A hearing aid parameter adaptability evaluation system and method based on multi-dimensional electroencephalogram index fusion

Abstract

The application discloses a hearing aid parameter adaptability evaluation system and method based on multi-dimensional electroencephalogram index fusion, and the natural speech flow is taken as the hearing stimulation material, the prefabricated natural speech is played under the condition that the subject wears the hearing aid with the target parameter configuration, the scalp electroencephalogram signal is synchronously collected, three types of neural indexes, namely, the perception clarity index, the cognitive load index and the emotional comfort index, are extracted from the same electroencephalogram data in parallel, the three types of indexes are input into the pre-training fusion model, and the 0-100 score parameter adaptation degree score representing the rationality of the current hearing aid parameter configuration is output; the score is taken as the feedback signal to drive the iterative optimization until the parameter adaptation degree score reaches the standard. The application replaces the evaluation link depending on the subjective feedback in the hearing aid fitting process into the objective, quantitative and multi-dimensional electroencephalogram physiological evaluation, and can cover the groups such as children, the aged and people with cognitive disorders, which are difficult to be effectively served by the traditional methods.

Description

Technical Field

[0001] This invention relates to the intersection of hearing rehabilitation technology and brain-computer interface technology, specifically to a hearing aid parameter fit assessment system and method based on the fusion of multi-dimensional EEG indicators. Background Technology

[0002] Traditional hearing aid fitting processes rely heavily on the wearer's subjective feedback and the audiologist's experience, lacking objective physiological indicators to assess whether the hearing aid settings are truly suitable for the wearer's auditory processing state. The traditional process mainly includes: pure-tone audiometry to determine the hearing threshold, calculating initial gain parameters using fitting formulas based on the audiogram, and fine-tuning through real-ear analysis. Each step of this process requires the wearer's active cooperation and verbal reporting of subjective feelings (such as "too big," "can't hear clearly," "harsh," etc.), resulting in the following drawbacks: strong subjectivity, inconsistent standards of expression among different wearers; poor applicability to groups that cannot effectively cooperate, such as children, the elderly, and those with cognitive impairments; lengthy processing time, often requiring multiple visits to the fitting center; and inability to reflect the true listening state in complex auditory scenarios (such as noisy environments or multi-speaker scenarios).

[0003] In recent years, some studies have attempted to introduce EEG technology into the field of hearing aids.

[0004] One approach is based on electroencephalography (EEG) for hearing threshold testing, which uses evoked potential techniques such as auditory brainstem response (ABR) and auditory steady-state response (ASSR) to objectively assess hearing thresholds. Related patents, such as US12520091B2 (T&W / Widex), propose a self-fitting method that determines hearing thresholds by analyzing the user's EEG response to test tones. However, these methods can only detect hearing values ​​at single frequencies, which differs significantly from the multi-parameter fitting scenarios in real-world hearing aids. Furthermore, they use pure tones or click sounds as stimulus signals, failing to reflect the wearer's auditory processing state in real-world language communication scenarios.

[0005] Secondly, there are EEG-based auditory attention decoding methods, which utilize EEG signals to decode the wearer's attentional targets in multi-speaker scenarios. For example, Oticon's patents US9432777B2 and CN107864440B disclose methods for reconstructing audio spectra and controlling audio processing of hearing aids based on EEG signals. MIT's US11630513B2 discloses an end-to-end deep neural network method for auditory attention decoding. These methods are all geared towards the daily real-time operation of hearing aids, i.e., using EEG signals to control the audio processing parameters of the hearing aid in real time, rather than addressing the objective assessment problems during the fitting stage. Furthermore, these methods typically only use a single dimension of EEG characteristics, failing to comprehensively reflect the wearer's diverse subjective experiences such as clarity, effort, and comfort during listening.

[0006] Thirdly, there are self-fitting methods based on purely subjective feedback. For example, Southeast University CN112653980B and Tsinghua University CN114339564B have disclosed self-fitting methods based on subjective evaluation and neural network prediction, respectively. These methods still rely heavily on the user's subjective feedback and do not introduce objective neurophysiological indicators.

[0007] In summary, current technologies still face three major bottlenecks: the lack of objective assessment methods leads to individualized fitting with significant blind spots; EEG methods only measure hearing sensitivity or attention, failing to simultaneously reflect comprehensive experiences such as clarity, fatigue, and comfort; and the stimulation materials are mostly pure tones or repetitive phonemes, which differ greatly from continuous speech, making it difficult to extrapolate assessment results to everyday communication scenarios. These shortcomings severely restrict the accuracy and universality of hearing aid fitting. Summary of the Invention

[0008] In view of the problems existing in the prior art, the purpose of this invention is to provide a hearing aid parameter fit assessment system and method based on multi-dimensional EEG index fusion, so as to improve the accuracy and universality of hearing aid fitting.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A hearing aid parameter fit assessment system based on multi-dimensional EEG index fusion, comprising: The speech playback module is used to play pre-made natural speech stimulation materials when the subject is wearing a hearing aid; The EEG acquisition module is used to simultaneously acquire the subject's scalp electroencephalogram (EEG) signals while playing natural speech stimulation materials; The feature extraction module is used to extract three types of neural indicators in parallel from EEG signals, including perceptual clarity indicators, cognitive load indicators, and emotional comfort indicators. The perceived clarity index is calculated based on the neural envelope tracking similarity between the EEG signal and the speech envelope. The cognitive load index was calculated based on the rate of change of the theta band power in the prefrontal region and the alpha band power in the parietal region relative to the subject's resting baseline. The emotional comfort index is calculated based on the asymmetry between the left and right frontal lobes of the EEG alpha wave. The fusion assessment module has a built-in pre-trained fusion model that receives perceptual clarity indicators, cognitive load indicators, and emotional comfort indicators, and outputs a parameter fit score from 0 to 100. The results output module is used to present the parameter fit score and the scores of each dimension to the audiologist or fitting system in a visual form.

[0010] The natural speech stream is a short Chinese sentence or paragraph containing complete semantic information, including but not limited to: clear version, light noise version, medium noise version and distorted version, to simulate listening conditions under different auditory scenarios.

[0011] The EEG acquisition module is a head-mounted multi-channel EEG cap, which covers at least the following electrode locations: the frontal region includes F3, Fz, and F4 electrodes; the central region includes C3, Cz, and C4 electrodes; and the parietal lobe region includes P3, Pz, and P4 electrodes.

[0012] The neural envelope tracking similarity is calculated as follows: a wide-spectrum time-domain envelope is extracted from the speech signal, and multi-band envelopes are extracted by Hilbert transform or gammatone filter bank and then weighted summed. The canonical correlation analysis coefficient or cross-correlation peak value between the envelope and the EEG signal is then calculated as a perceptual sharpness index. The cognitive load index is calculated as follows: the power spectral density (PSD) of each EEG channel is calculated using the Welch method, and the absolute power value is obtained by trapezoidal numerical integration of the PSD of the target frequency band. The power in the θ band, θ_Power, is the integral of the PSD over the range of 4Hz to 8Hz, i.e. And take the average value of the prefrontal electrodes F3, Fz, and F4; The α-band power α_Power is the integral of PSD in the range of 8Hz to 13Hz, i.e. And take the average value of the top leaf electrode P3, Pz, P4; Calculate the normalized rate of change of the above values ​​relative to the subject's resting baseline: Δθ = (θ_Power θ_Rest) / θ_Rest Δα = (α_Power α_Rest) / α_Rest Where Δθ is the normalized rate of change of the θ band power relative to the subject's resting baseline, Δα is the normalized rate of change of the α band power relative to the subject's resting baseline, and θ_Rest and α_Rest are the θ band power and α band power at the subject's resting baseline, respectively. Finally, Δθ and Δα are combined to form a scalar cognitive load index, which is defined as Load = Δθ Δα; The emotional comfort index is the asymmetry index of the α-band power of the left and right frontal lobes, and is defined as emotional comfort index α_Asymmetry=(R_alpha) L_alpha) / (R_alpha+L_alpha), where R_alpha is the power of the right frontal lobe F4 electrode in the α band, and L_alpha is the power of the left frontal lobe F3 electrode in the α band; The fusion model can be any of the following: a weighted linear fusion model, a logistic regression model, or a lightweight multilayer perceptron neural network.

[0013] The system also includes: The signal preprocessing module is used to perform bandpass filtering, notch filtering, independent component analysis for artifact removal, and average reference processing on the raw EEG signal. The feature extraction module extracts three types of neural indicators in parallel from the preprocessed EEG signal.

[0014] A method for assessing hearing aid parameter fit based on the fusion of multi-dimensional electroencephalogram (EEG) indicators includes the following steps: Step S1: With the subject wearing a hearing aid configured with the target parameters, play a pre-made natural speech stream as auditory stimulation to the subject; Step S2: Simultaneously collect the subject's scalp electroencephalogram (EEG) signals by wearing an EEG acquisition module on the subject's head; Step S3: Extract three types of neural indicators in parallel from the acquired EEG signals, including: The perceived clarity index is calculated based on the neural envelope tracking similarity between the EEG signal and the speech envelope. Cognitive load indices were calculated based on the rate of change of the theta band power in the prefrontal region and the alpha band power in the parietal region relative to the subject's resting baseline. The emotional comfort index was calculated based on the asymmetry between the left and right frontal lobes of the EEG alpha wave. Step S4: Input the three types of neural indicators into the pre-trained fusion model, and the fusion model outputs a parameter fit score that represents the rationality of the current hearing aid parameter configuration; Step S5: Output the parameter fit score to guide the adjustment of hearing aid parameters: Determine whether the preset threshold has been reached based on the parameter fit score; if not, determine the source of the problem based on the sub-scores of each dimension, and adjust one or more parameters of the hearing aid accordingly. The parameters include at least one of gain, compression ratio, compression inflection point, maximum output sound pressure level, frequency response curve, and noise reduction intensity. Repeat steps S1 to S5 until the parameter fit score reaches the preset threshold.

[0015] The natural speech stream in step S1 is a short Chinese sentence or paragraph containing complete semantic information, including but not limited to: clear version, light noise version, medium noise version and distorted version, to simulate listening conditions under different auditory scenarios.

[0016] In step S3 The neural envelope tracking similarity is calculated as follows: a wide-spectrum time-domain envelope is extracted from the speech signal, and multi-band envelopes are extracted by Hilbert transform or gammatone filter bank and then weighted summed. The canonical correlation analysis coefficient or cross-correlation peak value between the envelope and the EEG signal is then calculated as a perceptual sharpness index. The cognitive load index is calculated as follows: the power spectral density (PSD) of each EEG channel is calculated using the Welch method, and the absolute power value is obtained by trapezoidal numerical integration of the PSD of the target frequency band. The power in the θ band, θ_Power, is the integral of the PSD over the range of 4Hz to 8Hz, i.e. And take the average value of the prefrontal electrodes F3, Fz, and F4; The α-band power α_Power is the integral of PSD in the range of 8Hz to 13Hz, i.e. And take the average value of the top leaf electrode P3, Pz, P4; Calculate the normalized rate of change of the above values ​​relative to the subject's resting baseline: Δθ = (θ_Power θ_Rest) / θ_Rest Δα = (α_Power α_Rest) / α_Rest Where Δθ is the normalized rate of change of the θ band power relative to the subject's resting baseline, Δα is the normalized rate of change of the α band power relative to the subject's resting baseline, and θ_Rest and α_Rest are the θ band power and α band power at the subject's resting baseline, respectively. Finally, Δθ and Δα are combined to form a scalar cognitive load index, which is defined as Load = Δθ Δα; The emotional comfort index is the asymmetry index of the α-band power of the left and right frontal lobes, and is defined as emotional comfort index α_Asymmetry=(R_alpha) L_alpha) / (R_alpha+L_alpha), where R_alpha is the power of the right frontal lobe F4 electrode in the α band, and L_alpha is the power of the left frontal lobe F3 electrode in the α band.

[0017] The fusion model in step S4 can be any of the following forms: a weighted linear fusion model, a logistic regression model, or a lightweight multilayer perceptron neural network.

[0018] The EEG acquisition module mentioned in step S2 is a head-mounted multi-channel EEG cap, and the electrodes of the EEG cap cover at least the frontal region, central region and parietal lobe region.

[0019] By adopting the above solution, this invention replaces the subjective feedback-dependent assessment process in hearing aid fitting with an objective quantitative assessment based on electroencephalogram (EEG) signals, fundamentally solving the problem of traditional fitting methods' high dependence on the wearer's cooperation and expression abilities. EEG signals are the auditory system's automatic physiological response to external stimuli, unaffected by the wearer's subjective will, expression ability, or cognitive state. Therefore, this method can cover groups that traditional methods struggle to effectively serve, such as children, the elderly, and those with cognitive impairments.

[0020] Simultaneously, this invention organically integrates three independent dimensions of EEG indicators: perceptual clarity (neural envelope tracking), cognitive load (θ / α power), and emotional comfort (α asymmetry). These correspond to the three core aspects of the listening experience: "clear hearing," "easy listening," and "comfortable listening," respectively, thus providing a more complete picture of the wearer's overall listening experience. This invention replaces the pure tone or click-sound stimulation used in traditional methods with natural speech streams, upgrading the assessment environment from "laboratory pure tone testing" to "simulated real-life dialogue scenarios," which can induce a complete hierarchical auditory processing chain in the brain. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall system framework of the present invention; Figure 2 This is a flowchart of the method of the present invention; Figure 3 This is a schematic diagram of the multi-dimensional feature extraction of the present invention; Figure 4 This is a schematic diagram of the architecture of the integrated evaluation model; Figure 5 This is a diagram showing the electrode layout for a head-mounted multichannel EEG acquisition device. Figure 6 Power statistics for different resolution conditions in the frequency band; Figure 7 The diagram shows the spatial distribution and spectral characteristics of power in the θ band. Figure 8 This is a comparative diagram of the time-domain and frequency-domain neural responses. Detailed Implementation

[0022] To make the objectives, technical solutions, and beneficial effects of the embodiments of the present invention clearer, the specific implementation methods of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] like Figure 1As shown, this invention discloses a hearing aid parameter fit assessment system based on multi-dimensional EEG index fusion, which includes: a speech playback module, an EEG acquisition module, a signal preprocessing module, a feature extraction module, a fusion assessment module, and a result output module. The speech stimulation signal output by the speech playback module is sent to the hearing aid worn by the subject in one path, and is used as a reference signal input to the perceptual clarity extraction sub-module of the feature extraction module in another path. The EEG signal acquired by the EEG acquisition module is input to the signal preprocessing module. The preprocessed EEG signal is input to the three sub-modules of the feature extraction module respectively. The three types of feature indicators are converged and input to the fusion assessment module. The parameter fit score output by the fusion assessment module is fed back to the hearing aid parameter adjustment terminal through the result output module, forming a closed-loop iterative path.

[0024] Specifically, the speech playback module is used to play pre-prepared natural speech stimulation materials when the subject is wearing a hearing aid. This speech playback module can be a standalone audio playback device, such as a high-fidelity speaker or headphones, or it can be a software playback unit integrated into the hearing aid fitting system, outputting standardized speech stimuli through digital signal processing. The speech playback module and the feature extraction module have a synchronous trigger connection, used to send a trigger marker to the feature extraction module at the start of each speech stimulus to align the time axis of the speech stimulus with the EEG signal.

[0025] The EEG acquisition module is used to simultaneously acquire scalp EEG signals from subjects using a head-mounted multi-channel EEG cap. The EEG cap is equipped with electrodes arranged according to the international 10-20 system, covering at least the frontal, central, and parietal lobe regions. Specifically, the frontal region includes F3, Fz, and F4 electrodes; the central region includes C3, Cz, and C4 electrodes; and the parietal lobe includes P3, Pz, and P4 electrodes. The EEG cap can acquire EEG signals through contact with the scalp using conductive gel (in which case the electrode impedance must be less than 10kΩ) or dry electrodes. The signals are amplified by a differential amplifier and recorded digitally. The sampling rate of the EEG acquisition module is no less than 250Hz, preferably 500Hz or 1000Hz.

[0026] The signal preprocessing module is used to preprocess the raw EEG signal. It includes a bandpass filter unit, a notch filter unit, an independent component analysis (ICA) unit, a bad conductor interpolation unit, and an averaging reference unit. The bandpass filter unit has a filtering range of 1.6Hz to 40Hz and uses a fourth-order Butterworth filter for zero-phase bidirectional filtering. The notch filter unit is used to remove 50Hz power frequency interference. The ICA unit identifies and removes artifacts from electrooculography (EOG) and electromyography (EMG) signals. The bad conductor interpolation unit performs spherical spline interpolation on electrode channels with excessive impedance or abnormal signals. The averaging reference unit rereferences all electrode signals to a common averaging reference.

[0027] The feature extraction module is used to extract three types of neural indicators in parallel from the preprocessed EEG signal. It includes three parallel sub-modules: (1) The perceptual clarity extraction submodule is used to calculate the perceptual clarity index based on the neural envelope tracking similarity between the EEG signal and the speech envelope.

[0028] Specifically, the perceived sharpness extraction submodule filters the speech signal using a gammatone filter bank with a center frequency covering 80Hz to 8kHz, totaling 32 frequency bands. The output of each frequency band is subjected to a Hilbert transform to extract the amplitude envelope. Then, the envelopes of each frequency band are weighted and summed to obtain a wide-spectrum time-domain envelope. Alternatively, the speech signal can be directly subjected to a Hilbert transform to extract the time-domain envelope.

[0029] Then, the Pearson correlation coefficient between the speech temporal envelope and the signals of each EEG channel was calculated using a sliding time window method. The analysis time window was 0 to 2 seconds after the stimulus started, with a sliding step of 50 ms and a window length of 200 ms. The peak value of the correlation coefficient in all time windows was taken as the neural envelope tracking similarity of that channel (the candidate caliber is the mean value above the significance threshold). The significance threshold was set to p < 0.05 and corrected by FDR.

[0030] Finally, the average similarity of each electrode channel in the forehead and central areas was taken as the perceptual clarity index.

[0031] (2) The cognitive load extraction submodule is used to calculate the cognitive load index based on the rate of change of the power of the θ band in the prefrontal region and the power of the α band in the parietal region relative to the subject's resting baseline.

[0032] Specifically, the cognitive load extraction submodule uses the Welch method to calculate the power spectral density (PSD) of each EEG signal, with the following parameters: Hamming window, window length 2 seconds, overlap rate 50%, and FFT points equal to a power of 2 of the window length. Trapezoidal numerical integration is then performed on the PSD of the target frequency band to obtain the absolute power of that band, in μV². The power in the θ band, θ_Power, is the integral of the PSD over the range of 4Hz to 8Hz, i.e. The average values ​​of the prefrontal electrodes F3, Fz, and F4 were taken.

[0033] The α-band power α_Power is the integral of PSD in the range of 8Hz to 13Hz, i.e. The average values ​​of the top leaf electrodes P3, Pz, and P4 are taken.

[0034] Calculate the power change rate in the θ band and the power change rate in the α band respectively: Δθ = (θ_Power θ_Rest) / θ_Rest Δα = (α_Power α_Rest) / α_Rest Wherein, Δθ is the normalized rate of change of the θ band power relative to the subject's resting baseline, Δα is the normalized rate of change of the α band power relative to the subject's resting baseline, and θ_Rest and α_Rest are the θ band power and α band power at the subject's resting baseline, respectively. They were obtained by collecting 2 minutes of closed-eye resting EEG data before fitting began and calculated using the same method.

[0035] Finally, Δθ and Δα are combined to form a cognitive load index, which is a scalar Load = Δθ. Δα (a larger load indicates a higher cognitive load), an increase in θ power usually reflects an increase in the input of cognitive control resources ("listening fatigue"), and a decrease in α power in the task state relative to the resting state (desynchronization) usually reflects an increase in the degree to which the cortical area is called upon by the task.

[0036] (3) Emotional comfort extraction submodule, used to calculate emotional comfort index based on the asymmetry of the left and right frontal lobes of EEG alpha waves.

[0037] Specifically, this emotional comfort extraction submodule is used to calculate the asymmetry index of the power in the α band of the left and right frontal lobes. The α band ranges from 8Hz to 13Hz, and the formula for calculating the asymmetry index is as follows: α_Asymmetry=(R_alpha L_alpha) / (R_alpha+L_alpha) Where R_alpha is the power of the right frontal lobe F4 electrode in the α band, and L_alpha is the power of the left frontal lobe F3 electrode in the α band.

[0038] The fusion assessment module incorporates a pre-trained fusion model that receives the three types of indicators (perceptual clarity, cognitive load, and emotional comfort) and outputs a parameter fit score from 0 to 100. The fusion model can be any of a weighted linear fusion model, a logistic regression model, or a lightweight multilayer perceptron (MLP) neural network, with its three-dimensional input feature space fixed as [perceptual clarity, cognitive load, and emotional comfort]. The weights of the fusion model can be determined in two ways: (a) by directly deploying prior weights based on existing literature effect sizes; and (b) by supervised training optimization using a clinical dataset containing subjective rating annotations.

[0039] The results output module visually presents the parameter fit score and the scores for each dimension to the audiologist or fitting system. The scores for each dimension are obtained by mapping the normalized perceptual clarity, cognitive load, and emotional comfort indices to a score from 0 to 100, respectively, and are used to determine the direction for parameter adjustments. This results output module can output to the graphical user interface of the fitting software or transmit data to the parameter optimization unit of the automated fitting system via a data interface.

[0040] like Figure 2 As shown, based on the same inventive concept, this invention also discloses a method for assessing hearing aid parameter fit based on the fusion of multi-dimensional EEG indicators, including the following steps: Step S1: With the subject wearing a hearing aid configured with the target parameters, play a pre-made natural speech stream as auditory stimulation to the subject.

[0041] Subjects wear hearing aids configured with target parameters, the initial parameters of which can be derived from calculations using traditional pure-tone audiometry fitting formulas (such as NAL-NL2 or DSL v5.0). The speech playback module plays pre-prepared standardized natural speech stream stimulation materials, which are short Chinese sentences containing complete semantic information (such as "Can you hear me clearly now?" or "Please indicate the direction of the sound source") or short story passages (such as a story reading of about 60 seconds).

[0042] The natural speech stream is pre-generated into multiple versions via digital signal processing to simulate listening conditions under different auditory scenarios, including but not limited to: a clear version, a slightly noisy version, a moderately noisy version, and a distorted version. The clear version is the original high-fidelity recording with a signal-to-noise ratio (SNR) of at least 30 dB. The slightly noisy version uses superimposed speech-shaped noise with an SNR of approximately +5 dB. The moderately noisy version uses superimposed multi-speaker background noise (babble noise) with an SNR of approximately 0 dB. The distorted version simulates different frequency response imbalances through band filtering, such as a low-pass cutoff of 2 kHz or a high-pass cutoff of 1 kHz. Each version contains 20-40 short stimuli, each lasting 1.5-3 seconds, with randomized intervals of 2-4 seconds. The start time and version type label of each speech stimulus are recorded synchronously, and trigger markers are sent to the EEG acquisition system to align the data timeline.

[0043] Step S2: Simultaneously acquire the subject's scalp electroencephalogram (EEG) signals by wearing an EEG acquisition module on the subject's head.

[0044] During the stimulus presentation in step S1, the EEG acquisition module synchronously acquires scalp EEG signals through a multi-channel EEG cap worn on the subject's head. The EEG cap is configured with electrodes according to the international 10-20 system, covering at least the following three regions of interest: the frontal lobe (F3, Fz, F4, for theta-band cognitive load analysis), the central lobe (C3, Cz, C4, for auditory evoked potential analysis), and the parietal lobe (P3, Pz, P4, for alpha-band attentional processing analysis). The sampling rate is not less than 250 Hz, preferably 500 Hz or 1000 Hz. The electrode impedance is checked before the experiment and ensured to be less than 10 kΩ (when using conductive gel) or meets the operating requirements of dry electrodes.

[0045] Step S3: Preprocess the acquired EEG signal and extract multi-dimensional features in parallel.

[0046] First, the EEG signal is preprocessed using the following steps: bandpass filtering (1.6-40Hz, fourth-order Butterworth filter, zero-phase bidirectional filtering), notch filtering (50Hz, to remove power frequency interference), independent component analysis (ICA, to identify and remove artifacts from EEG and EMG), bad conduction interpolation (spherical spline interpolation for electrode channels with excessive impedance or abnormal signals), and averaging reference (re-referencing all electrode signals to a common averaging reference).

[0047] like Figure 3 As shown, after preprocessing, the following three types of neural indicators are extracted in parallel from the same EEG signal: Step S3.1: Extraction of perceived clarity indicators.

[0048] Perceptual sharpness index is calculated based on the neural envelope similarity between EEG signals and speech envelopes. The specific steps are as follows: First, a wide-spectrum time-domain envelope is extracted from the speech stimulus signal. The speech signal is filtered using a gammatone filter bank (center frequency covering 80Hz-8kHz, a total of 32 frequency bands). The amplitude envelope is extracted by performing a Hilbert transform on the output of each frequency band. Then, the envelopes of each frequency band are weighted and summed to obtain the wide-spectrum time-domain envelope. Alternatively, the time-domain envelope can also be extracted directly by performing a Hilbert transform on the speech signal.

[0049] Then, the Pearson correlation coefficient between the speech temporal envelope and the signals of each EEG channel was calculated using a sliding time window method. The analysis time window was 0-2 seconds after the stimulus started, with a sliding step of 50ms and a window length of 200ms. The peak value of the correlation coefficient in all time windows was taken as the neural envelope tracking similarity of that channel (the alternative caliber is the mean value above the significance threshold (p<0.05, after FDR correction)).

[0050] Finally, the average similarity of each electrode channel in the forehead and central regions is taken as the perceptual clarity index. The higher the index value, the higher the fidelity of the auditory system's encoding of the temporal structure of speech, that is, "the clearer it is heard".

[0051] As an alternative, neural envelope tracking similarity can also be achieved by calculating the canonical correlation analysis (CCA) coefficient or cross-correlation peak between the speech envelope and the EEG signal.

[0052] Step S3.2: Extraction of cognitive load indicators.

[0053] Cognitive load indices were calculated based on the rate of change of power in the theta band (4-8 Hz) of the prefrontal cortex and the alpha band (8-13 Hz) of the parietal cortex relative to the subject's resting baseline. The details are as follows: First, the theta band power θ_Rest and α band power α_Rest of the subject at resting baseline were recorded. Two minutes of closed-eye resting EEG were collected before the fitting began and calculated using the same method.

[0054] Then, the Welch method was used to calculate the power spectral density (PSD) of each EEG channel, with the following parameters: Hamming window, window length of 2 seconds (corresponding to 500 sampling points at a sampling rate of 250Hz), overlap rate of 50%, and FFT points equal to a power of 2 of the window length (padded with zeros to 2^nextpow2). The PSD of the target frequency band was then subjected to trapezoidal numerical integration to obtain the absolute power (unit: μV²) of that frequency band. The power in the θ band, θ_Power, is the integral of the PSD over the range of 4Hz to 8Hz, i.e. The average values ​​of the prefrontal electrodes F3, Fz, and F4 were taken.

[0055] The α-band power α_Power is the integral of PSD in the range of 8Hz to 13Hz, i.e. The average values ​​of the top leaf electrodes P3, Pz, and P4 are taken.

[0056] Calculate the normalized rate of change of the above values ​​relative to the subject's resting baseline: Δθ = (θ_Power θ_Rest) / θ_Rest Δα = (α_Power α_Rest) / α_Rest An increase in theta power typically reflects an increase in cognitive control resource input ("listening fatigue"), while a decrease in alpha power in the task state relative to the resting state (desynchronization) typically reflects an increase in the degree to which the cortical area is called upon by the task.

[0057] Finally, Δθ and Δα are combined to form a scalar cognitive load index, which is defined as Load = Δθ Δα. Since the α power in the task state is usually lower than that in the resting state (desynchronization, Δα is negative), a larger Load indicates a higher cognitive load.

[0058] Step S3.3: Extraction of emotional comfort index.

[0059] An emotional comfort index was calculated based on the asymmetry of the left and right frontal lobes in EEG alpha waves. Specifically, the asymmetry index of power in the alpha band (8-13Hz) of the left and right frontal lobes was calculated: α_Asymmetry = (R_alpha L_alpha) / (R_alpha + L_alpha) Where R_alpha represents the power of the right frontal lobe (F4) in the alpha band, and L_alpha represents the power of the left frontal lobe (F3) in the alpha band. Positive values ​​indicate right-side alpha dominance (usually associated with negative / avoidant emotions), while negative values ​​indicate left-side alpha dominance (usually associated with positive / approaching emotions). The signed asymmetry index is used as an indicator of emotional comfort; if only the intensity of emotional engagement is needed without regard to direction, the absolute value of the asymmetry index can be used as a supplementary indicator.

[0060] As an alternative, a pre-trained emotion EEG classifier can be used to extract higher-dimensional emotion feature vectors, and these vectors can be dimensionality reduced or their dominant components can be mapped to scalars to serve as an emotional comfort index, in order to maintain the consistency of the three-dimensional input feature space of the fusion model.

[0061] Step S4: Input the three-dimensional feature vector into the pre-trained fusion model, and the fusion model outputs a parameter fit score that represents the rationality of the current hearing aid parameter configuration.

[0062] like Figure 4 As shown, the fusion evaluation module receives the three-dimensional feature vector output by step S3: [perceptual clarity index, cognitive load index, emotional comfort index].

[0063] First, each indicator is normalized to a uniform numerical range (e.g., 0-1) to eliminate dimensional differences. Normalization can be performed using min-max normalization or z-score standardization, with reference values ​​based on population distributions in clinical databases or the subject's own resting-state baseline.

[0064] Then, the normalized 3D feature vectors are input into the pre-trained fusion model. The fusion model can take any of the following forms, and can be flexibly selected according to the computing power and data scale of the actual deployment scenario: (a) Weighted linear fusion model: P_suitability = w1× Clarity_norm + w2× (1 Load_norm) + w3 × Comfort_norm, where w1, w2, and w3 are the weights of each dimension, satisfying w1 + w2 + w3 = 1. The weights can be determined in two ways: (i) directly deploying the prior weights based on the effect sizes of each indicator in existing literature (such as Cohen's d or standardized effect size indicators such as η²), for example, using neural envelope tracking similarity as the main indicator (w1 = 0.5), and θ / α power (w2 = 0.3) and α asymmetry (w3 = 0.2) as auxiliary correction indicators; (ii) optimizing the weight allocation through supervised training using labeled data collected clinically (multi-dimensional subjective scores configured by subjects for each parameter plus corresponding EEG feature vectors).

[0065] (b) Logistic regression model: Maps the probability estimate of parameter fit to 0-100 points.

[0066] (c) Lightweight Multilayer Perceptron (MLP): 3 neurons in the input layer (corresponding to 3D features), 1-2 hidden layers (8-16 neurons per layer, ReLU activation), and 1 neuron in the output layer (Sigmoid activation, mapped to 0-100 points).

[0067] All three model types share the same three-dimensional input feature space [perceptual clarity, cognitive load, emotional comfort], and can be flexibly selected according to the actual amount of data, computing power, and deployment complexity.

[0068] The training data for the fusion model comes from a labeled dataset collected clinically. Each sample contains: a three-dimensional EEG feature vector collected under a specific hearing aid parameter configuration, and the wearer's multi-dimensional subjective rating of the configuration (clarity rating, effort rating, comfort rating, and overall satisfaction rating).

[0069] In addition, the appropriate hierarchical fusion model can be selected for parameter fit assessment based on the subject's age, type and degree of hearing loss.

[0070] Step S5: Output the parameter fit score to guide the adjustment of hearing aid parameters.

[0071] The parameter fit score (0-100 points) output by the fusion model and the sub-scores of each dimension are presented to the audiologist or automated fitting system in a visual form.

[0072] If the parameter fit score reaches the preset threshold, it indicates that the current hearing aid parameter configuration is reasonable and the fitting is complete; the preset threshold can be set according to the clinical calibration results, and 70 points is used as an example.

[0073] If the parameter fit score is lower than the preset threshold, the audiologist or automated fitting system will determine the source of the problem based on the sub-scores of each dimension and adjust the hearing aid parameters accordingly. If the perceived sharpness index is too low, adjust the gain (especially the high-frequency gain) and the frequency response curve; If the cognitive load index is too high (θ increases significantly / α decreases significantly), then adjust the noise reduction intensity, compression ratio, and compression inflection point. If the emotional comfort index is positive (i.e., the α asymmetry index is positive, reflecting right-side α dominance, negativity / discomfort), then adjust the maximum output sound pressure level (MPO) and overall volume. Here, gain is the overall or crossover amplification factor of the input signal (in dB); compression ratio is the ratio of the compressor's input dynamic range to its output dynamic range; compression inflection point is the input sound pressure level at which the compressor begins to function; maximum output sound pressure level (MPO) is the upper limit of the maximum sound pressure level allowed by the hearing aid; frequency response curve shows the distribution of gain at different frequencies; and noise reduction intensity is the degree of suppression of noise components.

[0074] After completing the parameter adjustment, return to step S1 and repeat the evaluation process of S1-S5 with the new parameter configuration to form a closed-loop iterative optimization of "evaluation → adjustment → re-evaluation" until the parameter fit score reaches the preset threshold.

[0075] To verify the technical feasibility of the above method and system, a proof-of-concept experiment was conducted.

[0076] Experimental objective: To verify the engineering feasibility of the technical route of "natural speech stream-induced EEG → parallel extraction of multidimensional features → fusion evaluation".

[0077] Experimental overview: such as Figure 5 As shown, a 19-channel standard 10-20 EEG cap with a sampling rate of 250Hz was used to conduct a proof-of-concept experiment on natural speech-evoked EEG in subjects. Subjects listened to 39 short natural Chinese sentences, categorized into three conditions based on speech intelligibility: High condition (clear speech, 15 sentences); Mid condition (medium intelligibility, 12 sentences); and Low condition (noisy or low intelligibility, 12 sentences). Each speech sentence lasted approximately 2.5 seconds, with randomized stimulus intervals. After preprocessing with bandpass filtering (1.6Hz to 40Hz), 50Hz notch filtering, ICA artifact removal, and average reference, the EEG data were extracted within a 0-2 second time window following the stimulus start for frequency band power analysis. Welch power spectral density estimation combined with trapezoidal numerical integration was used to calculate the absolute power of each frequency band.

[0078] Cognitive load dimension indicators observation as follows Figure 6 , Figure 7 As shown: The theta band power of electrodes F3, Fz, and F4 in the prefrontal region under three conditions are: 2.87µV² for High condition, 2.56µV² for Mid condition, and 2.59µV² for Low condition.

[0079] The α-band power of electrodes P3, Pz, and P4 in the top leaf region under the three conditions are: 7.30µV² for High condition, 7.19µV² for Mid condition, and 6.62µV² for Low condition.

[0080] The α power decreases by approximately 9.3% under Low conditions compared to High conditions.

[0081] This indicates that the cognitive load of the subjects increased significantly under low-resolution listening conditions.

[0082] Emotional comfort dimension indicators observation: left and right frontal alpha band power asymmetry index, i.e. (F4 power) F3 power) / (F4 power+F3 power), under three conditions, are: High condition 0.048, Mid condition 0.190, Low condition +0.007.

[0083] Time-domain response observation as Figure 8 As shown: The auditory event-related potential waveforms of the Cz electrode in the central region show that there are modulation differences in the N1-P2 components after stimulation under the High and Low conditions (N1 is the negative auditory cortex component about 100 ms after stimulation, and P2 is the positive component about 200 ms after stimulation), indicating that speech stimulation under different clarity conditions can induce distinguishable temporal EEG responses.

[0084] The above experiments confirm that the entire technical chain of natural speech stream stimulation, EEG acquisition, preprocessing, and multidimensional feature extraction is feasible in engineering, and each link can operate stably. The three-dimensional feature fusion evaluation architecture proposed in this invention has been verified at the technical implementation level.

[0085] In summary, this invention replaces the subjective feedback-based assessment in the hearing aid fitting process with an objective, quantitative assessment based on electroencephalogram (EEG) signals, fundamentally solving the problem of traditional fitting methods' high dependence on the wearer's cooperation and expressive abilities. EEG signals are the auditory system's automatic physiological response to external stimuli, unaffected by the wearer's subjective will, expressive ability, or cognitive state. Therefore, this method can cover groups that traditional methods struggle to effectively serve, such as children, the elderly, and those with cognitive impairments.

[0086] Simultaneously, this invention organically integrates three independent dimensions of EEG indicators: perceptual clarity (neural envelope tracking), cognitive load (θ / α power), and emotional comfort (α asymmetry). These correspond to the three core aspects of the listening experience: "clear hearing," "easy listening," and "comfortable listening," respectively, thus providing a more complete picture of the wearer's overall listening experience. This invention replaces the pure tone or click-sound stimulation used in traditional methods with natural speech streams, upgrading the assessment environment from "laboratory pure tone testing" to "simulated real-life dialogue scenarios," which can induce a complete hierarchical auditory processing chain in the brain.

[0087] Furthermore, the system of this invention can be independent of specific EEG acquisition hardware and can be expanded to all scenarios from professional clinical use to consumer-grade home fitting.

[0088] The above description is merely an embodiment of the present invention and does not constitute any limitation on the technical scope of the present invention. Therefore, any minor modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A hearing aid parameter fit assessment system based on multi-dimensional EEG index fusion, characterized in that, include: The speech playback module is used to play pre-made natural speech stimulation materials when the subject is wearing a hearing aid; The EEG acquisition module is used to simultaneously acquire the subject's scalp electroencephalogram (EEG) signals while playing natural speech stimulation materials; The feature extraction module is used to extract three types of neural indicators in parallel from EEG signals, including perceptual clarity indicators, cognitive load indicators, and emotional comfort indicators. The perceived clarity index is calculated based on the neural envelope tracking similarity between the EEG signal and the speech envelope. The cognitive load index was calculated based on the rate of change of the theta band power in the prefrontal region and the alpha band power in the parietal region relative to the subject's resting baseline. The emotional comfort index is calculated based on the asymmetry between the left and right frontal lobes of the EEG alpha wave. The fusion assessment module has a built-in pre-trained fusion model that receives perceptual clarity indicators, cognitive load indicators, and emotional comfort indicators, and outputs a parameter fit score from 0 to 100. The results output module is used to present the parameter fit score and the scores of each dimension to the audiologist or fitting system in a visual form.

2. The hearing aid parameter fit assessment system based on multi-dimensional EEG index fusion according to claim 1, characterized in that, The natural speech stream is a short Chinese sentence or paragraph containing complete semantic information, including but not limited to: clear version, light noise version, medium noise version and distorted version, to simulate listening conditions under different auditory scenarios.

3. The hearing aid parameter fit assessment system based on multi-dimensional EEG index fusion according to claim 1, characterized in that, The EEG acquisition module is a head-mounted multi-channel EEG cap, which covers at least the following electrode locations: the frontal region includes F3, Fz, and F4 electrodes; the central region includes C3, Cz, and C4 electrodes; and the parietal lobe region includes P3, Pz, and P4 electrodes.

4. The hearing aid parameter fit assessment system based on multi-dimensional EEG index fusion according to claim 1, characterized in that, The neural envelope tracking similarity is calculated as follows: a wide-spectrum time-domain envelope is extracted from the speech signal, and multi-band envelopes are extracted by Hilbert transform or gammatone filter bank and then weighted summed. The canonical correlation analysis coefficient or cross-correlation peak value between the envelope and the EEG signal is then calculated as a perceptual sharpness index. The cognitive load index is calculated as follows: the power spectral density (PSD) of each EEG channel is calculated using the Welch method, and the absolute power value is obtained by trapezoidal numerical integration of the PSD of the target frequency band. The power in the θ band, θ_Power, is the integral of the PSD over the range of 4Hz to 8Hz, i.e. And take the average value of the prefrontal electrodes F3, Fz, and F4; The α-band power α_Power is the integral of PSD in the range of 8Hz to 13Hz, i.e. And take the average value of the top leaf electrode P3, Pz, P4; Calculate the normalized rate of change of the above values ​​relative to the subject's resting baseline: Δθ = (θ_Power θ_Rest) / θ_Rest Δα = (α_Power a_Rest) / a_Rest Where Δθ is the normalized rate of change of the θ band power relative to the subject's resting baseline, Δα is the normalized rate of change of the α band power relative to the subject's resting baseline, and θ_Rest and α_Rest are the θ band power and α band power at the subject's resting baseline, respectively. Finally, Δθ and Δα are combined to form a scalar cognitive load index, which is defined as Load = Δθ Δα; The emotional comfort index is the asymmetry index of the α-band power of the left and right frontal lobes, and is defined as emotional comfort index α_Asymmetry=(R_alpha) L_alpha) / (R_alpha+L_alpha), where R_alpha is the power of the right frontal lobe F4 electrode in the α band, and L_alpha is the power of the left frontal lobe F3 electrode in the α band; The fusion model can be any of the following: a weighted linear fusion model, a logistic regression model, or a lightweight multilayer perceptron neural network.

5. The hearing aid parameter fit assessment system based on multi-dimensional EEG index fusion according to claim 1, characterized in that, The system also includes: The signal preprocessing module is used to perform bandpass filtering, notch filtering, independent component analysis for artifact removal, and average reference processing on the raw EEG signal. The feature extraction module extracts three types of neural indicators in parallel from the preprocessed EEG signal.

6. A method for assessing hearing aid parameter fit based on multi-dimensional EEG index fusion, characterized in that, Includes the following steps: Step S1: With the subject wearing a hearing aid configured with the target parameters, play a pre-made natural speech stream as auditory stimulation to the subject; Step S2: Simultaneously collect the subject's scalp electroencephalogram (EEG) signals by wearing an EEG acquisition module on the subject's head; Step S3: Extract three types of neural indicators in parallel from the acquired EEG signals, including: The perceived clarity index is calculated based on the neural envelope tracking similarity between the EEG signal and the speech envelope. Cognitive load indices were calculated based on the rate of change of the theta band power in the prefrontal region and the alpha band power in the parietal region relative to the subject's resting baseline. The emotional comfort index was calculated based on the asymmetry between the left and right frontal lobes of the EEG alpha wave. Step S4: Input the three types of neural indicators into the pre-trained fusion model, and the fusion model outputs a parameter fit score that represents the rationality of the current hearing aid parameter configuration; Step S5: Output the parameter fit score to guide the adjustment of hearing aid parameters: Determine whether the preset threshold has been reached based on the parameter fit score; if not, determine the source of the problem based on the sub-scores of each dimension, and adjust one or more parameters of the hearing aid accordingly. The parameters include at least one of gain, compression ratio, compression inflection point, maximum output sound pressure level, frequency response curve, and noise reduction intensity. Repeat steps S1 to S5 until the parameter fit score reaches the preset threshold.

7. The hearing aid parameter fit assessment method based on multi-dimensional EEG index fusion according to claim 6, characterized in that, The natural speech stream in step S1 is a short Chinese sentence or paragraph containing complete semantic information, including but not limited to: clear version, light noise version, medium noise version and distorted version, to simulate listening conditions under different auditory scenarios.

8. The hearing aid parameter fit assessment method based on multi-dimensional EEG index fusion according to claim 6, characterized in that, In step S3 The neural envelope tracking similarity is calculated as follows: a wide-spectrum time-domain envelope is extracted from the speech signal, and multi-band envelopes are extracted by Hilbert transform or gammatone filter bank and then weighted summed. The canonical correlation analysis coefficient or cross-correlation peak value between the envelope and the EEG signal is then calculated as a perceptual sharpness index. The cognitive load index is calculated as follows: the power spectral density (PSD) of each EEG channel is calculated using the Welch method, and the absolute power value is obtained by trapezoidal numerical integration of the PSD of the target frequency band. The power in the θ band, θ_Power, is the integral of PSD over the range of 4Hz to 8Hz, i.e. And take the average value of the prefrontal electrodes F3, Fz, and F4; The α-band power α_Power is the integral of PSD in the range of 8Hz to 13Hz, i.e. And take the average value of the top leaf electrode P3, Pz, P4; Calculate the normalized rate of change of the above values ​​relative to the subject's resting baseline: Δθ = (θ_Power θ_Rest) / θ_Rest Δα = (α_Power a_Rest) / a_Rest Where Δθ is the normalized rate of change of the θ band power relative to the subject's resting baseline, Δα is the normalized rate of change of the α band power relative to the subject's resting baseline, and θ_Rest and α_Rest are the θ band power and α band power at the subject's resting baseline, respectively. Finally, Δθ and Δα are combined to form a scalar cognitive load index, which is defined as Load = Δθ Δα; The emotional comfort index is the asymmetry index of the α-band power of the left and right frontal lobes, and is defined as emotional comfort index α_Asymmetry=(R_alpha) L_alpha) / (R_alpha+L_alpha), where R_alpha is the power of the right frontal lobe F4 electrode in the α band, and L_alpha is the power of the left frontal lobe F3 electrode in the α band.

9. The method for assessing hearing aid parameter suitability based on multi-dimensional EEG index fusion according to claim 6, characterized in that, The fusion model in step S4 can be any of the following forms: a weighted linear fusion model, a logistic regression model, or a lightweight multilayer perceptron neural network.

10. The method for evaluating hearing aid parameter suitability based on multi-dimensional EEG index fusion according to claim 6, characterized in that, The EEG acquisition module mentioned in step S2 is a head-mounted multi-channel EEG cap, and the electrodes of the EEG cap cover at least the frontal region, central region and parietal lobe region.

Images