Method for evaluating sound source localization ability on basis of EEG signal

Abstract

Disclosed in the present invention is a method for evaluating a sound source localization ability on the basis of an EEG signal. The method comprises the following steps: constructing a standard model; on the basis of an evaluation method for the standard model, performing prediction by means of the standard model to obtain the sound source localization status of a patient; and establishing a sound source localization ability impairment level evaluation standard. The step of constructing a standard model comprises: constructing a sound source localization test experimental paradigm; collecting large-scale EEG signal data from normal-hearing subjects; preprocessing an EEG signal to obtain channel data with the maximum ACC signal under each sound source angle change of the EEG signal to serve as a model input vector; and constructing, training and evaluating the standard model, and generating the standard model by means of big data training. In the present invention, the response of a subject to a sound source angle change is decoded on the basis of ACC, and the mapping of an EEG signal to the sound source angle change is established by means of a machine learning model, thereby providing a more objective and reliable method for evaluating the sound source localization ability of a patient.

Description

A method for evaluating sound source localization capability based on EEG signals Technical Field

[0001] This invention belongs to the field of medical device technology, and specifically relates to a method for evaluating the sound source localization capability based on EEG signals. Background Technology

[0002] Sound localization, a fundamental perceptual ability, aids in daily communication, target finding, and personal safety. Reduced or lost sound localization hinders speech comprehension and normal life, such as the inability to avoid oncoming vehicles. Currently, sound localization testing has been incorporated into hearing assessment protocols based on the WHO's new classification framework, providing important guidance for assessing hearing loss. Sound localization primarily includes horizontal and vertical localization. The interaural time difference (ITD) and interaural level difference (ILD) of sound signals arriving at both ears are the main cues relied upon for horizontal sound localization, while vertical sound localization is mainly related to the specific spectral signals generated by sound bypassing the head and auricle. In recent years, with increased attention to hearing health, improved auditory rehabilitation, and the widespread use of hearing aids and cochlear implants, researchers have begun to focus on assessing and researching sound localization abilities in pathological states or during auditory rehabilitation to restore patients' binaural hearing. However, currently, there is still a lack of medical institutions in China conducting such clinical tests, and the overall testing and evaluation of sound source localization is progressing slowly, which contradicts the growing clinical demand. Therefore, popularizing sound source localization ability testing and establishing objective evaluation technical standards are particularly important for promoting the development of the domestic audiology industry as a whole.

[0003] Currently, sound localization ability is primarily assessed clinically using auditory scales and behavioral tests. Scales related to sound localization include the Spatial Hearing Questionnaire (SHQ) and the Speech, Spatial and Qualities of Hearing Scale (SSQ). However, while these scales help understand the spatial auditory performance of patients, especially young children who cannot cooperate with other forms of assessment, in daily life, the results are subjective feedback and lack objectivity. Furthermore, there is currently a lack of translated or Chinese versions of these scales. Behavioral tests for sound localization mainly include two categories: source azimuth discrimination tests and source azimuth identification tests. The former tests the subject's spatial discrimination ability for sound sources, specifically the smallest angular interval between two sound sources that the subject can just distinguish, called the minimum audible angle (MAA). The latter tests the accuracy of the subject's sound source localization. Sound localization behavioral tests can reflect a subject's spatial hearing ability relatively objectively. However, these behavioral tests often rely on the subject's subjective participation and behavioral feedback. For infants, young children, or hearing-impaired individuals who have difficulty understanding the test rules, are unwilling or unable to concentrate on the test and provide correct feedback, it is particularly necessary to establish an objective and reliable clinical tool for diagnosing and assessing their sound localization ability.

[0004] Recent studies have shown that the Acoustic Change Complex (ACC) in electroencephalogram (EEG) signals has promising applications in auditory perception assessment. ACC can characterize the neural response of the auditory cortex to changes in sound signals (such as frequency, intensity, and duration), reflecting the auditory cortex's ability to distinguish sound changes and the brain's ability to process and extract features from speech and other sound signals. Furthermore, ACC responses show a strong correlation with behavioral discrimination abilities. Therefore, ACC holds promise as an effective means of objectively assessing sound discrimination or speech perception abilities in clinical practice. Some researchers have studied the relationship between ACC signals and changes in the horizontal direction of sound sources in normal adolescents. They found that changes in the angle of the sound source in the horizontal direction can successfully induce ACC signals. Moreover, as the angle of the sound source increases, the amplitude of the ACC complex N1-P2 increases, accompanied by a decrease in the delay of N1 and P2. This suggests that ACC could also be a method for assessing and diagnosing sound source localization abilities.

[0005] Summary of the Invention

[0006] To address the aforementioned issues, this invention provides a method for assessing sound source localization ability based on EEG signals. Based on ACC decoding of the subject's response to changes in sound source angle, a machine learning model is used to establish a mapping between EEG signals and changes in sound source angle. This provides an objective and reliable method for clinically assessing the sound source localization ability of hearing-impaired patients, further improving the rehabilitation training effect for these patients.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for evaluating sound source localization capability based on EEG signals, comprising the following steps:

[0008] S10: Standard model construction;

[0009] S20: A standard model-based assessment method that uses a standard model to predict and obtain the patient's sound source localization.

[0010] S30: Establish a standard for assessing the level of impairment in sound source localization ability;

[0011] Step S10 includes the following steps:

[0012] S101: Construction of experimental paradigm for sound source localization test;

[0013] S102: EEG signal big data acquisition, based on the sound source localization test experimental paradigm, to obtain a large dataset of EEG signals induced by the sound source angle change of normal hearing subjects.

[0014] S103: EEG signal preprocessing, used to remove artifacts from each EEG signal in the EEG signal big data set to obtain an EEG signal without artifacts; and to establish mapping data between the EEG signal without artifacts and the changes in sound source angle, to obtain the channel data with the largest ACC amplitude signal under each change in sound source angle of the EEG signal, and the channel data is used as the model input vector; all the obtained model input vectors constitute the preprocessed EEG signal big data set;

[0015] S104: Construction, training and evaluation of the standard model. Construct a machine learning model to predict changes in the sound source from the perspective of EEG signals. Train the machine learning model using the large dataset of EEG signals preprocessed in step S103, and generate a standard model corresponding to normal hearing subjects.

[0016] Preferably, the specific steps for constructing the sound source localization test paradigm in step S101 are as follows:

[0017] S1011: Construct an arc-shaped speaker array: Construct an arc-shaped speaker array centered on the subject. The arc-shaped speaker array includes multiple speakers arranged on both sides of the subject's front direction, extending within a 180-degree range. One of the speakers is located directly in front of the subject, and the remaining speakers are symmetrically placed along the subject's front direction, with unevenly distributed positions.

[0018] S1012: Set the speaker to play a stimulating sound;

[0019] S1013: After the subject is prepared, the subject's EEG signal is collected for several seconds as a baseline signal; then, a set of two speakers set at symmetrical angles is selected to play the stimulus sound from one side to the other, which is a sound source angle change; every 1000-2000ms, a set of speakers is changed to play the stimulus sound; the subject's EEG signal is continuously collected during the process, and then amplified and saved; there are a total of Y sound source angle changes, each sound source angle change is randomly performed X times, and the playback order of the X*Y sound source angle changes is randomly set by the program control.

[0020] Preferably, in step S1011, the arc-shaped speaker array includes 19 speakers, with a total of 10 different sound source angle changes, and each sound source angle change is performed at least 6 times.

[0021] Preferably, in step S101, the stimulus sound played by the speaker is a broadband stimulus signal obtained by bandpass filtering white noise at 125-6000Hz.

[0022] Preferably, in the EEG signal preprocessing of step S103, artifact removal of the EEG signal is performed using a method based on multidimensional spatiotemporal frequency correlation component decomposition and temporal convolutional network (TCN).

[0023] Preferably, the artifact removal in the EEG signal preprocessing of step S103 includes the following steps:

[0024] S1031: EEG signal input and decomposition: The N-channel EEG signal is decomposed using the fast multidimensional empirical mode decomposition method to obtain the corresponding multidimensional intrinsic mode function (IMF) and trend term. Assuming that each channel is decomposed into K IMFs, a set of IMFs is obtained, which contains N*K IMFs.

[0025] S1032: Intrinsic Mode Unmixing: Based on the canonical correlation analysis (CCA) method, the IMF obtained in S1031 is analyzed for relevant components to obtain N*K projection matrices;

[0026] S1033: Component screening: The TCN is used to automatically extract and classify the components in the projection matrix to obtain P brain power sources and Q pseudo-trace sources, satisfying P+Q=N*K;

[0027] S1034: Artifact Removal: Set the column values ​​of the CCA unmixing matrix corresponding to the Q artifact sources to 0, and then use inverse CCA to calculate the clean IMF corresponding to the EEG signal of each channel;

[0028] S1035: Signal reconstruction and output: Reconstruct the pure IMF corresponding to each channel to obtain an N-dimensional EEG signal without artifacts.

[0029] Preferably, step S103, which involves obtaining the channel data with the largest ACC signal amplitude for each change in the EEG signal source angle, specifically includes the following steps:

[0030] S1036: Bandpass filtering is performed on the N-dimensional EEG signal without artifacts, and the sampling rate is reduced;

[0031] S1037: Divide the N-dimensional, artifact-free EEG signal with reduced sampling rate into X*Y segments, with each segment corresponding to the angle change of each sound source;

[0032] S1038: The average of the segment data after baseline correction for X playback cycles of each sound source angle change is used to obtain the ACC signal of each EEG signal channel under each sound source angle change.

[0033] S1039: Select the channel data with the largest ACC signal amplitude under each change in sound source angle as the model input vector for model training.

[0034] Preferably, a long short-term memory network model is used as the machine learning model. The preprocessed EEG signal dataset obtained in step S103 is divided into a training set (80%) and a test set (20%). The machine learning model is trained using the training set data to obtain a standard model constructed from data of people with normal hearing. The standard model is evaluated through the test set. The output of the standard model is Y categories, corresponding to Y kinds of sound source angle changes between 0 and 180 degrees.

[0035] Preferably, step S20 specifically includes the following steps: by constructing the sound source localization test experimental paradigm, the patient's EEG signal is collected, and after EEG signal preprocessing, the channel data with the maximum ACC signal corresponding to each sound source angle change is obtained. The obtained multiple channel data are used as the model input vector of the standard model in sequence, and the patient's sound source localization ability is obtained by predicting through the standard model.

[0036] Preferably, in step S20, which involves obtaining the patient's sound source localization ability through standard model prediction, a standard curve is calculated with the standard model prediction value as the vertical axis and the actual sound source angle change value as the horizontal axis; and the patient's sound source localization curve is obtained by predicting the patient's model input vector through the standard model; the ratio of the AUC of the patient's sound source localization curve to the AUC of the standard curve is used as an indicator to measure the patient's sound source localization ability.

[0037] The beneficial effects of this invention are as follows:

[0038] 1. This invention provides a method for assessing sound source localization ability based on EEG signals. Based on ACC decoding of the subject's response to changes in sound source angle, a machine learning model is used to establish a mapping from EEG signals to changes in sound source angle. This provides an objective method for clinically assessing the sound source localization ability of hearing-impaired patients, yielding more objective and reliable results. It eliminates the need for hearing-impaired patients to be highly focused on the test and provide active feedback, making it more objective and efficient, and further improving the rehabilitation training effect for hearing-impaired patients.

[0039] 2. The present invention provides a method for assessing sound source localization ability based on EEG signals. This method establishes a sound source localization experimental paradigm, collects EEG signals from normal hearing subjects to form a large dataset, and uses this dataset to establish a standard model for normal hearing subjects. Based on this standard model, the sound source localization ability of patients is determined, which is more accurate, faster, and more objective.

[0040] 3. The present invention provides a method for evaluating sound source localization capability based on EEG signals. By constructing an effective experimental paradigm for sound source localization testing, the method limits the types of stimuli used and the types of changes in sound source angle, thereby further improving the accuracy of the test.

[0041] 4. The present invention provides a method for evaluating sound source localization capability based on EEG signals, which removes artifacts from EEG signals and reduces the impact of other internal electrophysiological noise on obtaining accurate ACC signals, thereby effectively improving the accuracy and reliability of model prediction.

[0042] 5. The present invention provides a method for evaluating sound source localization ability based on EEG signals. By establishing a standard curve and a patient sound source localization curve, and using the ratio of the AUC of the patient's sound source localization curve to the AUC of the standard curve as a standard for measuring the patient's sound source localization, the calculation is simple and convenient, and the sound source localization of hearing-impaired patients is obtained more accurately, quickly and objectively. Attached Figure Description

[0043] Figure 1 is a flowchart of a method for evaluating sound source localization capability based on EEG signals according to a specific embodiment of the present invention;

[0044] Figure 2 is a schematic diagram of step S101 in a method for evaluating sound source localization capability based on EEG signals according to a specific embodiment of the present invention.

[0045] Figure 3 is a schematic diagram of step S101 in a method for evaluating sound source localization capability based on EEG signals according to a specific embodiment of the present invention.

[0046] Figure 4 shows a standard curve and a patient sound source localization curve of a method for evaluating sound source localization ability based on EEG signals according to a specific embodiment of the present invention.

[0047] Figure 5 is a diagram of the sound source localization ability impairment level assessment standard of a method for assessing sound source localization ability based on EEG signals according to a specific embodiment of the present invention. Detailed Implementation

[0048] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0049] Referring to Figure 1, a method for evaluating sound source localization capability based on EEG signals is shown, which specifically includes the following steps:

[0050] S10: Standard model construction;

[0051] S20: A standard model-based assessment method that uses a standard model to predict and obtain the patient's sound source localization.

[0052] S30: Establish a standard for assessing the level of impairment in sound source localization ability.

[0053] Step S10 includes the following steps:

[0054] S101: Construction of experimental paradigm for sound source localization test;

[0055] S102: EEG signal big data acquisition, used to acquire EEG signals induced by sound stimulation of multiple normal hearing subjects under the influence of changes in the angle of spatial sound sources using the sound source localization test experimental paradigm, and to form a large dataset of EEG signals.

[0056] S103: EEG signal preprocessing, used to remove artifacts from each EEG signal in the EEG signal big data set to obtain an EEG signal without artifacts; and to establish mapping data between the EEG signal without artifacts and the changes in sound source angle, to obtain the channel data with the largest ACC signal amplitude under each change in sound source angle of the EEG signal, and the channel data is used as the model input vector; all the obtained model input vectors constitute the preprocessed EEG signal big data set;

[0057] S104: Construction, training and evaluation of the standard model. Construct a machine learning model to predict changes in the sound source angle from EEG signals. Train the machine learning model using the preprocessed EEG signal dataset obtained in step S103 to generate a standard model corresponding to normal hearing subjects.

[0058] Referring to Figures 1 and 2, the construction of the sound source localization test paradigm for S101 specifically includes the following steps:

[0059] S1011: Constructing an Arc-Shaped Speaker Array: In a soundproof room (with a noise floor below 30 dBA, and preferably with electromagnetic shielding), construct an arc-shaped speaker array centered on the subject. The arc-shaped speaker array can include multiple speakers positioned on either side of the subject's front, extending within a 180-degree range. One speaker is located directly in front of the subject, while the remaining speakers are symmetrically placed along the subject's front, with unevenly distributed positions. The number of speakers can be adjusted appropriately during the actual experiment; theoretically, more speakers are better, resulting in more accurate test results. However, this requires a larger test radius or a smaller speaker diameter. Further, the arc-shaped speaker array can include 19 speakers with angles of 0°, ±2°, ±5°, ±10°, ±15°, ±30°, ±45°, ±60°, ±75°, and ±90°. The radius of the arc-shaped speaker array is 1–2 m; in this specific embodiment, the radius is set to 1.2 m. In addition, to reduce the influence of visual aids, the soundproof room can be covered with black curtains to keep the testing environment dark so that the subjects can judge the location of the sound source by hearing alone.

[0060] S1012: Setting the Stimulus Sound Played by the Speaker: The stimulus sound can be a broadband stimulus signal obtained by bandpass filtering white noise at 125–6000 Hz, with a duration of 200–1000 ms; avoid selecting stimulus sounds between 1500–3000 Hz, which have the lowest horizontal azimuth localization accuracy. When playing the stimulus sound, control the sound pressure level at the center of the sound field to 65–70 dBA. Other stimulus sounds can also be used, such as pure tones, other filtered tones, white noise, speech, music, and complex tones, but in principle, it is best to use sounds that help the subject distinguish changes in the sound source angle. Generally speaking, complex tones are more effective than pure tones. Procedure for Collecting EEG Signals from Subjects: EEG signal collection follows standard operating procedures, using a 10–20 international standard lead system. Electrodes are installed on the subject's scalp to collect EEG signals, which are then amplified and stored. High-quality signal acquisition is crucial for obtaining accurate and reliable EEG data and for the final assessment of sound source localization ability. Therefore, various methods and experimental techniques are needed to ensure the quality of signal acquisition. For example, the EEG cap should be kept clean and maintained, and subjects should be asked to clean their scalp before wearing the cap. Furthermore, the number of leads can be determined based on specific circumstances; 64 or 128 leads are recommended. Fewer leads can also be used, but they must include electrodes for prefrontal and central brain regions where ACC signals are relatively strong, such as Cz, Fz, and FCz. Leads for electrooculography (EOG) are also recommended to be retained to facilitate the identification and removal of eye movement artifacts in subsequent EEG data analysis.

[0061] S1013: The subject sits in a chair in the center of the sound field, wearing an EEG cap and facing 0 degrees, keeping their eyes closed to reduce interference from eye movements on the EEG signal. Simultaneously, the body and head remain still to minimize motion artifacts. After preparation, the subject initiates the experiment using a keyboard, mouse, or gamepad. The timing of the sound stimulation and speaker selection throughout the experiment are controlled by an automated program.

[0062] After the subjects are prepared, the experiment begins. First, EEG signals from the subjects are collected for 1000–2000 ms as a baseline signal. Then, two symmetrically positioned speakers play stimulus sounds from one side to the other (each speaker's playback duration is set to 500 ms), representing a change in sound source angle (e.g., selecting two speakers at ±15° intervals results in a 30-degree change in sound source angle). Afterward, at 1000–2000 ms intervals, two other symmetrical speakers are randomly selected to play stimulus sounds. EEG signals from the subjects are continuously collected throughout the process. After all playback conditions are met, a voice message is played from the 0-degree speaker to indicate the end of the experiment, and EEG signal acquisition is stopped and automatically saved to the local computer. There are Y types of sound source angle changes, each randomly performed X times, and the playback order of the X*Y sound source angle changes is randomly set. According to the description in step S1011, Y can be 10. During the experiment, each sound source angle change can be performed at least 6 times (3 times from left to right, 3 times from right to left). In addition, to collect training data with a 0-degree change in sound source angle, six different speakers were randomly selected from 19 speakers (for the sake of balanced training data), and each speaker played a stimulus sound for 1000ms (i.e., no change in sound source angle). All experimental settings for sound playback conditions were controlled by a program and randomized. The saved EEG signal file automatically added corresponding event flags to the start time of the sound emitted by the speakers under each test condition (the flags were pre-defined to distinguish specific changes in sound source angle and remained consistent across all experiments) for subsequent data analysis. The total number of times the sound source angle changed was played was 6*9+6=60, including the randomly set silent sections. The entire experimental process took 2-3 minutes.

[0063] S102: EEG signal big data acquisition, specifically including the following steps:

[0064] At least 1,000 participants with normal hearing were recruited nationwide. These participants were generally aged 14-60 and had hearing thresholds of less than 20 dB HL in the range of 125-8000 Hz.

[0065] Each subject with normal hearing was instructed to perform the above-mentioned sound source localization test experimental paradigm. EEG signals induced by the change in the angle of the sound source were collected simultaneously and saved to form a large dataset of EEG signals.

[0066] S103: EEG signal preprocessing, specifically including the following steps:

[0067] First, the multidimensional EEG data is rereferenced using the average of two mastoid electrode data or the average of all electrodes (excluding the electrooculogram electrode) as reference data, followed by artifact removal. Because EEG records weak brain electrical activity on the scalp, it is easily contaminated by various noises, including internal electrophysiological noise such as electrooculogram and electromyogram, and noise generated by external equipment such as power line noise. Removing noise interference from the EEG signal is crucial for obtaining accurate ACC signals. To reduce the impact of interfering signals on normal EEG signals and improve the accuracy and reliability of standard model predictions, artifact removal of the EEG signal is first performed using a method based on multidimensional spatiotemporal frequency correlation component decomposition and a temporal convolutional network (TCN). Artifact removal of the EEG signal for each normally hearing subject includes the following steps:

[0068] S1031: EEG signal input and decomposition: The N-channel (lead) EEG signal is decomposed using Fast Multi-dimensional Empirical Mode Decomposition (FMEMD) to obtain the corresponding Intrinsic Mode Function (IMF) and trend term. Assuming that each channel is decomposed into K IMFs, a set of IMFs is obtained, which contains N*K IMFs.

[0069] S1032: Intrinsic Mode Unmixing: Based on the canonical correlation analysis (CCA) method, the IMF obtained in S1031 is analyzed for relevant components to obtain N*K projection matrices;

[0070] S1033: Component screening: The TCN is used to automatically extract and classify the components in the projection matrix to obtain P brain power sources and Q pseudo-trace sources, satisfying P+Q=N*K;

[0071] S1034: Artifact Removal: Set the column values ​​of the CCA unmixing matrix corresponding to the Q artifact sources to 0, and then use inverse CCA to calculate the pure IMF corresponding to the EEG signal of each channel;

[0072] S1035: Signal reconstruction and output: Reconstruct the pure IMF corresponding to each channel to obtain an N-dimensional EEG signal without artifacts.

[0073] The above-mentioned artifact removal methods can also be achieved through other means, such as the more commonly used Independent Component Analysis (ICA).

[0074] The specific steps for obtaining the maximum channel data of the ACC signal under each change in the sound source angle of the EEG signal in step S103 are as follows:

[0075] S1036: Perform 0.1-30Hz bandpass filtering on the N-dimensional EEG signal without artifacts obtained in step S1035, and reduce the sampling rate to 100Hz;

[0076] S1037: Based on the event marker at the start time of the stimulus sound, the continuous N-dimensional, artifact-free EEG signal with reduced sampling rate is divided into X*Y segments, each segment corresponding to the angle change of each sound source; the time window of each segment is -200 to 1000 ms relative to the start time of the stimulus sound (a total of 120 data points). Specifically, in one embodiment, there are 60 segments in total. Each small segment corresponds to one of 10 sound source angle change categories (0, 4, 10, 20, 30, 60, 90, 120, 150, 180).

[0077] S1038: The average of the segment data after baseline correction for 6 playback cycles of each sound source angle change is used to obtain the ACC signal for each EEG signal channel under each sound source angle change.

[0078] S1039: Select the channel data with the largest ACC signal amplitude under each change in sound source angle as the model input vector for model training.

[0079] In the construction, training and evaluation of the standard model in step S104, a machine learning model for predicting changes in the angle of the sound source from EEG signals is constructed. The machine learning model is trained using the preprocessed EEG signal dataset obtained in step S103 to generate a standard model corresponding to normal hearing subjects.

[0080] There are many types of machine learning models, such as decision trees, support vector machines, random forests, and artificial neural networks. The specific choice should be made through experimentation to select the model with the best predictive performance. Given that EEG signal data is time-series data, this embodiment of the solution preferably uses a Long Short-Term Memory (LSTM) model. To enable the model to focus on more important moments in the ACC waveform, such as the peaks and troughs of the ACC signal, and to improve the performance in predicting changes in the sound source angle, an LSTM model with an attention mechanism (Attention-LSTM) is constructed to predict changes in the sound source angle. The specific steps are as follows:

[0081] A Long Short-Term Memory (LSTM) network model was used as the machine learning model. The preprocessed EEG signal dataset obtained in step S103 was divided into a training set (80%) and a test set (20%). The machine learning model was trained using the training set data to obtain a standard model constructed from data of people with normal hearing. The predictive performance of the standard model was evaluated using the test set. The model was continuously optimized by adjusting the hyperparameter settings until the best predictive effect was achieved on the test set. The output of the standard model consists of 10 categories, corresponding to the 10 sound source angle variations between 0 and 180 degrees mentioned above.

[0082] Step S20 specifically includes the following steps:

[0083] By constructing a sound source localization test experimental paradigm, EEG signals from patients were acquired and preprocessed; channel data with the maximum ACC signal corresponding to each sound source angle change were obtained.

[0084] Step S201: Standard model prediction. The obtained data from multiple channels are used sequentially as the model input vector of the standard model. The patient's sound source localization ability is obtained through standard model prediction.

[0085] Step S202: Calculate the relevant feature values ​​of sound source localization ability, calculate the standard curve with the predicted value of the standard model on the vertical axis and the actual sound source angle change value on the horizontal axis; and obtain the patient's sound source localization curve by predicting the patient's model input vector through the standard model; use the ratio of the AUC of the patient's sound source localization curve to the AUC of the standard curve as an indicator to measure the patient's sound source localization ability.

[0086] Referring to Figure 4, the standard AUC is S0, and the AUC of the patient's sound source localization curve is S1. The value of S1 / S0 is used as an indicator to measure the patient's sound source localization ability.

[0087] Referring to Figures 4 and 5, in step S30, the ratio of the AUC of the patient's sound source localization curve to the AUC of the standard curve is used as an indicator to measure the patient's sound source localization ability. The characteristic value ranges from 0 to 1; the closer the value is to 1, the closer the subject's sound source localization ability is to that of people with normal hearing. Based on the characteristic value calculation results, sound source localization ability impairment is divided into 5 levels, with higher levels indicating more severe loss of sound source localization ability. Level I means that the sound source localization ability is relatively close to that of people with normal hearing, while Level V means that the sound source localization impairment is very severe.

[0088] In summary, X, Y, N, K, P, and Q mentioned above are all integers.

[0089] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.

Claims

1. A method for assessing sound source localization ability based on EEG signals, characterized in that, The method comprises the following steps: S10: standard model construction; S20: based on the evaluation method of the standard model, the sound source positioning of the patient is obtained by the standard model prediction; S30: establish the sound source positioning ability disorder grade evaluation standard; Wherein, step S10 includes the following steps: S101: sound source positioning test paradigm construction; S102: EEG signal big data acquisition, based on the sound source positioning test paradigm, the EEG signal big data set induced by the normal hearing subjects under the stimulation sound of the sound source angle change is obtained; S103: EEG signal preprocessing, for removing artifacts from each EEG signal in the EEG signal big data set to obtain an EEG signal without artifacts; and establishing a mapping data of the EEG signal without artifacts corresponding to the sound source angle change, obtaining the channel data with the maximum amplitude of the ACC signal of the EEG signal under each sound source angle change, the channel data as the model input vector; all the obtained model input vectors constitute the preprocessed EEG signal big data set; S104: construction, training and evaluation of the standard model, a machine learning model for predicting the sound source angle change from the EEG signal is constructed, the preprocessed EEG signal big data set obtained in step S103 is used to train the machine learning model, and a standard model corresponding to the normal hearing subject is generated.

2. The method for evaluating sound source localization ability based on EEG signals according to claim 1, characterized in that, The specific steps of the sound source positioning test paradigm construction of step S101 are: S1011: build an arc-shaped loudspeaker array: build an arc-shaped loudspeaker array around the subject, the arc-shaped loudspeaker array includes a plurality of loudspeakers arranged in a range of 180 degrees expanding on both sides of the subject's front direction, wherein one of the loudspeakers is located in front of the subject, and the remaining loudspeakers are symmetrically placed along the subject's front direction and are unevenly distributed; S1012: set the stimulus sound played by the loudspeaker; S1013: after the subject is ready, first collect several seconds of the subject's EEG signal as baseline signal; then select two angle-symmetric loudspeakers to play the stimulus sound from left to right or from right to left as a kind of sound source angle change; every 1000-2000ms, randomly change another set of symmetric loudspeakers for sound stimulation; continuously collect the subject's EEG signal in the process; there are Y kinds of sound source angle changes, each kind of sound source angle change is randomly performed X times, and the playing order of X*Y times of sound source angle changes is randomly set by program control.

3. The method of claim 2, wherein, In the arc-shaped loudspeaker array of step S1011, the arc-shaped loudspeaker array includes 19 loudspeakers, and there are 10 kinds of sound source angle changes, each of which is performed at least 6 times.

4. The method of claim 2, wherein, In the step S101 of setting the stimulus sound played by the loudspeaker, the stimulus sound is a wideband stimulus signal obtained by band-pass filtering white noise at 125-6000Hz.

5. The method of claim 2, wherein the method further comprises: In the step S103 of EEG signal preprocessing, the EEG signal is removed based on the multi-dimensional space-time frequency correlation component decomposition and time convolution network TCN.

6. The method for evaluating sound source localization ability based on EEG signals according to claim 5, characterized in that, The artifact removal in the step S103 of EEG signal preprocessing includes the following steps: S1031: EEG signal input and decomposition: the N-channel EEG signal is decomposed by using a fast multi-dimensional empirical mode decomposition method to obtain corresponding multi-dimensional intrinsic mode functions (IMFs) and a trend item. Assuming that each channel is decomposed into K IMFs, a set of IMFs is obtained, which contains N*K IMFs; S1032: Intrinsic mode demixing: the IMFs obtained in S1031 are analyzed by using a canonical correlation analysis method (CCA) to obtain N*K projection matrices; S1033: Component screening: the TCN is used to automatically extract and classify the components in the projection matrix to obtain P brain electrical sources and Q artifact sources, which satisfy P+Q=N*K; S1034: Artifact removal: the column values of the CCA demixing matrix corresponding to the Q artifact sources are set to 0, and then the pure IMF corresponding to each channel EEG signal is calculated by using inverse CCA; S1035: Signal reconstruction and output: the pure IMF corresponding to each channel is reconstructed to obtain N-dimensional EEG signals without artifacts.

7. The method of claim 6, wherein the method further comprises: In the step S103 of obtaining the channel data with the maximum ACC signal amplitude under each sound source angle change of the EEG signal, the following steps are specifically included: S1036: The N-dimensional EEG signal without artifacts is band-pass filtered and the sampling rate is reduced; S1037: The N-dimensional EEG signal without artifacts with reduced sampling rate is divided into X*Y segments, and each segment corresponds to each sound source angle change; S1038: The segment data of each sound source angle change after baseline correction is averaged to obtain the ACC signal of each EEG signal channel under each sound source angle change; S1039: The channel data with the maximum ACC signal amplitude under each sound source angle change is selected as the model input vector for model training.

8. The method of claim 2, wherein, The long short-term memory network model is used as the machine learning model, the preprocessed EEG signal big data set obtained in step S103 is divided into a training set (80%) and a test set (20%), the machine learning model is trained using the training set data, a standard model constructed by normal hearing population data is obtained, and the standard model is evaluated by the test set; wherein the output of the standard model is Y classifications, corresponding to Y sound source angle changes between 0-180 degrees.

9. The method of claim 1, wherein, The step S20 specifically includes the following steps: through the constructed sound source positioning test experimental paradigm, the EEG signal of the patient is collected, and after the EEG signal preprocessing, the channel data with the maximum ACC signal corresponding to each sound source angle change is obtained, the obtained multiple channel data are sequentially used as the model input vector of the standard model, and the sound source positioning ability of the patient is predicted by the standard model.

10. The method for evaluating sound source localization ability based on EEG signals according to claim 9, wherein, In the step S20 of predicting the sound source positioning ability of the patient by the standard model, a standard curve is calculated with the standard model prediction value as the vertical axis and the real sound source angle change value as the horizontal axis; and a patient sound source positioning curve is obtained by predicting the patient's model input vector by the standard model; the ratio of the AUC of the patient sound source positioning curve to the AUC of the standard curve is used as an index for measuring the sound source positioning ability of the patient.

Images