40Hz acousto-optic stimulation dynamic optimization system based on electroencephalogram closed-loop feedback

Abstract

The invention provides a 40Hz acousto-optic stimulation dynamic optimization system based on electroencephalogram closed loop feedback, belongs to the technical field of cognitive disorder treatment, and solves the technical problems that an existing stimulation system lacks a dynamic closed loop, ignores individual differences, lacks a cross-band coupling mechanism and the like. Comprising an electroencephalogram gamma oscillation power detection module, an embedded signal processing and analysis module, a stimulation parameter generation and execution adjustment module, a personalized strategy database module, a multiple safety protection module, an edge calculation and communication module, a user interaction module, a head-mounted acquisition and stimulation device, a scene adaptation engine module and a double-loop closed-loop feedback control module. And the processing algorithm module comprises user terminal equipment with application software and a cloud central processing unit. According to real-time electroencephalogram feedback, a personalized historical data strategy library and an algorithm, parameters of 40Hz acousto-optic stimulation can be dynamically adjusted by relying on a double-loop closed-loop feedback control mechanism, and a mature cloud-edge-end cooperative computing ecosystem is constructed.

Description

Technical Field

[0001] This invention belongs to the field of cognitive impairment treatment technology, and relates to an electroencephalogram (EEG) closed-loop feedback acoustic-optical stimulation system, particularly a 40Hz acoustic-optical stimulation dynamic optimization system based on EEG closed-loop feedback. Background Technology

[0002] In the field of cognitive impairment treatment, 40Hz audio-visual stimulation technology is based on a major neuroscience discovery: brain gamma oscillations (40Hz rhythm) are a key biomarker of cognitive function, and this rhythm is significantly attenuated in patients with Alzheimer's disease and other conditions. In 2016, an MIT team first demonstrated in Nature that 40Hz light stimulation can activate microglia to clear Aβ plaques, opening up a new pathway for non-pharmacological intervention.

[0003] Existing technical solutions: Current mainstream 40Hz audio-visual stimulation systems use fixed-parameter stimulation (such as constant sound pressure / light intensity) or limited adjustments based on static EEG characteristics. Their shortcomings include: lack of dynamic closed-loop: relying solely on baseline EEG data before stimulation, unable to respond in real-time to gamma power fluctuations; ignoring individual differences: failing to integrate biomarkers such as hippocampal atrophy, sleep comorbidities, and the APOE4 gene, resulting in low response rates in specific populations; and lack of cross-frequency coupling: failing to establish a gamma coupling feedback mechanism (gamma band modulation cannot be triggered when the boost value is <25%), weakening neuroplasticity. Furthermore, static stimulation protocols lead to poor neural oscillation adaptation; comorbid patients are prone to circadian rhythm disruption and have a high rate of abnormal melatonin secretion.

[0004] Therefore, a dynamic optimization system for 40Hz audio-visual stimulation based on EEG closed-loop feedback is proposed. Summary of the Invention

[0005] The purpose of this invention is to address the aforementioned problems in existing technologies by proposing a dynamic optimization system for 40Hz audio-visual stimulation based on EEG closed-loop feedback. The technical problem this invention aims to solve is: how to dynamically adjust the parameters of 40Hz audio-visual stimulation based on real-time EEG feedback, a personalized historical data strategy library, and algorithms, relying on a dual-loop closed-loop feedback control mechanism, thereby constructing a mature "cloud-edge-device" collaborative computing ecosystem.

[0006] The objective of this invention can be achieved through the following technical solutions: A dynamic optimization system for 40Hz audio-visual stimulation based on EEG closed-loop feedback includes an EEG gamma oscillation power detection module, an embedded signal processing and analysis module, a stimulation parameter generation and execution adjustment module, a personalized strategy database module, a multi-layered security protection module, an edge computing and communication module, a user interaction module, a head-mounted acquisition and stimulation device, a scene adaptation engine module, a dual-loop closed-loop feedback control module, a processing algorithm module, a user terminal device with application software, and a cloud-based central processing unit. The system forms a cloud-edge-device collaborative processing architecture that can dynamically and continuously optimize 40Hz stimulation parameters based on real-time EEG feedback, a personalized historical data strategy library, and algorithms.

[0007] The working principle of this invention: This system is a data-driven, personalized, and adaptive neuromodulation platform. Its core principle lies in constructing a cloud-edge-device collaborative intelligent closed loop centered on "perception-decision-execution-learning." By dynamically optimizing the 40Hz audio-visual stimulation parameters, it achieves precise, safe, and efficient enhancement of the user's brain gamma oscillations. Real-time perception and edge computing: Signal acquisition: The user wears a head-mounted device whose dry electrode array focuses on acquiring EEG signals from the occipital lobe region, which is most sensitive to visual stimulation. Feature extraction: The acquired raw EEG signals are immediately processed in real-time by the device's embedded DSP chip. It quickly calculates the power value of the 40Hz frequency band in a 200ms window. This step is completed locally, ensuring low latency and real-time performance. Intelligent decision-making and parameter generation: State comparison and strategy matching: The calculated gamma power value is compared with data in a personalized strategy database. The database not only stores the user's historical response data but also includes feature tags such as sleep quality and cognitive function scores, as well as optimization objectives defined by the clinical gold dataset. Multi-module collaborative decision-making: Dual-loop closed-loop control module: Based on real-time feedback... The difference between the feedback and the target value determines the adjustment direction and magnitude; Scene adaptation engine: invokes preset advanced optimization strategies based on the user's specific situation; Stimulus parameter dynamic adjustment module: receives instructions and generates specific, executable combinations of stimulation parameters; Precise execution and dynamic adjustment: the head-mounted device receives new instructions from the stimulation parameter generation and execution adjustment module; The sound pressure controller, light intensity modulator, and acoustic-optical phase difference generator on the device work together to precisely output the adjusted 40Hz acoustic-optical stimulation: Sound pressure: adjustable within the range of 0-75dB; Light intensity: within the range of 100-1000 dB. Adjustable within the lux range; Phase difference: Achieves precise time delay of 0-360° between acoustic and light stimuli to study or utilize cross-sensory integration effects; Continuous learning and safety safeguards: Cloud learning: All data is uploaded to the cloud central processor via edge computing and communication modules; The cloud utilizes more powerful computing power for deep analysis, model training, and algorithm iteration to continuously optimize system intelligence and other performance aspects; Comprehensive safety protection: Multiple safety protection modules monitor signal quality, physiological responses, and stimulation parameters throughout the process to ensure that any operation is within an absolutely safe range, and any abnormalities will be immediately terminated.

[0008] The EEG gamma oscillation power detection module includes an EEG signal acquisition unit, an auxiliary physiological signal acquisition unit, an EEG artifact compensation unit, an individual structural data import interface unit, an environmental perception unit, an embedded fNIRS monitoring unit, a stress hormone monitoring unit, and a multimodal data fusion engine unit.

[0009] Furthermore, the EEG signal acquisition unit employs a high-density dry electrode array, fitted to the occipital and prefrontal lobes, to acquire EEG signals from both lobes. The sampling rate is ≥250Hz, motion artifacts are reduced by 40%, and the signal-to-noise ratio is ≥20dB. Automatic electrode impedance calibration ensures that an alarm is triggered in case of poor contact, and EEG data is acquired. The auxiliary physiological signal acquisition unit includes a pupil monitoring camera, a heart rate variability sensor, a motion state sensor, and a cortisol sensor. The pupil monitoring camera uses a miniature infrared camera, extracting photoplethysmography (PPG) and heart rate variability based on a micro-motion amplification algorithm. It monitors the user's pupil diameter, blink frequency, facial micro-expressions, and head posture in real time, identifying states such as fatigue, discomfort, and inattention, and acquiring cerebral blood oxygenation signals and visual behavior data. The EEG artifact compensation unit uses motion data from a 6-axis IMU and a Kalman filter algorithm to perform real-time motion compensation on EEG signals in moving scenarios. The gamma power calculation error is ≤5%, suitable for mild activity scenarios; the individual structural data import interface unit supports importing individual user T1-weighted magnetic resonance images to generate personalized 3D brain models for accurate sound field simulation and targeted positioning; the environmental perception unit collects environmental parameters such as ambient light intensity and noise level; the embedded fNIRS monitoring unit uses an 8-channel flexible fNIRS sensor attached to the middle frontal gyrus and inferior parietal lobule to collect local cerebral blood oxygen saturation and deoxyhemoglobin concentration in real time at a sampling rate of 10Hz, used to verify neuronal metabolic activation corresponding to EEG gamma oscillations; the stress hormone monitoring unit integrates a wearable interstitial fluid sensor to detect cortisol concentration through electrochemical methods at a sampling period of 5 minutes, used to determine the user's stress state; the multimodal data fusion engine unit uses an attention mechanism neural network to weightedly fuse EEG, fNIRS, and HRV feature vectors to generate a comprehensive state assessment index.

[0010] The embedded signal processing and analysis module includes a preprocessing unit, a feature calculation unit, a feature extraction unit, a data temporary storage unit, and a state evaluation unit.

[0011] Furthermore, equipped with a TITMS320C6748DSP chip, it adopts a two-stage processing flow of "preprocessing-feature extraction": the preprocessing unit eliminates power frequency interference through a 50 / 60Hz notch filter, removes baseline drift through a 4th-order Butterworth high-pass filter, and removes artifacts from electrooculography and electromyography through independent component analysis; the feature calculation unit uses a 200ms sliding window to calculate the γ power in the 40Hz±2Hz frequency band using short-time Fourier transform, and extracts the γ phase amplitude coupling index through Hilbert transform; the feature extraction unit calculates the γ oscillation power, θ band energy, θ-γ phase amplitude coupling index, and high-frequency oscillation events >70Hz in real time; the data buffer unit has a built-in 8GB DDR3 cache to store the raw EEG data and γ power calculation results in real time, supporting local backtracking analysis; the state assessment unit evaluates the user's cognitive state and stimulus response in real time based on a deep learning model, with the model size compressed to 1 / 20 of the original size to meet embedded requirements.

[0012] The stimulation parameter generation and execution adjustment module includes an acoustic-optical stimulation signal generation unit, a sound field simulation and compensation unit, a dynamic frequency chirp unit, a hierarchical modulation unit, a parameter adjustment unit, a cross-modal synchronous tactile stimulation unit, and a cross-frequency band modulation unit.

[0013] Furthermore, based on γ power feedback, the sound pressure, light intensity, and phase difference are dynamically adjusted to output cross-modal synchronized acoustic, optical, and tactile stimulation. The acoustic-optical stimulation signal generation unit uses a quantum dot-enhanced optical system to provide high-purity optical stimulation and employs a magnetically controlled acoustic array to achieve focused propagation of sound waves. The sound pressure controller is adjustable from 0 to 75 dB with a step accuracy of 1 dB; the light intensity controller is adjustable from 100 to 1000 lux with linear adjustment; the acoustic-optical phase difference generator has a programmable delay of 0 to 360° with an accuracy of ±5°; the real-time optimized frequency ≥0.5Hz sound field simulation and compensation unit, based on the user's individual MRI skull model, simulates the path loss of sound waves propagating to the target brain region and pre-compensates the output sound waves to ensure precise and controllable sound energy reaching the target; the dynamic frequency chirp unit... The 40Hz fundamental frequency can dynamically change within the range of 38-42Hz to simulate natural neural rhythms and avoid neuronal desensitization—after 2 weeks, the gamma power enhancement rate drops from 85% to 70%; the graded modulation unit adjusts the theta modulation depth according to the gamma power enhancement rate, and the long-term stimulation effect decay rate is ≤5%; the parameter adjustment unit adjusts sound pressure, light intensity, and phase difference in real time; the cross-modal synchronous tactile stimulation unit integrates a tactile vibrator, whose vibration rhythm is precisely synchronized with the sound and light stimulation in time and phase, forming a synergistic effect of the whole body and multiple senses, enhancing the control effect; the cross-band modulation unit automatically generates a 4-7Hz theta band signal when the gamma oscillation power enhancement value is <25%, and superimposes it on the 40Hz sound and light carrier in an amplitude modulation manner to promote theta-gamma cross-band coupling.

[0014] The personalized strategy database module includes a user feature storage unit, a treatment efficacy prediction unit, a strategy online iteration unit, a strategy mapping unit, a golden data calibration unit, and an alternative algorithm unit.

[0015] Furthermore, the user feature storage unit stores multi-dimensional user data, including structural imaging features (hippocampal volume), genetic features (APOE4 genotype), and clinical scores (MMSE cognitive score and PSQI sleep quality score). The efficacy prediction unit uses a lightweight Transformer model combined with transfer learning, taking into input user static features and multimodal data from the first three stimuli, and outputting stimulus response probabilities. The model size is ≤3MB, and the inference time is <30ms. The online strategy iteration unit is based on reinforcement learning, using "increased γ oscillation power + increased HbO2 + user subjective score" as the reward function, and automatically updates the personalized strategy library daily without manual intervention. The strategy mapping unit establishes a three-level strategy mapping model, automatically matching the initial stimulus scheme based on user features, and dynamically correcting it with real-time γ power. The golden data calibration unit has a built-in clinically validated golden threshold. When real-time data deviates from the golden threshold by ±15%, it triggers forced strategy calibration. The alternative algorithm unit uses a support vector machine classifier to replace the traditional threshold judgment. The classifier is trained on 10,000 clinical samples and is suitable for nonlinear feature association scenarios.

[0016] The multiple security protection modules include a light intensity dynamic upper limit control unit, an epilepsy risk blocking unit, an autonomic nervous system balance monitoring unit, a blockchain data storage unit, a federated learning unit, a multimodal fusion security monitoring unit, a digital biomarker mining engine unit, an epilepsy risk early warning unit, a visual fatigue adjustment unit, a comfort closed-loop unit, a circadian rhythm protection unit, an emergency unit, a homomorphic encryption unit, a permission hierarchical subunit, and a data traceability unit.

[0017] Furthermore, the dynamic upper limit control unit for light intensity monitors the degree of pupil constriction in real time through infrared eye tracking. When pupil constriction is detected to be greater than 15%, the light intensity is automatically reduced by 30% to prevent eye fatigue and photosensitivity. The epilepsy risk blocking unit monitors high-frequency oscillations (>70Hz) in the EEG in real time. Once epileptiform discharges are detected, stimulation is immediately stopped and an alarm is issued. The autonomic nervous system balance monitoring unit assesses autonomic nervous system tension through heart rate variability analysis. When excessive activation of the sympathetic nervous system is detected, stimulation is automatically paused. The blockchain data storage unit stores core user data through a consortium blockchain. Every data modification is traceable and tamper-proof, complying with HIPAA and GDPR regulations. Requirements: The federated learning unit supports "data remains stationary, model moves" in cross-center clinical studies, where hospitals only upload model parameters and do not share raw data, protecting user privacy while achieving joint algorithm optimization; the multimodal fusion safety monitoring unit integrates EEG, fNIRS, and CV data to construct a multi-factor risk identification model, enabling risk warnings and proactive interventions earlier than users' subjective perceptions; the digital biomarker mining engine unit tracks users' multidimensional response data over a long period, using machine learning to mine digital biomarkers highly correlated with clinical efficacy, automatically generating efficacy trend reports and prognostic predictions; the epilepsy risk warning unit monitors high-frequency oscillations >70Hz in EEG in real time. Energy; when the energy exceeds the baseline by 2.5 standard deviations, the system does not immediately shut down, but first smoothly reduces the stimulation intensity by 50% and continues to observe; if the situation worsens, it stops completely to avoid frequent treatment interruptions; the visual fatigue adjustment unit integrates a miniature infrared camera to monitor pupil diameter changes in real time; when pupil constriction >12% is detected, it is determined to be visual fatigue, and the system automatically lowers the light intensity to a comfortable range and records the user's light intensity sensitivity preference; the comfort closed-loop unit assesses the user's autonomic nervous system excitability based on heart rate variability and body movement data; when tension or discomfort occurs, the system can fine-tune the stimulation rhythm to make it gentler; the circadian rhythm protection unit avoids the melatonin sensitive period. To avoid endocrine disruption; the emergency unit is equipped with a physical emergency stop button and supports remote pausing via the APP; the homomorphic encryption unit uses a partially homomorphic encryption algorithm, allowing direct γ power calculation and parameter optimization without decryption, thus preventing privacy leaks during calculations; the access control unit includes: Patients: can only view their own treatment reports and have no parameter modification rights; Nursing staff: can view status and perform emergency pauses, but require authorization to modify parameters; Doctors: have full access, and all operations are logged; Researchers: can only access anonymized data for model training; the data traceability unit adds "unique device identifier + operator ID + timestamp" to the blockchain notation, allowing for the tracing of every data modification.

[0018] The edge computing and communication module includes a local edge computing unit, a wireless communication unit, and a local storage unit.

[0019] Furthermore, the edge computing and communication module uses a lightweight deep learning model to achieve real-time analysis of EEG signals, with a model size ≤2MB and inference time <50ms; the local edge computing unit deploys a lightweight AI model to achieve millisecond-level signal processing and parameter adjustment; the wireless communication unit supports data interaction and remote monitoring with the cloud platform; and the local storage unit encrypts and stores raw data and stimulation parameters with a capacity ≥16GB.

[0020] The user interaction module includes a visual interface unit, a feedback input unit, an emergency stop button unit, and an elderly-friendly interaction unit.

[0021] Furthermore, the visualization interface unit displays real-time stimulation parameters and status indicators; the feedback input unit receives user subjective feeling scores; the emergency stop button unit provides physical emergency intervention methods; and the elderly-friendly interaction unit: the terminal interface adopts "large font + voice navigation", the operation steps are simplified to 3 steps, and "accidental touch protection" is added: the emergency stop is triggered only after pressing it 3 times in a row to avoid accidental operation.

[0022] The head-mounted acquisition and stimulation device includes a wireless dry electrode EEG cap, LED light stimulation glasses, bone conduction headphones with acoustic stimulation unit, and auxiliary sensors, including an EDA sensor, a miniature fNIRS sensor, an IMU sensor, a thermoelectric power generation sensor, and a portable control unit.

[0023] Furthermore, the wireless dry electrode EEG cap adopts a "modular design": the acquisition module and stimulation module are separated and connected magnetically, allowing users to easily disassemble and charge them separately; the LED light stimulation glasses use a 40Hz LED array with a wavelength of 470nm, an adjustable light intensity range of 50-300cd / m², a resolution of 3840×2160, and a refresh rate of 120Hz; the bone conduction earphone has two acoustic stimulation units, embedded behind the ear with a frequency response range of 20-20000Hz and a sound pressure level of 60-85dB; the EDA sensor uses the Swiss Sensirion sensor. The S-EDA-01 module features a 10Hz sampling rate, a measurement range of 0-200μS, and an error of <5%, communicating with the control terminal via an I2C interface. The miniature fNIRS sensor uses the FNIRS-Micro2 module from Taiwan's Superior Technologies, with wavelengths of 730nm / 850nm, a detection depth of 1-2cm, power consumption <10mW, and weight <5g. The IMU sensor is a Bosch BMI160, 6-axis, with a 100Hz sampling rate and an error of <1%, transmitting motion data via an SPI interface for EEG artifact compensation. The thermoelectric sensor uses the NextremeMX-2 module from the USA, measuring 10mm × 10mm, with an output power of 50μW at a temperature difference of 3℃, supplemented by a boost circuit to extend the lithium battery's lifespan. The portable control unit features a quad-core ARM Cortex-A73 processor with a 2.0GHz clock speed; 4GB LPDDR4 memory; 64GB eMMC storage; a 10000mAh lithium battery supporting 8 hours of continuous operation; and interfaces including USB-C, Bluetooth 5.0, and Wi-Fi 6.

[0024] The scene adaptation engine module includes a home scene unit, a hospital scene unit, an outdoor scene unit, an elderly and disabled-friendly unit, and a mobile scene unit.

[0025] Furthermore, system parameters are dynamically adjusted according to usage scenarios: Home scenario unit: Some professional monitoring functions are turned off to prioritize battery life and simplified operation; Hospital scenario unit: The "multi-device synchronization" function is enabled, supporting one host to control 10 stimulation devices and synchronously collect data to the hospital LIS system; Outdoor scenario unit: The "environmental anti-interference mode" is activated, which eliminates outdoor noise through adaptive filtering and adjusts light stimulation parameters to compensate for changes in ambient light intensity, maintaining an EEG signal-to-noise ratio >20dB; Elderly and disability-friendly unit: The headband device adopts a "magnetic, adjustment-free frame" that can be fastened with one hand by caregivers, replacing traditional straps; The APP adds a "remote management terminal for caregivers," which supports real-time viewing of stimulation status and modification of parameters, solving the problem of disabled patients being unable to operate the device; Mobile scenario unit: The portable control terminal adopts a "low-power mode," and with a thermoelectric power generation module, a single charge is extended from 8 hours to 12 hours, supporting multiple stimulations throughout the day.

[0026] The dual-loop closed-loop feedback control module includes an internal fast loop and an external slow loop, a data interaction unit, a decision-making unit, and a log recording unit.

[0027] Furthermore, the system includes: an internal fast loop that monitors the EEG gamma oscillation power in real time and automatically adjusts the intensity and phase difference of the audio-visual stimulation when the power deviates from the target range to ensure optimal neural oscillation entrainment; an external slow loop that comprehensively analyzes user fatigue levels and changes in the theta / gamma coupling index and adjusts the stimulation strategy every 5 minutes; a data interaction unit that uses a CAN bus to transmit data between modules, receiving gamma power data from the EEG module and the initial scheme from the strategy library in real time, and outputting control commands to the parameter adjustment module; a decision-making unit that executes a three-step closed-loop logic: real-time comparison: comparing the current gamma power with the target value in the strategy library; strategy matching: if the gamma power is not up to standard, calling the adjustment rules for the corresponding features in the strategy library; parameter output: generating adjustment commands for light intensity, sound pressure, and phase difference; and a log unit that automatically records the parameter adjustment history, gamma power change curve, and user characteristics for each stimulation, generating a daily treatment report that is synchronized to the cloud.

[0028] The processing algorithm module includes a feature extraction algorithm unit, a deep learning model algorithm unit, a closed-loop control algorithm unit, a predictive adaptive regulation algorithm unit, a γ power calculation algorithm unit, a PID adjustment algorithm unit, an SVM substitution algorithm unit, and an adaptive AI algorithm unit.

[0029] Furthermore, the feature extraction algorithm unit uses short-time Fourier transform to calculate the γ-band power; obtains the instantaneous phase of the θ-band and γ-band through Hilbert transform, and calculates the phase amplitude coupling index; and uses an adaptive threshold method to detect high-frequency oscillation events >70Hz. The deep learning model algorithm unit: constructs a stimulus response prediction model based on a lightweight CNN architecture; input: time-frequency features of a 2-second EEG segment; output: stimulus effect score and parameter adjustment suggestions; model optimization uses knowledge distillation technology to compress the model size to below 2MB, with an inference time <50ms; the closed-loop control algorithm unit: control cycle: 100ms / parameter evaluation, 500ms / parameter adjustment; adjustment strategy: based on the proportional-integral-derivative control principle, combined with reinforcement learning optimization; objective function: maximize γ-oscillation power, maintaining the θ / γ coupling index in the range of 0.3-0.5; the predictive adaptive control algorithm unit uses LSTM or Transformer models to predict user state, and performs feedforward control based on the prediction results; γ power calculation... Method Unit: Sliding window: 200ms, step size 100ms; STFT parameters: Hanning window, frequency resolution 1Hz; γ power extraction: power integral of the 40Hz±2Hz frequency band, unit μV² / Hz; PID adjustment algorithm unit: proportional coefficient: light intensity 0.8, sound pressure 0.5; integral coefficient: 0.2; derivative coefficient: 0.1; target value: user features based on policy library; SVM substitution algorithm unit: input features: γ power change rate, PSQI score, hippocampal atrophy level, encoded as 1-3; kernel function: radial basis function RBF; training dataset: 10,000 clinical samples, cross-validation accuracy 92.3%; Adaptive AI algorithm unit adopts an online federated reinforcement learning framework, with γ power increase rate, HbO2 change, and user comfort score as reward functions, updating the policy every 1 minute, model adaptation accuracy ≥95%, using homomorphic encryption + federated averaging algorithm, in multi-center collaboration only uploading encrypted model parameters, not sharing original data, the model accuracy is improved by ≥15% compared to single center.

[0030] The user terminal devices equipped with application software include smartphones or dedicated tablets, which can be used to bind user identities, schedule stimulation times, view real-time parameters, generate treatment reports, and pause work in an emergency via a user APP.

[0031] The cloud-based central processing unit integrates the NVIDIA Jetson edge computing platform with a computing power of 30 TOPS. It supports TensorRT to accelerate deep learning inference and runs deep learning algorithms to analyze EEG signals in real time. It adopts a federated learning framework to aggregate encrypted model parameters (rather than raw data) from various devices, generate a powerful global optimization model, and regularly distribute updates to each edge device to achieve continuous evolution of the overall intelligence of the system.

[0032] Compared with existing technologies, this 40Hz audio-visual stimulation dynamic optimization system based on EEG closed-loop feedback has the following advantages: Through high-density dry electrodes and an embedded DSP chip, it detects EEG gamma oscillation power in real time at millisecond speeds. Relying on a dual-loop closed-loop feedback control mechanism, it can dynamically adjust the parameters of the 40Hz audio-visual stimulation. This rapid "perception-decision-execution" closed loop ensures that neural stimulation always acts on the user's most sensitive and effective parameter points, improving the efficiency and accuracy of neural modulation.

[0033] The system constructs a multi-layered security protection system that spans the entire process and integrates multiple modalities. From the hardware level, which includes dynamic upper limits on light intensity and epilepsy risk blocking, to the data level, which uses homomorphic encryption and blockchain notarization, it comprehensively protects the physiological and privacy security of users.

[0034] The system has built a mature cloud-edge-device collaborative computing ecosystem. It rationally distributes the computing burden, balancing real-time performance with intelligence. It possesses excellent environmental adaptability and user-friendliness, significantly expanding the system's applicable user base and application scope. Attached Figure Description

[0035] Figure 1 This is the system topology diagram of the present invention.

[0036] Figure 2 This is the upper half of the flowchart of the present invention.

[0037] Figure 3 This is the lower half of the flowchart of the present invention.

[0038] Figure 4 This is a system topology diagram of the EEG gamma oscillation power detection module in this invention.

[0039] Figure 5 This is a system topology diagram of the embedded signal processing and analysis module in this invention.

[0040] Figure 6 This is a system topology diagram of the stimulus parameter generation and adjustment module in this invention.

[0041] Figure 7 This is a system topology diagram of the personalized strategy database module in this invention.

[0042] Figure 8 This is a system topology diagram of the multiple safety protection modules in this invention.

[0043] Figure 9 This is a system topology diagram of the edge computing and communication module in this invention.

[0044] Figure 10 This is a system topology diagram of the user interaction module in this invention.

[0045] Figure 11 This is a system topology diagram of the head-mounted acquisition and stimulation device in this invention.

[0046] Figure 12 This is a system topology diagram of the scene adaptation engine module in this invention.

[0047] Figure 13 This is a system topology diagram of the dual-loop closed-loop feedback control module in this invention.

[0048] Figure 14 This is a system topology diagram of the user terminal equipment in this invention.

[0049] Figure 15 This is a system topology diagram of the cloud-based central processing unit in this invention. Detailed Implementation

[0050] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments.

[0051] like Figures 1-15 As shown, this 40Hz audio-visual stimulation dynamic optimization system based on EEG closed-loop feedback includes an EEG gamma oscillation power detection module, an embedded signal processing and analysis module, a stimulation parameter generation and execution adjustment module, a personalized strategy database module, a multi-layered security protection module, an edge computing and communication module, a user interaction module, a head-mounted acquisition and stimulation device, a scene adaptation engine module, a dual-loop closed-loop feedback control module, a processing algorithm module, a user terminal device with application software, and a cloud-based central processing unit. This system forms a cloud-edge-device collaborative processing architecture that can dynamically and continuously optimize 40Hz stimulation parameters based on real-time EEG feedback, a personalized historical data strategy library, and algorithms.

[0052] In this embodiment, the system is a data-driven, personalized, and adaptive neuromodulation platform. Its core principle lies in constructing a cloud-edge-device collaborative intelligent closed loop with "perception-decision-execution-learning" as its core. By dynamically optimizing the 40Hz audio-visual stimulation parameters, it achieves precise, safe, and efficient enhancement of the user's brain gamma oscillation. Real-time perception and edge computing: Signal acquisition: The user wears a head-mounted device whose dry electrode array focuses on acquiring electroencephalogram (EEG) signals from the occipital lobe, which is most sensitive to visual stimuli. Feature extraction: The acquired raw EEG signal is immediately processed in real time by the device's built-in embedded DSP chip; it quickly calculates the power value of the 40Hz frequency band (gamma oscillation) in a 200ms window; this step is completed locally, ensuring low latency and real-time performance of the system. Intelligent decision-making and parameter generation: state comparison and strategy matching: the calculated γ power value is compared with the data in the personalized strategy database; the database not only stores the user's historical response data, but also includes feature labels such as sleep quality and cognitive function scores, as well as optimization targets defined by the clinical gold dataset (known to be a 25% increase in γ as an effective standard); Multi-module collaborative decision-making: Dual-loop closed-loop control module: determines the direction and magnitude of adjustment based on the difference between real-time feedback and target value; Scene adaptation engine: invokes preset advanced optimization strategies based on the user's specific situation (such as "hippocampal atrophy" or "sleep disorder"); Stimulus parameter dynamic adjustment module: receives instructions and generates specific, executable combinations of stimulation parameters. Precise execution and dynamic adjustment: The head-mounted device receives new instructions from the stimulus parameter generation and execution adjustment module; the sound pressure controller, light intensity modulator, and acoustic-optical phase difference generator on the device work together to precisely output the adjusted 40Hz acoustic-optical stimulus: sound pressure: adjustable in the range of 0-75dB; light intensity: adjustable in the range of 100-1000lux; phase difference: achieving a precise time delay of 0-360° between acoustic and optical stimuli to study or utilize cross-sensory integration effects; Continuous Learning and Security Safeguards: Cloud Learning: All data is uploaded to the cloud central processor via edge computing and communication modules; the cloud utilizes more powerful computing power for in-depth analysis, model training, and algorithm iteration to achieve continuous optimization of system intelligence and other performance aspects; Comprehensive Security Protection: Multiple security protection modules monitor signal quality, physiological responses, and stimulation parameters throughout the process to ensure that any operation is within an absolutely safe range (e.g., sound pressure level does not exceed 75dB), and will immediately stop if any abnormality occurs.

[0053] The EEG gamma oscillation power detection module includes an EEG signal acquisition unit, an auxiliary physiological signal acquisition unit, an EEG artifact compensation unit, an individual structural data import interface unit, an environmental perception unit, an embedded fNIRS monitoring unit, a stress hormone monitoring unit, and a multimodal data fusion engine unit.

[0054] In this embodiment, the EEG signal acquisition unit employs a high-density dry electrode array (≥64 channels), fitted to the occipital lobe (O1, O2, Oz sites) and prefrontal lobe (Fz), acquiring EEG signals from both the prefrontal and occipital lobes. The sampling rate is ≥250Hz, motion artifacts are reduced by 40%, and the signal-to-noise ratio is ≥20dB. Automatic electrode impedance calibration (range: 5kΩ-50kΩ) ensures that an alarm is triggered in case of poor contact, facilitating the acquisition of EEG data. The auxiliary physiological signal acquisition unit includes a pupil monitoring camera, a heart rate variability sensor, a motion state sensor, and a skin... The cerebral blood oxygen sensor and pupil monitoring camera utilize a miniature infrared camera, extracting photoplethysmography (PPG) and heart rate variability (HRV) based on a micro-motion amplification algorithm. It monitors the user's pupil diameter, blink frequency, facial micro-expressions, and head posture in real time, identifying states such as fatigue, discomfort, and inattention, thus acquiring cerebral blood oxygenation signals and visual behavior data. The EEG artifact compensation unit uses motion data from a 6-axis IMU and a Kalman filter algorithm to perform real-time motion compensation on the EEG signal, achieving a γ-power calculation error of ≤5% in mobile scenarios, suitable for light activity. The individual structural data import interface unit supports importing individual T1-weighted magnetic resonance (MRI) images to generate personalized 3D brain models for precise sound field simulation and targeted localization. The environmental perception unit collects environmental parameters such as ambient light intensity and noise level. The embedded fNIRS monitoring unit uses an 8-channel flexible fNIRS sensor (wavelength 730nm / 850nm) attached to the middle frontal gyrus and inferior parietal lobule to collect local brain oxygen saturation (HbO2) and deoxyhemoglobin (HbR) concentration in real time at a sampling rate of 10Hz. This is used to verify the neuronal metabolic activation corresponding to EEG γ oscillations (scientific basis: elevated HbO2 indicates increased local blood flow to neurons, which is positively correlated with γ oscillation power). The stress hormone monitoring unit integrates a wearable interstitial fluid sensor to detect cortisol concentration via electrochemical methods (detection range 0.1-50 nmol / L, error <5%), with a sampling period of 5 minutes, to determine the user's stress state. The multimodal data fusion engine unit uses an attention mechanism neural network to weightedly fuse EEG, fNIRS, and HRV feature vectors to generate a comprehensive state assessment index.

[0055] The embedded signal processing and analysis module includes a preprocessing unit, a feature calculation unit, a feature extraction unit, a data temporary storage unit, and a state evaluation unit.

[0056] In this embodiment, a TITMS320C6748 DSP chip is used, employing a two-stage processing flow of "preprocessing-feature extraction": the preprocessing unit eliminates power frequency interference using a 50 / 60Hz notch filter, removes baseline drift using a 4th-order Butterworth high-pass filter (cutoff frequency 0.5Hz), and removes electrooculography and electromyography artifacts using independent component analysis (ICA) (artifact removal rate ≥85%); the feature calculation unit uses a 200ms sliding window to calculate the γ power in the 40Hz±2Hz frequency band using short-time Fourier transform (STFT), and simultaneously extracts γ power using Hilbert transform. The γ phase amplitude coupling index is calculated (calculation delay < 200ms); the feature extraction unit calculates the γ oscillation (30-100Hz) power, θ band (4-7Hz) energy, θ-γ phase amplitude coupling index, and high-frequency oscillation events > 70Hz in real time; the data storage unit has a built-in 8GB DDR3 cache to store the original EEG data and γ power calculation results in real time, supporting local backtracking analysis; the state assessment unit evaluates the user's cognitive state and stimulus response in real time based on a deep learning model (lightweight CNN), and the model size is compressed to 1 / 20 of the original size to meet embedded requirements.

[0057] The stimulation parameter generation and execution adjustment module includes an acoustic-optical stimulation signal generation unit, a sound field simulation and compensation unit, a dynamic frequency chirping unit, a hierarchical modulation unit, a parameter adjustment unit, a cross-modal synchronous tactile stimulation unit, and a cross-frequency band modulation unit.

[0058] In this embodiment, sound pressure, light intensity, and phase difference are dynamically adjusted based on γ power feedback to output cross-modal synchronized acoustic, optical, and tactile stimulation. The acoustic-optical stimulation signal generation unit uses a quantum dot-enhanced optical system to provide high-purity optical stimulation and a magnetically controlled acoustic array to achieve focused propagation of sound waves. The sound pressure controller is adjustable from 0 to 75 dB with a step accuracy of 1 dB; the light intensity controller is adjustable from 100 to 1000 lux with linear adjustment; the acoustic-optical phase difference generator has a programmable delay of 0 to 360° with an accuracy of ±5°; and the real-time optimized frequency is ≥0.5 Hz (sampling interval ≤2 s). The sound field simulation and compensation unit simulates the path loss of sound waves propagating to the target brain region (such as the hippocampus) based on the user's individual MRI skull model, pre-compensating the output sound waves to ensure that the sound energy reaching the target is accurately controllable. The dynamic frequency chirp unit allows the 40 Hz fundamental frequency to dynamically change within the range of 38-42 Hz (with a change period of 10 seconds) to simulate natural neural networks. Rhythm, avoiding neuronal desensitization—after 2 weeks, the gamma power enhancement rate decreased from 85% to 70%; the graded modulation unit adjusts the theta modulation depth (10%-30%) according to the gamma power enhancement rate (excellent >30%, medium 15%-30%, poor <15%), with a long-term stimulation effect attenuation rate ≤5%; the parameter adjustment unit: real-time adjustment of sound pressure (range: 60-85dB), light intensity (range: 50-300cd / m²), and phase difference (0-360°); the cross-modal synchronous tactile stimulation unit integrates a tactile vibrator (such as a handle or back pad), whose vibration rhythm is precisely synchronized with the sound and light stimulation in time and phase, forming a whole-body, multi-sensory synergistic entrainment to enhance the modulation effect; the cross-frequency modulation unit: when the gamma oscillation power enhancement value is <25%, it automatically generates a 4-7Hz theta band signal, which is superimposed on a 40Hz sound and light carrier in an amplitude modulation manner (theta modulation depth is adjustable: 10%-30%) to promote theta-gamma cross-frequency coupling.

[0059] The personalized strategy database module includes a user feature storage unit, a treatment efficacy prediction unit, a strategy online iteration unit, a strategy mapping unit, a golden data calibration unit, and an alternative algorithm unit.

[0060] In this embodiment, the user feature storage unit stores multi-dimensional user data, including structural imaging features: hippocampal volume (obtained via MRI, categorized into mild / moderate / severe atrophy levels); genetic features: APOE4 genotype (positive / negative); clinical scores: MMSE cognitive score (0-30 points), PSQI sleep quality score (0-21 points, ≥8 points indicates sleep disorder); the efficacy prediction unit employs a lightweight Transformer model combined with transfer learning, inputting user static features and multimodal data from the first three stimuli, and outputting stimulus response probability (0-100%), with a model size ≤3MB and inference time <30ms; the online strategy iteration unit is based on reinforcement learning (PPO algorithm), using "increased γ oscillation power + increased HbO2 + user subjective score" as the reward. The system features an excitation function that automatically updates a personalized strategy library daily (e.g., if a user responds best to 5Hz θ modulation, the system will prioritize that parameter in the future) without manual intervention. The strategy mapping unit establishes a three-level strategy mapping model, automatically matching the initial stimulus scheme based on user characteristics and dynamically correcting it using real-time γ power. The gold data calibration unit incorporates clinically validated gold thresholds (γ oscillation power increase of 25%, morning light intensity safety threshold of 800 lux). When real-time data deviates from the gold threshold by ±15%, forced strategy calibration is triggered. The alternative algorithm unit uses a support vector machine (SVM) classifier to replace traditional threshold judgment (e.g., if γ increase < 25%, SVM predicts a probability of "θ modulation required" > 90%). The classifier is trained on 10,000 clinical samples (accuracy ≥ 92%) and is suitable for nonlinear feature association scenarios.

[0061] The multiple security protection modules include a light intensity dynamic upper limit control unit, an epilepsy risk blocking unit, an autonomic nervous balance monitoring unit, a blockchain data storage unit, a federated learning unit, a multimodal fusion security monitoring unit, a digital biomarker mining engine unit, an epilepsy risk early warning unit, a visual fatigue adjustment unit, a comfort closed-loop unit, a circadian rhythm protection unit, an emergency unit, a homomorphic encryption unit, a permission hierarchical subunit, and a data traceability unit.

[0062] In this embodiment, the dynamic upper limit control unit for light intensity monitors the degree of pupil constriction in real time through infrared eye tracking. When pupil constriction >15% is detected, the light intensity is automatically reduced by 30% to prevent eye fatigue and photosensitivity. The epilepsy risk blocking unit monitors high-frequency oscillations (HFOs) activity >70Hz in the EEG in real time. Once epileptiform discharges (HFOs power exceeding 3 standard deviations from baseline) are detected, stimulation is immediately stopped and an alarm is issued. The autonomic balance monitoring unit assesses autonomic tension through heart rate variability (HRV) analysis. When sympathetic overactivation is detected (LF / HF ratio >3.0), stimulation is automatically paused. The blockchain data storage unit stores core user data (genes, EEG) via a consortium blockchain (nodes include hospitals and research institutions). Every data modification is traceable and tamper-proof, complying with HIPAA and GDPR requirements. The federated learning unit supports "data remains still, model moves" in cross-center clinical research. Each hospital only uploads model parameters and does not share raw data, protecting user privacy while achieving joint algorithm optimization. The multimodal fusion safety monitoring unit integrates EEG (HFOs), fNIRS (abnormal blood flow), and CV (frowning, eye closing) data to construct a multifactor risk identification model, enabling risk identification earlier than the user's subjective perception. Early warning and proactive intervention (e.g., smooth attenuation of stimulus intensity rather than abrupt cessation); the digital biomarker mining engine unit tracks users' multidimensional response data over a long period, using machine learning to mine digital biomarkers highly correlated with clinical efficacy (e.g., MoCA score) (e.g., γ power rise slope, θ / γ coupling stability, etc.), automatically generating efficacy trend reports and prognostic predictions; the epilepsy risk warning unit monitors the energy of high-frequency oscillations (HFOs) >70Hz in the EEG in real time; when the energy exceeds 2.5 standard deviations from the baseline, the system does not immediately shut down, but first smoothly reduces the stimulus intensity by 50% and continues to observe; if the situation worsens, it stops completely. To avoid frequent treatment interruptions, the visual fatigue adjustment unit integrates a miniature infrared camera to monitor pupil diameter changes in real time. When pupil constriction >12% is detected, visual fatigue is identified, and the system automatically adjusts the light intensity to a comfortable range, recording the user's light intensity sensitivity preference. The comfort closed-loop unit assesses the user's autonomic nervous system excitability based on heart rate variability and body movement data. When tension or discomfort occurs, the system can fine-tune the stimulation rhythm to make it gentler. The circadian rhythm protection unit avoids melatonin-sensitive periods to prevent endocrine interference. The emergency unit is equipped with a physical emergency stop button (response time <100ms) and supports remote pause via APP (Bluetooth 5).(0 connections); The homomorphic encryption unit uses a partially homomorphic encryption algorithm (such as the PaiLier algorithm). Data can be directly used for gamma power calculation and parameter optimization while encrypted, without decryption, thus avoiding privacy leaks during the calculation process; Hierarchical access control unit: Patients: can only view their own treatment reports, no parameter modification rights; Nursing staff: can view status and perform emergency pauses, authorization is required to modify parameters; Doctors: full permissions (parameter adjustment, strategy optimization), operations are logged; Researchers: only receive anonymized data (removing identity information) for model training; Data traceability unit: Blockchain notarization adds "unique device identifier + operator ID + timestamp" to trace each data modification, complying with the traceability requirements of the Data Security Law for medical data.

[0063] The edge computing and communication module includes a local edge computing unit, a wireless communication unit, and a local storage unit.

[0064] In this embodiment, the edge computing and communication module uses a lightweight deep learning model to achieve real-time analysis of EEG signals, with a model size ≤2MB and inference time <50ms; the local edge computing unit deploys a lightweight AI model to achieve millisecond-level signal processing and parameter adjustment; the wireless communication unit supports data interaction and remote monitoring with the cloud platform; and the local storage unit encrypts and stores the original data and stimulation parameters with a capacity ≥16GB.

[0065] The user interaction module includes a visual interface unit, a feedback input unit, an emergency stop button unit, and an elderly-friendly interaction unit.

[0066] In this embodiment, the visualization interface unit displays real-time stimulation parameters and status indicators; the feedback input unit receives the user's subjective feeling score; the emergency stop button unit provides physical emergency intervention measures; the elderly-friendly interaction unit: the terminal interface adopts "large font + voice navigation" (supports dialect recognition, such as Cantonese and Sichuan dialect), the operation steps are simplified to 3 steps (power on - confirm identity - start stimulation), and a new "accidental touch protection" is added: the emergency stop is triggered only after pressing it 3 times in a row to avoid accidental operation.

[0067] The head-mounted acquisition and stimulation device includes a wireless dry electrode EEG cap, LED light stimulation glasses, bone conduction headphones, and auxiliary sensors, including EDA sensors, miniature fNIRS sensors, IMU sensors, thermoelectric sensors, and a portable control unit.

[0068] In this embodiment, the wireless dry electrode EEG cap adopts a "modular design": the acquisition module (weight <150g) and the stimulation module (weight <200g) are separated and connected magnetically, making it convenient for users to disassemble and charge them separately; the LED light stimulation glasses use a 40Hz LED array with a wavelength of 470nm, an adjustable light intensity range of 50-300cd / m², a resolution of 3840×2160, and a refresh rate of 120Hz; the bone conduction earphone has two acoustic stimulation units, embedded behind the ear with a frequency response range of 20-20000Hz and a sound pressure level of 60-85dB; the EDA sensor uses the Swiss Sensirion S-EDA-01 module with a sampling rate of 10Hz, a measurement range of 0-200μS, and an error of <5%, communicating with the control terminal via an I2C interface; the miniature fNIRS sensor uses the Taiwan-based Superior Technologies FNIRS-Micro2 module .... The sensor has a length of 730nm / 850nm, a detection depth of 1-2cm (covering the apical cortex), power consumption of <10mW, and a weight of <5g. The IMU sensor is a Bosch BMI160, 6-axis (accelerometer + gyroscope), with a sampling rate of 100Hz and an error of <1%. Motion data is transmitted via an SPI interface for EEG artifact compensation. The thermoelectric sensor uses a US-made NextremeMX-2 module, measuring 10mm × 10mm, with an output power of 50μW at a temperature difference of 3℃. Power is supplied by a boost circuit (3.3V output) to supplement the lithium battery. The portable control unit features a quad-core ARM Cortex-A73 processor with a 2.0GHz clock speed; 4GB LPDDR4 memory; 64GB eMMC storage; a 10000mAh lithium battery supporting 8 hours of continuous operation; and interfaces including USB-C, Bluetooth 5.0, and Wi-Fi 6.

[0069] The scene adaptation engine module includes home scene units, hospital scene units, outdoor scene units, elderly and disabled friendly units, and mobile scene units.

[0070] In this embodiment, system parameters are dynamically adjusted according to the usage scenario: Home scenario unit: Some professional monitoring functions are turned off to prioritize battery life and simplified operation; Hospital scenario unit: The "multi-device synchronization" function is enabled, supporting one host to control 10 stimulation devices and synchronously collect data to the hospital LIS system; Outdoor scenario unit: The "environmental anti-interference mode" is activated, which eliminates outdoor noise (20-8000Hz) through adaptive filtering and adjusts light stimulation parameters to compensate for changes in ambient light intensity, maintaining an EEG signal-to-noise ratio >20dB; Elderly and disabled-friendly unit: The head-mounted device adopts a "magnetic, adjustment-free frame," which can be fastened with one hand by caregivers (wearing completed in 3 seconds), replacing traditional straps; The APP adds a "remote management terminal for caregivers," which supports real-time viewing of stimulation status and modification of parameters (requires doctor authorization), solving the problem of disabled patients being unable to operate; Mobile scenario unit: The portable control terminal adopts a "low-power mode" (power consumption reduced by 30%), combined with a thermoelectric power generation module (utilizing the 3°C temperature difference of the head to supplement 10% of battery life), extending a single charge from 8 hours to 12 hours, supporting multiple stimulations throughout the day.

[0071] The dual-loop closed-loop feedback control module includes an internal fast loop and an external slow loop, a data interaction unit, a decision-making unit, and a log recording unit.

[0072] In this embodiment, the internal fast loop (response time <100ms) monitors the EEG gamma oscillation power in real time. When the power deviates from the target range (0.5-1.2μV²), it automatically adjusts the intensity of the audio-visual stimulation (in increments of ±5%) and the phase difference (±5°) to ensure optimal neural oscillation entrainment effect. The external slow loop (response time >5min) comprehensively analyzes the user's fatigue level (assessed by HRV heart rate variability) and changes in the θ / γ coupling index, adjusting the stimulation strategy every 5 minutes. If an increase in fatigue level is detected (HRV below 20% of baseline), stimulation is automatically paused. The data interaction unit enables data transmission between modules via a CAN bus (transmission rate 500kbps). The system receives gamma power data from the EEG module and the initial scheme from the strategy library in real time, and outputs control commands to the parameter adjustment module; Decision unit: executes "three-step closed-loop logic": real-time comparison: compares the current gamma power with the target value of the strategy library (such as the base value × 1.25); strategy matching: if the gamma power does not meet the standard, calls the adjustment rules of the corresponding feature in the strategy library; parameter output: generates adjustment commands for light intensity, sound pressure, and phase difference (such as light intensity +50 lux, phase difference +30°); Log unit: automatically records the parameter adjustment history, gamma power change curve, and user characteristics for each stimulus, generates a daily treatment report (including gamma enhancement rate and number of parameter adjustments), and supports cloud synchronization (compatible with HL7 medical data standard).

[0073] The processing algorithm module includes a feature extraction algorithm unit, a deep learning model algorithm unit, a closed-loop control algorithm unit, a predictive adaptive regulation algorithm unit, a γ power calculation algorithm unit, a PID adjustment algorithm unit, an SVM substitution algorithm unit, and an adaptive AI algorithm unit.

[0074] In this embodiment, the feature extraction algorithm unit uses Short Time Fourier Transform (STFT) to calculate the power of the γ band (30-100Hz); obtains the instantaneous phase of the θ band (4-7Hz) and γ band through Hilbert Transform, and calculates the phase amplitude coupling index; and uses an adaptive threshold method to detect high-frequency oscillation events >70Hz. The deep learning model algorithm unit constructs a stimulus response prediction model based on a lightweight CNN architecture; input: time-frequency features (256×128) of a 2-second EEG segment; output: stimulus effect score (0-100) and parameter adjustment suggestions; model optimization uses... The distillation technique compresses the model size to below 2MB and the inference time to <50ms; the closed-loop control algorithm unit has the following characteristics: control cycle: 100ms / parameter evaluation, 500ms / parameter adjustment; adjustment strategy: based on the proportional-integral-derivative (PID) control principle, combined with reinforcement learning optimization; objective function: maximize γ oscillation power and maintain the θ / γ coupling exponent in the range of 0.3-0.5; the predictive adaptive regulation algorithm unit uses LSTM or Transformer models to predict user states and performs feedforward control based on the prediction results; and the γ power calculation algorithm unit... Sliding window: 200ms, step size 100ms; STFT parameters: window function is Hanning window, frequency resolution 1Hz; γ power extraction: power integral of the 40Hz±2Hz frequency band, unit μV² / Hz; PID adjustment algorithm unit: proportional coefficient (Kp): light intensity 0.8, sound pressure 0.5; integral coefficient (Ki): 0.2; derivative coefficient (Kd): 0.1; target value: user features based on the policy library, such as the target value for APOE4 positive patients is based on γ power × 1.28; SVM substitution algorithm unit: input features: γ power change rate, PSQI Scoring, hippocampal atrophy level (coded as 1-3); kernel function: radial basis function (RBF); training dataset: 10,000 clinical samples (including 5,000 patients with sleep disorders), cross-validation accuracy of 92.3%; the adaptive AI algorithm unit adopts an online federated reinforcement learning framework, with γ power enhancement rate, HbO2 change, and user comfort score as reward functions, and updates the policy every 1 minute, with an adaptation accuracy of ≥95%. It uses homomorphic encryption + federated averaging algorithm, and only uploads encrypted model parameters when collaborating with multiple centers, without sharing the original data. The model accuracy is improved by ≥15% compared to single-center models.

[0075] User terminal devices equipped with application software include smartphones or dedicated tablets, which can be used to bind user identities, schedule stimulation times, view real-time parameters, generate treatment reports, and pause work in an emergency via a user APP.

[0076] The cloud-based central processing unit integrates the NVIDIA Jetson edge computing platform, with a computing power of 30 TOPS. It supports TensorRT to accelerate deep learning inference and runs deep learning algorithms to analyze EEG signals in real time. It adopts a federated learning framework to aggregate encrypted model parameters (rather than raw data) from various devices, generate a powerful global optimization model, and regularly distribute updates to various edge devices to achieve continuous evolution of the overall system intelligence.

[0077] In summary, by using high-density dry electrodes and an embedded DSP chip, the power of brainwave gamma oscillations can be detected in real time at millisecond speeds. Relying on a dual-loop closed-loop feedback control mechanism, the parameters of 40Hz audio-visual stimulation can be dynamically adjusted. This rapid "perception-decision-execution" closed loop ensures that neural stimulation always acts on the user's most sensitive and effective parameter points, improving the efficiency and accuracy of neural modulation.

[0078] The system constructs a multi-layered security protection system that integrates multiple modalities throughout the entire process. From the dynamic upper limit of light intensity and epilepsy risk blocking at the hardware level, to homomorphic encryption and blockchain evidence storage at the data level, it comprehensively protects the physiological and privacy security of users. More groundbreakingly, its cloud-based central processor adopts a federated learning framework, which aggregates encrypted model parameters from users worldwide without collecting the original data, continuously trains and generates a more powerful globally optimized model, and then distributes it to various terminal devices.

[0079] This system constructs a mature "cloud-edge-device" collaborative computing ecosystem. The head-mounted device on the device side integrates lightweight algorithms for real-time processing and response; the edge terminal handles interaction and relay; and the cloud focuses on massive data analysis and model iteration. This architecture cleverly distributes the computational burden rationally, balancing real-time performance with intelligence. Furthermore, a unique scene adaptation engine module enables the system to intelligently switch working modes and security policies with a single click based on different usage scenarios such as home, hospital, outdoor, and elderly settings, demonstrating excellent environmental adaptability and user-friendliness, significantly broadening the system's applicable user base and application scope.

[0080] Example 1 (APOE4 positive + sleep disorder patient) I. Subjects: Patient information: 68-year-old male, APOE4 genotype positive, PSQI score 14 (moderate sleep disorder), MMSE score 22 (mild cognitive impairment), MRI showed mild hippocampal atrophy (volume reduced by 12% compared to healthy individuals of the same age).

[0081] II. Implementation Steps: 1. System Initialization (10 minutes before treatment): 1. The patient wears a head-mounted device, and the identity is bound to the APP. The system automatically loads the patient's characteristics (APOE4+, PSQI=14, mild hippocampal atrophy). The strategy library matches the initial plan: morning (7:00-7:30) stimulation, light intensity 800 lux, sound pressure 55 dB, phase difference 90°, 30 minutes per session, with 10% depth theta modulation superimposed throughout.

[0082] 2. Real-time closed-loop adjustment (during treatment): 2-minute mark: The EEG module calculates the gamma power as the base value × 1.18 (< the target value 1.28), calls the rules from the policy library, increases the light intensity to 880 lux, and maintains the sound pressure level at 55 dB; 8th minute: The pupil camera detected a contraction rate of 18% (>15%), the security module triggered the first level of protection, and the light intensity dropped to 748 lux (880×0.85). 15 minutes: γ power rises to baseline value × 1.25 (close to target), system maintains parameters, and HRV sensor shows SDNN = 65ms (normal); 25 minutes: γ power drops to base value × 1.22 (< target), triggering the θ modulation depth to increase from 10% to 15%.

[0083] Treatment End and Feedback (5 minutes after treatment): The system generates a report: the average gamma enhancement rate is 22%, parameters were adjusted 3 times, and there were no safety anomalies; the strategy library is updated: based on this data, the initial light intensity for the next treatment is adjusted to 820 lux (optimized initial plan); the report is pushed to the doctor via the APP, and the doctor can adjust the strategy remotely. Implementation effect

[0084] 1. Short-term effects: Two hours after treatment, gamma power remained at 1.15 times the baseline value, and patients' subjective cognitive clarity scores improved by 30%; 2. Long-term effects: After 4 weeks of continuous treatment, the PSQI score dropped to 9 (mild sleep disorder), the MMSE score increased to 24, and the rate of abnormal melatonin secretion decreased to 8%.

[0085] IV. Security Verification Test 1. Epilepsy Risk Test: Input a simulated EEG signal containing 75Hz high-frequency oscillation into the system. If the system detects an abnormality within 280ms, it will immediately stop stimulation and issue an alarm. The response time meets the requirements. 2. Light safety test: 10 healthy volunteers (aged 20-60) participated. When the light intensity was increased from 800 lux to 1000 lux, the pupils of 8 volunteers contracted by more than 15%. The system reduced the light intensity to 850 lux within 2 seconds, and no discomfort was reported. 3. Circadian rhythm interference test: 30 patients with sleep disorders (PSQI≥10) were accidentally stimulated at night (after 21:00). The system automatically limited the light intensity to below 300 lux. The next day, the melatonin test showed an abnormality rate of only 7%, which was significantly lower than the 40% of the existing system.

[0086] Example 2: (Child with Down syndrome) I. Subject: An 8-year-old male with Down syndrome, MMSE-TR score of 45 (mild cognitive impairment), PSQI score of 6 (normal sleep), and sensitivity to bright light (previous squinting response to stimuli).

[0087] II. Implementation Steps: 1. Initialization phase: Special population strategy: Stimulation frequency 38-40Hz, light intensity limit 400lux, sound pressure 50-60dB, intermittent mode (10 minutes of stimulation + 5 minutes of rest). Hardware compatibility: The head-mounted device is designed for children (head circumference 48-52cm), and the light stimulation angle is adjusted to 15° (avoid direct light into the eyes). 2. Real-time adjustment phase: At the 10th minute: EDA monitoring showed that the skin conductivity increased from 10μS to 13μS (+30%, indicating discomfort), and the light intensity decreased from 350 lux to 315 lux within 1 second, while the sound pressure decreased from 55dB to 50dB; 15 minutes (rest period): fNIRS showed that parietal HbO2 increased from 65% to 68% (metabolic improvement), and the current parameter was automatically maintained for the next stimulation; 25 minutes: γ power increased by 28% (medium efficiency), triggering α band preprocessing for 3 minutes (8Hz, light intensity 250 lux); III. Feedback on Results: Short-term: One hour after stimulation, gamma power remained at a 25% increase, with no squinting reaction; Long-term: After two consecutive weeks of stimulation, the gamma power enhancement rate remained at 92% (without significant fatigue), and the MMSE-TR score increased to 48.

[0088] IV. Security Verification Test: Second-level stress response test: 10 healthy volunteers simulated discomfort (sudden shortness of breath). The EDA of all volunteers increased by more than 20% within 1 second, and the system adjusted the parameters within 1.2 seconds. The response time met the requirements. Child stimulation safety test: 20 children aged 3-12 participated, using the 38-40Hz intermittent mode, no cases of eye discomfort occurred, and the intolerance rate was 0% (original plan was 35%). Homomorphic encryption test: Encrypted calculations were performed on data from 100 patients. The gamma power calculation error was less than 3%, consistent with the plaintext calculation results, demonstrating effective privacy protection.

[0089] In summary, this system can improve the digital marker of "hippocampal gamma power response speed" by up to 40%, and the MoCA score will steadily improve after 3 months.

[0090] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A 40Hz audio-visual stimulation dynamic optimization system based on EEG closed-loop feedback, characterized in that, It includes an EEG gamma oscillation power detection module, an embedded signal processing and analysis module, a stimulation parameter generation and execution adjustment module, a personalized strategy database module, a multi-layer security protection module, an edge computing and communication module, a user interaction module, a head-mounted acquisition and stimulation device, a scene adaptation engine module, a dual-loop closed-loop feedback control module, a processing algorithm module, a user terminal device with application software, and a cloud-based central processing unit. The EEG gamma oscillation power detection module collects and accurately calculates and analyzes the gamma band oscillation characteristics in the EEG signal in real time. The embedded signal processing and analysis module immediately processes the acquired raw EEG signals in real time and calculates the power value of the gamma oscillation; the stimulation parameter generation and execution adjustment module receives instructions and generates specific, executable combinations of stimulation parameters; the personalized strategy database module performs state comparison and strategy matching on the calculated gamma power value; a multi-layered safety protection module monitors signal quality, physiological responses, and stimulation parameters throughout the process to ensure that all operations are within an absolutely safe range; the edge computing and communication module uses a lightweight deep learning model to achieve real-time analysis of EEG signals and uploads all data to the cloud central processing unit; the user interaction module displays real-time stimulation parameters and status indicators simultaneously. The system provides feedback input intervention; a head-mounted acquisition and stimulation device is worn by the user to receive new instructions from the stimulation parameter generation and adjustment module; the scene adaptation engine module calls preset advanced optimization strategies according to the user's specific situation; the dual-loop closed-loop feedback control module determines the adjustment direction and magnitude based on the difference between real-time feedback and target values; the processing algorithm module provides an algorithm database; the user terminal device performs user identity binding, reservation, parameter viewing, report generation, and emergency pause; the cloud central processor receives all data uploaded by the edge computing and communication module, and uses powerful computing power for deep analysis, model training, and algorithm iteration to achieve continuous optimization of system intelligence and other performance aspects. The system forms a cloud-edge-device collaborative processing architecture that can dynamically and continuously optimize the 40Hz stimulation parameters based on real-time EEG feedback, a personalized historical data strategy library, and algorithms.

2. The 40Hz audio-visual stimulation dynamic optimization system based on EEG closed-loop feedback according to claim 1, characterized in that, The EEG gamma oscillation power detection module includes an EEG signal acquisition unit, an auxiliary physiological signal acquisition unit, an EEG artifact compensation unit, an individual structural data import interface unit, an environmental perception unit, an embedded fNIRS monitoring unit, a stress hormone monitoring unit, and a multimodal data fusion engine unit. The EEG signal acquisition unit uses a high-density dry electrode array, which is attached to the occipital lobe and the prefrontal lobe, to acquire EEG signals. The auxiliary physiological signal acquisition unit includes a pupil monitoring camera, a heart rate variability sensor, a motion state sensor, and a cortisol sensor, which monitor the user in real time, identify their status, and acquire cerebral blood oxygenation signals and visual behavior data. The EEG artifact compensation unit uses motion data from a 6-axis IMU and a Kalman filter algorithm to perform real-time motion compensation on EEG signals, adapting to mild activity scenarios. The individual structural data import interface unit supports importing individual MRI images of the user to generate a 3D brain model for precise sound field simulation and targeted positioning. The environmental perception unit collects environmental parameters. The embedded fNIRS monitoring unit uses an 8-channel flexible fNIRS sensor to collect local cerebral blood oxygen saturation and deoxyhemoglobin concentration in real time. The stress hormone monitoring unit integrates a wearable interstitial fluid sensor to determine the user's stress state. The multimodal data fusion engine unit uses an attention mechanism neural network to generate a comprehensive state assessment index. The embedded signal processing and analysis module includes a preprocessing unit, a feature calculation unit, a feature extraction unit, a data storage unit, and a state evaluation unit. The preprocessing unit eliminates power frequency interference using a 50 / 60Hz notch filter, removes baseline drift using a 4th-order Butterworth high-pass filter, and removes electrooculography and electromyography artifacts using independent component analysis. The feature calculation unit uses a 200ms sliding window and employs short-time Fourier transform to calculate the γ power in the 40Hz±2Hz frequency band, while simultaneously extracting the γ phase amplitude coupling index using Hilbert transform. The feature extraction unit calculates the γ oscillation power, θ band energy, θ-γ phase amplitude coupling index, and high-frequency oscillation events >70Hz in real time. The data storage unit has a built-in 8GB DDR3 cache to store raw EEG data and gamma power calculation results in real time; the state assessment unit evaluates the user's cognitive state and stimulus response in real time based on a deep learning model.

3. The 40Hz audio-visual stimulation dynamic optimization system based on EEG closed-loop feedback according to claim 2, characterized in that, The stimulation parameter generation and execution adjustment module includes an acoustic-optical stimulation signal generation unit, a sound field simulation and compensation unit, a dynamic frequency chirp unit, a hierarchical modulation unit, a parameter adjustment unit, a cross-modal synchronous tactile stimulation unit, and a cross-frequency modulation unit. The acoustic-optical stimulation signal generation unit uses a quantum dot-enhanced optical system to provide high-purity optical stimulation and a magnetically controlled acoustic array to achieve focused propagation of sound waves. The sound field simulation and compensation unit simulates the path loss of sound waves propagating to the target brain region based on the user's individual MRI skull model and pre-compensates the output sound waves. The dynamic frequency chirp unit can dynamically change within the range of 38-42Hz with a 40Hz fundamental frequency to simulate natural neural rhythms and avoid neuronal desensitization. The graded modulation unit adjusts the θ modulation depth according to the γ power boost rate; the parameter adjustment unit adjusts the sound pressure, light intensity and phase difference in real time; the cross-modal synchronous tactile stimulation unit integrates a tactile vibrator, and the vibration rhythm is precisely synchronized with the acoustic and optical stimulation in time and phase; when the γ oscillation power boost value is <25%, the cross-band modulation unit automatically generates a 4-7Hz θ band signal, which is superimposed on a 40Hz acoustic and optical carrier in an amplitude modulation manner to promote θ-γ cross-band coupling; The personalized strategy database module includes a user feature storage unit, an efficacy prediction unit, a strategy online iteration unit, a strategy mapping unit, a golden data calibration unit, and an alternative algorithm unit. The user feature storage unit stores multi-dimensional user data. The efficacy prediction unit uses a lightweight Transformer combined with transfer learning, taking user features as input and outputting stimulus response probabilities. The strategy online iteration unit automatically updates the personalized strategy library daily. The strategy mapping unit establishes a three-level strategy mapping model, combined with real-time dynamic correction of gamma power. The golden data calibration unit incorporates clinically validated golden thresholds to trigger mandatory strategy calibration. The alternative algorithm unit uses a support vector machine classifier for threshold determination.

4. The 40Hz audio-visual stimulation dynamic optimization system based on EEG closed-loop feedback according to claim 3, characterized in that, The multiple security protection modules include a dynamic upper limit control unit for light intensity, an epilepsy risk blocking unit, an autonomic nervous system balance monitoring unit, a blockchain data storage unit, a federated learning unit, a multimodal fusion security monitoring unit, a digital biomarker mining engine unit, an epilepsy risk early warning unit, a visual fatigue adjustment unit, a comfort closed-loop unit, a circadian rhythm protection unit, an emergency unit, a homomorphic encryption unit, a permission-level subunit, and a data traceability unit. The dynamic upper limit control unit for light intensity monitors the degree of pupil constriction in real time through infrared eye tracking to prevent visual fatigue and photosensitivity. The epilepsy risk blocking unit monitors high-frequency oscillatory activity in real time, and if epileptiform discharges are detected, it immediately stops stimulation and issues an alarm. Autonomy The neural balance monitoring unit assesses autonomic nerve tension through heart rate variability analysis and automatically pauses stimulation when sympathetic nerve overactivation is detected; the blockchain data storage unit stores user core data through a consortium blockchain, leaving a trace of every data modification; the federated learning unit supports cross-center clinical research, where hospitals only upload model parameters without sharing raw data, protecting user privacy while achieving joint optimization of algorithms. The multimodal fusion safety monitoring unit constructs a multi-factor risk identification model to achieve risk early warning and proactive intervention; The digital biomarker mining engine unit tracks users' multidimensional response data over a long period, using machine learning to mine digital biomarkers highly correlated with clinical efficacy, and automatically generates efficacy trend reports and prognostic predictions; the epilepsy risk warning unit monitors high-frequency oscillation energy in real time. When the energy exceeds 2.5 standard deviations from the baseline, the system does not immediately shut down, but first smoothly reduces the stimulation intensity by 50% and continues to observe; if the situation worsens, it stops completely to avoid frequent treatment interruptions; the visual fatigue adjustment unit integrates a miniature infrared camera to monitor pupil diameter changes in real time; it automatically adjusts the light intensity to a comfortable range and records the user's light intensity sensitivity preference; the comfort closed-loop unit assesses the user's autonomic nervous system excitability based on heart rate variability and body movement data; the circadian rhythm protection unit avoids melatonin sensitive periods to prevent endocrine interference; the emergency unit is equipped with a physical emergency stop button and supports remote pausing via the APP; the homomorphic encryption unit uses a partially homomorphic encryption algorithm, allowing direct γ power calculation and parameter optimization operations on encrypted data; Access control hierarchy unit: Patients: Can only view their own treatment reports, have no right to modify parameters; Nursing staff: Can view status and perform emergency pause, require authorization to modify parameters; Doctor: Full access, all operations are recorded; Researchers: Only anonymized data is obtained for model training; Data traceability unit: Blockchain evidence is used to ensure traceability of every data modification.

5. The 40Hz audio-visual stimulation dynamic optimization system based on EEG closed-loop feedback according to claim 4, characterized in that, The edge computing and communication module includes a local edge computing unit, a wireless communication unit, and a local storage unit. The edge computing and communication module employs a lightweight deep learning model to achieve real-time analysis of EEG signals. The local edge computing unit deploys a lightweight AI model to achieve millisecond-level signal processing and parameter adjustment. The wireless communication unit supports data interaction and remote monitoring with the cloud platform. The local storage unit encrypts and stores raw data and stimulation parameters. The user interaction module includes a visual interface unit, a feedback input unit, an emergency stop button unit, and an elderly-friendly interaction unit; the visual interface unit displays real-time stimulation parameters and status indicators; the feedback input unit receives user subjective feeling scores. The emergency stop button unit provides a physical emergency intervention method; the elderly-friendly interaction unit: the terminal interface uses large fonts and voice navigation.

6. The 40Hz audio-visual stimulation dynamic optimization system based on EEG closed-loop feedback according to claim 5, characterized in that, The head-mounted acquisition and stimulation device includes a wireless dry electrode EEG cap, LED light stimulation glasses, a bone conduction earphone acoustic stimulation unit, and auxiliary sensors. The auxiliary sensors include an EDA sensor, a miniature fNIRS sensor, an IMU sensor, a thermoelectric sensor, and a portable control unit. The wireless dry electrode EEG cap is modular: the acquisition module and stimulation module are separate and connected magnetically, allowing for easy disassembly and charging by the user. The LED light stimulation glasses use a 40Hz LED array. The bone conduction earphone acoustic stimulation unit consists of two units, embedded behind the ear within the frequency response range. 20-20000Hz, sound pressure level 60-85dB; EDA sensor sampling rate 10Hz, measurement range 0-200μS; miniature fNIRS sensor wavelength 730nm / 850nm, detection depth 1-2cm, power consumption <10mW, weight <5g; IMU sensor is 6-axis, sampling rate 100Hz, error <1%; thermoelectric sensor size 10mm×10mm, output power 50μW at a temperature difference of 3℃; portable control unit processor is quad-core ARM Cortex-A73, main frequency 2.0GHz; Memory: 4GB LPDDR4, Storage: 64GB eMMC; Power supply: 10000mAh lithium battery, supporting 8 hours of continuous operation; Interfaces: USB-C, Bluetooth 5.0, Wi-Fi 6; The scenario adaptation engine module includes a home scenario unit, a hospital scenario unit, an outdoor scenario unit, an elderly and disabled-friendly unit, and a mobile scenario unit. The system parameters are dynamically adjusted according to the usage scenario: the home scenario unit disables some professional monitoring functions, prioritizing battery life and simplified operation; the hospital scenario unit enables multi-device synchronization, supporting one host to control 10 stimulation devices and synchronously collecting data to the hospital's LIS system; the outdoor scenario unit activates an environmental anti-interference mode, eliminating outdoor noise through adaptive filtering and adjusting light stimulation parameters to compensate for changes in ambient light intensity, maintaining an EEG signal-to-noise ratio >20dB; the elderly and disabled-friendly unit's headband uses a magnetic, adjustment-free frame, allowing caregivers to fasten it with one hand; the mobile scenario unit's portable control terminal uses a low-power mode, coupled with a thermoelectric power generation module, extending a single charge from 8 hours to 12 hours, supporting multiple stimulations throughout the day.

7. The 40Hz audio-visual stimulation dynamic optimization system based on EEG closed-loop feedback according to claim 6, characterized in that, The dual-loop closed-loop feedback control module includes an internal fast loop and an external slow loop, a data interaction unit, a decision-making unit, and a log recording unit; The internal fast loop monitors the EEG gamma oscillation power in real time. When the power deviates from the target range, it automatically adjusts the intensity and phase difference of the audio-visual stimulation to ensure optimal neural oscillation entrainment. The external slow loop comprehensively analyzes the user's fatigue level and changes in the theta / gamma coupling index, adjusting the stimulation strategy every 5 minutes. The data interaction unit realizes inter-module data transmission through the CAN bus, receiving gamma power data and the initial scheme of the strategy library in real time, and outputting control commands to the parameter adjustment module. The decision unit executes a three-step closed-loop logic, comparing the current gamma power with the target value of the strategy library in real time. The strategy matching unit calls the adjustment rules of the corresponding feature in the strategy library if the gamma power does not meet the target. The parameter output generates adjustment commands for light intensity, sound pressure, and phase difference. The log unit automatically records the parameter adjustment history, gamma power change curve, and user characteristics for each stimulation, generating a daily treatment report that supports cloud synchronization.

8. The 40Hz audio-visual stimulation dynamic optimization system based on EEG closed-loop feedback according to claim 7, wherein the processing algorithm module includes a feature extraction algorithm unit, a deep learning model algorithm unit, a closed-loop control algorithm unit, a predictive adaptive regulation algorithm unit, a γ power calculation algorithm unit, a PID adjustment algorithm unit, an SVM substitution algorithm unit, and an adaptive AI algorithm unit; The feature extraction algorithm unit uses short-time Fourier transform to calculate the γ-band power, obtains the instantaneous phase of the θ-band and γ-band through Hilbert transform, calculates the phase amplitude coupling index, and uses an adaptive threshold method to detect high-frequency oscillation events >70Hz; the deep learning model algorithm unit constructs a stimulus-response prediction model based on a lightweight CNN architecture; input: time-frequency features of a 2-second EEG segment; Output: Stimulus effect score and parameter adjustment suggestions; model optimization uses knowledge distillation technology to compress the model size to below 2MB, with inference time <50ms; control cycle of the closed-loop control algorithm unit: 100ms / parameter evaluation, 500ms / parameter adjustment; adjustment strategy: based on the proportional-integral-derivative control principle, combined with reinforcement learning optimization; objective function: maximize γ oscillation power, maintain θ / γ coupling index in the range of 0.3-0.5; predictive adaptive control algorithm unit uses LSTM or Transformer model to predict user state, and performs feedforward control based on the prediction results; sliding window of the γ power calculation algorithm unit: 200ms, step size 100ms; STFT parameters: window function is Hanning window, frequency resolution 1Hz; γ power extraction: power integral of the 40Hz±2Hz frequency band, unit μV² / Hz; proportional coefficient of the PID adjustment algorithm unit: light intensity 0.8, sound pressure 0.5; integral Coefficient: 0.2; Differential coefficient: 0.1; Target values: User features based on the policy library; Input features of the SVM alternative algorithm unit: not limited to γ ​​power change rate, PSQI score, hippocampal atrophy level; Kernel function: radial basis function; The adaptive AI algorithm unit adopts an online federated reinforcement learning framework, with γ power increase rate, HbO2 change, and user comfort score as reward functions. The policy is updated every 1 minute, and the model adaptation accuracy is ≥95%. Homomorphic encryption + federated averaging algorithm is used. When collaborating with multiple centers, only encrypted model parameters are uploaded, and the original data is not shared. The model accuracy is improved by ≥15% compared to single center.

9. A 40Hz audio-visual stimulation dynamic optimization system based on EEG closed-loop feedback according to claim 8, characterized in that, The user terminal devices equipped with application software include smartphones or dedicated tablets, which can be used to bind user identities, schedule stimulation times, view real-time parameters, generate treatment reports, and pause work in an emergency via a user APP.

10. A 40Hz audio-visual stimulation dynamic optimization system based on EEG closed-loop feedback according to claim 9, characterized in that, The cloud-based central processing unit integrates the NVIDIA Jetson edge computing platform, runs deep learning algorithms to analyze EEG signals in real time, aggregates encrypted model parameters from various devices, generates a powerful global optimization model, and regularly distributes updates to each edge device to achieve continuous evolution of the system's overall intelligence.

Images