Hearing threshold prediction method based on frequency sweep OAEs and deep learning model

Abstract

The invention relates to a hearing threshold prediction method based on frequency sweep OAEs and a deep learning model, and the method comprises the steps: employing a frequency sweep stimulation signal normal form with a unified mode to extract an SFOAEs signal and a DPOAEs signal, the stimulation sound normal form of the SFOAEs signal is a four-segment frequency sweep signal, and the stimulation sound normal form of the DPOAEs signal is a single-segment frequency sweep signal. And carrying out interference signal filtering on a residual signal of the single DPOAEs signal or the SFOAEs signal through a tracking filter with a dynamically changed zero pole so as to extract frequency sweep OAEs spectrum information. And constructing a data set of a deep learning model based on the sweep frequency OAEs spectrum information, and training and evaluating the deep learning model to obtain a hearing threshold prediction model. And calling the hearing threshold prediction model to perform hearing threshold prediction on the SFOAEs signals and the DPOAEs signals under different test intensities, and outputting hearing thresholds corresponding to the frequencies, thereby ensuring the accuracy of diagnosing the hearing loss degree.

Description

Technical Field

[0001] The present invention relates to the field of medical assistance technology, and in particular to a hearing threshold prediction method based on frequency sweeping OAEs and a deep learning model. Background Art

[0002] Otoacoustic emissions (OAEs) are weak audio energy generated in the cochlea of ​​the inner ear and released into the external auditory canal through the ossicular chain and the eardrum. They are part of the normal function of the human ear. The discovery of OAEs confirmed that the cochlea, as a peripheral auditory receptor, not only passively converts external sound signals into bioelectrical signals that are transmitted to the central nervous system to induce hearing, but also actively releases energy, thus establishing the theory that the cochlea is a bidirectional transducer.

[0003] Hearing loss is often accompanied by an increase in the minimum audible sound intensity, or hearing threshold. Clinically, hearing loss is primarily caused by cochlear hair cell lesions. Current diagnostic methods rely on behavioral hearing threshold testing, which is susceptible to subjective factors such as attention deficit disorder. However, otoacoustic emissions (OAEs), a byproduct of active cochlear processes, can characterize changes in the absolute hearing threshold in hearing-impaired ears.

[0004] Currently, clinically used OAE detection systems primarily assess a subject's hearing by detecting transient evoked otoacoustic emissions (TEOAEs) and / or distortion product otoacoustic emissions (DPOAEs). Existing research has revealed that the mechanisms of OAE generation primarily include linear coherent reflections and nonlinear distortion. These OAEs carry different information characterizing the active mechanisms of the cochlea. Stimulus frequency otoacoustic emissions (SFOAEs) and TEOAEs are believed to arise primarily from linear coherent reflections, with SFOAEs exhibiting significantly higher frequency specificity than TEOAEs. DPOAEs are believed to arise primarily from nonlinear distortion. Therefore, given the subjectivity of current clinical diagnosis of hearing loss (hearing threshold), accurate predictions are difficult to ensure with existing technologies. Summary of the Invention

[0005] Based on this, it is necessary to provide a hearing threshold prediction method based on swept frequency OAEs and deep learning models that can ensure the accuracy of the current clinical diagnosis of hearing loss degree (hearing threshold) to address the above technical problems.

[0006] A first aspect of the present invention provides a method for predicting hearing thresholds based on frequency-sweep OAEs and a deep learning model, the method comprising: The SFOAEs and DPOAEs signals were extracted using a unified frequency sweep stimulation signal paradigm, wherein the SFOAEs signal stimulation sound paradigm was a four-segment frequency sweep signal, and the DPOAEs signal stimulation sound paradigm was a single-segment frequency sweep signal. The residual signal of a single DPOAEs signal or SFOAEs signal is filtered through a tracking filter with dynamically changing zeros and poles to remove interference signals and extract the swept frequency OAEs spectrum information. Constructing a data set of a deep learning model based on the swept frequency OAEs spectrum information, training and evaluating the deep learning model to obtain a hearing threshold prediction model; The hearing threshold prediction model is called to predict the hearing threshold of the SFOAEs signal and the DPOAEs signal at different test intensities to output the hearing threshold corresponding to each frequency.

[0007] In one embodiment, the extraction of SFOAEs and DPOAEs signals using a unified frequency sweep stimulation signal paradigm includes: The DPOAEs signal is extracted using a single-segment sweep signal stimulus paradigm, wherein the DPOAEs signal includes a first fundamental frequency sweep signal and a second fundamental frequency sweep signal, and the first fundamental frequency sweep signal and the second fundamental frequency sweep signal are respectively transmitted to the human ear through two independent sound tubes; The SFOAEs signal is extracted through a stimulation sound paradigm of a four-segment sweep signal, wherein the SFOAEs signal includes a stimulation sound and an inhibition sound, and the stimulation sound and the inhibition sound are respectively transmitted to the human ear through two independent sound tubes; The frequencies of the first pitch sweep signal and the second pitch sweep signal are respectively a first frequency and a second frequency, and a ratio between the first frequency and the second frequency is any value within a first preset range; The four-segment sweep frequency signal includes a first-segment sweep frequency signal, a second-segment sweep frequency signal, a third-segment sweep frequency signal, and a fourth-segment sweep frequency signal. The stimulation sound exists in each segment of the four-segment sweep frequency signal and presents an initial phase alternating between positive and negative. The suppression sound exists in the third-segment sweep frequency signal and the fourth-segment sweep frequency signal with the same initial phase. The frequencies of the stimulation sound and the suppression sound are the third frequency and the fourth frequency, respectively. The ratio between the third frequency and the fourth frequency is any value within a second preset range.

[0008] In one embodiment, filtering out interference signals from the residual signal of a single DPOAEs signal or a SFOAEs signal through a tracking filter with dynamically changing zeros and poles to extract the swept frequency OAEs spectrum information includes: For the DPOAEs signal, the dynamic poles of the tracking filter are set to the instantaneous frequency of the swept-frequency DPOAEs signal induced by the DPOAEs signal swept-frequency paradigm, and the dynamic zeros are set to the first frequency and the second frequency respectively; For the residual signal of the SFOAEs signal, the dynamic pole of the tracking filter is set to the third frequency induced by the SFOAEs signal sweep paradigm, and the dynamic zero is set to the fourth frequency.

[0009] In one embodiment, filtering out interference signals from the residual signal of a single DPOAEs signal or SFOAEs signal through a tracking filter with dynamically changing zeros and poles to extract the swept frequency OAEs spectrum information further includes: The filtered multiple time domain signals are stacked and averaged, and the odd buffer and the even buffer are stored one by one in an odd-even order, wherein the average value of the time domain signal difference between the odd buffer and the even buffer is the noise signal of the OAEs signal; The amplitudes of the OAEs signal and noise at each induced moment were estimated by time-frequency analysis, and TFA was performed using a continuous Hanning window to calculate the resolution points by performing a discrete Fourier transform on the window; The amplitude at the swept frequency corresponding to the center moment of the current window in the DFT results of each window is taken as the swept OAEs signal and noise result at the center moment of the current window to determine the amplitude spectrum, noise spectrum and signal-to-noise ratio spectrum of the swept DPOAEs signal and SFOAEs signal under a single test intensity.

[0010] In one embodiment, the dataset of the deep learning model includes a feature set and a label set, wherein the feature set includes a feature matrix consisting of amplitude spectra and signal-to-noise ratio spectra of DPOAEs signals and SFOAEs signals at multiple test intensities; the label set is the hearing threshold of each frequency to be predicted; The architecture of the deep learning model includes an input layer, a self-feature extractor and an output layer. The input layer is used to intercept a local feature matrix from the feature matrix. The self-feature extractor adopts a CNN framework to learn a high-dimensional feature vector from the feature matrix. The output layer is used to convert the high-dimensional feature vector into a fixed random seed through a fully connected layer as a regressor.

[0011] In one embodiment, the dataset of the deep learning model is constructed based on the swept frequency OAEs spectrum information, and the deep learning model is trained and evaluated to obtain a hearing threshold prediction model, including: The feature set is constructed by selecting amplitude spectra and signal-to-noise ratio spectra of SFOAEs signals at a first number of test intensities and DPOAEs signals at a second number of test intensities, wherein the feature set has a third number of spectrum sequences; the third number is twice the sum of the first number and the second number; Converting the hearing threshold of each frequency to be predicted into a hearing threshold with a label, and setting rules for upper and lower hearing thresholds to construct the label set; The rule for the upper and lower hearing thresholds is to set the hearing threshold label value below the lowest hearing threshold as the lowest hearing threshold, and to set the hearing threshold label value above the highest hearing threshold as the highest hearing threshold, while normalizing all hearing thresholds to a fixed step size.

[0012] In one embodiment, the data set of the deep learning model is constructed based on the frequency sweep OAEs spectrum information, and the deep learning model is trained and evaluated to obtain a hearing threshold prediction model, further comprising: Dividing the dataset into 6 equal folds by stratified sampling, and training and evaluating the deep learning model using 6-fold cross validation; Among them, the deep learning model is a regression model built with a CNN architecture, the activation layer function of the deep learning model adopts the Leaky ReLU activation function, and the pooling layer adopts global average pooling.

[0013] A second aspect of the present invention provides a hearing threshold prediction device based on swept frequency OAEs and a deep learning model, for implementing the hearing threshold prediction method based on swept frequency OAEs and a deep learning model described in any one of the first aspects, the device comprising: A signal acquisition module, configured to extract SFOAEs and DPOAEs signals using a unified frequency sweep stimulation signal paradigm, wherein the SFOAEs signal stimulation sound paradigm is a four-segment frequency sweep signal, and the DPOAEs signal stimulation sound paradigm is a single-segment frequency sweep signal; An information extraction module is used to filter out interference signals from the residual signal of a single DPOAEs signal or SFOAEs signal through a tracking filter with dynamically changing zeros and poles, so as to extract the swept frequency OAEs spectrum information; A model training module is used to construct a data set of a deep learning model based on the swept frequency OAEs spectrum information, train and evaluate the deep learning model, and obtain a hearing threshold prediction model; The hearing threshold prediction module is used to call the hearing threshold prediction model to predict the hearing threshold of SFOAEs signals and DPOAEs signals under different test intensities to output the hearing threshold corresponding to each frequency.

[0014] The third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the hearing threshold prediction method based on swept frequency OAEs and deep learning model described in the first aspect.

[0015] A fourth aspect of the present invention provides a computer storage medium storing a computer program, which, when executed by a processor, implements the hearing threshold prediction method based on swept frequency OAEs and a deep learning model described in the first aspect.

[0016] The above-mentioned hearing threshold prediction method based on frequency sweep OAEs and deep learning model has the following advantages over the existing technology: (1) This invention addresses the inaccurate problem caused by the subjectivity of current clinical diagnosis of hearing loss (hearing threshold). It can maximize the accuracy of OAEs, an objective physiological indicator, in quantitatively assessing hearing loss and provides a means to objectively and quantitatively predict hearing thresholds.

[0017] (2) The present invention aims at the joint acquisition of high-frequency resolution SFOAEs and DPOAEs, which increases the demand for information complementarity. The swept frequency paradigm is adopted to extract SFOAEs and DPOAEs at multiple frequency points. For the hearing threshold prediction task, a deep learning model based on the convolutional neural network architecture (CNN) is used to learn the characterization information of hearing loss from the acquired multi-test intensity OAEs spectrum, thereby achieving objective quantitative prediction of the hearing threshold and further improving the accuracy of hearing threshold prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 This is a flow chart of the hearing threshold prediction method based on swept frequency OAEs and deep learning model provided by the present invention; Figure 2 This is a test paradigm diagram of swept-frequency DPOAEs and SFOAEs in the hearing threshold prediction method based on swept-frequency OAEs and a deep learning model in a specific embodiment provided by the present invention; Figure 3 A schematic diagram of the structure of a CNN feature extractor for a hearing threshold prediction method based on swept frequency OAEs and a deep learning model in a specific embodiment of the present invention; Figure 4This is the second flow chart of the hearing threshold prediction method based on swept frequency OAEs and deep learning model provided by the present invention; Figure 5 This is the third flow chart of the hearing threshold prediction method based on swept frequency OAEs and deep learning model provided by the present invention; Figure 6 This is a fourth flow chart of the hearing threshold prediction method based on swept frequency OAEs and deep learning model provided by the present invention; Figure 7 This is a fifth flow chart of the hearing threshold prediction method based on swept frequency OAEs and deep learning model provided by the present invention; Figure 8 A schematic diagram of the structure of a hearing threshold prediction device based on swept frequency OAEs and a deep learning model provided by the present invention; Figure 9 This is a diagram of the internal structure of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings 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 making creative efforts shall fall within the scope of protection of the present invention.

[0021] The following combination Figures 1 to 9 The present invention describes the hearing threshold prediction method based on swept frequency OAEs and deep learning model.

[0022] like Figure 1 As shown, in one embodiment, a method for predicting hearing thresholds based on swept frequency OAEs and a deep learning model includes the following steps: In step S110 , a unified sweep frequency stimulation signal paradigm is used to extract SFOAEs and DPOAEs signals. The stimulation sound paradigm of the SFOAEs signal is a four-segment sweep frequency signal, and the stimulation sound paradigm of the DPOAEs signal is a single-segment sweep frequency signal.

[0023] Specifically, a unified sweep frequency stimulation signal paradigm was used to extract SFOAEs and DPOAEs. The DPOAE stimulation sound paradigm is a single-segment sweep frequency signal with a segment duration of 1 second, including two sweep frequency signals, the first fundamental tone (primary tone) and the second fundamental tone, which are played to the human ear through two independent sound tubes. The frequencies of the two are and ,in, The sweep frequency is 0.5-8 kHz, and the instantaneous frequency relationship between the two is maintained as =M, the value of M can be set to any value within the range of 1.0~1.4 according to the needs. The stimulus paradigm of SFOAEs is a four-segment sweep signal with a segment length of 1 second, including a probe tone and a suppressor tone, which are played to the human ear through two independent sound tubes. The stimulus appears in the four segments A, B, C and D, and presents an initial phase alternating between positive and negative. Its frequency The suppression sound only appears in the C and D segments with the same initial phase, and its frequency and Time maintenance =N constant proportional relationship, N can be set to any value within the range of 0.9~1.1 according to requirements.

[0024] It should be noted that the instantaneous frequency of the DPOAEs induced by the DPOAEs sweep paradigm is 2 - , which is contained in the retrieved time domain signal and needs to be subsequently obtained by filtering. The swept frequency SFOAEs signal induced by the swept frequency paradigm of SFOAEs needs to be extracted by residual calculation, and the formula is , and then the SFOAEs signal is extracted from the residual signal by filtering.

[0025] Combine Figure 2 As shown, in a specific embodiment, the present invention provides a method for predicting hearing thresholds based on swept frequency OAEs and a deep learning model. DPOAEs is a single-segment stimulation paradigm, and the stimulation sound includes a pair of fundamental tones. and , the instantaneous frequency ratio of the two is =M( Figure 2 The DPOAEs signal induced by this is taken as 1.22. The instantaneous frequency is 2 - , which is included in the total recollection signal r with the pitch recollection artifact, and is subsequently filtered by tracking. Extraction. SFOAEs is a four-segment stimulation paradigm. The stimulus sound appears in all four segments with alternating initial phases. The inhibitory sound appears only in segments C and D with the same phase. The instantaneous frequency ratio of the inhibitory sound to the stimulus sound is =N (N is taken as 1.05 as an example in the figure). The SFOAEs signal derived from this needs to be obtained through the residual signal calculated from the four-segment signal. The calculation formula is: ,in, , , and These are the echo signals corresponding to the four stimulus sounds.

[0026] Step S120 , filtering out interference signals from the residual signal of a single DPOAEs signal or a SFOAEs signal through a tracking filter with dynamically changing zeros and poles, so as to extract the swept frequency OAEs spectrum information.

[0027] Specifically, during the sound-giving and acquisition process, the single DPOAEs signal or the residual signal of SFOAEs is passed through a tracking filter with dynamically changing zero poles to filter out interference signals such as the acquisition artifact. For DPOAEs, the dynamic pole of the filter is set to 2 - frequency, and the two dynamic zero points are set to the fundamental frequency and Frequency, for the SFOAEs residual signal, the dynamic pole of the filter is set to the stimulus frequency induced by SFOAEs , and the dynamic zero is set to suppress the sound Frequency. Multiple filtered time-domain signals are stacked and averaged to improve the signal-to-noise ratio (S / N). Simultaneously, the signals are stored sequentially in odd- and even-numbered buffers. The noise corresponding to the OAE signal is calculated as the average of the difference between the time-domain signals in the two buffers. Time-frequency analysis (TFA) is used to estimate the amplitudes of the OAE signal and noise at each extraction time. TFA is performed using a continuous Hanning window with a window length of 50 ms and a 50% overlap between windows. The discrete Fourier transform (DFT) of the window is calculated with a resolution of 48,000 points, consistent with the sampling rate. The amplitude at the swept frequency corresponding to the center of each window in the DFT result is taken as the swept OAE and noise result at that time. After the above calculations, the amplitude spectrum and noise spectrum of the swept-frequency SFOAEs and DPAOEs at a single test intensity can be obtained. By subtracting the two, their signal-to-noise ratio spectrum can be further obtained. The TFA analysis parameters are only for reference. In actual process, the analysis parameters can be flexibly changed according to the number of points and resolution requirements.

[0028] In a specific embodiment, the present invention provides a method for predicting hearing thresholds based on swept frequency OAEs and a deep learning model. The essence of the tracking filter is the dynamically changing zeros and poles. For the extraction of DPOAEs, the poles are set to the signal frequency of the DPOAEs. - , set the zero point to the frequency of the echo artifact and , to filter out the echo artifacts and extract the DPOAEs signal. For the extraction of SFOAEs, the pole is set to the signal frequency of SFOAEs. , set the zero point to the signal frequency of the suppressed sound , in order to filter out the artifacts induced by the suppressive sound in the residual signal and extract the SFOAEs signal induced by the stimulus sound.

[0029] Step S130: construct a data set for a deep learning model based on the swept frequency OAEs spectrum information, train and evaluate the deep learning model, and obtain a hearing threshold prediction model.

[0030] Specifically, the deep learning (DL) model method includes dataset construction, model construction, and model evaluation. The dataset consists of two parts: a feature set and a label set. The feature set is a feature matrix composed of the amplitude spectrum and signal-to-noise ratio spectrum of SFOAEs and DPOAEs at multiple test intensities; the label set is the hearing threshold of each frequency to be predicted. Model construction is carried out independently for each hearing threshold frequency. The model architecture includes an input layer, a self-feature extractor, and an output layer. The input layer extracts the required local feature matrix from the original feature matrix. The feature extractor uses a CNN framework to learn high-dimensional feature vectors from the original feature matrix. The output layer uses a fully connected layer as a regressor to convert the feature vector into the hearing threshold to be predicted. Six-fold cross-validation is used for model evaluation, combined with 20 fixed random seeds. The mean absolute error (MAE) of the test set is used as the generalization error evaluation result of the model.

[0031] During the dataset construction, the feature set includes SFOAEs at A test intensities ( selected from the range of 20 to 60 dB SPL) and DPOAEs at B test intensities ( The amplitude spectrum and signal-to-noise ratio spectrum of the OAEs were selected within the range of 20 to 60 dB SPL, resulting in a total of (A+B) × 2 spectral sequences. The frequency range of each sequence is 0.5-8 kHz, with a frequency interval of 1 / 40 octave. The feature sequence is interpolated from the original OAE spectrum, so the number of points in each feature sequence is 161. The size of the original feature matrix is ​​2(A+B) × 161. The label set is the hearing threshold of the test subject at each frequency. The original hearing threshold is converted to obtain the labeled hearing threshold. The rule is to set the upper and lower limit hearing thresholds. The label value below the 0 dB HL hearing threshold is set to 0 dB HL, and the label value above the 65 dB HL hearing threshold is set to 65 dB HL. All thresholds are normalized to integers with a step size of 5 dB. The values ​​of the above settings can also be flexibly adjusted according to changes in the actual model prediction needs.

[0032] In the process of model construction, in order to give full play to the potential of joint hearing threshold prediction of multi-frequency SFOAEs and DPOAEs, the most commonly used CNN architecture in the field of deep learning was selected to build a regression model, which includes three parts: input layer, feature extractor and output layer: the input layer is centered on the hearing threshold frequency to be predicted, extracts a local sequence of 1 octave frequency range from the original feature matrix, and standardizes the matrix spectrum type of the local sequence spectrum; the feature extractor is mainly composed of convolutional neural networks, which is used to automatically learn the high-dimensional features representing hearing loss in the feature matrix. Abstract: The CNN network consists of one convolutional layer, one activation layer, and one pooling layer. The convolutional layer has 600 convolution kernels, a one-dimensional kernel with a height of 2 (A + B), equal to the number of sequences, and a width of 3. The activation layer uses LeakyReLU, and the pooling layer uses global average pooling. The output layer consists of two fully connected layers, which convert the high-dimensional feature vector output by the feature extractor into the hearing threshold to be predicted. The first fully connected layer has 100 nodes, and the second has 1 node. The hyperparameters involved in the above model architecture can be flexibly adjusted as the dataset content and scale change.

[0033] During model evaluation, 6-fold cross-validation is used for model training and evaluation. The dataset is divided into 6 roughly equal folds using stratified sampling. One of the folds is rotated as the test set, and one of the remaining 5 folds is randomly stratified and sampled as the validation set, leaving the remaining 4 folds as the training set. The training set is used to optimize model parameters to fit the relationship between data and labels. The validation set is used to determine the optimal combination of hyperparameters for model training. The test set is used to ultimately determine the model's generalized predictive performance. A fixed layer of 20 random seed cycles is set outside of cross-validation during model training to ensure that the model's predictive performance is evaluated under 20 fixed but internally different data partitioning methods, resulting in more robust model evaluation results.

[0034] Combine Figure 3As shown in a specific embodiment, the present invention provides a hearing threshold prediction method based on swept-frequency OAEs and a deep learning model. The deep learning model construction includes three parts: a dataset, a model architecture, and a model evaluation. The dataset consists of a feature set and a label set. The feature set includes the amplitude spectrum and signal-to-noise ratio spectrum of two (A+B) spectra, including A-intensity swept-frequency DPOAEs and B-intensity swept-frequency SFOAEs. The label set consists of hearing thresholds for five octave bands from 0.5-8 kHz, normalized by upper and lower limits and rounded to 5 dB. The model architecture consists of an input layer, a feature extractor, and an output layer. The input layer truncates the feature set according to the hearing threshold frequency to be predicted. The feature extractor uses a CNN architecture, and the output layer is a regressor composed of fully connected layers. Model evaluation primarily involves model training and generalization error estimation using 6-fold cross-validation. Twenty random seeds are set in addition to the cross-validation to fully evaluate model performance under different dataset partitioning methods.

[0035] In this embodiment, the local spectrum sequence group obtained from the input layer (size = number of spectra * number of frequency points) is calculated through the convolution layer to obtain multiple feature sequence groups (size = number of convolution kernels * number of sequence points). This is then compressed into a one-dimensional feature vector (size = number of convolution kernels) through a pooling layer after a nonlinear activation (Leaky ReLU). The number of frequency points is 40 (corresponding to a 1-octave spectrum with 1 / 40 octave intervals), and the number of convolution kernels is 600. Zero padding is used to keep the number of sequence points consistent with the number of frequency points. However, in actual operation, these hyperparameters can be adjusted within an appropriate range based on changes in the data structure.

[0036] It should be noted that during the 6-fold cross-validation process, the analysis data is stratified and sampled using the label set. Each fold of data is used as the test set in turn. One fold of the remaining data is randomly stratified and sampled as the validation set, and the remaining folds are used for training. The training set updates the model parameters to fit the relationship between the data and the label. The validation set is used to select the model and training hyperparameters, such as the number of iterative updates. The test set is used to evaluate the generalization error of the model after training. The generalization error of each cross-validation is the average of the errors of the 6-fold test set. The overall generalization error is calculated as the average result of 20 cross-validations with a fixed random seed setting.

[0037] Step S140 , calling a hearing threshold prediction model to predict the hearing threshold of the SFOAEs signal and the DPOAEs signal at different test intensities, so as to output the hearing threshold corresponding to each frequency.

[0038] Specifically, after obtaining the hearing threshold prediction model, the trained hearing threshold prediction model can be called to take the hearing threshold of SFOAEs signals and DPOAEs signals under different test intensities as input, and finally output the hearing threshold prediction results corresponding to each frequency through model processing and analysis.

[0039] The above-mentioned hearing threshold prediction method based on swept-frequency OAEs and deep learning models increases the demand for information complementarity for the joint acquisition of high-frequency resolution SFOAEs and DPOAEs. It adopts a swept-frequency tone paradigm to extract SFOAEs and DPOAEs at multiple frequency points. For the hearing threshold prediction task, a deep learning model based on a convolutional neural network architecture (CNN) is used to learn the characterization information of hearing loss from the acquired multi-test intensity OAEs spectra, thereby achieving objective quantitative prediction of the hearing threshold and further improving the accuracy of hearing threshold prediction.

[0040] like Figure 4 As shown, in one embodiment, the hearing threshold prediction method based on swept frequency OAEs and deep learning model provided by the present invention, step S110 specifically includes the following steps: Step S111 , extracting a DPOAEs signal through a single-segment sweep signal stimulus paradigm, wherein the DPOAEs signal includes a first pitch sweep signal and a second pitch sweep signal, and the first pitch sweep signal and the second pitch sweep signal are respectively transmitted to the human ear through two independent sound tubes.

[0041] In step S112 , a SFOAEs signal is extracted using a stimulus sound paradigm of four-segment sweep frequency signals. The SFOAEs signal includes a stimulus sound and an inhibitory sound, and the stimulus sound and the inhibitory sound are respectively transmitted to the human ear through two independent sound tubes.

[0042] The frequencies of the first fundamental frequency sweep signal and the second fundamental frequency sweep signal are respectively the first frequency and the second frequency, and the ratio between the first frequency and the second frequency is any value within a first preset range. The four-segment sweep signal includes a first-segment sweep signal, a second-segment sweep signal, a third-segment sweep signal, and a fourth-segment sweep signal. The stimulus sound is present in each segment of the four-segment sweep signal and exhibits an initial phase that alternates between positive and negative. The inhibitory sound is present in the third-segment sweep signal and the fourth-segment sweep signal with the same initial phase. The frequencies of the stimulus sound and the inhibitory sound are respectively the third frequency and the fourth frequency, and the ratio between the third frequency and the fourth frequency is any value within a second preset range.

[0043] like Figure 5 As shown, in one embodiment, the hearing threshold prediction method based on swept frequency OAEs and deep learning model provided by the present invention, step S120 specifically includes the following steps: Step S121 , for the DPOAEs signal, the dynamic poles of the tracking filter are set to the instantaneous frequency of the swept-frequency DPOAEs signal induced by the DPOAEs signal swept-frequency paradigm, and the dynamic zeros are set to the first frequency and the second frequency respectively.

[0044] Step S122 : For the residual signal of the SFOAEs signal, the dynamic pole of the tracking filter is set to the third frequency induced by the SFOAEs signal sweeping paradigm, and the dynamic zero is set to the fourth frequency.

[0045] like Figure 6 As shown, in one embodiment, the hearing threshold prediction method based on swept frequency OAEs and deep learning model provided by the present invention, step S120 specifically further includes the following steps: In step S123, the filtered multiple time domain signals are stacked and averaged, and the odd buffer and the even buffer are stored one by one in an odd-even order. The average value of the time domain signal difference between the odd buffer and the even buffer is the noise signal of the OAEs signal.

[0046] In step S124 , the amplitudes of the OAEs signal and noise at each induced moment are estimated by time-frequency analysis, and TFA is performed using a continuous Hanning window to perform discrete Fourier transform on the window to calculate the resolution points.

[0047] Step S125 , taking the amplitude at the sweep frequency corresponding to the center moment of the current window in the DFT results of each window as the swept OAEs signal and noise result at the center moment of the current window, to determine the amplitude spectrum, noise spectrum, and signal-to-noise ratio spectrum of the swept DPOAEs signal and SFOAEs signal under a single test intensity.

[0048] In some embodiments, the present invention provides a method for predicting hearing thresholds based on swept frequency OAEs and a deep learning model, wherein the data set of the deep learning model includes a feature set and a label set, wherein the feature set includes a feature matrix consisting of the amplitude spectrum and the signal-to-noise ratio spectrum of the DPOAEs signal and the SFOAEs signal at multiple test intensities. The label set is the hearing threshold of each frequency to be predicted. The architecture of the deep learning model includes an input layer, a self-feature extractor, and an output layer. The input layer is used to intercept a local feature matrix from the feature matrix. The self-feature extractor uses a CNN framework to learn a high-dimensional feature vector from the feature matrix. The output layer is used to convert the high-dimensional feature vector into a fixed random seed through a fully connected layer as a regressor.

[0049] like Figure 7 As shown, in one embodiment, the hearing threshold prediction method based on swept frequency OAEs and deep learning model provided by the present invention, step S130 specifically includes the following steps: Step S131, selecting the amplitude spectra and signal-to-noise ratio spectra of the SFOAEs signals under a first number of test intensities and the DPOAEs signals under a second number of test intensities to construct a feature set, wherein the feature set has a third number of spectrum sequences; the third number is twice the sum of the first number and the second number.

[0050] Step S132 : Convert the hearing threshold of each frequency to be predicted into a hearing threshold with a label, and set the rules of the upper and lower hearing thresholds to construct a label set.

[0051] The rule for the upper and lower hearing thresholds is to set the hearing threshold label value below the lowest hearing threshold as the lowest hearing threshold, and set the hearing threshold label value above the highest hearing threshold as the highest hearing threshold, while normalizing all hearing thresholds to a fixed step size.

[0052] Step S133: Divide the dataset into 6 equal folds through stratified sampling, and use 6-fold cross validation to train and evaluate the deep learning model.

[0053] Among them, the deep learning model is a regression model built with CNN architecture. The activation layer function of the deep learning model adopts Leaky ReLU activation function, and the pooling layer adopts global average pooling.

[0054] The following describes the hearing threshold prediction device based on swept-frequency OAEs and a deep learning model provided by the present invention. The hearing threshold prediction device based on swept-frequency OAEs and a deep learning model described below and the hearing threshold prediction method based on swept-frequency OAEs and a deep learning model described above can be referenced to each other.

[0055] like Figure 8 As shown, in one embodiment, a hearing threshold prediction device based on swept frequency OAEs and a deep learning model includes a signal acquisition module 810, an information extraction module 820, a model training module 830 and a hearing threshold prediction module 840.

[0056] The signal acquisition module 810 is used to extract SFOAEs signals and DPOAEs signals using a unified sweep frequency stimulation signal paradigm. The stimulation sound paradigm of the SFOAEs signal is a four-segment sweep frequency signal, and the stimulation sound paradigm of the DPOAEs signal is a single-segment sweep frequency signal.

[0057] The information extraction module 820 is used to filter out interference signals from the residual signal of a single DPOAEs signal or a SFOAEs signal through a tracking filter with dynamically changing zeros and poles, so as to extract the swept frequency OAEs spectrum information.

[0058] The model training module 830 is used to construct a data set for a deep learning model based on the swept frequency OAEs spectrum information, train and evaluate the deep learning model, and obtain a hearing threshold prediction model.

[0059] The hearing threshold prediction module 840 is used to call the hearing threshold prediction model to predict the hearing threshold of the SFOAEs signal and the DPOAEs signal under different test intensities, so as to output the hearing threshold corresponding to each frequency.

[0060] In this embodiment, the present invention provides a hearing threshold prediction device based on swept frequency OAEs and a deep learning model, wherein the signal acquisition module 810 is specifically configured to: The DPOAEs signal is extracted through a stimulation sound paradigm of a single-segment sweep signal. The DPOAEs signal includes a first fundamental frequency sweep signal and a second fundamental frequency sweep signal, and the first fundamental frequency sweep signal and the second fundamental frequency sweep signal are respectively transmitted to the human ear through two independent sound tubes.

[0061] The SFOAEs signal is extracted through a stimulation sound paradigm of four-segment sweep frequency signals. The SFOAEs signal includes stimulation sound and inhibition sound, and the stimulation sound and inhibition sound are transmitted to the human ear through two independent sound tubes respectively.

[0062] The frequencies of the first fundamental frequency sweep signal and the second fundamental frequency sweep signal are respectively a first frequency and a second frequency, and a ratio between the first frequency and the second frequency is any value within a first preset range.

[0063] The four-segment sweep frequency signal includes a first sweep frequency signal, a second sweep frequency signal, a third sweep frequency signal, and a fourth sweep frequency signal. The stimulating sound is present in each segment of the four sweep frequency signals and exhibits an initial phase that alternates between positive and negative. The suppressing sound is present in the third sweep frequency signal and the fourth sweep frequency signal with the same initial phase. The frequencies of the stimulating sound and the suppressing sound are the third frequency and the fourth frequency, respectively. The ratio between the third frequency and the fourth frequency is any value within a second preset range.

[0064] In this embodiment, the hearing threshold prediction device based on swept frequency OAEs and deep learning models provided by the present invention, the information extraction module 820 is specifically used to: For the DPOAEs signal, the dynamic poles of the tracking filter are set to the instantaneous frequency of the swept-frequency DPOAEs signal induced by the DPOAEs signal swept-frequency paradigm, and the dynamic zeros are set to the first frequency and the second frequency respectively.

[0065] For the residual signal of the SFOAEs signal, the dynamic pole of the tracking filter is set to the third frequency induced by the SFOAEs signal sweep paradigm, and the dynamic zero is set to the fourth frequency.

[0066] In this embodiment, the hearing threshold prediction device based on swept frequency OAEs and deep learning models provided by the present invention, the information extraction module 820 is further configured to: The multiple filtered time domain signals are stacked and averaged, and the odd buffer and the even buffer are stored one by one in odd-even order. The average value of the time domain signal difference between the odd buffer and the even buffer is the noise signal of the OAEs signal.

[0067] The amplitudes of OAEs signals and noise at each induced moment were estimated by time-frequency analysis, and TFA was performed using a continuous Hanning window to calculate the resolution points by performing discrete Fourier transform on the window.

[0068] The amplitude at the swept frequency corresponding to the center moment of the current window in the DFT results of each window is taken as the swept OAEs signal and noise result at the center moment of the current window to determine the amplitude spectrum, noise spectrum and signal-to-noise ratio spectrum of the swept DPOAEs signal and SFOAEs signal under a single test intensity.

[0069] In this embodiment, the present invention provides a hearing threshold prediction device based on swept-frequency OAEs and a deep learning model. The deep learning model's dataset includes a feature set and a label set. The feature set contains a feature matrix consisting of the amplitude spectra and signal-to-noise ratio spectra of DPOAE and SFOAE signals at multiple test intensities. The label set contains the hearing thresholds for each frequency to be predicted.

[0070] The architecture of the deep learning model includes an input layer, a self-feature extractor, and an output layer. The input layer is used to extract the local feature matrix from the feature matrix. The self-feature extractor uses the CNN framework to learn high-dimensional feature vectors from the feature matrix. The output layer is used to convert the high-dimensional feature vector into a fixed random seed through a fully connected layer as a regressor.

[0071] In this embodiment, the present invention provides a hearing threshold prediction device based on frequency sweep OAEs and a deep learning model, wherein the model training module 830 is specifically configured to: A feature set is constructed by selecting amplitude spectra and signal-to-noise ratio spectra of SFOAEs signals under a first number of test intensities and DPOAEs signals under a second number of test intensities, wherein the feature set has a third number of spectrum sequences; the third number is twice the sum of the first number and the second number.

[0072] The hearing thresholds of each frequency to be predicted are converted into hearing thresholds with labels, and the rules of upper and lower hearing thresholds are set to construct a label set.

[0073] The rule for the upper and lower hearing thresholds is to set the hearing threshold label value below the lowest hearing threshold as the lowest hearing threshold, and set the hearing threshold label value above the highest hearing threshold as the highest hearing threshold, while normalizing all hearing thresholds to a fixed step size.

[0074] In this embodiment, the hearing threshold prediction device based on swept frequency OAEs and deep learning model provided by the present invention, the model training module 830 is further configured to: The dataset was divided into 6 equal folds through stratified sampling, and 6-fold cross-validation was used to train and evaluate the deep learning model.

[0075] Among them, the deep learning model is a regression model built with CNN architecture. The activation layer function of the deep learning model adopts Leaky ReLU activation function, and the pooling layer adopts global average pooling.

[0076] Figure 9 The following is a schematic diagram of the physical structure of an electronic device. The electronic device may be a smart terminal, and its internal structure diagram may be as follows: Figure 9 As shown. The electronic device includes a processor, an internal memory, and a network interface connected via a system bus. The processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the electronic device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a hearing threshold prediction method based on swept frequency OAEs and a deep learning model is implemented, the method comprising: The SFOAEs and DPOAEs signals were extracted using a unified frequency sweep stimulus paradigm. The SFOAEs stimulus paradigm was a four-segment frequency sweep signal, while the DPOAEs stimulus paradigm was a single-segment frequency sweep signal. The residual signal of a single DPOAEs signal or SFOAEs signal is filtered through a tracking filter with dynamically changing zeros and poles to remove interference signals and extract the swept frequency OAEs spectrum information. A dataset of deep learning models was constructed based on the frequency-sweep OAEs spectrum information. The deep learning models were trained and evaluated to obtain a hearing threshold prediction model. The hearing threshold prediction model is called to predict the hearing threshold of SFOAEs and DPOAEs signals under different test intensities to output the hearing threshold corresponding to each frequency.

[0077] Those skilled in the art will understand that Figure 9 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the electronic device to which the solution of the present invention is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement. In another aspect, the present invention further provides a computer storage medium storing a computer program, which, when executed by a processor, implements a method for predicting hearing thresholds based on swept frequency OAEs and a deep learning model, the method comprising: The SFOAEs and DPOAEs signals were extracted using a unified frequency sweep stimulus paradigm. The SFOAEs stimulus paradigm was a four-segment frequency sweep signal, while the DPOAEs stimulus paradigm was a single-segment frequency sweep signal. The residual signal of a single DPOAEs signal or SFOAEs signal is filtered through a tracking filter with dynamically changing zeros and poles to remove interference signals and extract the swept frequency OAEs spectrum information. A dataset of deep learning models was constructed based on the frequency-sweep OAEs spectrum information. The deep learning models were trained and evaluated to obtain a hearing threshold prediction model. The hearing threshold prediction model is called to predict the hearing threshold of SFOAEs and DPOAEs signals under different test intensities to output the hearing threshold corresponding to each frequency. In yet another aspect, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium, and when the processor executes the computer instructions, implements a method for predicting hearing thresholds based on swept frequency OAEs and a deep learning model, the method comprising: The SFOAEs and DPOAEs signals were extracted using a unified frequency sweep stimulus paradigm. The SFOAEs stimulus paradigm was a four-segment frequency sweep signal, while the DPOAEs stimulus paradigm was a single-segment frequency sweep signal. The residual signal of a single DPOAEs signal or SFOAEs signal is filtered through a tracking filter with dynamically changing zeros and poles to remove interference signals and extract the swept frequency OAEs spectrum information. A dataset of deep learning models was constructed based on the frequency-sweep OAEs spectrum information. The deep learning models were trained and evaluated to obtain a hearing threshold prediction model. The hearing threshold prediction model is called to predict the hearing threshold of SFOAEs and DPOAEs signals under different test intensities to output the hearing threshold corresponding to each frequency. Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory.

[0078] By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0079] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0080] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A hearing threshold prediction method based on frequency sweep OAEs and deep learning model, characterized in that: The method comprises: The SFOAEs and DPOAEs signals were extracted using a unified frequency sweep stimulation signal paradigm, wherein the SFOAEs signal stimulation sound paradigm was a four-segment frequency sweep signal, and the DPOAEs signal stimulation sound paradigm was a single-segment frequency sweep signal. The residual signal of a single DPOAEs signal or SFOAEs signal is filtered through a tracking filter with dynamically changing zeros and poles to remove interference signals and extract the swept frequency OAEs spectrum information. Constructing a data set of a deep learning model based on the swept frequency OAEs spectrum information, training and evaluating the deep learning model to obtain a hearing threshold prediction model; The hearing threshold prediction model is called to predict the hearing threshold of the SFOAEs signal and the DPOAEs signal at different test intensities to output the hearing threshold corresponding to each frequency.

2. The hearing threshold prediction method based on swept frequency OAEs and deep learning model according to claim 1, characterized in that: The method of extracting SFOAEs and DPOAEs signals using a unified frequency sweep stimulation signal paradigm includes: The DPOAEs signal is extracted using a single-segment sweep signal stimulus paradigm, wherein the DPOAEs signal includes a first fundamental frequency sweep signal and a second fundamental frequency sweep signal, and the first fundamental frequency sweep signal and the second fundamental frequency sweep signal are respectively transmitted to the human ear through two independent sound tubes; The SFOAEs signal is extracted through a stimulation sound paradigm of a four-segment sweep signal, wherein the SFOAEs signal includes a stimulation sound and an inhibition sound, and the stimulation sound and the inhibition sound are respectively transmitted to the human ear through two independent sound tubes; The frequencies of the first pitch sweep signal and the second pitch sweep signal are respectively a first frequency and a second frequency, and a ratio between the first frequency and the second frequency is any value within a first preset range; The four-segment sweep frequency signal includes a first-segment sweep frequency signal, a second-segment sweep frequency signal, a third-segment sweep frequency signal, and a fourth-segment sweep frequency signal. The stimulation sound exists in each segment of the four-segment sweep frequency signal and presents an initial phase alternating between positive and negative. The suppression sound exists in the third-segment sweep frequency signal and the fourth-segment sweep frequency signal with the same initial phase. The frequencies of the stimulation sound and the suppression sound are the third frequency and the fourth frequency, respectively. The ratio between the third frequency and the fourth frequency is any value within a second preset range.

3. The hearing threshold prediction method based on swept frequency OAEs and deep learning model according to claim 2, characterized in that: The residual signal of a single DPOAEs signal or a SFOAEs signal is filtered out through a tracking filter with dynamically changing zeros and poles to extract the swept frequency OAEs spectrum information, including: For the DPOAEs signal, the dynamic poles of the tracking filter are set to the instantaneous frequency of the swept-frequency DPOAEs signal induced by the DPOAEs signal swept-frequency paradigm, and the dynamic zeros are set to the first frequency and the second frequency respectively; For the residual signal of the SFOAEs signal, the dynamic pole of the tracking filter is set to the third frequency induced by the SFOAEs signal sweep paradigm, and the dynamic zero is set to the fourth frequency.

4. The hearing threshold prediction method based on swept frequency OAEs and deep learning model according to claim 3, characterized in that: The method further comprises filtering out interference signals from the residual signal of a single DPOAEs signal or a SFOAEs signal through a tracking filter with dynamically changing zeros and poles to extract the swept frequency OAEs spectrum information. The filtered multiple time domain signals are stacked and averaged, and the odd buffer and the even buffer are stored one by one in an odd-even order, wherein the average value of the time domain signal difference between the odd buffer and the even buffer is the noise signal of the OAEs signal; The amplitudes of the OAEs signal and noise at each induced moment were estimated by time-frequency analysis, and TFA was performed using a continuous Hanning window to calculate the resolution points by performing a discrete Fourier transform on the window; The amplitude at the swept frequency corresponding to the center moment of the current window in the DFT results of each window is taken as the swept OAEs signal and noise result at the center moment of the current window to determine the amplitude spectrum, noise spectrum and signal-to-noise ratio spectrum of the swept DPOAEs signal and SFOAEs signal under a single test intensity.

5. The hearing threshold prediction method based on frequency sweep OAEs and deep learning model according to claim 4, characterized in that: The dataset of the deep learning model includes a feature set and a label set. The feature set includes a feature matrix composed of amplitude spectra and signal-to-noise ratio spectra of DPOAEs signals and SFOAEs signals at multiple test intensities; the label set is the hearing threshold of each frequency to be predicted; The architecture of the deep learning model includes an input layer, a self-feature extractor and an output layer. The input layer is used to intercept a local feature matrix from the feature matrix. The self-feature extractor adopts a CNN framework to learn a high-dimensional feature vector from the feature matrix. The output layer is used to convert the high-dimensional feature vector into a fixed random seed through a fully connected layer as a regressor.

6. The hearing threshold prediction method based on swept frequency OAEs and deep learning model according to claim 5, characterized in that: The data set of the deep learning model is constructed based on the frequency sweep OAEs spectrum information, and the deep learning model is trained and evaluated to obtain a hearing threshold prediction model, including: The feature set is constructed by selecting amplitude spectra and signal-to-noise ratio spectra of SFOAEs signals at a first number of test intensities and DPOAEs signals at a second number of test intensities, wherein the feature set has a third number of spectrum sequences; the third number is twice the sum of the first number and the second number; Converting the hearing threshold of each frequency to be predicted into a hearing threshold with a label, and setting rules for upper and lower hearing thresholds to construct the label set; The rule for the upper and lower hearing thresholds is to set the hearing threshold label value below the lowest hearing threshold as the lowest hearing threshold, and to set the hearing threshold label value above the highest hearing threshold as the highest hearing threshold, while normalizing all hearing thresholds to a fixed step size.

7. The hearing threshold prediction method based on swept frequency OAEs and deep learning model according to claim 6, characterized in that: The data set of the deep learning model is constructed based on the frequency-sweep OAEs spectrum information, and the deep learning model is trained and evaluated to obtain a hearing threshold prediction model, further comprising: Dividing the dataset into 6 equal folds by stratified sampling, and training and evaluating the deep learning model using 6-fold cross validation; Among them, the deep learning model is a regression model built with a CNN architecture, the activation layer function of the deep learning model adopts a Leaky ReLU activation function, and the pooling layer adopts global average pooling.

8. A hearing threshold prediction device based on frequency sweep OAEs and deep learning model, characterized in that: A method for predicting hearing thresholds based on swept frequency OAEs and a deep learning model according to any one of claims 1 to 7, the device comprising: A signal acquisition module, configured to extract SFOAEs and DPOAEs signals using a unified frequency sweep stimulation signal paradigm, wherein the SFOAEs signal stimulation sound paradigm is a four-segment frequency sweep signal, and the DPOAEs signal stimulation sound paradigm is a single-segment frequency sweep signal; An information extraction module is used to filter out interference signals from the residual signal of a single DPOAEs signal or SFOAEs signal through a tracking filter with dynamically changing zeros and poles, so as to extract the swept frequency OAEs spectrum information; A model training module is used to construct a data set of a deep learning model based on the swept frequency OAEs spectrum information, train and evaluate the deep learning model, and obtain a hearing threshold prediction model; The hearing threshold prediction module is used to call the hearing threshold prediction model to predict the hearing threshold of SFOAEs signals and DPOAEs signals under different test intensities to output the hearing threshold corresponding to each frequency.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the hearing threshold prediction method based on swept frequency OAEs and a deep learning model are implemented as described in any one of claims 1 to 7.

10. A computer storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the hearing threshold prediction method based on swept frequency OAEs and a deep learning model according to any one of claims 1 to 7 are implemented.

Images