EEG-triggered adaptive acoustic drive system for sleep enhancement and tinnitus suppression

Abstract

This invention discloses an EEG-triggered adaptive acoustic drive system for sleep enhancement and tinnitus suppression, relating to the fields of brain-computer interface and neuroacoustic modulation technology. It includes: an EEG signal acquisition and preprocessing module, a sleep stage and tinnitus neural feature joint decoding module, a stage adaptation strategy matching module, a dynamic parameter optimization module, and an acoustic drive output module. This invention constructs an adaptive mapping framework between sleep state and tinnitus suppression strategy by jointly decoding sleep stage and tinnitus neural features using EEG signals. This enables automatic switching of differentiated acoustic drive schemes under different sleep stages, adapting to the differences in neural plasticity levels and auditory perception thresholds at each stage. It dynamically optimizes acoustic parameters and sets a smooth transition mechanism through a reinforcement learning algorithm driven by EEG feedback, continuously suppressing tinnitus while avoiding stimulating sound interference with normal sleep structure. It integrates dual-target regulation of sleep enhancement and tinnitus suppression based on a closed-loop brain-computer interface pathway.

Description

Technical Field

[0001] This invention relates to the field of brain-computer interface and neuroacoustic modulation technology, specifically to an EEG-triggered adaptive acoustic drive system for sleep enhancement and tinnitus suppression. Background Technology

[0002] Sleep is a core physiological process for maintaining homeostasis, cognitive function, and emotional regulation. Sleep enhancement refers to optimizing sleep structure, increasing the proportion of deep sleep, and improving sleep continuity through neural modulation and physical intervention. This has significant health value and clinical significance for promoting nerve repair, improving daytime cognitive function, and maintaining overall metabolic homeostasis. Tinnitus refers to abnormal auditory symptoms subjectively perceived by an individual in the absence of external sound sources. Chronic tinnitus is often associated with abnormal synchronous oscillations of the auditory cortex and dysfunction of the thalamic-cortical circuits. Long-term tinnitus can severely interfere with a patient's sleep, attention, and emotional state, significantly reducing their quality of life. Tinnitus suppression aims to reduce or eliminate the subjective perception of tinnitus through acoustic intervention and neural modulation, and is an important research direction in the fields of otology and neuromodulation.

[0003] EEG-triggered acoustic drive systems represent a typical application of brain-computer interface (BCI) technology in health intervention. By collecting scalp EEG signals and decoding brain neural states, these systems trigger and regulate the output of acoustic stimuli, achieving a closed-loop interactive regulation of brain neural activity and external acoustic input. Compared to traditional open-loop acoustic devices, EEG-triggered drive systems can achieve precise and personalized stimulus output based on real-time brain states, significantly improving the targeting and effectiveness of acoustic interventions. They possess significant technological advantages in scenarios such as sleep regulation and hearing disease intervention. Sleep disorders and tinnitus have a clear bidirectional pathological coupling relationship. Tinnitus symptoms can induce sleep fragmentation and difficulty falling asleep, while decreased sleep quality further increases the excitability of the auditory cortex and intensifies the subjective perception of tinnitus. Therefore, integrating sleep enhancement and tinnitus suppression into the same EEG-triggered acoustic drive system, relying on BCI technology to achieve dual-target synergistic regulation, is a key technological path to break the vicious cycle between the two and improve the overall intervention effect.

[0004] However, in existing technologies, acoustic modulation devices for sleep enhancement and acoustic therapy devices for tinnitus suppression are mostly designed independently, without establishing a synergistic regulation mechanism to address the bidirectional pathological coupling between the two. Some integrated intervention devices use fixed acoustic therapy plans, which cannot dynamically switch tinnitus suppression strategies based on sleep staging results decoded in real time by EEG. This makes it difficult to adapt to the differences in neural plasticity levels and auditory perception thresholds at different sleep stages, easily leading to problems such as insufficient tinnitus suppression intensity during wakefulness and stimulation sound interference with normal sleep structure during deep sleep. At the same time, most existing systems are open-loop or semi-closed-loop designs, with insufficient interaction depth between EEG signals and acoustic outputs. They cannot dynamically optimize stimulation parameters by combining the intensity of tinnitus neural activity and sleep stability indicators from EEG feedback, making it difficult to ensure and improve sleep quality while continuously suppressing tinnitus. This restricts the clinical intervention effect and the long-term user experience. Therefore, developing an EEG-triggered adaptive acoustic drive system for sleep enhancement and tinnitus suppression is of great significance. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an EEG-triggered adaptive acoustic drive system for sleep enhancement and tinnitus suppression. It can solve the problems of existing devices being unable to coordinate the bidirectional pathological relationship between sleep and tinnitus, fixed acoustic schemes being difficult to adapt to the differences in neural states throughout the sleep cycle, and insufficient depth of closed-loop interaction, thereby achieving the dual goals of continuous tinnitus suppression and enhanced sleep quality.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an EEG-triggered adaptive acoustic drive system for sleep enhancement and tinnitus suppression, the system comprising: an EEG signal acquisition and preprocessing module, a sleep staging and tinnitus neural feature joint decoding module, a staging adaptation strategy matching module, a dynamic parameter optimization module, and an acoustic drive output module;

[0007] The EEG signal acquisition and preprocessing module is used to acquire multi-channel scalp EEG signals, complete signal amplification, filtering and artifact removal processing, and output standardized EEG signals to the sleep stage and tinnitus nerve feature joint decoding module.

[0008] The sleep staging and tinnitus neural feature joint decoding module is used to extract sleep feature waveforms and abnormal oscillation features of the auditory cortex from the electroencephalogram (EEG) signal, and output sleep staging labels and tinnitus neural activity intensity indicators to the staging adaptation strategy matching module and the dynamic parameter optimization module.

[0009] The staged adaptation strategy matching module is used to match the corresponding acoustic intervention strategy according to the sleep stage label and output it to the dynamic parameter optimization module;

[0010] The dynamic parameter optimization module is used to adjust the parameters of the acoustic intervention strategy in conjunction with the tinnitus nerve activity intensity index and output them to the acoustic drive output module.

[0011] The acoustic drive output module is used to generate and play acoustic stimulus signals based on the received parameters.

[0012] Furthermore, the sleep stage and tinnitus neural feature joint decoding module performs the following operations when performing feature decoding and state discrimination:

[0013] The preprocessed multichannel EEG signal was decomposed into frequency bands to extract the energy distribution data of five frequency bands: delta, theta, alpha, beta, and gamma. At the same time, the time domain parameters of three characteristic waveforms, namely slow waves, sleep spindles, and K-complexes, were identified.

[0014] Based on the frequency of frequency band energy proportion and the frequency of occurrence of characteristic waveforms, four sleep stage labels are output according to the sleep stage discrimination rules: wakefulness, light sleep, deep sleep, and REM sleep.

[0015] Signals from the corresponding EEG channels in the temporal lobe were selected, and the phase synchronization index and energy ratio of the gamma band were calculated to generate an index of tinnitus neural activity intensity.

[0016] Based on the degree of fluctuation in sleep stages and the frequency band energy variation coefficient within a continuous time window, a sleep stability index is calculated and generated.

[0017] The sleep stage labels, tinnitus neural activity intensity indicators, and sleep stability indicators are simultaneously output to the downstream modules.

[0018] Furthermore, the sleep staging and tinnitus neural feature joint decoding module uses a sliding time window for real-time decoding. The window length and step size of the sliding time window are fixed values, and feature extraction and state discrimination are completed once within each time window. When calculating the gamma band phase synchronization index, the sleep staging and tinnitus neural feature joint decoding module selects two EEG channels corresponding to the bilateral temporal lobes as calculation channels, uses Hilbert transform to extract the instantaneous phase of the single-channel signal, and calculates the quantitative index using the weighted phase synchronization index formula, which is expressed as: ,in The weighted phase synchronization index, used as a quantitative indicator of the intensity of tinnitus neural activity, ranges from 0 to 1 and is dimensionless. This refers to the number of EEG signal sampling points within a single sliding time window. for The signal-to-noise weighting coefficients at each time step, dimensionless. , The gamma band signals of the two EEG channels corresponding to the bilateral temporal lobes are respectively located in... The instantaneous phase at time t, in radians, and the signal-to-noise weighting coefficient. The signal-to-noise ratio (SNR) of the preprocessed EEG signal is calculated. A higher SNR corresponds to a larger weighting coefficient. The weighting coefficients for all moments within a single time window are normalized, and the sum of the coefficients is proportional to the number of sampling points. Similarly, when the sleep staging and tinnitus neural feature joint decoding module calculates the sleep stability index, it counts the number of times the sleep staging label is switched within a set time period and the standard deviation of the energy of each frequency band, and obtains the quantitative value of sleep stability through weighted calculation.

[0019] Furthermore, the EEG signal acquisition and preprocessing module includes a multi-channel dry electrode acquisition unit, a signal amplification unit, and a digital filtering unit. The electrode points of the multi-channel dry electrode acquisition unit cover the prefrontal cortex, central region, parietal lobe, and bilateral temporal lobes, corresponding to the acquisition of EEG signals related to the whole brain and auditory cortex. The signal amplification unit performs fixed-multiplier programmable amplification on the raw EEG signals, amplifying the microvolt-level raw EEG signals to a voltage range suitable for analog-to-digital conversion. The digital filtering unit uses bandpass filtering to retain EEG signals in a specified frequency band and uses independent component analysis to separate and remove EEG artifacts and EMG artifacts. The EEG signal acquisition and preprocessing module outputs standardized EEG signals with a fixed sampling rate.

[0020] Furthermore, the phased adaptation strategy matching module has a built-in acoustic intervention strategy library, which includes the following four groups of basic acoustic intervention strategies.

[0021] The first group corresponds to the broadband acclimatization sound strategy during the waking period. The strategy parameters include the spectral range of broadband noise and the temporal fluctuation modulation mode.

[0022] The second group corresponds to the narrowband phase cancellation acoustic strategy for the light sleep period. The strategy parameters include the center frequency and phase offset of the narrowband signal.

[0023] The third group corresponds to the slow-wave modulation masking sound strategy during deep sleep. The strategy parameters include the intensity of the background masking sound and the slow-wave frequency modulation depth.

[0024] The fourth group corresponds to the low-intensity white noise strategy during rapid eye movement (REM) phase. The strategy parameters include the output intensity and spectral flatness of the white noise.

[0025] The staged adaptation strategy matching module calls the basic acoustic intervention strategy for the corresponding group and outputs a complete parameter set based on the received sleep stage tags.

[0026] Furthermore, the dynamic parameter optimization module performs the following operations when adjusting the acoustic intervention strategy parameters:

[0027] It receives the basic acoustic intervention strategy output by the staged adaptation strategy matching module, and at the same time receives the tinnitus nerve activity intensity index and sleep stability index output by the sleep staged and tinnitus nerve feature joint decoding module.

[0028] The two types of indicators are input into the built-in pre-trained reinforcement learning model to calculate the adjustment amount of the acoustic parameters;

[0029] It detects whether there is a difference between the current sleep stage and the previous sleep stage, and generates a parameter smooth transition curve when the stage changes;

[0030] The adjusted acoustic intervention strategy parameters and transition curves are integrated into a real-time output command and sent to the acoustic drive output module.

[0031] Furthermore, the reinforcement learning model built into the dynamic parameter optimization module adopts a deep Q-network architecture. The state space of the reinforcement learning model includes three types of input features: sleep stage labels, tinnitus neural activity intensity indicators, and sleep stability indicators. The action space of the reinforcement learning model includes three types of output parameters: acoustic intensity adjustment, spectral range offset, and temporal modulation depth adjustment. The immediate reward value of the reinforcement learning model is calculated using a dual-objective weighted reward formula, which is expressed as follows: ,in The immediate reward value for a single round of intervention. , These are the tinnitus nerve activity intensity indices before and after acoustic stimulation intervention. , These are sleep stability indicators before and after acoustic stimulation intervention. These are the weighting coefficients for the tinnitus suppression dimension. These are the weighting coefficients for the sleep stability dimension; both are dimensionless constants. and The initial values ​​are determined by prior experience in the field of clinical neuromodulation, and the initial state is... and The sum of the values ​​is 1. In subsequent use, the closed-loop efficacy assessment and strategy iteration module will adaptively adjust the values ​​based on the long-term intervention effect. The training process of the reinforcement learning model adopts a combination of offline pre-training and online fine-tuning. Offline pre-training is completed using public sleep datasets and tinnitus nerve data. Online fine-tuning is completed based on real-time EEG feedback data of a single user. The reinforcement learning model completes parameter inference once within each decoding time window and outputs the corresponding acoustic parameter adjustment values.

[0032] Furthermore, the acoustic drive output module includes a digital-to-analog converter, a power amplifier, and a dual-path sound generation unit. The digital-to-analog converter receives acoustic parameter commands in digital format and converts them into analog time-domain audio signals. The power amplifier adjusts the power gain of the analog audio signals and outputs audio signals that meet the drive level requirements. The dual-path sound generation unit includes two types of sound generation devices: air-conduction headphones and bone-conduction vibrators. It supports two working modes: single-path output and dual-path synchronous output. It outputs corresponding acoustic stimuli according to the received audio signals. The output parameters of the dual-path sound generation unit correspond one-to-one with the real-time commands output by the dynamic parameter optimization module.

[0033] Furthermore, the system also includes a closed-loop efficacy assessment and strategy iteration module. This module incorporates a strategy update model, while the staged adaptation strategy matching module incorporates an acoustic intervention strategy library. The dynamic parameter optimization module incorporates a reinforcement learning model. The input of the closed-loop efficacy assessment and strategy iteration module is connected to the output of the sleep stage and tinnitus neural feature joint decoding module. The output of the closed-loop efficacy assessment and strategy iteration module is connected to the inputs of both the staged adaptation strategy matching module and the dynamic parameter optimization module. The closed-loop efficacy assessment and strategy iteration module collects EEG feedback data after a set duration of acoustic stimulation output, extracts the changes in the phase synchronization index of the auditory cortex gamma band and the slow wave energy ratio before and after stimulation, and uses these two types of changes as the basis for parameter updates. It then adjusts the reward weight of the reinforcement learning model in the dynamic parameter optimization module and the baseline parameters of the acoustic intervention strategy library in the staged adaptation strategy matching module.

[0034] Furthermore, the closed-loop efficacy assessment and strategy iteration module performs the following operations when updating the strategy iteration:

[0035] Two segments of EEG data, one before and one after the acoustic stimulus output, are extracted according to a fixed time window. The tinnitus neural activity intensity and sleep stability values ​​for the corresponding time periods are calculated. The difference between the two values ​​before and after the stimulus is calculated to generate a neural response feature vector.

[0036] The neural response feature vector is input into the built-in policy update model, and the parameter correction value is calculated using the adaptive learning rate parameter correction formula, which is expressed as follows: ,in This is the parameter correction vector for the basic acoustic intervention strategy, which has the same dimensions as the baseline parameter vector. The adaptive learning rate coefficient is dimensionless. The difference in reward values ​​between two consecutive interventions is dimensionless. The adaptive learning rate coefficients represent the baseline parameter vector of the basic acoustic intervention strategy for the current cycle. The initial value is determined by the model convergence speed during the offline training phase. During online use, it is dynamically adjusted according to the fluctuation range of efficacy over multiple consecutive rounds. The larger the fluctuation range of efficacy, the smaller the value of the learning rate coefficient. The output is the reward weight adjustment value of the reinforcement learning model and the parameter correction value of the basic acoustic intervention strategy.

[0037] The adjustment and correction values ​​are distributed to the corresponding modules to complete a single round of strategy iteration update.

[0038] Compared with existing technologies, this EEG-triggered adaptive acoustic drive system for sleep enhancement and tinnitus suppression has the following advantages:

[0039] This invention constructs an adaptive mapping framework between sleep states and tinnitus suppression strategies by jointly decoding sleep stages and tinnitus neural characteristics using EEG signals. This framework enables automatic switching of differentiated acoustic driving schemes for different sleep stages, adapting to the differences in neural plasticity levels and auditory perception thresholds at each stage. Through reinforcement learning algorithms driven by EEG feedback, acoustic parameters are dynamically optimized and a smooth transition mechanism is set up. While continuously suppressing tinnitus, it avoids stimulating sounds from interfering with normal sleep structure. Relying on the closed-loop pathway of brain-computer interaction, it integrates the dual-target regulation of sleep enhancement and tinnitus suppression, blocking the pathological cycle of bidirectional deterioration of both. This improves tinnitus symptoms while optimizing sleep structure, effectively enhancing the overall intervention effect and user comfort in the long term.

[0040] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0042] Figure 1 A schematic diagram of an EEG-triggered adaptive acoustic drive system for sleep enhancement and tinnitus suppression.

[0043] Figure 2 A flowchart of an EEG-triggered adaptive acoustic drive system for sleep enhancement and tinnitus suppression.

[0044] Figure 3 This is a flowchart of the sleep stage and tinnitus neural feature joint decoding module during feature decoding and state discrimination. Detailed Implementation

[0045] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0046] This invention relates to an EEG-triggered adaptive acoustic drive system for sleep enhancement and tinnitus suppression. Based on a closed-loop brain-computer interface architecture, the system comprises a complete neuromodulation chain consisting of an EEG signal acquisition and preprocessing module, a sleep staging and tinnitus neural feature joint decoding module, a staging adaptation strategy matching module, a dynamic parameter optimization module, an acoustic drive output module, and a closed-loop efficacy evaluation and strategy iteration module. The system acquires real-time neural activity data from the user through a multi-channel scalp EEG acquisition unit. After signal amplification, filtering, and artifact removal, it simultaneously decodes three core neural indicators: the user's current sleep stage, tinnitus neural activity intensity, and sleep stability. Based on the sleep stage results, it matches and calls corresponding basic acoustic intervention strategies. Combining real-time neural activity indicators, it dynamically adjusts acoustic intervention parameters through a built-in reinforcement learning model. When a sleep stage transition is detected, a smooth transition curve is generated to avoid acoustic abrupt changes that could disturb the sleep structure. Finally, adaptive acoustic stimulation is applied to the user through the acoustic output unit. Simultaneously, the system continuously acquires EEG feedback data after stimulation, continuously optimizing the strategy baseline parameters and reinforcement learning reward weights through a closed-loop iteration mechanism to achieve synergistic regulation of both sleep enhancement and tinnitus suppression. The following is a detailed description with reference to specific embodiments.

[0047] This embodiment is applied to a nighttime home sleep intervention scenario for users with chronic tinnitus accompanied by sleep structure disorders. The user wears a head-mounted intervention device integrating a dry electrode acquisition unit and a bone conduction sound generation unit. The system operates automatically throughout the entire nighttime sleep cycle, from the user's awake state before falling asleep until the user's awake state upon waking in the morning. Before system operation, the device is worn and parameters are initialized. The electrode positions are placed according to the international 10-20 EEG system standard to ensure that the acquired signal quality meets the requirements for subsequent decoding.

[0048] The overall system architecture adopted in this embodiment is as follows: Figure 1 As shown, the functional modules are connected sequentially according to the flow from EEG signal input to acoustic stimulus output and then to feedback iteration, ultimately forming a complete closed-loop brain-computer interface pathway. Specifically, the EEG signal acquisition and preprocessing module provides the raw neural signal input for the entire system; the sleep staging and tinnitus neural feature joint decoding module performs neural state analysis; the staging adaptation strategy matching module completes the initial mapping from sleep state to intervention plan; the dynamic parameter optimization module is the core decision-making unit of the system; the acoustic drive output module realizes the conversion of control commands into physical stimuli; and the closed-loop efficacy evaluation and strategy iteration module is responsible for the long-term autonomous optimization of the system.

[0049] In this embodiment, the EEG signal acquisition and preprocessing module adopts a 6-channel dry electrode acquisition architecture. The electrode positions correspond to the Cz position in the central area of ​​the prefrontal cortex (FP1 and FP2), the Pz position in the parietal lobe, and the T7 and T8 positions in the bilateral temporal lobes, respectively. This covers the whole-brain sleep feature acquisition area and the corresponding acquisition area of ​​the auditory cortex, enabling the simultaneous acquisition of sleep structure-related EEG signals and tinnitus-related auditory cortical neural activity signals. The raw sampling rate of the acquisition unit is set to 250Hz. The signal amplification unit uses a programmable amplifier circuit to apply a fixed gain amplification of 1000 times to the raw microvolt-level EEG signal, adjusting the signal amplitude to the adaptation voltage range of the analog-to-digital conversion unit to ensure the accuracy and dynamic range of signal sampling.

[0050] After signal amplification, the signal enters the digital filtering unit for processing. First, a bandpass filter (0.5Hz-45Hz) is used to retain the effective EEG frequency band, filtering out DC drift and high-frequency electromagnetic interference. Then, independent component analysis is used for artifact separation. A blind source separation algorithm identifies the independent components corresponding to EEG and EMG artifacts, removing these components from the original signal to complete artifact removal. After preprocessing, the module outputs standardized multi-channel EEG signals, which are transmitted to the sleep staging and tinnitus neural feature joint decoding module for subsequent feature decoding. During preprocessing, the module monitors the signal quality of each channel in real time, marking channels with poor contact and supplementing data using adjacent channel interpolation to ensure input stability in the subsequent decoding process.

[0051] The feature decoding and state discrimination process of the sleep stage and tinnitus neurological feature joint decoding module is as follows: Figure 3 As shown, the module uses a sliding time window for real-time decoding. The window length of the sliding time window is set to 30s and the step size is set to 10s. Within each time window, a complete feature extraction and state discrimination calculation is completed, and the neural state index corresponding to the current time window is output.

[0052] The decoding process first performs bandgap decomposition on the input standardized EEG signal. Bandpass filtering is used to extract signal components from the delta, theta, alpha, beta, and gamma bands, corresponding to frequency ranges of 0.5Hz-4Hz, 4Hz-8Hz, 8Hz-13Hz, 13Hz-30Hz, and 30Hz-45Hz, respectively. The energy percentage of each band is calculated. Simultaneously, the module performs feature waveform recognition on the time-domain EEG signal. By using waveform amplitude and duration thresholds, it identifies three types of characteristic waveforms: slow waves, sleep spindles, and K-complexes. The frequency and time-domain parameters of each type of waveform are then statistically analyzed.

[0053] After extracting frequency band energy and feature waveforms, the module performs sleep stage classification. Based on the energy proportion of each frequency band and the frequency of feature waveforms, it outputs the sleep stage label corresponding to the current time window according to the sleep stage classification rules. The stage categories include wakefulness, light sleep, deep sleep, and REM sleep. Wakefulness corresponds to the state with the highest proportion of alpha frequency band energy and accompanied by a large amount of electromyographic activity; light sleep corresponds to the state with an increasing proportion of theta frequency band energy and the appearance of sleep spindles and K-complexes; deep sleep corresponds to the state with the dominant proportion of delta frequency band energy; and REM sleep corresponds to the state with low-amplitude mixed-frequency EEG and accompanied by rapid eye movement features. During the classification process, the stage results of continuous time windows are combined for smoothing to avoid stage jumps caused by single-window errors and improve the continuity and accuracy of the stage results.

[0054] While completing sleep stage identification, the module simultaneously calculates the intensity index of tinnitus neural activity, selecting the T7 and T8 EEG channels corresponding to the bilateral temporal lobes as calculation channels, and performing phase synchronization analysis on the gamma band signal. In the specific implementation of this embodiment, the weighted phase synchronization index is used as the quantitative index of tinnitus neural activity intensity, and the calculation formula is as follows: ,in The weighted phase synchronization index, used as a quantitative indicator of the intensity of tinnitus neural activity, ranges from 0 to 1 and is dimensionless. This refers to the number of EEG signal sampling points within a single sliding time window. for The signal-to-noise weighting coefficients at each time step, dimensionless. , The gamma band signals of the two EEG channels corresponding to the bilateral temporal lobes are respectively located in... The instantaneous phase at a given moment, expressed in radians.

[0055] Among them, the signal-to-noise weighting coefficient The signal-to-noise ratio (SNR) of the preprocessed EEG signals is calculated in real time. Specifically, for each sampling point, 20 neighboring sampling points are selected before and after it. The ratio of effective EEG energy to noise energy within the neighboring area is used as the SNR for that point; a higher SNR corresponds to a larger weighting coefficient. The weighting coefficients for all moments within a single time window are normalized so that the sum of all weighting coefficients equals the number of sampling points. Equality ensures that the weighted calculation result has the same numerical scale as the traditional phase-locked value, facilitating subsequent index comparison and threshold setting. Instantaneous phase extraction is achieved using Hilbert transform. A Hilbert transform is performed on the single-channel gamma band signal to obtain an analytical signal, and the argument of this analytical signal is taken as the instantaneous phase at the corresponding moment. A higher weighted phase synchronization index indicates stronger neural synchronicity in the bilateral auditory cortex gamma band, corresponding to higher intensity of tinnitus neural activity.

[0056] After calculating the intensity of tinnitus neural activity, the module further calculates a sleep stability index. It statistically analyzes the number of sleep stage label switching times and the standard deviation of each frequency band energy over the past 5 minutes, and calculates a quantitative value for sleep stability through a weighted summation. The number of stage switching times corresponds to the degree of sleep fragmentation; more switching times indicate poorer sleep stability. The standard deviation of frequency band energy corresponds to the degree of fluctuation in brain electrical activity; a larger standard deviation indicates a more unstable sleep state. Both indicators are multiplied by their respective weights and then summed to obtain the final sleep stability index; a higher value indicates a more stable sleep state.

[0057] After the module completes all calculations, it simultaneously outputs the sleep stage label, tinnitus neural activity intensity index, and sleep stability index to the downstream stage adaptation strategy matching module and dynamic parameter optimization module, providing a decision-making basis for subsequent strategy matching and parameter optimization.

[0058] The phased adaptation strategy matching module has a built-in acoustic intervention strategy library, which contains four sets of basic acoustic intervention strategies, each corresponding to one of the four sleep phased scenarios. After receiving the sleep phased tags output from the upstream, the module calls the basic acoustic intervention strategy for the corresponding category and outputs the complete strategy parameter set to the dynamic parameter optimization module.

[0059] The first set of strategies corresponds to the wideband habituation sound strategy during the waking period. The strategy parameters include the frequency range of the wideband noise and the temporal fluctuation modulation mode. The frequency range is set to 200Hz-8000Hz, covering the main tinnitus perception frequency band. The temporal fluctuation adopts a sine wave modulation mode with a modulation frequency of 0.5Hz. By slowly fluctuating the intensity, the user's attention is guided to relax, which helps to enter a sleep state and at the same time plays a role in tinnitus habituation intervention.

[0060] The second set of strategies corresponds to the narrowband phase cancellation sound strategy during light sleep. The strategy parameters include the center frequency and phase offset of the narrowband signal. The center frequency matches the user's tinnitus dominant frequency, and the phase offset is set to a value that is out of phase with the abnormal oscillation of the auditory cortex. The narrowband sound stimulation applies phase cancellation to the abnormal neural oscillation, reducing the excessive synchronization of the auditory cortex.

[0061] The third strategy corresponds to the slow-wave modulation masking sound strategy during deep sleep. The strategy parameters include the intensity of the background masking sound and the slow-wave frequency modulation depth. The background masking sound uses low-intensity broadband noise with an intensity set to 30 dBSPL. At the same time, the masking sound is amplitude modulated with a slow-wave frequency of 0.8 Hz. The slow-wave rhythm modulated sound stimulation enhances sleep slow-wave activity and improves the quality of deep sleep.

[0062] The fourth strategy corresponds to the low-intensity white noise strategy during REM sleep. The strategy parameters include the output intensity and spectral flatness of the white noise. The output intensity is set to 20 dBSPL, and the spectral flatness deviation is controlled within 1 dB. The strategy maintains a mild masking effect with a very low-intensity, stable background sound, while avoiding interference with the dream sleep and normal sleep structure during REM sleep.

[0063] After the module completes the strategy matching, it packages all the parameters of the corresponding strategy into a basic parameter set and transmits it to the dynamic parameter optimization module for subsequent parameter fine-tuning.

[0064] The overall system workflow is as follows: Figure 2 As shown, the dynamic parameter optimization module is the core decision-making link in the process. After receiving the basic acoustic intervention strategy and real-time neural indicators, it completes the parameter optimization calculation and outputs the final acoustic output command. When adjusting the acoustic intervention strategy parameters, the dynamic parameter optimization module first receives the basic acoustic intervention strategy output by the staged adaptation strategy matching module, and simultaneously receives the tinnitus neural activity intensity index and sleep stability index output by the sleep stage and tinnitus neural feature joint decoding module. It then uses these two types of neural indicators as state inputs to feed into the built-in pre-trained reinforcement learning model.

[0065] In this embodiment, the reinforcement learning model employs a deep Q-network architecture. The model's state space includes three types of input features: sleep stage labels, tinnitus neural activity intensity indicators, and sleep stability indicators. The model's action space includes three types of output parameters: acoustic intensity adjustment, spectral range offset, and temporal modulation depth adjustment. The model evaluates the expected rewards corresponding to different actions in the current state and selects the optimal action parameters as the adjustment amounts for the acoustic parameters.

[0066] In the specific implementation of this embodiment, the immediate reward value of the reinforcement learning model is calculated using a dual-objective weighted reward formula, which is as follows: ,in The immediate reward value for a single round of intervention. , These are the tinnitus nerve activity intensity indices before and after acoustic stimulation intervention. , These are sleep stability indicators before and after acoustic stimulation intervention. These are the weighting coefficients for the tinnitus suppression dimension. , where are the weighting coefficients for the sleep stability dimension, and both are dimensionless constants.

[0067] Among them, the weighting coefficient and The initial values ​​were determined by prior experience in the field of clinical neuromodulation, based on contribution data of the dual effects of tinnitus and sleep in large-sample clinical studies. The value is 0.6. The value is set to 0.4, and the sum of the two values ​​is 1. Tinnitus suppression is the primary control target by default, while also protecting sleep stability. In subsequent long-term use, the weighting coefficient is adaptively adjusted by the closed-loop efficacy assessment and strategy iteration module based on the user's long-term intervention effects. If the user's sleep stability declines significantly, the weighting coefficient will be increased. The proportion of tinnitus suppression should be increased if the tinnitus suppression effect is poor. The proportion of the two always remains at a sum of 1.

[0068] The reinforcement learning model is trained using a combination of offline pre-training and online fine-tuning. Offline pre-training utilizes publicly available sleep datasets and tinnitus neural data, simulating intervention responses under different sleep states and tinnitus intensities to equip the model with basic parameter optimization capabilities. Online fine-tuning is based on real-time EEG feedback data from a single user, gradually adapting to individual neural characteristics as usage time increases, thus improving personalized control effects. The model performs parameter inference once within each decoding time window, outputting the corresponding adjusted acoustic parameters.

[0069] After calculating the parameter adjustments, the module checks for differences between the current sleep stage and the previous sleep stage. If a stage change is detected, a smooth transition curve is generated with a transition time of 10 seconds. This linear transition gradually changes the acoustic parameters from the old stage values ​​to the new stage values, preventing sleep disturbances caused by sudden changes in acoustic parameters. If no stage change is detected, no transition curve is generated, and the adjusted parameters are output directly.

[0070] Finally, the module integrates the adjusted acoustic intervention strategy parameters and transition curves into a real-time output command, which is then sent to the acoustic drive output module.

[0071] The acoustic drive output module includes a digital-to-analog converter (DAC), a power amplifier, and a dual-path sound-generating unit. After receiving digital acoustic parameter commands from the dynamic parameter optimization module, it first converts the digital parameters into analog time-domain audio signals via the DAC. The DAC sampling rate is set to 44.1kHz to ensure high-quality audio output and accurate detail reproduction. The converted analog audio signal is then fed into the power amplifier, which adjusts the power gain to output an audio signal that meets the driving level requirements of the sound-generating device.

[0072] In this embodiment, the dual-path sound unit includes two types of sound-generating devices: air-conduction headphones and bone conduction vibrators. It supports two working modes: single-path output and dual-path synchronous output. The default mode is single-path output via bone conduction, transmitting sound stimulation to the inner ear through the skull, eliminating the need for insertion into the ear canal, thus improving nighttime wearing comfort and preventing sound leakage from the air conduction system from disturbing bed partners. Users can switch output modes according to their own usage habits. The dual-path synchronous output mode is suitable for scenarios with severe tinnitus, enhancing the intervention intensity through combined air conduction and bone conduction stimulation. The output parameters of the dual-path sound unit correspond one-to-one with the real-time commands output by the dynamic parameter optimization module, updating synchronously with the decoding time window to ensure real-time matching between stimulation and neural state.

[0073] During system operation, the closed-loop efficacy assessment and strategy iteration module continuously monitors the intervention effect, performing a strategy iteration update operation every 30 minutes. The module has a built-in strategy update model, with its input end connected to the output end of the sleep staging and tinnitus neural feature joint decoding module, and its output end connected to the input ends of the staging adaptation strategy matching module and the dynamic parameter optimization module, respectively, driving the continuous optimization of system parameters through EEG feedback data.

[0074] When the module performs iterative updates, it first extracts two segments of EEG data before and after the acoustic stimulus output according to a fixed time window. Each segment of data lasts for 5 minutes. It then calculates the tinnitus neural activity intensity value and sleep stability value for the corresponding time period, calculates the difference between the two values ​​before and after the stimulus, and generates a neural response feature vector.

[0075] In the specific implementation of this embodiment, the policy update model calculates the parameter correction value using the adaptive learning rate parameter correction formula. The calculation formula is as follows: ,in This is the parameter correction vector for the basic acoustic intervention strategy, which has the same dimensions as the baseline parameter vector. The adaptive learning rate coefficient is dimensionless. The difference in reward values ​​between two consecutive interventions is dimensionless. This is the baseline parameter vector for the basic acoustic intervention strategy in the current cycle.

[0076] Where the adaptive learning rate coefficient The initial value is determined by the model's convergence speed during offline training. Multiple sets of learning rates are used for comparative training during offline training to monitor the rate of decrease in validation set loss and the final convergence accuracy. A learning rate value with moderate convergence speed and stable final accuracy is selected as the initial value; in this embodiment, the initial value is set to 0.01. During online use, the learning rate coefficient is dynamically adjusted based on the fluctuation range of efficacy over five consecutive rounds. The fluctuation range is calculated using the standard deviation of reward values ​​over multiple rounds. A larger fluctuation range results in a smaller learning rate coefficient, preventing system oscillations caused by large parameter adjustments; conversely, a smaller fluctuation range results in a larger learning rate coefficient, accelerating parameter convergence. In this embodiment, the upper limit of the learning rate coefficient adjustment range is 0.05, and the lower limit is 0.001, ensuring that parameter updates remain within a stable range.

[0077] After calculating the parameter correction values, the module outputs the reward weight adjustment values ​​for the reinforcement learning model and the parameter correction values ​​for the basic acoustic intervention strategy. These adjusted values ​​are then distributed to the dynamic parameter optimization module and the phased adaptation strategy matching module, respectively, completing a single round of strategy iteration updates. As the number of uses increases, the system gradually converges to the optimal parameter combination for the user, continuously improving the intervention effect.

[0078] In summary, through the complete implementation of this embodiment, the system can achieve dual-target synergistic regulation of sleep enhancement and tinnitus suppression based on brain-computer interface technology. It automatically adapts to the user's neural state changes throughout the entire sleep cycle, applying differentiated acoustic intervention schemes at different sleep stages. Combined with real-time EEG feedback, it dynamically adjusts stimulation parameters and avoids acoustic abrupt changes from disrupting sleep structure through a smooth transition mechanism. The system continuously optimizes its regulation strategy based on a closed-loop iterative mechanism, gradually adapting to the user's individual neural characteristics. While suppressing abnormal neural activity related to tinnitus, it optimizes sleep structure, improves sleep continuity and the proportion of deep sleep, alleviates the bidirectional worsening relationship between tinnitus and sleep disorders, and improves the user's nighttime sleep quality and daytime mental state. It possesses good clinical application value and a positive long-term user experience.

[0079] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An EEG-triggered adaptive acoustic actuation system for sleep enhancement and tinnitus suppression, characterized in that, The system includes: an EEG signal acquisition and preprocessing module, a sleep staging and tinnitus neural feature joint decoding module, a staging adaptation strategy matching module, a dynamic parameter optimization module, and an acoustic drive output module; The EEG signal acquisition and preprocessing module is used to acquire multi-channel scalp EEG signals, complete signal amplification, filtering and artifact removal processing, and output standardized EEG signals to the sleep stage and tinnitus nerve feature joint decoding module. The sleep staging and tinnitus neural feature joint decoding module is used to extract sleep feature waveforms and abnormal oscillation features of the auditory cortex from the electroencephalogram (EEG) signal, and output sleep staging labels and tinnitus neural activity intensity indicators to the staging adaptation strategy matching module and the dynamic parameter optimization module. The staged adaptation strategy matching module is used to match the corresponding acoustic intervention strategy according to the sleep stage label and output it to the dynamic parameter optimization module; The dynamic parameter optimization module is used to adjust the parameters of the acoustic intervention strategy in conjunction with the tinnitus nerve activity intensity index and output them to the acoustic drive output module. The acoustic drive output module is used to generate and play acoustic stimulus signals based on the received parameters.

2. The EEG-triggered adaptive acoustic drive system for sleep enhancement and tinnitus suppression according to claim 1, characterized in that, The sleep stage and tinnitus neural feature joint decoding module performs the following operations when performing feature decoding and state discrimination: The preprocessed multichannel EEG signal was decomposed into frequency bands to extract the energy distribution data of five frequency bands: delta, theta, alpha, beta, and gamma. At the same time, the time domain parameters of three characteristic waveforms, namely slow waves, sleep spindles, and K-complexes, were identified. Based on the frequency of frequency band energy proportion and the frequency of occurrence of characteristic waveforms, four sleep stage labels are output according to the sleep stage discrimination rules: wakefulness, light sleep, deep sleep, and REM sleep. Signals from the corresponding EEG channels in the temporal lobe were selected, and the phase synchronization index and energy ratio of the gamma band were calculated to generate an index of tinnitus neural activity intensity. Based on the degree of fluctuation in sleep stages and the frequency band energy variation coefficient within a continuous time window, a sleep stability index is calculated and generated. The sleep stage labels, tinnitus neural activity intensity indicators, and sleep stability indicators are simultaneously output to the downstream modules.

3. The EEG-triggered adaptive acoustic drive system for sleep enhancement and tinnitus suppression according to claim 2, characterized in that, The sleep staging and tinnitus neurological feature joint decoding module uses a sliding time window for real-time decoding. The window length and step size of the sliding time window are fixed values. Feature extraction and state discrimination are completed once within each time window. When calculating the gamma band phase synchronization index, the sleep staging and tinnitus neurological feature joint decoding module selects two EEG channels corresponding to the bilateral temporal lobes as calculation channels. Hilbert transform is used to extract the instantaneous phase of the single-channel signal, and the quantification index is calculated using the weighted phase synchronization index formula, which is expressed as follows: ,in The weighted phase synchronization index. This refers to the number of EEG signal sampling points within a single sliding time window. for Signal-to-noise weighting coefficients at any given time. , The gamma band signals of the two EEG channels corresponding to the bilateral temporal lobes are respectively located in... When calculating the sleep stability index, the sleep stage and tinnitus nerve feature joint decoding module calculates the instantaneous phase of the sleep stage label within a set time period and the standard deviation of the energy of each frequency band. The quantitative value of sleep stability is obtained by weighted calculation.

4. The EEG-triggered adaptive acoustic drive system for sleep enhancement and tinnitus suppression according to claim 1, characterized in that, The EEG signal acquisition and preprocessing module includes a multi-channel dry electrode acquisition unit, a signal amplification unit, and a digital filtering unit. The electrode points of the multi-channel dry electrode acquisition unit cover the prefrontal cortex, central region, parietal lobe, and bilateral temporal lobes, corresponding to the acquisition of EEG signals related to the whole brain and auditory cortex. The signal amplification unit performs fixed-fold programmable amplification on the raw EEG signals, amplifying the microvolt-level raw EEG signals to a voltage range suitable for analog-to-digital conversion. The digital filtering unit uses bandpass filtering to retain EEG signals in a specified frequency band and uses independent component analysis to separate and remove EEG artifacts and EMG artifacts. The EEG signal acquisition and preprocessing module outputs standardized EEG signals with a fixed sampling rate.

5. The EEG-triggered adaptive acoustic drive system for sleep enhancement and tinnitus suppression according to claim 1, characterized in that, The phased adaptation strategy matching module has a built-in acoustic intervention strategy library, which includes the following four groups of basic acoustic intervention strategies. The first group corresponds to the broadband acclimatization sound strategy during the waking period. The strategy parameters include the spectral range of broadband noise and the temporal fluctuation modulation mode. The second group corresponds to the narrowband phase cancellation acoustic strategy for the light sleep period. The strategy parameters include the center frequency and phase offset of the narrowband signal. The third group corresponds to the slow-wave modulation masking sound strategy during deep sleep. The strategy parameters include the intensity of the background masking sound and the slow-wave frequency modulation depth. The fourth group corresponds to the low-intensity white noise strategy during rapid eye movement (REM) phase. The strategy parameters include the output intensity and spectral flatness of the white noise. The staged adaptation strategy matching module calls the basic acoustic intervention strategy for the corresponding group and outputs a complete parameter set based on the received sleep stage tags.

6. The EEG-triggered adaptive acoustic drive system for sleep enhancement and tinnitus suppression according to claim 1, characterized in that, The dynamic parameter optimization module performs the following operations when adjusting the acoustic intervention strategy parameters: It receives the basic acoustic intervention strategy output by the staged adaptation strategy matching module, and at the same time receives the tinnitus nerve activity intensity index and sleep stability index output by the sleep staged and tinnitus nerve feature joint decoding module. The two types of indicators are input into the built-in pre-trained reinforcement learning model to calculate the adjustment amount of the acoustic parameters; It detects whether there is a difference between the current sleep stage and the previous sleep stage, and generates a parameter smooth transition curve when the stage changes; The adjusted acoustic intervention strategy parameters and transition curves are integrated into a real-time output command and sent to the acoustic drive output module.

7. The EEG-triggered adaptive acoustic drive system for sleep enhancement and tinnitus suppression according to claim 6, characterized in that, The reinforcement learning model built into the dynamic parameter optimization module adopts a deep Q-network architecture. The state space of the reinforcement learning model includes three types of input features: sleep stage labels, tinnitus neural activity intensity indicators, and sleep stability indicators. The action space of the reinforcement learning model includes three types of output parameters: acoustic intensity adjustment, spectral range offset, and temporal modulation depth adjustment. The immediate reward value of the reinforcement learning model is calculated using a bi-objective weighted reward formula, expressed as follows: ,in The immediate reward value for a single round of intervention. , These are the tinnitus nerve activity intensity indices before and after acoustic stimulation intervention. , These are sleep stability indicators before and after acoustic stimulation intervention. These are the weighting coefficients for the tinnitus suppression dimension. The weighting coefficients for the sleep stability dimension are used in the training process of the reinforcement learning model, which employs a combination of offline pre-training and online fine-tuning. Offline pre-training is completed using publicly available sleep datasets and tinnitus neural data, while online fine-tuning is completed based on real-time EEG feedback data from a single user. The reinforcement learning model completes parameter inference once within each decoding time window and outputs the corresponding adjustment values ​​for acoustic parameters.

8. The EEG-triggered adaptive acoustic drive system for sleep enhancement and tinnitus suppression according to claim 1, characterized in that, The acoustic drive output module includes a digital-to-analog converter, a power amplifier, and a dual-channel sound generation unit. The digital-to-analog converter receives acoustic parameter commands in digital format and converts them into analog time-domain audio signals. The power amplifier adjusts the power gain of the analog audio signals and outputs audio signals that meet the drive level requirements. The dual-channel sound generation unit includes two types of sound generation devices: air-conduction headphones and bone-conduction vibrators. It supports two working modes: single-channel output and dual-channel synchronous output. It outputs corresponding acoustic stimuli according to the received audio signals. The output parameters of the dual-channel sound generation unit correspond one-to-one with the real-time commands output by the dynamic parameter optimization module.

9. The EEG-triggered adaptive acoustic drive system for sleep enhancement and tinnitus suppression according to claim 1, characterized in that, The system also includes a closed-loop efficacy assessment and strategy iteration module. This module has a built-in strategy update model, a staged adaptation strategy matching module with a built-in acoustic intervention strategy library, and a dynamic parameter optimization module with a built-in reinforcement learning model. The input of the closed-loop efficacy assessment and strategy iteration module is connected to the output of the sleep staging and tinnitus neural feature joint decoding module. The output of the closed-loop efficacy assessment and strategy iteration module is connected to the inputs of the staged adaptation strategy matching module and the dynamic parameter optimization module, respectively. The closed-loop efficacy assessment and strategy iteration module is used to collect EEG feedback data after a set duration of acoustic stimulation output, extract the changes in the phase synchronization index of the auditory cortex gamma band and the changes in the proportion of slow wave energy before and after stimulation, and use these two types of changes as the basis for parameter updates, adjusting the reward weight of the reinforcement learning model built into the dynamic parameter optimization module and the baseline parameters of the acoustic intervention strategy library built into the staged adaptation strategy matching module.

10. The EEG-triggered adaptive acoustic drive system for sleep enhancement and tinnitus suppression according to claim 9, characterized in that, The closed-loop efficacy assessment and strategy iteration module performs the following operations during strategy iteration updates: Two segments of EEG data, one before and one after the acoustic stimulus output, are extracted according to a fixed time window. The tinnitus neural activity intensity and sleep stability values ​​for the corresponding time periods are calculated. The difference between the two values ​​before and after the stimulus is calculated to generate a neural response feature vector. The neural response feature vector is input into the built-in policy update model, and the parameter correction value is calculated using the adaptive learning rate parameter correction formula, which is expressed as follows: ,in This is the parameter correction vector for the basic acoustic intervention strategy, which has the same dimensions as the baseline parameter vector. For adaptive learning rate coefficients, The difference in reward values ​​between two consecutive interventions. Given the baseline parameter vector of the basic acoustic intervention strategy for the current cycle, output the reward weight adjustment value of the reinforcement learning model and the parameter correction value of the basic acoustic intervention strategy; The adjustment and correction values ​​are distributed to the corresponding modules to complete a single round of strategy iteration update.

Images