Method for predicting cochlear hearing loss based on pure tone induced SFOAE

Abstract

The invention relates to a method for predicting cochlear hearing loss based on pure tone induced SFOAE, which comprises the following steps of: extracting a stimulation frequency otoacoustic emission signal through a six-section stimulation signal normal form to obtain a stimulation sound playing sequence diagram in single stimulation frequency otoacoustic emission signal acquisition, and extracting a time domain residual signal; and low-frequency background noise in the time domain residual signals is removed through a zero phase shift high-pass filter with cut-off frequency. And performing stack average processing on the plurality of filtered time-domain signals, and storing the time-domain signals in an odd-order buffer area and an even-order buffer area one by one according to an odd-even order. The method comprises the following steps: performing dimension reduction feature extraction on a single stimulation frequency otoacoustic emission signal through principal component analysis, training a machine learning algorithm based on principal component features, and constructing a hearing loss prediction model; and calling the hearing loss prediction model to compare the feature input with the representation of the single stimulation frequency otoacoustic emission signal without dimension reduction in the original data space, and quantitatively predicting the cochlea hearing loss threshold.

Description

Technical Field

[0001] The present invention relates to the field of intelligent medical assistance technology, and in particular to a method for predicting cochlear hearing loss based on pure tone induced SFOAE. Background Art

[0002] Otoacoustic Emissions (OAE) refers to a weak audio energy generated in the inner ear cochlea, transmitted and released to the external auditory canal through the ossicular chain and the eardrum, and is part of the normal function of the human ear. The discovery of otoacoustic emissions confirms that the cochlea, as a peripheral auditory receptor, can not only passively convert external sound signals into bioelectric signals transmitted to the central nervous system to cause hearing, but also has an active energy release process, thus establishing the theory that the cochlea is a two-way transducer.

[0003] Cochlear hearing loss (CHL) is a common hearing loss caused by lesions of cochlear hair cells. The current behavioral hearing threshold test for quantitative diagnosis of CHL in clinical practice is easily interfered by subjective factors such as attention. As a byproduct of the active process of the cochlea, the stimulation frequency otoacoustic emission signal (SFOAEs) can characterize the change of absolute hearing threshold in CHL. At present, the otoacoustic emission detection system used in clinical practice can only detect and analyze transient evoked otoacoustic emission (TEOAE) signals and distorted otoacoustic emission (DPOAE) signals. SFOAEs are an audio energy generated in the cochlea with exactly the same frequency as the stimulus sound. Compared with DPOAEs and TEOAEs, SFOAEs can provide a more frequency-specific response. Its advantage is that the generation of SFOAEs mainly involves the contribution of a mechanism of reflection source, which is simpler and more objective to explain the hearing function of the cochlea.

[0004] Therefore, the existing otoacoustic emission detection methods have limitations in detecting and analyzing transient evoked otoacoustic emission (TEOAE) signals and distorted otoacoustic emission (DPOAE) signals, which makes the interpretation of cochlear hearing function more complicated and difficult to quantitatively and objectively predict the cochlear hearing threshold. Summary of the invention

[0005] Based on this, it is necessary to provide a prediction method for cochlear hearing loss based on pure tone-induced SFOAEs, which can achieve quantitative and objective prediction of cochlear hearing thresholds based on stimulation frequency otoacoustic emissions (SFOAEs) detection and has a relatively simple interpretation of cochlear hearing function, in order to address the above technical problems.

[0006] The present invention provides a method for predicting cochlear hearing loss based on pure tone induced SFOAE, the method comprising:

[0007] Extracting the stimulation frequency otoacoustic emission signal through a six-segment stimulation signal paradigm, obtaining a stimulation sound playback timing diagram during the single stimulation frequency otoacoustic emission signal acquisition process, and extracting a time domain residual signal of the single stimulation frequency otoacoustic emission signal based on the stimulation sound playback timing diagram;

[0008] Using a zero-phase-shift high-pass filter with a cutoff frequency to remove background noise with a frequency lower than a first threshold in the time-domain residual signal, to obtain a plurality of filtered time-domain signals;

[0009] Performing stacking and averaging processing on the filtered multiple time domain signals, and storing the time domain signals one by one in an odd-order buffer and an even-order buffer in an odd-even order;

[0010] Performing dimensionality reduction feature extraction on the single stimulation frequency otoacoustic emission signals of the normal hearing ear and the hearing loss ear by principal component analysis to obtain principal component features, and training a machine learning algorithm based on the principal component features to construct a hearing loss prediction model;

[0011] Calling the hearing loss prediction model to compare the model principal component feature input and the representation of the single stimulation frequency otoacoustic emission signal without dimensionality reduction in the original data space, so as to quantitatively predict the cochlear hearing loss threshold of the single stimulation frequency otoacoustic emission signal;

[0012] The background noise is the average value of the time domain signal difference in the odd-order buffer and the even-order buffer, and the principal component characteristics include the proportion of original data variance explained by each principal component and the cumulative data variance, which are used to characterize the characteristic differences in the stimulation frequency otoacoustic emission signals between the normal hearing ear and the hearing-impaired ear.

[0013] In one embodiment, the filtering of the plurality of time domain signals is performed with stacked average processing to improve the signal-to-noise ratio of the stimulation frequency otoacoustic emission signal, and the time domain signals are stored one by one in an odd-order buffer and an even-order buffer in an odd-even order, before which the method further includes:

[0014] Acquire a sound signal recorded in the ear canal, and store the sound signal in a plurality of separate buffers in sequence, wherein the sound signal includes a stimulus sound artifact, a suppression sound artifact, a stimulus sound induced signal, a suppression sound induced signal, and a residual signal induced by the stimulus sound and suppressed by the suppression sound;

[0015] calculating the signal amplitude of the suppressed stimulation frequency otoacoustic emission signal according to the types of the sound signals stored in the plurality of separate buffers;

[0016] Wherein, twice the amplitude of the suppressed stimulation frequency otoacoustic emission signal is equal to the difference between the amplitudes of the buffer signal containing the stimulation sound artifact and the stimulation sound induced signal.

[0017] In one embodiment, when the signal amplitude of the suppressed acoustic artifact exceeds a second threshold, the stimulation frequency otoacoustic emission signal will be completely suppressed, the signal amplitude of the residual signal induced by the stimulation sound and suppressed by the suppressed sound is zero, and the signal amplitude of the suppressed stimulation frequency otoacoustic emission signal is equal to twice the signal amplitude of the stimulation frequency otoacoustic emission signal induced by the stimulation sound.

[0018] In one embodiment, the extracting dimensionality reduction features of the single stimulation frequency otoacoustic emission signals of the normal hearing ear and the hearing loss ear by principal component analysis to obtain principal component features, and training a machine learning algorithm based on the principal component features to construct a hearing loss prediction model includes:

[0019] Acquire a pure tone stimulation frequency otoacoustic emission signal original data set based on the single stimulation frequency otoacoustic emission signal of the normal hearing ear and the hearing-impaired ear;

[0020] Performing data cleaning on the original data set to divide the original data set into a binary classification task data set and a regression task data set according to the prediction task;

[0021] Among them, the binary classification task data set includes all stimulation frequency otoacoustic emission signals participating in the test task, and is used for the binary classification task of qualitatively distinguishing hearing loss. The regression task data set excludes stimulation frequency otoacoustic emission signals whose hearing frequency threshold exceeds the third threshold, and is used for the regression task of quantitatively predicting the hearing threshold.

[0022] In one embodiment, the method further comprises: extracting dimensionality reduction features of the single stimulation frequency otoacoustic emission signals of the normal hearing ear and the hearing loss ear by principal component analysis to obtain principal component features, and training a machine learning algorithm based on the principal component features to construct a hearing loss prediction model;

[0023] Performing independent principal component analysis on the original data set of each frequency, and converting the original data matrix corresponding to the original data set into a score matrix in the principal component space by orthogonal transformation in each principal component analysis;

[0024] According to the principal component scores in the score matrix, the principal component features for explaining the variance of the original data are selected as inputs of the machine learning algorithm to achieve feature dimensionality reduction;

[0025] Among them, the principal component scores in the score matrix are the projections of the original data set in the principal component space, which are used to characterize the principal component characteristics, and the principal component characteristics selected to explain the variance of the original data are used to characterize the characteristic differences in the stimulation frequency otoacoustic emission signals between the normal hearing ear and the hearing-impaired ear.

[0026] In one embodiment, the single stimulation frequency otoacoustic emission signals of the normal hearing ear and the hearing loss ear are subjected to dimensionality reduction feature extraction by principal component analysis to obtain principal component features, and a machine learning algorithm is trained based on the principal component features to construct a hearing loss prediction model, and then the method further includes:

[0027] In the outer loop, the original data set is divided into multiple first subsets with equal sample numbers, and one first subset is selected from the multiple first subsets as a test set, and other first subsets except the one first subset corresponding to the test set are used as training sets; and

[0028] In the inner loop, the training set obtained in the outer loop is further divided into multiple second subsets with equal sample numbers, and one second subset is selected from the multiple second subsets as the inner loop test set, and the other second subsets except the one second subset corresponding to the inner loop test set are used as the inner loop training set;

[0029] Hyperparameter tuning is performed based on the inner loop test set and the inner loop training set by a grid search method to search for and provide optimal hyperparameters for an outer loop, and the hearing loss model is trained and tested based on the training set and the test set of the outer loop by using the optimal hyperparameters.

[0030] In one embodiment, the method further comprises:

[0031] When the prediction task is the regression task, extracting features of the single-point stimulation frequency otoacoustic emission signal induced under multiple continuous stimulation intensities input by connecting a series of nonlinear processing layers through the feature parameters of the deep learning algorithm, so as to convert the extracted features into abstract features;

[0032] The abstract features are transferred from the lower network layer to the higher network layer to obtain the corresponding high-dimensional features, so as to learn the characteristic associations of the stimulation frequency otoacoustic emission signals between different stimulation intensities.

[0033] The present invention also provides a device for predicting cochlear hearing loss based on pure tone-induced SFOAE, the device comprising:

[0034] A signal extraction module, used to extract the stimulation frequency otoacoustic emission signal through a six-segment stimulation signal paradigm, obtain a stimulation sound playback timing diagram during the single stimulation frequency otoacoustic emission signal acquisition process, and extract a time domain residual signal of the single stimulation frequency otoacoustic emission signal based on the stimulation sound playback timing diagram;

[0035] A filtering processing module, used for removing background noise with a frequency lower than a first threshold in the time domain residual signal through a zero-phase shift high-pass filter with a cutoff frequency, so as to obtain a plurality of filtered time domain signals;

[0036] A buffer processing module, used for performing stacking and averaging processing on the filtered multiple time domain signals, and storing the time domain signals one by one into an odd-order buffer and an even-order buffer in an odd-even order;

[0037] A model building module, for extracting dimensionality reduction features of the single stimulation frequency otoacoustic emission signals of the normal hearing ear and the hearing loss ear through principal component analysis to obtain principal component features, and training a machine learning algorithm based on the principal component features to build a hearing loss prediction model;

[0038] A hearing loss prediction module, used to call the hearing loss prediction model to compare the model principal component feature input and the representation of the single stimulation frequency otoacoustic emission signal without dimension reduction in the original data space, so as to quantitatively predict the cochlear hearing loss threshold of the single stimulation frequency otoacoustic emission signal;

[0039] The background noise is the average value of the time domain signal difference in the odd-order buffer and the even-order buffer, and the principal component characteristics include the proportion of original data variance explained by each principal component and the cumulative data variance, which are used to characterize the characteristic differences in the stimulation frequency otoacoustic emission signals between the normal hearing ear and the hearing-impaired ear.

[0040] The present invention also provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the method for predicting cochlear hearing loss based on pure tone induced SFOAE as described in any one of the above.

[0041] The present invention also provides a computer storage medium storing a computer program, wherein when the computer program is executed by a processor, the method for predicting cochlear hearing loss based on pure tone induced SFOAE as described in any one of the above is implemented.

[0042] The above-mentioned method for predicting cochlear hearing loss based on pure tone-induced SFOAE extracts the stimulation frequency otoacoustic emission signal through a six-segment stimulation signal paradigm, obtains the stimulation sound playback timing diagram during the single stimulation frequency otoacoustic emission signal acquisition process, and extracts the time domain residual signal of the single stimulation frequency otoacoustic emission signal based on the stimulation sound playback timing diagram. Subsequently, the low-frequency background noise in the time domain residual signal is removed by a zero-phase shift high-pass filter with a cutoff frequency to obtain multiple filtered time domain signals. The multiple filtered time domain signals are then stacked and averaged, and the time domain signals are stored one by one in an odd-order buffer and an even-order buffer in an odd-even order. Then, the single stimulation frequency otoacoustic emission signals of normal hearing ears and hearing-impaired ears are subjected to dimensionality reduction feature extraction through principal component analysis to obtain the principal component features, and the machine learning algorithm is trained based on the principal component features to construct a hearing loss prediction model. Finally, the hearing loss prediction model is called to compare the model principal component feature input and the representation of the single stimulation frequency otoacoustic emission signal without dimensionality reduction in the original data space to quantitatively predict the cochlear hearing loss threshold of the single stimulation frequency otoacoustic emission signal. This method extracts the principal component from the stimulation frequency otoacoustic emission signal collected at a single point through principal component analysis technology and machine learning algorithm, and quantitatively predicts the cochlear hearing loss threshold through the hearing loss prediction model constructed by the machine learning algorithm, which improves the accuracy and objectivity of hearing loss prediction, and the prediction results are also relatively simple to interpret the cochlear hearing function. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces 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 creative work.

[0044] Figure 1 This is a schematic diagram of a method for predicting cochlear hearing loss based on pure tone-induced SFOAE provided by the present invention;

[0045] Figure 2 A timing diagram of the stimulation sound playback during the single SFOAEs acquisition process of the method for predicting cochlear hearing loss based on pure tone-induced SFOAE in a specific embodiment provided by the present invention;

[0046] Figure 3 A schematic diagram of a machine learning algorithm flow chart of a method for predicting cochlear hearing loss based on pure tone-induced SFOAE in a specific embodiment provided by the present invention;

[0047] Figure 4A schematic diagram of a 5×5-fold nested cross-validation process of a method for predicting cochlear hearing loss based on pure tone-induced SFOAE in a specific embodiment provided by the present invention;

[0048] Figure 5 The second flow chart of the method for predicting cochlear hearing loss based on pure tone-induced SFOAE provided by the present invention;

[0049] Figure 6 The third flow chart of the method for predicting cochlear hearing loss based on pure tone-induced SFOAE provided by the present invention;

[0050] Figure 7 The fourth flow chart of the method for predicting cochlear hearing loss based on pure tone-induced SFOAE provided by the present invention;

[0051] Figure 8 The fifth flow chart of the method for predicting cochlear hearing loss based on pure tone-induced SFOAE provided by the present invention;

[0052] Fig. 9 The sixth flow chart of the method for predicting cochlear hearing loss based on pure tone-induced SFOAE provided by the present invention;

[0053] Fig.10 A schematic diagram of the structure of a device for predicting cochlear hearing loss based on pure tone-induced SFOAE provided by the present invention;

[0054] Fig.11 This is a diagram of the internal structure of the computer device provided by the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are 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.

[0056] Combine the following Figure 1-Figure 11 The present invention describes a method for predicting cochlear hearing loss based on pure tone induced SFOAE.

[0057] like Figure 1 As shown, in one embodiment, a method for predicting cochlear hearing loss based on pure tone induced SFOAE comprises the following steps:

[0058] Step S110, extracting the stimulation frequency otoacoustic emission signal through the six-segment stimulation signal paradigm, obtaining the stimulation sound playback timing diagram in the process of collecting the single stimulation frequency otoacoustic emission signal, and extracting the time domain residual signal of the single stimulation frequency otoacoustic emission signal based on the stimulation sound playback timing diagram.

[0059] Specifically, the server extracts the stimulation frequency otoacoustic emission signal through a six-segment stimulation signal paradigm, obtains a stimulation sound playback timing diagram during the single stimulation frequency otoacoustic emission signal acquisition process, and extracts a time domain residual signal corresponding to the single stimulation frequency otoacoustic emission signal based on the stimulation sound playback timing diagram.

[0060] Combination Figure 2 As shown, in a specific embodiment, the present invention provides a method for predicting cochlear hearing loss based on pure tone-induced SFOAE, and extracts SFOAEs (stimulation frequency otoacoustic emission signals) using a six-segment stimulation signal paradigm, that is, the stimulation sound playback timing diagram in the process of collecting a single SFOAE is divided into six segments, A, B, C, D, M, and N, excluding the last 5ms, to eliminate the influence of system delay and SFOAEs latency. The duration of the first segment M is 2T d , where T d The system delay is measured in advance, which is about 14.5ms. The duration of segments A, B, C, D, and N is T w The stimulation signal includes a probe tone and a suppressor tone, which are played to the human ear through two independent sound tubes. p It is a continuous pure tone with the same phase at segments A, B, C, D and N. s It is a short pure tone, that is, a pure tone multiplied by a cosine length of 5ms for both the rising and falling edges. The suppression sound is not played in the M, A and B segments, and the suppression sound in the C and D segments is inverted.

[0061] Step S120: removing background noise with a frequency lower than a first threshold in the time domain residual signal through a zero-phase-shift high-pass filter with a cutoff frequency, to obtain a plurality of filtered time domain signals.

[0062] Specifically, the server removes background noise with a frequency lower than a set threshold in the time domain residual signal through a zero-phase shift high-pass filter with a cutoff frequency, and obtains a plurality of filtered time domain signals.

[0063] Step S130 , performing stacking and averaging processing on the filtered multiple time domain signals, and storing the time domain signals one by one in an odd-order buffer and an even-order buffer in an odd-even order.

[0064] The background noise is calculated as the average value of the time domain signal difference in the odd-order buffer and the even-order buffer.

[0065] Specifically, the server performs stacking and averaging processing on the multiple filtered time domain signals obtained in step S120, and stores the time domain signals one by one into the odd-order buffer and the even-order buffer in the odd-even order.

[0066] In a specific embodiment, the method for predicting cochlear hearing loss based on pure tone induced SFOAE provided by the present invention is to, after the sound giving-collection process, pass the single SFOAEs time domain residual signal through a cutoff frequency of f c A zero-phase shift high-pass filter is used to remove low-frequency background noise. When extracting the SFOAEs signal at 1 to 8 kHz, f c =500Hz, when extracting the SFOAEs signal at 0.5kHz, f c =350Hz.

[0067] In this embodiment, a real-time artifact rejection method can also be used to reduce transient noise: the difference between the recorded signal amplitudes corresponding to the stimulus sound segments A and B (i.e., AB) only contains background noise. If the root mean square amplitude of AB exceeds the set threshold (0.0008), the trial is eliminated in real time, and additional trials are performed until the predetermined number of trials is completed.

[0068] Subsequently, the filtered multiple SFOAEs time domain signals are stacked and averaged to improve the signal-to-noise ratio (S / N) of the SFOAEs signal. At the same time, they are stored one by one in the odd and even order in two buffers, and the background noise is calculated as the average value of the difference between the time domain signals in the two buffers.

[0069] In this embodiment, the power spectral density (PSD) can also be used to estimate the amplitude of the SFOAEs signal and background noise in the frequency domain. For example, 10 consecutive overlapping Kaiser windows are used, the window length is 42.7ms (2048 points), the overlap between adjacent windows is half the window length, and the number of discrete Fourier transform (DFT) points is equal to the sampling rate, that is, 48000 points. Power spectrum estimation can be performed in MATLAB (MATLAB 2020a, MathWorks Inc., Natick, MA, USA) using the function "Pwelch". After the above calculation process, the power spectrum of the SFOAEs signal and background noise can be obtained.

[0070] It should be noted that in the process of removing background noise, the sound signal recorded in the ear canal includes the stimulus sound artifact R p , suppress acoustic artifacts R s , by the stimulus sound Rp The induced SFOAEs signal SFE, which is generated by the suppression of sound R s The induced SFOAEs signal SFEs and the residual SFOAEs signal SFE' induced by the stimulus and suppressed by the suppression sound. The sound signals of segments A to D collected in the ear canal are stored in four separate buffers A to D in sequence. Since segments A and B only contain stimulus sounds, buffers A and B only contain R p and SFE, buffer C contains R p , R s , SFEs and SFE', due to the suppression sound in segments C and D, the buffer D contains R p , -R s , -SFEs and SFE'. The twice of the suppressed SFOAEs signal amplitude is equal to the difference between the buffer (A+B) and (C+D), and the calculation formula is:

[0071] Suppressed SFOAEs=(A+B)-(C+D)

[0072] =(R p +SFE+R p +SFE)-(R p +R s +SFE'+SFEs)

[0073] -(R p -R s +SFE'-SFEs)=2SFE-2SFE'.

[0074] It should be further explained that if the sound R is suppressed s If the amplitude is large enough, then in the C and D segments, SFOAEs will be completely suppressed, so SFE' = 0, and the suppressed SFOAEs signal is equal to twice the amplitude of the SFOAEs signal induced by the stimulus, that is, 2SFE. Therefore, the SFOAEs time domain residual signal to be extracted is equal to [(A+B)-(C+D)] / 2.

[0075] Step S140, performing dimensionality reduction feature extraction on the single stimulation frequency otoacoustic emission signals of the normal hearing ear and the hearing loss ear through principal component analysis to obtain principal component features, and training a machine learning algorithm based on the principal component features to construct a hearing loss prediction model.

[0076] Specifically, the server performs dimensionality reduction feature extraction on the single stimulation frequency otoacoustic emission signals of the normal hearing ear and the hearing-impaired ear through principal component analysis to obtain the principal component features, and trains the machine learning algorithm based on the extracted principal component features to complete the construction of the hearing loss prediction model.

[0077] The principal component characteristics include the proportion of original data variance explained by each principal component and the cumulative data variance, which are used to characterize the characteristic differences in the stimulation frequency otoacoustic emission signals between normal hearing ears and hearing-impaired ears.

[0078] Combination Figure 3 As shown, in a specific embodiment, the present invention provides a method for predicting cochlear hearing loss based on pure tone-induced SFOAE, and the machine learning (ML) algorithm includes data sets, data cleaning, feature engineering, model construction, model training and evaluation. In addition, the key to the success of the ML algorithm is to select appropriate feature engineering and model algorithms. PCA (principal component analysis) is introduced in the feature extraction process to maximize the extraction of information related to the hearing threshold from SFOAEs under multiple stimulation intensities. In the model algorithm selection, by comparing a variety of traditional ML algorithms and deep learning techniques, in order to maximize the potential of SFOAEs to predict hearing loss (CHL), the average error of 20 nested cross-validations is used as the generalization error estimate of the model during the model training and evaluation process to obtain a more stable model evaluation result.

[0079] In the process of model building, the data needs to be cleaned first. For example, the data set of 0.5-4kHz contains the stimulus intensity L p = SFOAEs signal induced by 5-60dB SPL. The amplitude of SFOAEs signal at 8kHz is relatively small, so the stimulation intensity L is not eliminated at this frequency. p =65, 70dB SPL induced SFOAEs signals. In addition, for the binary classification task involving qualitative discrimination of hearing loss status and the regression task involving quantitative prediction of hearing threshold, the data set can be divided according to different prediction tasks. The unrestricted data set includes the SFOAEs data of all ears involved in the test and is used for the binary classification task. For more complex regression tasks, the data set can be restricted according to the generation mechanism of otoacoustic emission signals (OAEs). For example, OHCs lesions (when IHCs are completely normal) can generally explain hearing loss of ≤50 (or ≤60) dB HL [43,94]. The data set of the regression task is restricted to the hearing loss range that can be explained by OHCs lesions (≤60dB HL), that is, the data at the frequency of hearing threshold >60dB HL are excluded from the regression task.

[0080] In this embodiment, an independent PCA analysis needs to be performed for the original data set at each frequency to reduce the correlation between the characteristic variables under each stimulation intensity. In each PCA analysis, the original data matrix is ​​converted into a score matrix in the principal component space by an orthogonal transformation. The principal component score (PC score) in the score matrix is ​​the projection of the original data set in the principal component space and is used as the PC (principal component feature). Through PCA analysis, the individual differences in SFOAEs characteristics between subjects can be expressed by the principal component score. Since the first few principal components can generally explain most of the variance of the original data, other principal components can be abandoned to simplify the representation of the data and achieve feature dimensionality reduction.

[0081] In addition, under the premise of ensuring that sufficient information related to the hearing threshold has been extracted, the first three PCs are retained as the feature input of the model, that is, a total of 73.9%, 84.8%, 88.0%, 87.8%, and 80.0% of the data variance are explained in 0.5-8kHz, respectively. Among them, the first two principal components can already explain the differences in SFOAEs characteristics between normal hearing ears and hearing loss (CHL) ears. It is worth noting that according to the analysis process of machine learning (ML), the data set will be divided into a training set and a test set. In the actual feature extraction process, only the PCA dimensionality reduction is used for the training set to obtain the projection matrix of the principal component space, and then the projection matrix of the training set is directly used for the feature dimensionality reduction of the test set, which can effectively avoid information leakage.

[0082] In this embodiment, in order to give full play to the potential of single-point SFOAEs signals to predict hearing loss (CHL), a variety of commonly used machine learning algorithms can be selected to construct regression and classification models respectively, and their performance on the SFOAEs data set can be compared to select the optimal prediction algorithm. For the regression task of quantitatively predicting the hearing threshold, deep learning technology can be used to predict the hearing threshold. The feature parameters of deep learning can be extracted by connecting nonlinear processing layers in series, and then the single-point SFOAEs signals induced by multiple continuous stimulation intensities of the input are converted into more abstract features, and passed from the low network layer to the higher network layer to generate high-dimensional features.

[0083] For example, using convolutional neural networks (CNN) to automatically learn L p =5~60dB SPL (or at 8kHz: L p =5~70dB SPL) stimulus intensities to achieve feature-independent end-to-end hearing threshold prediction task.

[0084] Specifically, the CNN regression model includes an input layer, a convolution layer, a pooling layer and two fully connected layers. The two features of SFOAEs amplitude and S / N under k stimulus intensities (0.5-4kHz: k=12, 8kHz: k=14) are encoded into a k×2 matrix as the input of CNN. CNN realizes automatic feature extraction and selection through convolution layer and pooling layer respectively. The convolution layer uses 3 convolution kernels of different sizes, where the width of the convolution kernel is equal to the width 2 of the input matrix, and the convolution kernel performs convolution operation on the input matrix in the dimension of stimulus intensity. This embodiment does not use zero padding technology, and the sliding step size S=1 of the convolution kernel, so the convolution operation converts the input matrix into a series of feature maps of size (k-h+1)×1, where h is the variable convolution kernel height, providing hyperparameters to be tuned for the CNN model, and each feature map shares weight parameters, reducing CNN parameters. The convolution operation not only considers the SFOAEs features under each stimulus intensity, but also considers the SFOAEs feature associations between adjacent stimulus intensities and the SFOAEs I / O curve morphology of the local area. The combination of convolution kernels of different sizes can learn the SFOAEs feature associations of different local areas. A pooling layer is added after the convolution layer, and the average value can be simply extracted from the feature map obtained by the convolution layer through global average pooling (GAP), effectively reducing the size of the feature map. In addition, the use of global average pooling instead of maximum pooling can potentially smooth the existence of local valleys in the I / O curve of the SFOAEs signal and reduce the prediction error of the model. Finally, the various feature parameters after the pooling operation are spliced ​​together, and the hearing threshold is output after two fully connected layers. In this embodiment, the Leaky ReLU function is selected as the activation function of the convolution layer and the fully connected layer, which can realize the nonlinear mapping of features.

[0085] Step S150, calling the hearing loss prediction model to compare the model principal component feature input and the representation of the single stimulation frequency otoacoustic emission signal without dimensionality reduction in the original data space, so as to quantitatively predict the cochlear hearing loss threshold of the single stimulation frequency otoacoustic emission signal.

[0086] Specifically, the server calls the hearing loss prediction model constructed in step S140 to compare the model principal component feature input and the representation of the single stimulation frequency otoacoustic emission signal without dimensionality reduction in the original data space to quantitatively predict the cochlear hearing loss threshold of the single stimulation frequency otoacoustic emission signal.

[0087] The above-mentioned method for predicting cochlear hearing loss based on pure tone-induced SFOAE extracts the stimulation frequency otoacoustic emission signal through a six-segment stimulation signal paradigm, obtains the stimulation sound playback timing diagram during the single stimulation frequency otoacoustic emission signal acquisition process, and extracts the time domain residual signal of the single stimulation frequency otoacoustic emission signal based on the stimulation sound playback timing diagram. Subsequently, the low-frequency background noise in the time domain residual signal is removed by a zero-phase shift high-pass filter with a cutoff frequency to obtain multiple filtered time domain signals. The multiple filtered time domain signals are then stacked and averaged, and the time domain signals are stored one by one in an odd-order buffer and an even-order buffer in an odd-even order. Then, the single stimulation frequency otoacoustic emission signals of normal hearing ears and hearing-impaired ears are subjected to dimensionality reduction feature extraction through principal component analysis to obtain the principal component features, and the machine learning algorithm is trained based on the principal component features to construct a hearing loss prediction model. Finally, the hearing loss prediction model is called to compare the model principal component feature input and the representation of the single stimulation frequency otoacoustic emission signal without dimensionality reduction in the original data space to quantitatively predict the cochlear hearing loss threshold of the single stimulation frequency otoacoustic emission signal. This method extracts the principal component from the stimulation frequency otoacoustic emission signal collected at a single point through principal component analysis technology and machine learning algorithm, and quantitatively predicts the cochlear hearing loss threshold through the hearing loss prediction model constructed by the machine learning algorithm, which improves the accuracy and objectivity of hearing loss prediction, and the prediction results are also relatively simple to interpret the cochlear hearing function.

[0088] like Figure 5 As shown, in one embodiment, the method for predicting cochlear hearing loss based on pure tone induced SFOAE provided by the present invention performs stacking and averaging processing on the filtered multiple time domain signals, and stores the time domain signals one by one in an odd order buffer and an even order buffer in an odd-even order, and also includes:

[0089] Step S510, obtaining the sound signal recorded in the ear canal, and storing the sound signal in a plurality of separate buffers in sequence, wherein the sound signal includes stimulation sound artifacts, suppression sound artifacts, stimulation sound induced signals, suppression sound induced signals, and residual signals induced by stimulation sounds and suppressed by suppression sounds.

[0090] Step S520: Calculate the signal amplitude of the suppressed stimulation frequency otoacoustic emission signal according to the types of the sound signals stored in the plurality of separate buffers.

[0091] Wherein, twice the amplitude of the suppressed stimulation frequency otoacoustic emission signal is equal to the difference between the amplitudes of the buffer signal containing the stimulation sound artifact and the stimulation sound evoked signal.

[0092] It should be noted that when the signal amplitude of the suppressed acoustic artifact exceeds the second threshold, the stimulation frequency otoacoustic emission signal will be completely suppressed, the signal amplitude of the residual signal induced by the stimulation sound and suppressed by the suppressed sound is zero, and the signal amplitude of the suppressed stimulation frequency otoacoustic emission signal is equal to twice the signal amplitude of the stimulation frequency otoacoustic emission signal induced by the stimulation sound.

[0093] like Figure 6 As shown, in one embodiment, the method for predicting cochlear hearing loss based on pure tone induced SFOAE provided by the present invention performs dimensionality reduction feature extraction on single stimulation frequency otoacoustic emission signals of normal hearing ears and hearing loss ears through principal component analysis to obtain principal component features, and trains a machine learning algorithm based on the principal component features to construct a hearing loss prediction model, specifically comprising the following steps:

[0094] Step S142: acquiring an original data set of pure tone stimulation frequency otoacoustic emission signals based on the single stimulation frequency otoacoustic emission signals of the normal hearing ear and the hearing impaired ear.

[0095] Step S144, performing data cleaning on the original data set to divide the original data set into a binary classification task data set and a regression task data set according to the prediction task.

[0096] Among them, the binary classification task data set includes all the stimulation frequency otoacoustic emission signals involved in the test task, which is used for the binary classification task of qualitatively distinguishing hearing loss. The regression task data set excludes the stimulation frequency otoacoustic emission signals whose hearing frequency threshold exceeds the third threshold, and is used for the regression task of quantitatively predicting the hearing threshold.

[0097] like Figure 7 As shown, in one embodiment, the method for predicting cochlear hearing loss based on pure tone induced SFOAE provided by the present invention performs dimensionality reduction feature extraction on single stimulation frequency otoacoustic emission signals of normal hearing ears and hearing loss ears through principal component analysis to obtain principal component features, and trains a machine learning algorithm based on the principal component features to construct a hearing loss prediction model, which specifically also includes the following steps:

[0098] Step S146, performing independent principal component analysis on the original data sets of each frequency, and in each principal component analysis, converting the original data matrix corresponding to the original data set into a score matrix in the principal component space through orthogonal transformation.

[0099] Step S148, based on the principal component scores in the score matrix, select the principal component features used to explain the variance of the original data as the input of the machine learning algorithm to achieve feature dimensionality reduction.

[0100] The principal component scores in the score matrix are the projections of the original data set in the principal component space, which are used to characterize the principal component characteristics. The principal component characteristics selected to explain the variance of the original data are used to characterize the characteristic differences in the stimulation frequency otoacoustic emission signals between normal hearing ears and hearing-impaired ears.

[0101] like Figure 8 As shown, in one embodiment, the method for predicting cochlear hearing loss based on pure tone induced SFOAE provided by the present invention performs dimensionality reduction feature extraction on single stimulation frequency otoacoustic emission signals of normal hearing ears and hearing loss ears through principal component analysis to obtain principal component features, and trains a machine learning algorithm based on the principal component features to construct a hearing loss prediction model, and then further includes the following steps:

[0102] Step S810, in the outer loop, the original data set is divided into multiple first subsets with equal sample numbers, and one first subset is selected from the multiple first subsets as a test set, and the other first subsets except the one first subset corresponding to the test set are used as training sets.

[0103] Step S820, in the inner loop, the training set obtained in the outer loop is further divided into multiple second subsets with equal sample numbers, and one second subset is selected from the multiple second subsets as the inner loop test set, and the other second subsets except the one second subset corresponding to the inner loop test set are used as the inner loop training sets.

[0104] Step S830, hyperparameter tuning is performed based on the inner loop test set and the inner loop training set by using a grid search method to provide optimal hyperparameters for outer loop search, and the hearing loss model is trained and tested based on the outer loop training set and test set by using the optimal hyperparameters.

[0105] Combination Figure 4As shown, in a specific embodiment, the prediction method for cochlear hearing loss based on pure tone-induced SFOAE provided by the present invention uses 5×5 fold nested CV for model training and evaluation. In the outer loop, the data set is divided into 5 folds with roughly equal sample numbers, of which 1 fold subset is used as a test set in turn, and the remaining 4 fold subsets are used as training sets. In the inner loop, the 4 fold training set data obtained in the outer loop is further divided into 5 folds with roughly equal sample numbers (of which 1 fold is the inner loop test set and the remaining 4 folds are the inner loop training set). The 5-fold CV of the inner loop implements hyperparameter tuning based on the grid search method, searches for and provides the optimal hyperparameters for the outer loop, and uses the optimal hyperparameters to train the training set provided by the outer loop to obtain the final model, and at the same time, the model is tested on the test set provided by the outer loop. The test set of the outer loop does not participate in the training of the model, which can effectively prevent information leakage and thus obtain a relatively low model performance deviation. The nested cross-validation provides an error estimate that is almost unbiased for the true error, and the model error of a single nested CV is the average value of the test set error of the 5-fold outer loop. In addition, in order to improve the stability of the model results, the average error of 20 nested CVs with fixed random seeds can be used as the final generalization error estimate of the model. By fixing the random seed, the way of data division in each nested CV is consistent, and irrelevant factors are removed for performance comparison between models.

[0106] like Fig. 9 As shown, in one embodiment, the method for predicting cochlear hearing loss based on pure tone-induced SFOAE provided by the present invention further includes the following steps:

[0107] Step S910, when the prediction task is a regression task, feature extraction is performed on the single-point stimulation frequency otoacoustic emission signal induced under multiple continuous stimulation intensities input by connecting a series of nonlinear processing layers through the feature parameters of the deep learning algorithm, so as to convert the extracted features into abstract features.

[0108] Step S920, transferring the abstract features from the low network layer to the high network layer to obtain corresponding high-dimensional features, so as to learn the characteristic association of the stimulation frequency otoacoustic emission signals between different stimulation intensities.

[0109] The following is a description of the device for predicting cochlear hearing loss based on pure tone-induced SFOAE provided by the present invention. The device for predicting cochlear hearing loss based on pure tone-induced SFOAE described below and the method for predicting cochlear hearing loss based on pure tone-induced SFOAE described above can be referenced to each other.

[0110] like Fig.10As shown, in one embodiment, a device for predicting cochlear hearing loss based on pure tone induced SFOAE includes a signal extraction module 1010, a filtering processing module 1020, a buffer processing module 1030, a model building module 1040 and a hearing loss prediction module 1050.

[0111] The signal extraction module 1010 is used to extract the stimulation frequency otoacoustic emission signal through the six-segment stimulation signal paradigm, obtain the stimulation sound playback timing diagram during the single stimulation frequency otoacoustic emission signal acquisition process, and extract the time domain residual signal of the single stimulation frequency otoacoustic emission signal based on the stimulation sound playback timing diagram.

[0112] The filtering processing module 1020 is used to remove background noise with a frequency lower than a first threshold in the time domain residual signal through a zero-phase shift high-pass filter with a cutoff frequency, so as to obtain a plurality of filtered time domain signals.

[0113] The buffer processing module 1030 is used to perform stacking and averaging processing on the filtered multiple time domain signals, and store the time domain signals one by one into the odd-order buffer and the even-order buffer in the odd-even order.

[0114] The model building module 1040 is used to perform dimensionality reduction feature extraction on the single stimulation frequency otoacoustic emission signals of the normal hearing ear and the hearing loss ear through principal component analysis to obtain the principal component features, and train the machine learning algorithm based on the principal component features to build a hearing loss prediction model.

[0115] The hearing loss prediction module 1050 is used to call the hearing loss prediction model to compare the model principal component feature input and the representation of the single stimulation frequency otoacoustic emission signal without dimensionality reduction in the original data space, so as to quantitatively predict the cochlear hearing loss threshold of the single stimulation frequency otoacoustic emission signal.

[0116] Among them, the background noise is the average value of the time domain signal difference in the odd-order buffer and the even-order buffer, and the principal component characteristics include the proportion of original data variance explained by each principal component and the cumulative data variance, which are used to characterize the characteristic differences in the stimulation frequency otoacoustic emission signals between the normal hearing ear and the hearing-impaired ear.

[0117] In this embodiment, the prediction device for cochlear hearing loss based on pure tone induced SFOAE provided by the present invention further includes a signal amplitude calculation module for:

[0118] Acquire the sound signal recorded in the ear canal and store it in a plurality of separate buffers in sequence, wherein the sound signal includes a stimulation sound artifact, a suppression sound artifact, a stimulation sound induced signal, a suppression sound induced signal, and a residual signal induced by the stimulation sound and suppressed by the suppression sound.

[0119] The signal amplitude of the suppressed stimulation frequency otoacoustic emission signal is calculated according to the types of the sound signals stored in the plurality of separate buffers.

[0120] Wherein, twice the amplitude of the suppressed stimulation frequency otoacoustic emission signal is equal to the difference between the amplitudes of the buffer signal containing the stimulation sound artifact and the stimulation sound evoked signal.

[0121] In this embodiment, the present invention provides a prediction device for cochlear hearing loss based on pure tone induced SFOAE. When the signal amplitude of the suppressed acoustic artifact exceeds the second threshold, the stimulation frequency otoacoustic emission signal will be completely suppressed, and the signal amplitude of the residual signal induced by the stimulation sound and suppressed by the suppressed sound is zero. The signal amplitude of the suppressed stimulation frequency otoacoustic emission signal is equal to twice the signal amplitude of the stimulation frequency otoacoustic emission signal induced by the stimulation sound.

[0122] In this embodiment, the model building module of the prediction device for cochlear hearing loss based on pure tone induced SFOAE provided by the present invention is specifically used for:

[0123] The original data set of pure tone stimulation frequency otoacoustic emission signals is obtained based on the single stimulation frequency otoacoustic emission signals of the normal hearing ear and the hearing loss ear.

[0124] The original data set is cleaned to divide it into a binary classification task data set and a regression task data set according to the prediction task.

[0125] Among them, the binary classification task data set includes all the stimulation frequency otoacoustic emission signals involved in the test task, which is used for the binary classification task of qualitatively distinguishing hearing loss. The regression task data set excludes the stimulation frequency otoacoustic emission signals whose hearing frequency threshold exceeds the third threshold, and is used for the regression task of quantitatively predicting the hearing threshold.

[0126] In this embodiment, the model building module of the prediction device for cochlear hearing loss based on pure tone induced SFOAE provided by the present invention is specifically used for:

[0127] Independent principal component analysis is performed on the original data sets of each frequency, and in each principal component analysis, the original data matrix corresponding to the original data set is converted into a score matrix in the principal component space by orthogonal transformation.

[0128] According to the principal component scores in the score matrix, the principal component features used to explain the variance of the original data are selected as the input of the machine learning algorithm to achieve feature dimensionality reduction.

[0129] The principal component scores in the score matrix are the projections of the original data set in the principal component space, which are used to characterize the principal component characteristics. The principal component characteristics selected to explain the variance of the original data are used to characterize the characteristic differences in the stimulation frequency otoacoustic emission signals between normal hearing ears and hearing-impaired ears.

[0130] In this embodiment, the prediction device for cochlear hearing loss based on pure tone induced SFOAE provided by the present invention further includes a nested cross-validation module for:

[0131] In the outer loop, the original data set is divided into multiple first subsets with equal sample numbers, and one first subset is selected from the multiple first subsets as the test set, and the other first subsets except the one first subset corresponding to the test set are used as the training set.

[0132] In the inner loop, the training set obtained in the outer loop is further divided into multiple second subsets with equal sample numbers, and one second subset is selected from the multiple second subsets as the inner loop test set, and the other second subsets except the one second subset corresponding to the inner loop test set are used as the inner loop training sets.

[0133] Hyperparameters are tuned based on the inner loop test set and the inner loop training set through the grid search method to search and provide the optimal hyperparameters for the outer loop, and the hearing loss model is trained and tested based on the outer loop training set and test set through the optimal hyperparameters.

[0134] In this embodiment, the prediction device for cochlear hearing loss based on pure tone induced SFOAE provided by the present invention further includes a deep learning processing module for:

[0135] When the prediction task is a regression task, the feature parameters of the deep learning algorithm are connected in series with a nonlinear processing layer to extract the features of the single-point stimulation frequency otoacoustic emission signal induced under multiple continuous stimulation intensities of the input, so as to convert the extracted features into abstract features.

[0136] The abstract features are passed from the low network layer to the high network layer to obtain the corresponding high-dimensional features to learn the characteristic associations of the stimulation frequency otoacoustic emission signals between different stimulation intensities.

[0137] Fig.11 An example of a physical structure diagram of an electronic device is shown. The electronic device may be a smart terminal, and its internal structure diagram may be as follows: Fig.11As shown. The electronic device includes a processor, a 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 the 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 method for predicting cochlear hearing loss based on pure tone-induced SFOAE is implemented, and the method includes:

[0138] Extracting the stimulation frequency otoacoustic emission signal through the six-segment stimulation signal paradigm, obtaining the stimulation sound playback timing diagram in the process of collecting the single stimulation frequency otoacoustic emission signal, and extracting the time domain residual signal of the single stimulation frequency otoacoustic emission signal based on the stimulation sound playback timing diagram;

[0139] Using a zero-phase-shift high-pass filter with a cutoff frequency to remove background noise with a frequency lower than a first threshold in the time-domain residual signal, thereby obtaining a plurality of filtered time-domain signals;

[0140] Performing stacking and averaging processing on the filtered multiple time domain signals, and storing the time domain signals one by one in an odd-order buffer and an even-order buffer in an odd-even order;

[0141] Principal component analysis was used to extract the dimensionality reduction features of the single stimulation frequency otoacoustic emission signals of the normal hearing ear and the hearing loss ear, and the principal component features were obtained. The machine learning algorithm was trained based on the principal component features to build a hearing loss prediction model.

[0142] The hearing loss prediction model is called to compare the model principal component feature input and the representation of the single stimulation frequency otoacoustic emission signal without dimensionality reduction in the original data space, so as to quantitatively predict the cochlear hearing loss threshold of the single stimulation frequency otoacoustic emission signal;

[0143] Among them, the background noise is the average value of the time domain signal difference in the odd-order buffer and the even-order buffer, and the principal component characteristics include the proportion of original data variance explained by each principal component and the cumulative data variance, which are used to characterize the characteristic differences in the stimulation frequency otoacoustic emission signals between the normal hearing ear and the hearing-impaired ear.

[0144] Those skilled in the art will understand that Fig.11 The structure shown in the figure is only a block diagram of a part of the structure related to the scheme of the present invention, and does not constitute a limitation on the electronic device to which the scheme of the present invention is applied. The specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0145] On the other hand, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, implements a method for predicting cochlear hearing loss based on pure tone-induced SFOAE, the method comprising:

[0146] Extracting the stimulation frequency otoacoustic emission signal through the six-segment stimulation signal paradigm, obtaining the stimulation sound playback timing diagram in the process of collecting the single stimulation frequency otoacoustic emission signal, and extracting the time domain residual signal of the single stimulation frequency otoacoustic emission signal based on the stimulation sound playback timing diagram;

[0147] Using a zero-phase-shift high-pass filter with a cutoff frequency to remove background noise with a frequency lower than a first threshold in the time-domain residual signal, thereby obtaining a plurality of filtered time-domain signals;

[0148] Performing stacking and averaging processing on the filtered multiple time domain signals, and storing the time domain signals one by one in an odd-order buffer and an even-order buffer in an odd-even order;

[0149] Principal component analysis was used to extract the dimensionality reduction features of the single stimulation frequency otoacoustic emission signals of the normal hearing ear and the hearing loss ear, and the principal component features were obtained. The machine learning algorithm was trained based on the principal component features to build a hearing loss prediction model.

[0150] The hearing loss prediction model is called to compare the model principal component feature input and the representation of the single stimulation frequency otoacoustic emission signal without dimensionality reduction in the original data space, so as to quantitatively predict the cochlear hearing loss threshold of the single stimulation frequency otoacoustic emission signal;

[0151] Among them, the background noise is the average value of the time domain signal difference in the odd-order buffer and the even-order buffer, and the principal component characteristics include the proportion of original data variance explained by each principal component and the cumulative data variance, which are used to characterize the characteristic differences in the stimulation frequency otoacoustic emission signals between the normal hearing ear and the hearing-impaired ear.

[0152] In another aspect, a computer program product or a computer program is provided, the computer program product or the computer program comprising computer instructions, the computer instructions being 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, a method for predicting cochlear hearing loss based on pure tone-induced SFOAE is implemented, the method comprising:

[0153] Extracting the stimulation frequency otoacoustic emission signal through the six-segment stimulation signal paradigm, obtaining the stimulation sound playback timing diagram in the process of collecting the single stimulation frequency otoacoustic emission signal, and extracting the time domain residual signal of the single stimulation frequency otoacoustic emission signal based on the stimulation sound playback timing diagram;

[0154] Using a zero-phase-shift high-pass filter with a cutoff frequency to remove background noise with a frequency lower than a first threshold in the time-domain residual signal, thereby obtaining a plurality of filtered time-domain signals;

[0155] Performing stacking and averaging processing on the filtered multiple time domain signals, and storing the time domain signals one by one in an odd-order buffer and an even-order buffer in an odd-even order;

[0156] Principal component analysis was used to extract the dimensionality reduction features of the single stimulation frequency otoacoustic emission signals of the normal hearing ear and the hearing loss ear, and the principal component features were obtained. The machine learning algorithm was trained based on the principal component features to build a hearing loss prediction model.

[0157] The hearing loss prediction model is called to compare the model principal component feature input and the representation of the single stimulation frequency otoacoustic emission signal without dimensionality reduction in the original data space, so as to quantitatively predict the cochlear hearing loss threshold of the single stimulation frequency otoacoustic emission signal;

[0158] Among them, the background noise is the average value of the time domain signal difference in the odd-order buffer and the even-order buffer, and the principal component characteristics include the proportion of original data variance explained by each principal component and the cumulative data variance, which are used to characterize the characteristic differences in the stimulation frequency otoacoustic emission signals between the normal hearing ear and the hearing-impaired ear.

[0159] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing related 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.

[0160] 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 (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0161] The technical features of the above embodiments may 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.

[0162] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for predicting cochlear hearing loss based on pure tone induced SFOAE, characterized in that: The method comprises: Extracting the stimulation frequency otoacoustic emission signal through a six-segment stimulation signal paradigm, obtaining a stimulation sound playback timing diagram during the single stimulation frequency otoacoustic emission signal acquisition process, and extracting a time domain residual signal of the single stimulation frequency otoacoustic emission signal based on the stimulation sound playback timing diagram; Using a zero-phase-shift high-pass filter with a cutoff frequency to remove background noise with a frequency lower than a first threshold in the time-domain residual signal, to obtain a plurality of filtered time-domain signals; Performing stacking and averaging processing on the filtered multiple time domain signals, and storing the time domain signals one by one in an odd-order buffer and an even-order buffer in an odd-even order; Performing dimensionality reduction feature extraction on the single stimulation frequency otoacoustic emission signals of the normal hearing ear and the hearing loss ear by principal component analysis to obtain principal component features, and training a machine learning algorithm based on the principal component features to construct a hearing loss prediction model; Calling the hearing loss prediction model to compare the model principal component feature input and the representation of the single stimulation frequency otoacoustic emission signal without dimensionality reduction in the original data space, so as to quantitatively predict the cochlear hearing loss threshold of the single stimulation frequency otoacoustic emission signal; The background noise is the average value of the time domain signal difference in the odd-order buffer and the even-order buffer, and the principal component characteristics include the proportion of original data variance explained by each principal component and the cumulative data variance, which are used to characterize the characteristic differences in the stimulation frequency otoacoustic emission signals between the normal hearing ear and the hearing-impaired ear.

2. The method for predicting cochlear hearing loss based on pure tone induced SFOAE according to claim 1, characterized in that: The method further comprises: performing stacking and averaging processing on the filtered multiple time domain signals to improve the signal-to-noise ratio of the stimulation frequency otoacoustic emission signal, and storing the time domain signals one by one in an odd-order buffer and an even-order buffer in an odd-even order. Acquire a sound signal recorded in the ear canal, and store the sound signal in a plurality of separate buffers in sequence, wherein the sound signal includes a stimulus sound artifact, a suppression sound artifact, a stimulus sound induced signal, a suppression sound induced signal, and a residual signal induced by the stimulus sound and suppressed by the suppression sound; calculating the signal amplitude of the suppressed stimulation frequency otoacoustic emission signal according to the types of the sound signals stored in the plurality of separate buffers; Wherein, twice the amplitude of the suppressed stimulation frequency otoacoustic emission signal is equal to the difference between the amplitudes of the buffer signal containing the stimulation sound artifact and the stimulation sound induced signal.

3. The method for predicting cochlear hearing loss based on pure tone induced SFOAE according to claim 2, characterized in that: When the signal amplitude of the suppressed acoustic artifact exceeds the second threshold, the stimulation frequency otoacoustic emission signal will be completely suppressed, the signal amplitude of the residual signal induced by the stimulation sound and suppressed by the suppressed sound is zero, and the signal amplitude of the suppressed stimulation frequency otoacoustic emission signal is equal to twice the signal amplitude of the stimulation frequency otoacoustic emission signal induced by the stimulation sound.

4. The method for predicting cochlear hearing loss based on pure tone induced SFOAE according to claim 1, characterized in that: The method comprises extracting dimensionality reduction features of the single stimulation frequency otoacoustic emission signals of the normal hearing ear and the hearing loss ear by principal component analysis to obtain principal component features, and training a machine learning algorithm based on the principal component features to construct a hearing loss prediction model, including: Acquire a pure tone stimulation frequency otoacoustic emission signal original data set based on the single stimulation frequency otoacoustic emission signal of the normal hearing ear and the hearing-impaired ear; Performing data cleaning on the original data set to divide the original data set into a binary classification task data set and a regression task data set according to the prediction task; Among them, the binary classification task data set includes all stimulation frequency otoacoustic emission signals participating in the test task, and is used for the binary classification task of qualitatively distinguishing hearing loss. The regression task data set excludes stimulation frequency otoacoustic emission signals whose hearing frequency threshold exceeds the third threshold, and is used for the regression task of quantitatively predicting the hearing threshold.

5. The method for predicting cochlear hearing loss based on pure tone induced SFOAE according to claim 4, characterized in that: The method further comprises extracting dimensionality reduction features of the single stimulation frequency otoacoustic emission signals of the normal hearing ear and the hearing loss ear by principal component analysis to obtain principal component features, and training a machine learning algorithm based on the principal component features to construct a hearing loss prediction model, and further comprising: Performing independent principal component analysis on the original data set of each frequency, and converting the original data matrix corresponding to the original data set into a score matrix in the principal component space by orthogonal transformation in each principal component analysis; According to the principal component scores in the score matrix, the principal component features for explaining the variance of the original data are selected as inputs of the machine learning algorithm to achieve feature dimensionality reduction; Among them, the principal component scores in the score matrix are the projections of the original data set in the principal component space, which are used to characterize the principal component characteristics, and the principal component characteristics selected to explain the variance of the original data are used to characterize the characteristic differences in the stimulation frequency otoacoustic emission signals between the normal hearing ear and the hearing-impaired ear.

6. The method for predicting cochlear hearing loss based on pure tone induced SFOAE according to claim 5, characterized in that: The method further comprises extracting dimensionality reduction features of the single stimulation frequency otoacoustic emission signals of the normal hearing ear and the hearing loss ear by principal component analysis to obtain principal component features, and training a machine learning algorithm based on the principal component features to construct a hearing loss prediction model, and then further comprising: In the outer loop, the original data set is divided into multiple first subsets with equal sample numbers, and one first subset is selected from the multiple first subsets as a test set, and other first subsets except the one first subset corresponding to the test set are used as training sets; and In the inner loop, the training set obtained in the outer loop is further divided into multiple second subsets with equal sample numbers, and one second subset is selected from the multiple second subsets as the inner loop test set, and the other second subsets except the one second subset corresponding to the inner loop test set are used as the inner loop training set; Hyperparameter tuning is performed based on the inner loop test set and the inner loop training set by a grid search method to search for and provide optimal hyperparameters for an outer loop, and the hearing loss model is trained and tested based on the training set and the test set of the outer loop by using the optimal hyperparameters.

7. The method for predicting cochlear hearing loss based on pure tone induced SFOAE according to claim 5, characterized in that: The method further comprises: When the prediction task is the regression task, extracting features of the single-point stimulation frequency otoacoustic emission signal induced under multiple continuous stimulation intensities input by connecting a series of nonlinear processing layers through the feature parameters of the deep learning algorithm, so as to convert the extracted features into abstract features; The abstract features are transferred from the lower network layer to the higher network layer to obtain the corresponding high-dimensional features, so as to learn the characteristic associations of the stimulation frequency otoacoustic emission signals between different stimulation intensities.

8. A prediction device for cochlear hearing loss based on pure tone induced SFOAE, characterized in that: The device comprises: A signal extraction module, used to extract the stimulation frequency otoacoustic emission signal through a six-segment stimulation signal paradigm, obtain a stimulation sound playback timing diagram during the single stimulation frequency otoacoustic emission signal acquisition process, and extract a time domain residual signal of the single stimulation frequency otoacoustic emission signal based on the stimulation sound playback timing diagram; A filtering processing module, used for removing background noise with a frequency lower than a first threshold in the time domain residual signal through a zero-phase-shift high-pass filter with a cutoff frequency, so as to obtain a plurality of filtered time domain signals; A buffer processing module, used for performing stacking and averaging processing on the filtered multiple time domain signals, and storing the time domain signals one by one into an odd-order buffer and an even-order buffer in an odd-even order; A model building module, for extracting dimensionality reduction features of the single stimulation frequency otoacoustic emission signals of the normal hearing ear and the hearing loss ear through principal component analysis to obtain principal component features, and training a machine learning algorithm based on the principal component features to build a hearing loss prediction model; A hearing loss prediction module, used to call the hearing loss prediction model to compare the model principal component feature input and the representation of the single stimulation frequency otoacoustic emission signal without dimension reduction in the original data space, so as to quantitatively predict the cochlear hearing loss threshold of the single stimulation frequency otoacoustic emission signal; The background noise is the average value of the time domain signal difference in the odd-order buffer and the even-order buffer, and the principal component characteristics include the proportion of original data variance explained by each principal component and the cumulative data variance, which are used to characterize the characteristic differences in the stimulation frequency otoacoustic emission signals between the normal hearing ear and the hearing-impaired ear.

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 method according to any one of claims 1 to 7 are implemented.

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 method according to any one of claims 1 to 7 are implemented.

Images