dual thresholds. Neumann et al. (2024, Science Translational Medicine) reported a multi-target DBS loop selection method based on ECoG and machine learning to decode symptoms.

[0004] While such systems achieve a closed loop of "sensory-stimulus", they have the following fundamental limitations: - Univariate driven: Relies only on power in a single frequency band as a feedback indicator, without utilizing the multi-level predictive structure information contained in neural signals; - Ignoring precision regulation: The dynamic regulatory effects of neuromodulation systems such as acetylcholine (ACh), norepinephrine (NE), dopamine (DA), serotonin (5-HT), and histamine (Hist) on perceptual precision are not considered. - Fixed stimulation target: Unable to adaptively select stimulation target channels based on the error source domain.

[0005] Predictive coding theory (Rao & Ballard, 1999; Friston, 2005, 2010) provides a unified framework for understanding brain information processing, arguing that the brain continuously predicts sensory input through hierarchical generative models, with prediction errors (residuals) propagating up the hierarchy and weighted by accuracy. However, there is currently no mature technical solution to translate this theoretical framework into a clinically deployable DBS closed-loop control method.

[0006] Non-invasive neuromodulation techniques (tDCS, TMS, tFUS) also suffer from drawbacks such as fixed parameters, lack of real-time feedback, and independent operation of multiple modes.

[0007] Therefore, there is an urgent need for a neural closed-loop modulation method and system that can simultaneously utilize hierarchical prediction residuals and multi-neural conditioning accuracy gains to achieve adaptive optimization of stimulation parameters and stimulation target channels. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing closed-loop neural modulation techniques, such as single feedback variable, disregard for precision control, insufficient dimensionality of stimulus parameter adjustment, and lack of online adaptive capability, and to provide a neural closed-loop modulation method and system based on predictive coding residuals and multi-precision tuning joint mapping.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A neural closed-loop modulation method based on joint mapping of predictive coding residuals and multi-quality tuning precision includes the following steps: S1. Collect the subject's neural electrical signals; S2. Perform hierarchical predictive coding analysis on the neural electrical signal to separate at least two levels of prediction error signals δ(t)={δ1(t), δ2(t), ..., δ n δᵢ(t) = Oᵢ(t) − Ôᵢ(t), where Oᵢ(t) is the observation signal of the i-th level and Ôᵢ(t) is the prediction output of the generative model of the i-th level. The level includes at least a sensory input layer, a feature extraction layer, and a category inference layer. S3. Extract from the neural electrical signals at least three alternative precision indices Π(t) = {Π1(t), Π2(t), ..., Π...} for the neuromodulation systems. m (t)}; S4. Calculate the stimulus vector S(t) = Σ using the joint mapping function. k ₌1ᴷ w k ·g(|δᵢ k (t)|)·h(Πⱼ k (t))·e_target k Where K is the number of stimulation channels activated simultaneously, wk Let g(·) be the preset weighting coefficient for the k-th channel, g(·) be the nonlinear compression function of the residual amplitude, h(·) be the normalized gating function of the accuracy gain, and i k j k These are the residual level index and precision index corresponding to the k-th channel, respectively, e_target k Let be the target point spatial pointing vector of the k-th channel; S5. Drive the electrical stimulation device to output an electrical signal to the target area according to the stimulation vector S(t), wherein each component of S(t) corresponds to a preset set of multiple electrode contact channels, and independently modulates at least two of the stimulation amplitude, frequency and pulse width of each channel. S6. Continuously estimate the moving average F(t) of the precision-weighted sum of squares of the system prediction error. When the moving average of F(t) increases for N consecutive cycles, trigger the adaptive update of the parameters of the hierarchical prediction coding generation model.

[0010] Furthermore, the nonlinear compression function g(x) = x / (1+|x| / τ), where τ is the compression time constant; the accuracy normalization gating function h(π) = max(0, (π−π0) / (π_max−π0)), where π0 is the accuracy baseline threshold and π_max is the accuracy saturation upper limit.

[0011] Furthermore, the method also includes adaptive target switching: when the residual amplitude exceeds a first threshold and a specific quality control accuracy index is lower than a second threshold, the target stimulation channel corresponding to the quality control system is activated preferentially.

[0012] Furthermore, the neural electrical signal is one or more combinations of local field potential (LFP), electrocorticography (ECoG), electroencephalography (EEG), and surface electromyography (sEMG); the electrical stimulation device is at least one of implantable deep brain stimulation device, transcutaneous transcranial direct current stimulation device, transcranial magnetic stimulation device, and transcranial focused ultrasound device.

[0013] This invention also provides a neural closed-loop control system based on predictive coding residual and multi-quality tuning precision joint mapping, for implementing the above method, comprising: - Signal acquisition module, used to acquire the subject's neural electrical signals; - Feature extraction module, connected to the signal acquisition module, is used to calculate the hierarchical prediction error δ(t) and extract at least three types of quality-tuning substitution accuracy Π(t); - A joint mapping module, connected to the feature extraction module, is used to execute the joint mapping function to calculate the stimulus vector S(t); - A stimulus execution module, connected to the joint mapping module, is used to output an independently modulated electrical signal to the target area; - Model monitoring module, used to calculate F(t) and trigger online model updates.

[0014] The signal acquisition module and the stimulus execution module are configured in one of the following two modes: Mode A: The signal acquisition module is an implantable deep electrode, and the stimulation execution module is an implantable pulse generator and a multi-contact electrode; Mode B: The signal acquisition module is a scalp EEG electrode or wearable sensor, and the stimulation execution module is a transcutaneous electrical stimulation headband, magnetic stimulation coil, or ultrasonic transducer array.

[0015] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements all the steps of the aforementioned method. Attached Figure Description

[0016] Figure 1 Overall flowchart of the method of the present invention Figure 2 Signal processing flowchart of the hierarchical predictive codec module Figure 3 Frequency band-tuning mapping diagram of the precision bus estimation module Figure 4 Schematic diagram of the input-output surface of the joint mapping function Figure 5 Closed-loop control timing and free energy monitoring curves Figure 6 Adaptive target switching logic diagram Figure 7 System modular structure diagram Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. The following embodiments are merely illustrative and do not constitute a limitation on the scope of protection of this invention.

[0018] Example 1: Parkinson's Disease DBS Scenario (Computer Simulation) A multichannel neural signal model of the subthalamic nucleus (STN) and primary motor cortex (M1) in Parkinson's disease state was constructed. The sampling frequency was 2 kHz, and the signal was bandpass filtered from 0.5 to 500 Hz.

[0019] Construct a three-layer predictive coding generation model: - Layer 1 (Sensory Input Layer): Observe the local field potential of STN and predict the LFP waveform at the next moment; - Layer 2 (feature extraction layer): Observe the time-frequency characteristics of the prediction error of Layer 1 to predict the motion preparation state; - Layer 3 (Category Inference Layer): Infers the current motion intention category by combining the errors of Layer 2.

[0020] The prediction errors δ1(t), δ2(t), and δ3(t) of each layer are obtained by updating the model parameters through online gradient descent.

[0021] Simultaneously extract the following precision indicators: - Π_NE(t): Power in the M1 region β band (15-25Hz), after logarithmic transformation and z-score normalization; - Π_DA(t): θ (4-8Hz) - γ (60-80Hz) phase amplitude coupling index of STN-M1, calculated by the Tort method; - Π_ACh(t): Power variation coefficient of the apical leaf α band (10-12Hz).

[0022] Joint mapping parameters: The nonlinear compression function g(x) = x / (1+|x| / τ), where τ is a selected value within the range of 50ms to 500ms, specifically determined by clinical calibration; the precision normalization gating function h(π) = (π−π0) / (π_max−π0), where π0 is the baseline value of each precision channel, and π_max is the upper limit of precision saturation, both π0 and π_max being determined by calibration of the patient's resting-state signal.

[0023] Set up three stimulation channels: - Channel 1: STN double-sided contacts, corresponding to δ2 (motion preparation error) × Π_NE; - Channel 2: Pale sphere inner side (GPi) contact point, corresponding to δ3 (motion intention error) × Π_DA; - Channel 3: M1 cortical surface contact point, corresponding to δ1 (sensory prediction error) × Π_ACh.

[0024] When the patient enters the "frozen gait" state, δ3 increases significantly and Π_DA is below the threshold. The system automatically increases the stimulation amplitude of the GPi channel and reduces the weight of the STN channel.

[0025] Stimulation frequency range: 60-185Hz, pulse width: 60-120μs, amplitude: 0-5V. Free energy monitoring window: 30-second moving average. When F(t) increases for three consecutive windows, retraining of the generative model parameters is triggered.

[0026] Simulation comparison results: Conclusion: The present invention significantly reduces symptom residuals within the frozen gait event window compared to the open-loop system, and consumes only 6% of the energy of the open-loop system.

[0027] Example 2: Treatment-resistant depression DBS scenario (computer simulation) The patient's target area was changed to the nucleus accumbens (NAc) / medial prefrontal cortex (mPFC) / basolateral amygdala (BLA), and the precision indicators focused on Π_5-HT and Π_DA.

[0028] A three-layer predictive coding generation model is constructed to extract Π_5-HT (slow wave power coherence synchronization index), Π_DA (θ-γPAC index), and Π_NE (β / γ power ratio).

[0029] Adaptive target switching logic: - When the δ amplitude is greater than the first threshold and Π_DA is lower than the second threshold, the NAc stimulation channel is activated preferentially; - When the δ spatial source points to the visual cortex and Π_ACh is below the third threshold, the occipital α-blocking channel is preferentially activated; - When the δ space source points to BLA and Π_5-HT is below the fourth threshold, the mPFC slow wave modulation channel is activated preferentially.

[0030] Simulation results show that the accuracy of emotion state classification is significantly improved compared with univariate β feedback, and the system energy consumption is significantly reduced.

[0031] Example 3: Hardware Resource Estimation (FPGA Example) Taking Xilinx XC7Z020-CLG484 as an example, the resources required to implement the complete closed-loop logic of this invention are as follows: Conclusion: The entire closed-loop control logic can be integrated into a single XC7Z020 chip, meeting the requirements for implantable low-power applications. The nonlinear compression function g(x) is implemented using a lookup table ROM and a linear interpolation circuit.

[0032] Example 4: Non-invasive tDCS scenario for depression A 64-lead EEG cap was used to acquire resting and task-mode EEG data at a sampling rate of 500 Hz.

[0033] A three-layer predictive coding model is constructed: the first layer δ1 is the difference between the original EEG and the AR(6) prediction; the second layer δ2 is the difference between the power of the θ band (4-8Hz) and the predicted value; the third layer δ3 is the difference between the HMM predicted probability of the emotional state and the actual label.

[0034] Extract three levels of precision: Π_ACh (θ / α ratio), Π_NE (β / γ ratio), and Π_5HT (α power variation coefficient).

[0035] Joint mapping parameters: τ is a selected value from 50ms to 500ms, specifically determined by clinical calibration; the amplitude, frequency, and pulse width of e_target are individually set by the physician based on patient tolerance and treatment response; the weights w for each channel...k Determined by multi-channel calibration procedure.

[0036] Stimulation output: Anode F3 position [Trade secret: current value] mA, cathode F4 position [Trade secret: current value] mA, 20 minutes daily for 4 consecutive weeks.

[0037] Safety monitoring: The built-in accelerometer detects head movements and automatically pauses when the movement exceeds a threshold; the temperature sensor monitors the skin temperature at the electrodes and automatically reduces the current when it exceeds 38°C.

[0038] Example 5: Alzheimer's Disease tFUS Scenario EEG data was acquired using an 8-channel EEG headband at a sampling rate of 256 Hz. A three-layer predictive coding model was constructed to extract Π_ACh (θ / γ ratio), Π_DA (β / γ ratio), and Π_Hist (γ power variation coefficient).

[0039] Joint mapping parameters: τ is a selected value within the range of 50ms to 500ms, specifically determined by clinical calibration; the sound pressure level, frequency, and pulse width of e_target are determined by ultrasound safety dosage standards and the acoustic transmission characteristics of the patient's skull; the weights w for each channel... k Determined by multi-channel focused ultrasound calibration. Stimulation output: target area is bilateral hippocampus (via temporal window), frequency [trade secret: kHz] kHz, pulse width 10 ms, duty cycle 10%, 3 times a week for 12 consecutive weeks.

[0040] Safety monitoring: The ultrasonic transducer integrates a thermistor on the back to monitor scalp temperature in real time and automatically reduces power when it exceeds 41°C; the passive cavitation detector monitors cavitation signals and suspends stimulation when they exceed the threshold.

[0041] The above embodiments are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0042] Technical Definitions and Terminology of this Invention The "predictive residual δ", "quality-tuning substitution accuracy Π", "joint mapping function S(t)", and "accuracy-weighted sum of squares F(t)" mentioned in this invention are all custom naming systems for signal processing and control systems specific to neural closed-loop modulation scenarios. They belong to engineering algorithm modeling concepts and are not quantum field theory or cosmological field theory in the physical sense. The technical solution of this invention uses the joint mapping function S(t) to perform real-time physical modulation of the hardware parameters (amplitude, frequency, pulse width) of the electrostimulation device, producing a technical effect of adaptive optimization of stimulation parameters, which does not belong to purely intellectual activity rules.

Claims

1. A neural closed-loop modulation method based on joint mapping of predictive coding residuals and multi-precision tuning, characterized in that, Includes the following steps: S1. Collect the subject's neural electrical signals; S2. Perform hierarchical predictive coding analysis on the neural electrical signal to separate at least two levels of prediction error signals δ(t)={δ1(t), δ2(t), ..., δ n δᵢ(t) = Oᵢ(t) − Ôᵢ(t), where Oᵢ(t) is the observation signal of the i-th level and Ôᵢ(t) is the prediction output of the generative model of the i-th level. The level includes at least a sensory input layer, a feature extraction layer, and a category inference layer. S3. Extract from the neural electrical signals at least three alternative precision indices Π(t) = {Π1(t), Π2(t), ..., Π...} for the neuromodulation systems. m (t)}; S4. Calculate the stimulus vector S(t) = Σ using the joint mapping function. k ₌1ᴷ w k ·g(|δᵢ k (t)|)·h(Πⱼ k (t))·e_target k Where K is the number of stimulation channels activated simultaneously, w k Here, g(x) = x / (1+|x| / τ) is the preset weighting coefficient for the k-th channel, g(x) = x / (1+|x| / τ) is the nonlinear compression function for the residual amplitude, h(π) = max(0, (π−π0) / (π_max−π0)) is the accuracy normalization gating function, τ is the compression time constant, π0 is the accuracy baseline threshold, and π_max is the accuracy saturation upper limit. k j k These are the residual level index and precision index corresponding to the k-th channel, respectively, e_target k Let be the target point spatial pointing vector of the k-th channel; S5. Drive the electrical stimulation device to output an electrical signal to the target area according to the stimulation vector S(t), wherein each component of S(t) corresponds to a preset set of multiple electrode contact channels, and independently modulates at least two of the stimulation amplitude, frequency and pulse width of each channel. S6. Continuously estimate the moving average F(t) of the precision-weighted sum of squares of the system prediction error. When the moving average of F(t) increases for N consecutive cycles, trigger the adaptive update of the parameters of the hierarchical prediction coding generation model.

2. The method according to claim 1, characterized in that, The at least three alternative precision indices for neuromodulation systems include at least three of the following: a cholinergic precision index estimated by alpha band power, a norepinephrine precision index estimated by beta band power, a dopaminergic precision index estimated by the θ-γ phase amplitude coupling index, a serotonergic precision index estimated by the slow wave power coherence synchronization index, and a histaminergic precision index estimated by the tonic baseline fluctuation variance.

3. The method according to claim 1, characterized in that, The method further includes adaptive target switching: when the residual amplitude exceeds a first threshold and a specific quality control accuracy index is lower than a second threshold, the target stimulation channel corresponding to the quality control system is activated preferentially; wherein the first threshold is 2 to 3 times the sliding window standard deviation of the residual amplitude, and the second threshold is 0.3 to 0.5 times the daily periodic mean of the corresponding quality control accuracy index.

4. The method according to claim 1, characterized in that, The target switching rules include: when the residual amplitude is greater than the first threshold and the dopaminergic accuracy index is lower than the second threshold, the nucleus accumbens stimulation channel is activated preferentially; when the residual spatial source points to the visual cortex and the cholinergic accuracy index is lower than the third threshold, the occipital lobe stimulation channel is activated preferentially; when the residual spatial source points to the basolateral amygdala and the serotonergic accuracy index is lower than the fourth threshold, the medial prefrontal cortex stimulation channel is activated preferentially.

5. The method according to claim 1, characterized in that, The neural electrical signal is one or more combinations of local field potential (LFP), electrocorticography (ECoG), electroencephalography (EEG), and surface electromyography (sEMG); the electrical stimulation device is at least one of implantable deep brain stimulation device, transcutaneous transcranial direct current stimulation device, transcranial magnetic stimulation device, and transcranial focused ultrasound device.

6. A neural closed-loop control system based on joint mapping of predictive coding residuals and multi-quality tuning precision, characterized in that, To implement the method according to any one of claims 1 to 5, comprising: The signal acquisition module is used to acquire the subject's neural electrical signals; The feature extraction module, connected to the signal acquisition module, is used to calculate the hierarchical prediction error δ(t) and extract at least three types of quality-tuning substitution accuracy Π(t); The joint mapping module, connected to the feature extraction module, is used to execute the joint mapping function to calculate the stimulus vector S(t); The stimulus execution module, connected to the joint mapping module, is used to output independently modulated electrical signals to the target area; The model monitoring module is used to calculate F(t) and trigger online updates of the generated model; The signal acquisition module and the stimulus execution module are configured in one of the following two modes: Mode A: The signal acquisition module is an implantable deep electrode, and the stimulation execution module is an implantable pulse generator and a multi-contact electrode; Mode B: The signal acquisition module is a scalp EEG electrode or wearable sensor, and the stimulation execution module is a transcutaneous electrical stimulation headband, magnetic stimulation coil, or ultrasonic transducer array.

7. The system according to claim 6, characterized in that, The feature extraction module includes a digital signal processor (DSP) or a field-programmable gate array (FPGA) chip. The firmware of the FPGA chip is configured to execute the following operational logic: perform hierarchical predictive coding on the neural electrical signal, calculate the hierarchical predictive residual δ(t), extract at least three neuromodulation substitution precisions Π(t), and perform joint mapping using the mapping function S(t)=Σw k The stimulus vector is calculated by ·g(|δ|)·h(Π)·e_target, and F(t) is continuously estimated and the model is updated when it rises continuously; the nonlinear compression function g(x)=x / (1+|x| / τ) is implemented by lookup table ROM and linear interpolation circuit.

8. The system according to claim 6, characterized in that, The stimulus execution module includes at least one of the following three types of actuators: tDCS actuator: includes one or more pairs of Ag / AgCl stimulation electrodes and a constant current source circuit, with an output current of 0.5-4.0mA; TMS actuator: includes magnetic stimulation coil and pulse capacitor bank, output pulse magnetic field strength 0.5-2.5T; tFUS actuator: includes an ultrasonic transducer array and a power amplifier, with an output sound pressure level of 0.1-2.0 MPa.

9. The system according to claim 6, characterized in that, It also includes an external programmable device, which is connected to the signal acquisition module and the stimulation execution module via a wireless communication module. This device is used to display real-time δ value, Π value, S(t) matrix and F(t) monitoring value, configure the default value of the reference stimulation vector e_target for each channel, start / stop / pause stimulation output, and export stored data for offline analysis.

10. 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 method described in any one of claims 1 to 5.

Images