40Hz auditory related potential waveform recognition method based on artificial intelligence

Abstract

The invention discloses a 40Hz auditory related potential waveform recognition method based on artificial intelligence, and belongs to the technical field of artificial intelligence. Comprising the following steps: S1, acquiring 40Hz auditory related potential test data under different frequencies by auditory evoked potential test equipment, storing the 40Hz auditory related potential test data as a first format file, and performing cleaning and data conversion on the first format file to obtain a second format file; s2, carrying out preprocessing and continuous wavelet transformation on the waveform data of the second format file; s3, performing data set division and labeling on the preprocessed waveform data to obtain a labeled data set; s4, constructing a waveform classification model, and training the waveform classification model through the labeled data set to obtain a trained model; and S5, recognizing the 40Hz auditory related potential reaction threshold based on the trained model. The technical scheme has the beneficial effects that automatic detection is realized, subjective interpretation deviation caused by human factors is reduced, the detection accuracy is improved, the labor cost and the time cost are reduced, and the detection efficiency is high.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a 40 Hz auditory-related potential waveform recognition method. Background Art

[0002] Auditory dysfunction, as a common sensory disorder, profoundly affects human language communication and learning ability. In the forensic clinical identification of auditory dysfunction, it is crucial to objectively and accurately evaluate the subject's hearing threshold. With the advancement of science and technology, a variety of audiometric methods have been introduced in the field of forensic medicine. Among them, the 40Hz auditory-related potential technology is often used to assess the degree of forensic auditory dysfunction because of its frequency specificity, stable waveform, easy recognition and reaction threshold close to the behavioral hearing threshold. The 40Hz auditory-related potential is a bioelectric response induced from the auditory nerve pathway after a sound stimulus of 40 times / second. It is collected through a combination of electrodes on the surface of the human skull and the ear, and then through analog-to-digital conversion and subsequent superposition analysis to form a group of potentials consisting of 4 quasi-sine waves with an interval of approximately 25ms within a 100ms scanning time, such as Figure 1 The figure shows the response present waveform (RP). In judicial identification, forensic audiologists determine the hearing function level of the subject by observing whether a specific response waveform appears under different stimulus sound intensities.

[0003] Although the existing hearing diagnosis technology system has been widely used in clinical practice and can meet daily diagnostic needs to a certain extent, it still faces some significant challenges and shortcomings in the field of forensic identification. First, in most cases, the person being identified cannot cooperate well with pure tone audiometry to obtain true behavioral hearing threshold results, and needs to be inferred through objective testing. Secondly, although the 40Hz auditory related potential is an objective audiometric method, the response waveform still requires subjective identification and judgment by professionals, and there are certain interpretation errors. In addition, this subjectivity not only affects the reliability and consistency of the diagnosis, but also increases labor costs and time costs, affecting diagnostic efficiency. Summary of the invention

[0004] The purpose of the present invention is to provide a 40 Hz auditory-related potential waveform recognition method based on artificial intelligence to solve the above technical problems;

[0005] A 40 Hz auditory-related potential waveform recognition method based on artificial intelligence, comprising:

[0006] Step S1, the auditory evoked potential test device collects 40 Hz auditory related potential test data at different frequencies and saves them as a first format file, and cleans and converts the first format file to obtain a second format file;

[0007] Step S2, preprocessing and continuous wavelet transforming the waveform data of the second format file to obtain a two-dimensional time-frequency diagram;

[0008] Step S3, dividing and labeling the preprocessed waveform data into data sets to obtain labeled data sets;

[0009] Step S4, constructing a waveform classification model, and training the waveform classification model using the annotated data set to obtain a trained model;

[0010] Step S5, identifying the response threshold of the 40 Hz auditory-related potential based on the trained model.

[0011] Preferably, step S1 comprises,

[0012] Step S11, collecting the 40 Hz auditory-related potential response waveforms of the subject at four frequencies of 0.5 kHz, 1 kHz, 2 kHz, and 4 kHz through the auditory evoked potential test device to obtain the test data, and saving it as the first format file;

[0013] Step S12, removing redundant information in the first format file by a data cleaning script to obtain a cleaned file;

[0014] Step S13, using a composite naming strategy to determine the data column titles in the cleaned file;

[0015] Step S14, filtering out valid sampling points in the cleaned file and converting them into the second format file.

[0016] Preferably, in step S11, the first format file is a txt file, and the second format file is a csv file;

[0017] The composite naming strategy in step S13 is to combine the subject's number, name abbreviation, left and right ear identification, test frequency and stimulus sound intensity to form the data column title.

[0018] Preferably, step S2 comprises,

[0019] Step S21, converting the waveform data in the second format file into a time series;

[0020] Step S22, combining the time series and a preset scale array, performing continuous wavelet transform based on a preset wavelet mother function, and obtaining wavelet coefficients;

[0021] Step S23: generating the two-dimensional time-frequency diagram using the wavelet coefficients.

[0022] Preferably, the preset wavelet mother function in step S22 includes complex Morlet wavelet and Morl wavelet;

[0023] The preset scale array ranges from 1 to 500.

[0024] Preferably, step S3 comprises,

[0025] Step S31, dividing the preprocessed waveform data into data sets to obtain divided data sets;

[0026] Step S32, annotating the divided data set according to waveform features to obtain the annotated data set.

[0027] Preferably, step S4 comprises,

[0028] Step S41, constructing the waveform classification model;

[0029] Step S42, selecting 20% ​​of the annotated data set as a test set for evaluating the waveform classification model;

[0030] Step S43, optimizing the waveform classification model according to the binary cross entropy loss value to obtain the trained model.

[0031] Preferably, the input of the waveform classification model is the two-dimensional time-frequency diagram, and the output of the waveform classification model is a reaction presence waveform or a non-reaction waveform.

[0032] Preferably, the waveform classification model in step S4 includes a residual network, a VGG network and a self-attention network.

[0033] Preferably, step S5 comprises,

[0034] Step S51, the trained model sorts the waveform data of the test signal according to the stimulus intensity, analyzes the waveform characteristics under each stimulus intensity in turn, and outputs the reaction waveform or the no reaction waveform;

[0035] Step S52, searching for the minimum stimulus sound intensity in the reaction existence waveform, determining the minimum stimulus sound intensity as the reaction threshold and then outputting it.

[0036] The beneficial effects of the present invention are: realizing automated detection, reducing subjective interpretation deviation caused by human factors, improving detection accuracy, reducing labor costs and time costs, and achieving high detection efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a waveform diagram of the 40Hz auditory-related potential response in the prior art;

[0038] Figure 2 It is a step diagram of the 40 Hz auditory-related potential waveform recognition method of the present invention;

[0039] Figure 3 is a schematic diagram of step S1 of the present invention;

[0040] Figure 4 is a schematic diagram of step S2 of the present invention;

[0041] Figure 5 is a schematic diagram of step S3 of the present invention;

[0042] Figure 6 is a schematic diagram of step S4 of the present invention;

[0043] Figure 7 is a schematic diagram of step S5 of the present invention;

[0044] Figure 8 It is a flow chart of 40Hz auditory-related potential waveform recognition and reaction threshold discrimination of the present invention;

[0045] Fig. 9 The present invention is an experimental flow chart of 40 Hz auditory-related potential waveform recognition and reaction threshold discrimination in an embodiment. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0048] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0049] A 40Hz auditory-related potential waveform recognition method based on artificial intelligence, such as Figure 2 As shown, including

[0050] Step S1, the auditory evoked potential test device collects 40 Hz auditory related potential test data at different frequencies and saves them as a first format file, cleans and converts the first format file to obtain a second format file;

[0051] Step S2, preprocessing and continuous wavelet transforming the waveform data of the second format file to obtain a two-dimensional time-frequency diagram;

[0052] Step S3, dividing and labeling the preprocessed waveform data into data sets to obtain labeled data sets;

[0053] Step S4, constructing a waveform classification model, and training the waveform classification model using the labeled data set to obtain a trained model;

[0054] Step S5, identifying the response threshold of the 40 Hz auditory-related potential based on the trained model.

[0055] Specifically, the present invention provides a 40Hz auditory-related potential waveform recognition method based on artificial intelligence, which reduces manual intervention based on a waveform classification model. Through a deep learning model trained with a large amount of data, it can effectively identify complex hearing loss characteristics and reduce the impact of human factors on the diagnosis results. It can quickly process large quantities of data, reduce manual operation time, and reduce operating costs. Improve the accuracy, efficiency and objectivity of recognition, achieve accurate judgment of auditory dysfunction, and provide a reliable basis for forensic identification.

[0056] In a preferred embodiment, referring to Figure 3 , step S1 comprises,

[0057] Step S11, collecting the 40 Hz auditory-related potential response waveforms of the subject at four frequencies of 0.5 kHz, 1 kHz, 2 kHz, and 4 kHz by means of an auditory evoked potential test device to obtain test data, and saving the data as a first format file;

[0058] Step S12, removing redundant information in the first format file by a data cleaning script to obtain a cleaned file;

[0059] Step S13, using a composite naming strategy to determine the data column titles in the cleaned file;

[0060] Step S14, filtering out valid sampling points in the cleaned file and converting them into a second format file;

[0061] In step S11, the first format file is a txt file (a plain text file, usually stored in UTF-8 or ASCII encoding, suitable for storing unstructured or simple text data), and the second format file is a csv file (a text file with commas as delimiters, mainly used to store tabular data, suitable for storing and exchanging structured data, and convenient for data analysis and processing);

[0062] The composite naming strategy in step S13 is to combine the subject's number, name abbreviation, left and right ear identification, test frequency, and stimulus sound intensity to form a data column title.

[0063] Specifically, a 40 Hz auditory-related potential testing device was used to measure the waveforms and reaction thresholds of the subjects at four frequencies: 0.5 kHz, 1 kHz, 2 kHz, and 4 kHz.

[0064] The 40Hz auditory related potential test equipment was used to measure the waveform and response threshold of the subjects at four frequencies: 0.5kHz, 1kHz, 2kHz, and 4kHz. During the test, if the waveform was four quasi-sine waves with an interval of about 25ms, it was considered to be a response present (RP), otherwise it was considered to be no response (RA).

[0065] The reaction threshold of the subjects was tested by the "down ten and up five" method. The stimulation sound intensity started at 40dB nHL above the threshold and was reduced in steps of 10dB nHL until no reaction waveform was induced. The intensity was then increased in steps of 5dB nHL until a reaction waveform was induced again. When the waveform was poorly differentiated close to the reaction threshold, the stimulation was repeated to observe whether there was repeatability. The minimum stimulation sound intensity that could induce a reaction waveform was taken as the reaction threshold.

[0066] The data after each test was exported as a txt file, recording the waveforms of the subject's left and right ears at different frequencies and the corresponding stimulus sound intensity information.

[0067] To ensure the validity of the data, the txt file was cleaned to remove redundant information that was not related to waveform analysis. The redundant information included parameters and special symbols of the test equipment. A composite naming strategy was adopted to combine key elements such as subject number, name abbreviation, ear identification, test frequency, and stimulus intensity to ensure the uniqueness and traceability of the data column headers.

[0068] After data cleaning, the device sampling interval is 0.2ms, and the sampling starting point is 513 and the 500 valid sampling points thereafter are retained as core data to ensure the pertinence and efficiency of subsequent analysis.

[0069] Finally, the processed data is converted into CSV format to improve its compatibility and operability in subsequent processing.

[0070] In a preferred embodiment, referring to Figure 4 , step S2 comprises,

[0071] Step S21, converting the waveform data in the second format file into a time series;

[0072] Step S22, combining the time series and the preset scale array, performing continuous wavelet transform based on a preset wavelet mother function, and obtaining wavelet coefficients;

[0073] Step S23, generating a two-dimensional time-frequency diagram through wavelet coefficients;

[0074] The wavelet mother function preset in step S22 includes complex Morlet wavelet and Morl wavelet;

[0075] The default scale array range is 1 to 500.

[0076] Specifically, to ensure the integrity and accuracy of the signal, the waveform signal in each csv file is first converted into a time series with a step size of 0.2ms. Next, a scale array containing integers from 1 to 500 is generated for adjusting the scale parameters in the wavelet transform.

[0077] The selection of the wavelet mother function is crucial to the accuracy of time-frequency analysis. In this embodiment, the complex Morlet wavelet (cmor0.7-1.5) with good time-frequency smoothing characteristics is selected to better capture the detailed features of the signal.

[0078] Based on the generated scale array and complex Morlet wavelet, CWT (Continuous Wavelet Transform, a tool for time-frequency analysis that can capture local frequency changes in signals) is performed to generate a set of wavelet coefficients that reflect the energy distribution of the signal at different frequencies and time points. These coefficients can effectively capture the local characteristics of the signal and are used to generate a two-dimensional time-frequency diagram.

[0079] The horizontal axis of the two-dimensional time-frequency image represents time, the vertical axis represents frequency, and the color intensity represents the energy distribution of the signal at that frequency and time point. To ensure the accuracy of the auditory waveform recognition task, the system limits the Y-axis frequency range to ensure that the key information features near 40Hz are fully preserved.

[0080] Other wavelet mother functions (such as Morl wavelet) can also be used to adjust the analysis accuracy according to the requirements of different signal characteristics.

[0081] In a preferred embodiment, referring to Figure 5 , step S3 comprises,

[0082] Step S31, dividing the preprocessed waveform data into data sets to obtain divided data sets;

[0083] Step S32, annotating the divided data set according to the waveform features to obtain annotated data set.

[0084] Specifically, after preprocessing, the system divides the waveform data into training set, validation set and test set. Using five-fold cross validation, each fold of the data set division ensures that all data can be used in the training, validation and testing stages to avoid model overfitting.

[0085] The data were annotated based on the waveform features to distinguish between waveforms with and without responses. The annotated data sets were used for subsequent model training through five-fold cross validation to ensure that each type of data could be fully utilized.

[0086] The processed waveform data are divided into training set, validation set and test set, and five-fold cross validation is used to ensure that each data can fully participate in model training and testing.

[0087] The cross-validation method can enhance the robustness of the model, avoid overfitting of the model, and ensure the generalization ability of the model on real data. The training set is used to train the model, the validation set is used to measure the performance of the model during the training process, and the test set is used to evaluate the actual performance of the model after the training is completed.

[0088] In a preferred embodiment, referring to Figure 6 , step S4 comprises,

[0089] Step S41, constructing a waveform classification model;

[0090] Step S42, selecting 20% ​​of the annotated data set as a test set for evaluating the waveform classification model;

[0091] Step S43, optimizing the waveform classification model according to the binary cross entropy loss value to obtain a trained model.

[0092] Specifically, this embodiment adopts the ResNet classification network, which uses its deep residual structure to effectively extract complex features in waveform signals and is particularly suitable for processing complex 40Hz auditory-related potential waveforms.

[0093] For binary classification tasks, the system selects the binary cross entropy loss function for optimization, and continuously improves the classification accuracy of the model through round-by-round training.

[0094] After each round of cross-validation, the system evaluates the model performance through multiple key indicators, including accuracy, precision, sensitivity, specificity, F1 score, Youden's Index, and the area under the receiver operating characteristic curve (AUC, the area under the receiver operating characteristic curve, is an indicator for evaluating the performance of the classification model, ranging from 0 to 1, and the higher the value, the better the model performance). These indicators fully reflect the actual performance of the model in complex waveform classification, ensuring the reliability and accuracy of the final judgment results.

[0095] In a preferred embodiment, the input of the waveform classification model is a two-dimensional time-frequency diagram of the waveform, and the output of the waveform classification model is a response presence waveform or a non-response waveform.

[0096] In a preferred embodiment, the waveform classification model in step S4 includes a residual network, a VGG network and a self-attention network.

[0097] Specifically, the ResNet classification network can also be replaced by other deep network models (such as VGG or Transformer) to adapt to different data scales or feature extraction requirements.

[0098] VGG (Visual Geometry Group Network): It uses multiple small convolution kernels to build a deep network with deeper layers and larger parameters, which is suitable for image classification tasks.

[0099] Transformer is based on the self-attention mechanism, does not rely on sequential processing, has efficient parallelization capabilities, and is widely used in natural language processing and image fields.

[0100] In a preferred embodiment, referring to Figure 7 , step S5 comprises,

[0101] Step S51, the trained model sorts the waveform data of the test signal according to the stimulus intensity, analyzes the waveform characteristics under each stimulus intensity in turn, and outputs a response waveform or a non-response waveform;

[0102] Step S52, searching for the minimum stimulus sound intensity in the reaction existence waveform, determining the minimum stimulus sound intensity as the reaction threshold and then outputting it.

[0103] Specifically, in the 40Hz AERP reaction threshold judgment process, the system sorts the waveform judgment results of the test signal of each ear of each subject from high to low according to the stimulus sound intensity, and searches for the judgment result of the reaction waveform under the minimum stimulus sound intensity, and judges the minimum stimulus sound intensity as the reaction threshold.

[0104] Fig. 9 The guidance process of the yellow arrow in the figure is Figure 8 Schematic flow, follow the green arrows:

[0105] The first process is data cleaning and data format conversion. First, use the 40Hz auditory-related potential test equipment to collect the subject's response waveforms at four frequencies of 0.5kHz, 1kHz, 2kHz, and 4kHz, and test the subject's reaction threshold using the "drop ten and raise five" method.

[0106] After the test, the test data is exported to a txt file to record the data of the left and right ears. The device parameters and redundant symbols are removed through the written data cleaning script to ensure that only the waveform and sound intensity information related to the classification task are retained.

[0107] Due to the diversity of test data of the subjects, a composite naming strategy is set in advance: the subject's number, name abbreviation, left and right ear identification, test frequency and stimulus intensity are combined into a column name, such as 001WBLM55B.2_500_55. This naming method can avoid duplication of column names and ensure the traceability of data.

[0108] The waveform data in each record is collected by the device at a period of 0.2ms, starting at point 513, and then extracting 500 valid sampling points as the core data. The selected data is used for subsequent analysis to ensure that only the most representative waveform features are used.

[0109] After data cleaning and sampling point screening, the system converts it into CSV format, which is widely used in data analysis tasks to facilitate subsequent wavelet transform analysis and processing.

[0110] The second process is to refine the data preprocessing. The waveform data in each csv file is read line by line from the beginning, and the system generates a time series with a step size of 0.2ms. The generated time series will be input into the CWT module for time-frequency analysis.

[0111] To adapt to the time-frequency analysis of different waveform characteristics, the system will generate a scale array with values ​​ranging from 1 to 500 to control the analysis range of different resolutions in CWT. The scale array determines the level of detail of signal analysis and captures changes in different periodicities.

[0112] The scale array can be generated programmatically or mathematically, for example, using linear spacing to generate an array of integers from 1 to 500, or using logarithmic spacing to provide denser scales in certain frequency ranges.

[0113] After experiments, the present invention selected the complex Morlet wavelet (cmor0.7-1.5) as the mother function of CWT. This wavelet function can maintain good smoothness in both time and frequency domains, and is particularly suitable for analyzing complex auditory waveform data.

[0114] The third process is CWT analysis, which inputs the generated time series and scale array into the CWT algorithm, and the system calculates the wavelet coefficients of each signal through wavelet transform. These coefficients reflect the energy changes of the signal at different times and frequencies.

[0115] The system uses wavelet coefficients to generate a two-dimensional time-frequency diagram of the waveform (i.e., the picture above the CWT analysis process), with the horizontal axis representing time, the vertical axis representing frequency, and the color intensity representing signal energy. To enhance the accuracy of subsequent waveform classification, the image features of the frequency range around 40Hz of the time-frequency diagram are fully retained.

[0116] The generated time-frequency graph is cropped to remove irrelevant image elements (such as coordinate axes and legends) to improve the readability and effectiveness of the image. The final cropped image is saved and used for subsequent model training and analysis.

[0117] The fourth process is model training and image classification. The ResNet classification network is selected for waveform classification. This network can avoid the gradient vanishing problem through its residual connection mechanism and ensure effective feature extraction of complex waveform data. The input of the model is the two-dimensional time-frequency graph of each waveform, and the output is the classification result of the waveform (reaction exists or does not exist).

[0118] Using the five-fold cross validation method, 20% of the data is used for testing each time. In each iteration, the system calculates the binary cross entropy loss and optimizes the model parameters based on the loss value.

[0119] After each training session, the system evaluates the model using indicators such as accuracy, sensitivity, specificity, F1 score, Uden index, and AUC. The evaluation results will be used to optimize the model parameters to ensure that the final model has high accuracy and robustness.

[0120] The fifth process is the reaction threshold judgment. Each waveform judgment result of the test signal of each ear of each subject is sorted from high to low according to the stimulus sound intensity, and the minimum stimulus sound intensity for which the waveform judgment result exists is found, and the minimum stimulus sound intensity is judged as the reaction threshold.

[0121] The system ultimately outputs the subject's response threshold and uses the result to assess auditory dysfunction.

[0122] Based on the above logical judgment results, the prediction accuracy of the reaction threshold was calculated to ensure the comprehensiveness of the performance evaluation. The results of the five-fold cross-validation test showed that the accuracy of waveform recognition was 0.93 and the accuracy of 40Hz auditory related potential reaction threshold judgment was 0.84.

[0123] The above description is only a preferred embodiment of the present invention, and does not limit the implementation mode and protection scope of the present invention. For those skilled in the art, it should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the description and illustrations of the present invention should be included in the protection scope of the present invention.

Claims

1. A 40 Hz auditory-related potential waveform recognition method based on artificial intelligence, characterized in that: include, Step S1, the auditory evoked potential test device collects 40 Hz auditory related potential test data at different frequencies and saves them as a first format file, and cleans and converts the first format file to obtain a second format file; Step S2, preprocessing and continuous wavelet transforming the waveform data of the second format file to obtain a two-dimensional time-frequency diagram; Step S3, dividing and labeling the preprocessed waveform data into data sets to obtain labeled data sets; Step S4, constructing a waveform classification model, and training the waveform classification model using the annotated data set to obtain a trained model; Step S5, identifying the response threshold of the 40 Hz auditory-related potential based on the trained model.

2. The 40 Hz auditory-related potential waveform recognition method based on artificial intelligence according to claim 1 is characterized in that: Step S1 comprises, Step S11, collecting the 40 Hz auditory-related potential response waveforms of the subject at four frequencies of 0.5 kHz, 1 kHz, 2 kHz, and 4 kHz through the auditory evoked potential test device to obtain the test data, and saving it as the first format file; Step S12, removing redundant information in the first format file by a data cleaning script to obtain a cleaned file; Step S13, using a composite naming strategy to determine the data column titles in the cleaned file; Step S14, filtering out valid sampling points in the cleaned file and converting them into the second format file.

3. The 40 Hz auditory-related potential waveform recognition method based on artificial intelligence according to claim 2 is characterized in that: In step S11, the first format file is a txt file, and the second format file is a csv file; The composite naming strategy in step S13 is to combine the subject's number, name abbreviation, left and right ear identification, test frequency and stimulus sound intensity to form the data column title.

4. The 40 Hz auditory-related potential waveform recognition method based on artificial intelligence according to claim 1 is characterized in that: Step S2 comprises, Step S21, converting the waveform data in the second format file into a time series; Step S22, combining the time series and a preset scale array, performing continuous wavelet transform based on a preset wavelet mother function, and obtaining wavelet coefficients; Step S23: generating the two-dimensional time-frequency diagram using the wavelet coefficients.

5. The 40 Hz auditory-related potential waveform recognition method based on artificial intelligence according to claim 4 is characterized in that: The preset wavelet mother function in step S22 includes complex Morlet wavelet and Morl wavelet; The preset scale array ranges from 1 to 500.

6. The 40 Hz auditory-related potential waveform recognition method based on artificial intelligence according to claim 1 is characterized in that: Step S3 comprises, Step S31, dividing the preprocessed waveform data into data sets to obtain divided data sets; Step S32, annotating the divided data set according to waveform features to obtain the annotated data set.

7. The 40 Hz auditory-related potential waveform recognition method based on artificial intelligence according to claim 1 is characterized in that: Step S4 comprises, Step S41, constructing the waveform classification model; Step S42, selecting 20% ​​of the annotated data set as a test set for evaluating the waveform classification model; Step S43, optimizing the waveform classification model according to the binary cross entropy loss value to obtain the trained model.

8. The 40 Hz auditory-related potential waveform recognition method based on artificial intelligence according to claim 3 is characterized in that: The input of the waveform classification model is the two-dimensional time-frequency diagram, and the output of the waveform classification model is a response presence waveform or a non-response waveform.

9. The 40 Hz auditory-related potential waveform recognition method based on artificial intelligence according to claim 1 is characterized in that: The waveform classification model in step S4 includes a residual network, a VGG network and a self-attention network.

10. The 40 Hz auditory-related potential waveform recognition method based on artificial intelligence according to claim 8, characterized in that: Step S5 comprises, Step S51, the trained model sorts the waveform data of the test signal according to the stimulus intensity, analyzes the waveform characteristics under each stimulus intensity in turn, and outputs the reaction waveform or the no reaction waveform; Step S52, searching for the minimum stimulus sound intensity in the reaction existence waveform, determining the minimum stimulus sound intensity as the reaction threshold and then outputting it.

Images