Automatic pure tone audiometry method based on bayesian active learning and gaussian process regression

Abstract

This invention provides an automated pure-tone audiometry method based on Bayesian active learning and Gaussian process regression, comprising: performing a pre-test and establishing a regression model using a Gaussian process, and constructing an audiogram using the regression data; using the principle of Bayesian active learning to screen the iteration point with the greatest uncertainty, and obtaining the audiometric signal frequency and hearing level; performing audiometry on the subject to obtain data and adding it to the dataset; performing Gaussian process regression and active learning iterations until the target audiogram is obtained. This method replaces the fixed steps of traditional audiometry methods with the predictive power of machine learning, which can shorten the test time, reduce the test inaccuracy that may be caused by long test times, and provide a smooth and accurate audiogram with a much higher accuracy than traditional audiometry methods.

Description

Technical Field

[0001] This invention relates to the field of pure-tone audiometry technology, specifically to an automatic pure-tone audiometry method, system, medium, and terminal based on Bayesian active learning and Gaussian process regression. Background Technology

[0002] Pure-tone audiometry is a widely used hearing assessment tool that evaluates an individual's auditory sensitivity at different frequencies and volume levels. The test typically uses headphones to transmit pure-tone sounds and asks the test subject to indicate whether they can hear each sound. This test is often used to diagnose hearing impairment, assess the degree of hearing loss, and provide a basis for hearing therapy and rehabilitation. However, traditional pure-tone audiometry has many limitations, such as long testing time, insufficient information output, and the need for human guidance; these limitations are becoming increasingly prominent in modern society.

[0003] According to a World Health Organization report, 466 million people worldwide suffer from pathological hearing loss, and it is projected that nearly 1.5 billion people globally will have varying degrees of hearing loss by 2050, leading to a growing demand for hearing-related medical services. Conversely, there is a severe shortage of hearing-related professionals worldwide, exacerbating the inadequacy of these services. Therefore, an automated and efficient pure-tone audiometry method is needed to meet the needs of both patients and healthcare providers. This method would help improve testing efficiency, reduce testing costs, and increase testing reliability, while also improving the accessibility and efficiency of hearing-related medical services.

[0004] Traditional pure-tone audiometry uses the Hughson-Westlake method proposed in 1944. When the subject hears the sound signal, they press a button, the signal stops, the sound intensity decreases by 10 dB, and then the sound is presented again. If no response is heard, the intensity is increased by 5 dB until the subject responds again. The threshold is usually defined as the lowest intensity at which a response occurs twice out of three increases. The GB / T 16403 standard used in my country is an improvement upon this. For decades, pure-tone audiometry relied on traditional manual testing. However, based on its regular characteristics and testing procedures, it is very suitable for automation. In fact, automated pure-tone audiometry (AA) has been researched for many years, especially with the rise of artificial intelligence and the development of telemedicine in recent years. AA has seen increasing development, but it has not yet been widely used clinically. The Bekesy audiometer, which appeared in the 1940s, was the earliest automated pure-tone audiometry device. It used an automated adjustment method, allowing the listener to control an attenuator to determine the intensity of the presented stimulus that they could not hear. Subsequently, from the 1970s onwards, research on the use of limiting and adaptive methods in automated pure-tone testing became more widespread. Automated limiting involves using a computer-automated audiometry process based on the Hughson-Westlake method. The automated limiting method, building upon the Hughson-Westlake method, records the hearing threshold during both the rising and falling phases, resulting in higher testing efficiency.

[0005] However, these automated methods are basically within the existing framework of traditional hearing tests, and they still face many of the problems inherent in traditional methods, such as...

[0006] 1. Each test frequency requires multiple audio stimuli with either high or low detection probabilities.

[0007] 2. Repeated or nearly identical sound stimuli will be applied multiple times near the detection threshold.

[0008] 3. The choice of test sound is related to its frequency and intensity, which may affect the test results if the test is conducted by uncooperative subjects.

[0009] These problems have led to longer detection times, insufficient curve information, and lower detection accuracy. Summary of the Invention

[0010] To address the shortcomings of existing technologies, the purpose of this invention is to provide an automatic pure-tone audiometry method, system, medium, and terminal based on Bayesian active learning and Gaussian process regression.

[0011] According to one aspect of the present invention, an automatic pure-tone audiometry method based on Bayesian active learning and Gaussian process regression is provided, comprising:

[0012] Pure tones of different frequencies at the same hearing level (dB HL) were taken and benchmark tests were performed on the left and right ears respectively. Based on the benchmark tests and Bayesian active learning, a Gaussian process model was established.

[0013] Based on the dataset D obtained from the benchmark test, Gaussian process regression is performed based on the Gaussian process model to obtain the benchmark audiograms for the left and right ears;

[0014] Analyze the difference between the expected response entropy of the left and right ear benchmark audiograms and the average information entropy under a given distribution, and take the maximum value of the difference to maximize the mutual information between the expected response and the Gaussian process regression. The signal x corresponding to the maximum value is then... * As the signal for the next pure tone test;

[0015] The test signal x is determined based on the characteristics of the human ear. * Feasibility, adjust the test signal x * Intensity range, according to the masking rules, add masking to the non-test ear, according to the adjusted x * Play the corresponding test audio to the subjects and observe their feedback to obtain a new test set y. * ;

[0016] x * ,y * Add dataset D, and use the updated dataset D to perform the Gaussian process regression to generate a new audiogram. Repeat the acquisition of x. * and y * The loop continues until the regression result exceeds a set threshold, at which point it stops; the final binaural audiogram is generated based on the regression result.

[0017] Preferably, pure tones of different frequencies at the same hearing level are used to perform benchmark tests on the left and right ears respectively. Based on the benchmark tests and Bayesian active learning, a Gaussian process model is established, including:

[0018] Take 40dB HL pure tones at 250Hz, 500Hz, 1000Hz, 2000Hz, 4000Hz, and 8000Hz and perform air conduction pure tone tests on the subjects; or take 20dB HL pure tones at 250Hz, 500Hz, 1000Hz, 2000Hz, 3000Hz, and 4000Hz and perform bone conduction pure tone tests on the subjects.

[0019] Test the left and right ears separately and record the test results, including the frequency of the audio, the hearing level, and the corresponding audiometry results of whether the audio was heard;

[0020] Using the test results as the basic input data, a Gaussian process model is established based on Bayesian active learning.

[0021] Preferably, the step of using the test results as basic input data to establish a Gaussian process model based on Bayesian active learning includes:

[0022] According to Bayes' theorem Where D = (X, y) is the existing training set, X is the current input dataset, containing two dimensional components x = (v, i) of signal frequency and hearing level, and y is the label of the hearing test result; ω represents the model weight component, p(ω) is the prior probability, p(y|X, ω) represents the conditional likelihood function of obtaining the y label from the input dataset X under the weight of ω, and p(y|X) is the marginal probability;

[0023] Since p(y|X)=∫p(y|X,ω)p(ω)dω is a constant, we get p(ω|D)∝p(y|X,ω)p(ω);

[0024] Assuming the noise follows a Gaussian distribution, the likelihood function is transformed. The result shows that p(y|X,ω) follows a Gaussian distribution.

[0025] From the properties of Gaussian processes, we know that ω follows a normal distribution N(0,∑p). Substituting the likelihood function p(y|X,ω) and the prior probability p(ω) into the equation, we get p(ω|D)~N(μω,∑ω).

[0026] The posterior probability results conform to a multidimensional Gaussian distribution. Based on the posterior distribution of the weights, the result y corresponding to a given test point z is obtained. * Predicted probability distribution

[0027] Using the kernel function K(x,x')=φ(x) T • φ(x')=<φ(x),φ(x')> performs a nonlinear transformation on the input data;

[0028] For the objective function y = f(x) = φ(x) T Substituting the objective function into the input z in the posterior distribution, we obtain that the mean and covariance of the Gaussian process after nonlinear transformation are both related to the kernel function. The Gaussian process f(x) model under the kernel function condition is established as f(x)~N(μ(x),k(x,x')), where μ(x) is the mean of this Gaussian process distribution, and k(x,x') is both the kernel function and the covariance matrix of this Gaussian process distribution.

[0029] Preferably, the step of using the benchmark dataset D as a basis and performing Gaussian process regression based on the Gaussian process model to obtain the left and right ear benchmark audiograms includes:

[0030] The kernel function parameters for the input data of the Gaussian process regression model are designed as follows:

[0031] For the audio listening level dimension input in the test data, a kernel function k is established using a linear kernel function as input. L (i,i')=ii'; For the audio frequency input in the test data, a radial base kernel is selected. As input to the kernel function for the frequency dimension of the test data, d = (v - v'), and l is the Lengthscal that controls its fluctuation.

[0032] Construct a covariance matrix kernel function that is an additive combination of linear kernel functions and radial basis kernel functions. Where v is the frequency component of the input data and i is the intensity component of the input data;

[0033] Based on the established Gaussian process model f(x)~N(μ(x),k(x,x')), we know in the function space that μ(x) is the mean matrix of this Gaussian process and k(x,x') is the variance matrix.

[0034] For the input data, the training function is y = f(x) + ε. and the objective function f * =f(Z) establishes the relation

[0035] Using the properties of the multivariate Gaussian conditional distribution, the given predicted value f is calculated. * posterior probability distribution

[0036]

[0037] The normal cumulative distribution function φ(x)=∫N(x|0,1)dx is used as the classification fitting function, and the Expectation Propagation approximation method is used as the likelihood function. Based on the Gaussian process model, Gaussian process classification is performed to generate a three-dimensional probability distribution map with frequency v as the abscissa, hearing level i as the ordinate, and audiometry feedback probability f(x) as the ordinate.

[0038] In the three-dimensional probability distribution graph, the point on each horizontal axis that has the feedback probability closest to the set value is selected, the vertical coordinate of the point is extracted, and the point is plotted as a line. The line is then projected onto a plane with frequency v as the horizontal axis and hearing level i as the vertical axis. The image corresponding to this plane is the audiogram of this test.

[0039] Preferably, the analysis involves taking the maximum value of the difference between the expected response entropy and the average information entropy under a given distribution to maximize the mutual information between the expected response and the Gaussian process regression, and then assigning the signal x corresponding to the maximum value to this maximum value. * As the signal for the next pure tone test, it includes:

[0040] Using the Shannon entropy calculation method, we can find the test data point that maximizes the decrease in expected posterior entropy.

[0041] I = argmax x H[y|x,D]-E θ～p(θ|D) [H[y|x,θ]]

[0042] In the formula, the first part is to calculate the information entropy of the predicted value p(y|x,D) of the input point x given the existing training set D, and the second part is to represent the overall average information entropy after inputting x given the distribution θ. The difference between the two represents the difference between the information entropy of the input sample and the average information entropy of the population.

[0043] Transform the first part of the formula based on the mathematical expectation formula of the composite function of random variables. h represents the entropy under the Bernoulli distribution, φ(f) x f is the probability density function of the standard Gaussian distribution. x Let x be the probability value corresponding to x in the existing Gaussian process model, and step represent f. x The step size for discretizing the corresponding normal probability density function is determined by the number of sampling points nbins in the Gaussian distribution probability, i.e., the entire interval is divided into nbins smaller intervals according to step. The probability distribution matrix ∑f of the Gaussian process regression is then used. x The information entropy corresponding to the hearing test at this point is calculated using I1;

[0044] Transform the latter half into This formula discretizes the expected entropy and calculates the average information entropy of the distribution p(y|x,θ) for each input x under the discrete standard Gaussian probability density.

[0045] From I = max(I1 - I2), we can calculate f. x The corresponding coordinates (v, i) are used to temporarily determine the audio frequency and hearing level test signal x* to be used in the next test.

[0046] Preferably, the step of determining the signal x based on the characteristics of the human ear... * Feasibility, adjust the signal x * Intensity range, according to the masking rules, add masking to the non-test ear, according to the adjusted x *Play the corresponding test audio to the subjects and observe their feedback to obtain a new test set y. * ,include:

[0047] If a bone conduction hearing threshold baseline test is not performed, and the middle ear is not understood and only air conduction audiometry is performed, if the generated test audio i1 >= 40 dB HL, masking noise needs to be added to the non-test ear;

[0048] If a preliminary bone conduction test has been performed, and the test audio i1 of the test ear exceeds the bone conduction hearing threshold of the non-test ear by 40 dB HL, the non-test ear should be masked.

[0049] The noise level is the frequency corresponding to the baseline air conduction hearing threshold of the non-tested ear plus 15 dB EM.

[0050] If the test audio i1 is more than 60dB HL greater than the same frequency in the previous test, ignore this test point and select a point with slightly smaller entropy in I = max(I1-I2); if the test conditions are still not met, continue to the next test point.

[0051] Play the selected test audio x to the subject * If masking is required, masking noise should be added to the non-test ear.

[0052] Observe the feedback from the subjects and record whether the subjects heard the test audio. * Record it.

[0053] Preferably, the step of x * ,y * Add D, and use the updated dataset D to perform Gaussian process regression to generate new audiograms. Repeat this process to obtain x. * and y * The loop continues until the regression result exceeds a set threshold, at which point it stops. Based on the regression result, a final binaural hearing map is generated, including:

[0054] The frequency of the new test audio, hearing level data x * and feedback results y * Add the test dataset D, and based on the Gaussian process model, perform Gaussian process regression with the updated dataset as input to obtain a new posterior probability distribution map, which is the audiogram of the corresponding test ear. Then, according to the entropy calculation method, select the test point with the maximum uncertainty, select the test point to determine the test audio frequency and hearing level intensity, adjust the test audio and masking strategy, and conduct pure tone audio tests on the subjects; repeat the above steps to continuously obtain new test audio and new audiogram distributions until the threshold is reached.

[0055] After the test is stopped, the points corresponding to the posterior probability of each frequency in the final obtained posterior probability distribution map are selected and fitted and connected to generate the pure tone audiometry chart of the air conduction or bone conduction of the corresponding test ear.

[0056] Preferably, setting the threshold includes:

[0057] The maximum number of hearing test cycles cannot exceed the set number of cycles. That is, after the set number of hearing tests, the test will automatically stop and the current audiogram will be selected as the final audiogram for the corresponding ear being tested.

[0058] If the mean Euclidean distance between the mean matrix of the current iteration and the mean matrix of the previous iteration is less than a set value, it can be determined that there is no significant change between the two iterations, and the test can be stopped.

[0059] According to a second aspect of the present invention, a terminal is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, can be used to perform the method described herein.

[0060] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, can be used to perform the method described thereon.

[0061] Compared with the traditional Hughson-Westlake-based automatic pure-tone hearing threshold testing method, this invention, starting from the testing principle, completely abandons the testing process centered on "raising by 5 and lowering by 10". It changes the selection of test audio in pure-tone testing from fixed rules to prediction using machine learning methods, and has at least one advantage over traditional methods:

[0062] In this embodiment of the invention, the frequency and intensity range of the entire audiogram are used as training input, resulting in audiometry data across the entire frequency band from 125Hz to 16000Hz, with a maximum frequency resolution accurate to 1Hz. Traditional audiometry methods, however, can only obtain audiometry results at fixed frequencies such as 250Hz, 500Hz, 1000Hz, 2000Hz, 4000Hz, 8000Hz, and 16000Hz. Therefore, the output audiometry results of this embodiment are several orders of magnitude larger than those of traditional methods, achieving more accurate and smoother audiometry curve plotting.

[0063] This invention employs a Bayesian-based active learning method, selecting Gaussian process regression data with listening test results as input as samples. It then selects specific frequencies and intensities of audio data with the greatest uncertainty as audio sampling points, ensuring that each selected listening frequency point contains sufficient information. In contrast, traditional methods perform multiple similar listening tests near a certain listening frequency threshold, resulting in insufficient information in each test and low listening efficiency.

[0064] In this invention, the selection of frequency points during testing is unpredictable for the test subject. Traditional methods require gradually increasing or decreasing the intensity and frequency of the test audio according to specific rules, with clearly defined rules, typically the "5 dB increase / decrease" method. That is, if the subject responds to the test audio, the audio is played at a 10 dB decrease; if the subject does not respond, the audio is played at a 5 dB increase. This method is highly predictable for the subject, who can predict the test audio based on the pattern. This may lead to non-cooperative subjects affecting the audio measurement, resulting in inaccurate results. This invention, however, uses a method of selecting test audio based on uncertainty. Each test audio is given by the regression result of the previous method, making it impossible for the subject to predict the audio information, thus ensuring effective and accurate audio measurement results.

[0065] The embodiments of this invention focus on the selection of test audio and the generation of audiograms, and can be applied to air conduction pure tone audiometry and bone conduction pure tone audiometry. In bone conduction pure tone audiometry, the frequency range and hearing level range can be reduced according to specific circumstances. Attached Figure Description

[0066] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0067] Figure 1 This is a flowchart of an automatic pure-tone audiometry method based on Bayesian active learning and Gaussian process regression in one embodiment of the present invention.

[0068] Figure 2 This is a framework diagram of an automatic pure-tone audiometry method based on Bayesian active learning and Gaussian process regression in a preferred embodiment of the present invention.

[0069] Figure 3 This is a flowchart illustrating the implementation of the Gaussian process regression method based on the audiometry frequency and sound pressure level input during the audiometry process in a preferred embodiment of the present invention.

[0070] Figure 4 This is a flowchart of a preferred embodiment of the present invention, which describes a selection strategy method for choosing the next audiometry data using a Bayesian active learning strategy after generating the posterior probability distribution. Detailed Implementation

[0071] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0072] See Figure 1 This invention provides an embodiment of an automatic pure-tone audiometry method based on Bayesian active learning and Gaussian process regression, the specific process of which includes:

[0073] S100: Take pure tones of different frequencies at the same hearing level and perform benchmark tests on the left and right ears respectively. Based on the benchmark tests and Bayesian active learning, establish a Gaussian process model.

[0074] S200, based on the dataset D obtained from the benchmark test in S100, performs Gaussian process regression based on the Gaussian process model established in S100 to obtain the benchmark audiograms of the left and right ears.

[0075] S300 analyzes the difference between the expected response entropy of the left and right ear baseline audiograms obtained from S200 and the average information entropy under a given distribution. The maximum value of this difference is taken to maximize the mutual information between the expected response and the Gaussian process regression. The test signal x corresponding to the maximum value is then used. * As the signal for the next pure tone test;

[0076] S400, determine the test signal x based on the characteristics of the human ear. * Feasibility, adjust test signal x * Intensity range and according to the masking rules, add masking to the non-test ear, using the adjusted test signal x * Play the corresponding test audio to the subjects and observe their feedback to obtain a new test set y. * ;

[0077] S500, x * ,y * Add dataset D, repeat step S200 using the updated dataset D, generate new audiograms using Gaussian process regression, and repeat steps S300-S400 to obtain x. * and y * The dataset D is updated until the regression result is greater than the set threshold, at which point the loop stops; the final left and right ear baseline audiograms are generated based on the regression results.

[0078] This embodiment combines the interactivity of active learning with the interactive process of pure tone testing, aiming to automate both pure tone testing and audiogram generation.

[0079] In a preferred embodiment of the present invention, see [link to previous document]. Figure 2 Before the hearing test, the subject needs to undergo a familiarization test to ensure they can respond to pure tone audio signals and are familiar with the audio playback procedure. Specifically, the test process is explained to the patient manually, including:

[0080] 1. Explanation of the audio testing equipment: Inform the test subject of the audio equipment used in the test and how to use it, such as headphones, speakers, etc.

[0081] 2. Provide feedback on the audio test, informing the subject to record whether they heard the test audio and to provide feedback on the audio they could hear.

[0082] 3. Audio feedback test: Play pure tone at 1000Hz in 10dB increments to test whether the subject can provide feedback on the audio, observe the subject's level of cooperation, and provide necessary explanations.

[0083] After familiarization testing to determine whether the subject can comply with the basic testing rules, the pure tone benchmark test is then conducted.

[0084] In a preferred embodiment of the present invention, step S100 involves benchmark testing and the establishment of a Gaussian process regression model. Pure tones of 40 dB HL at 250 Hz, 500 Hz, 1000 Hz, 2000 Hz, 4000 Hz, and 8000 Hz are used to perform air conduction pure tone testing on the subject. If bone conduction pure tone testing is required, pure tones of 20 dB HL at 250 Hz, 500 Hz, 1000 Hz, 2000 Hz, 3000 Hz, and 4000 Hz are used for testing. The left and right ears are tested separately, and the frequency of the test audio, hearing level, and corresponding audiometry results (whether the audio was heard or not) are recorded. These test results will be used as the basic input data to establish a Gaussian process model. The model establishment process is as follows:

[0085] According to Bayes' theorem:

[0086]

[0087] Where D = (X, y) represents the existing training set, and X represents the current input dataset. Figure 3 As shown, x = (v, i) contains two dimensions: signal frequency and hearing level, and y is the label of the hearing test result. ω represents the model weight component, and its specific value is unknown during the calculation due to the characteristics of the Gaussian process. p(ω) is the prior probability, p(y|X, ω) represents the conditional likelihood function of obtaining the label y from the input dataset X under the weight of ω, and p(y|X) is the marginal probability.

[0088] Since p(y|X)=∫p(y|X,ω)p(ω)dω is a constant, therefore:

[0089] p(ω|D)∝p(y|X,ω)p(ω)

[0090] Furthermore, since the likelihood function can be reduced to the following under the assumption that the noise follows a Gaussian distribution:

[0091]

[0092] It can be seen that p(y|X,ω) follows a Gaussian distribution:

[0093]

[0094] Furthermore, according to the properties of Gaussian processes, ω follows a normal distribution N(0,∑p). Substituting the likelihood function p(y|X,ω) and the prior probability p(ω) into the equation, we can see that:

[0095] p(ω|D)~N(μ ω ,∑ ω )

[0096] The results show that the posterior probability conforms to a multidimensional Gaussian distribution. Based on the posterior distribution of the weights, the result y corresponding to a given test point z can be obtained. * Predicted probability distribution:

[0097]

[0098] Since the relationship between the input data and the distribution is not necessarily linear, it is necessary to use the kernel function K(x,x')=φ(x). T The input is transformed nonlinearly by φ(x') = <φ(x),φ(x')>, for the objective function y = f(x) = φ(x). T Substituting the objective function into the input z in the posterior distribution, we can see that the mean and covariance of the Gaussian process after nonlinear transformation are both related to the kernel function. Therefore, the Gaussian process f(x) model under kernel function conditions can be established as follows:

[0099] f(x)~N(μ(x),k(x,x'))

[0100] Where μ(x) is the mean of this Gaussian process distribution, and k(x,x') is both the kernel function and the covariance matrix of this Gaussian process distribution.

[0101] In this embodiment, the test data at 250Hz, 500Hz, 1000Hz, 2000Hz, 4000Hz, and 8000Hz used for initialization are modeled using a Gaussian process to obtain a continuous probability model. This model has an infinitely large value range. These initial data points are then input into the model, causing changes in its features. The more data input, the closer the model is to the target features. Although the model's features change, it remains a probability model based on a Gaussian process, so its value range is still infinitely large. By limiting the value range to the frequency range of the audiogram, audiometric data across the entire frequency band from 125Hz to 16000Hz can be obtained, with a maximum frequency resolution accurate to 1Hz.

[0102] In a preferred embodiment of the present invention, step S200 is implemented by using benchmark test data as a training set D = (X, y), where X is the set of input signals x, x is the signal frequency and intensity components, and y is the signal label fed back by the subject, and Gaussian process regression is performed on D to obtain the signal probability distribution.

[0103] Based on the Gaussian process regression model established in step S100 using a kernel function, this embodiment designs the kernel function parameters for the input data as follows: For the audio hearing level dimension input in the test data, a linear kernel function is used as the input, and the kernel function is established as follows:

[0104] k L (i,i')=ii'

[0105] Because in the frequency response characteristics of the human ear to audio, the probability of hearing audio increases linearly with the increase of audio intensity.

[0106] For the audio frequency input in the test data, since the frequency response curve distribution of otoacoustics differs among different subject populations, and there is no significant correlation between otoacoustics in different hearing-impaired patients, this embodiment selects the radial basis kernel as the kernel function input for the frequency dimension of the test data:

[0107]

[0108] Where d = (v - v'), and l is the lengthscale controlling the degree of fluctuation. The radial basis function (RBF) kernel can effectively approximate the nonlinear characteristics of the audio frequency dimension and exhibits high smoothness. Finally, this embodiment constructs the covariance matrix kernel function as an additive combination of a linear kernel function and a radial basis function kernel, as follows:

[0109]

[0110] Where v is the frequency component of the input data and i is the intensity component of the input data.

[0111] Based on the established Gaussian process model f(x)~N(μ(x),k(x,x')), in the function space, μ(x) is the mean matrix of this Gaussian process, and k(x,x') is the variance matrix. For the input data, the training function is y=f(x)+ε. and the objective function f * =f(Z) can be related as follows:

[0112]

[0113] By utilizing the properties of the multivariate Gaussian conditional distribution, the given predicted value f can be calculated.* posterior probability distribution

[0114] On the other hand, this invention uses the normal cumulative distribution function φ(x)=∫N(x|0,1)dx as the classification fitting function, the Expectation Propagation approximation method as the likelihood function, and performs Gaussian process classification based on the Gaussian process model to generate a three-dimensional probability distribution map with frequency v as the abscissa, hearing level i as the ordinate, and audiometry feedback probability f(x) as the ordinate.

[0115] In the probability distribution graph, take the point on each horizontal axis whose feedback probability is closest to 0.707, extract the vertical coordinate of the point, draw a line from the point, and project the line onto a plane with frequency v as the horizontal axis and hearing level i as the vertical axis. The image corresponding to this plane is the audiogram of this test.

[0116] In a preferred embodiment of the present invention, S300 is implemented to analyze the difference between the expected response entropy and the average information entropy under a given distribution, and to take the maximum value in order to maximize the mutual information between the expected response and the Gaussian process regression estimate. Let the maximum value be I, and use the signal corresponding to the maximum value as the next pure tone test signal.

[0117] See Figure 4 The specific steps are as follows: The core objective of finding iterative test points is to minimize the number of possible hypotheses as quickly as possible, that is, to minimize the uncertainty of parameters using Shannon entropy. In this embodiment, this means finding the test data points that maximize the decrease in expected posterior entropy. The Shannon entropy calculation method can be used as follows:

[0118] I = argmax x H[y|x,D]-E θ～p(θ|D) [H[y|x,θ]]

[0119] The first part of the formula calculates the information entropy of the predicted value p(y|x,D) of the input point x given the existing training set D. The second part represents the overall average information entropy after inputting x given the distribution θ. The difference between the two represents the difference between the information entropy of a certain input sample and the average information entropy of the population. The larger this difference is, the more it indicates that the input sample contains more information relative to the average level.

[0120] In actual testing, the first part of the above equation can be transformed into the following formula based on the expected value of a composite function of random variables:

[0121]

[0122] Here, h represents the entropy under the Bernoulli distribution. Since the input signal corresponds to frequency and intensity components in the audiogram, and these are discrete points, it is calculated in a discrete form. In the formula, φ(f x f is the probability density function of the standard Gaussian distribution. x Let x be the probability value corresponding to the existing Gaussian process model. step represents f. x The step size is used to discretize the corresponding normal probability density function. This step size is determined by the number of sampling points (nbins) in the Gaussian distribution probability, which is equivalent to dividing the entire interval into nbins smaller intervals according to the step. The probability distribution matrix ∑f of the Gaussian process regression is then used. x The information entropy corresponding to the hearing test at this point can be obtained by calculating using the above formula.

[0123] The second part of this expression can be simplified to:

[0124]

[0125] This formula discretizes the expected entropy and calculates the average information entropy of the distribution p(y|x,θ) for each input x under the discrete standard Gaussian probability density.

[0126] From I = max(I1 - I2), we can calculate f. x The corresponding coordinates (v, i) are used to temporarily determine the audio frequency and hearing level test signal x* to be used in the next test.

[0127] In a preferred embodiment of the present invention, S400 is implemented, the feasibility of the x* signal is determined based on the characteristics of the human ear, the intensity range of x* is appropriately adjusted, and a masking is added to the non-test ear according to the masking rules. The corresponding test audio is played to the subject according to the adjusted x*, and the subject's feedback is observed to obtain a new test set y*.

[0128] When measuring air conduction hearing thresholds in clinical practice using over-ear headphones, due to the influence of bone conduction, when the test sound in the test ear is amplified to a certain extent, it will be transmitted to the cochlea of ​​the non-test ear through skull vibration. This results in the phenomenon where the raw stimulus given by the test ear is heard by the non-test ear, known as peripheral hearing ("shadow response"). To measure the hearing thresholds of the left and right ears separately, it is essential to avoid mistaking the shadow response for the hearing threshold of the test ear. Therefore, for a newly selected test tone, it is necessary to determine whether a shadow response will occur. If a shadow response occurs, the selection of the test audio frequency x* needs to be adjusted. The judgment of the shadow response and the masking noise rules are specified in ISO 8253-1 (GB / T 16403-1990) as follows:

[0129] If no basic bone conduction hearing threshold test has been performed, and the middle ear is unknown or only air conduction audiometry is performed, if the generated test audio frequency i1 is greater than or equal to 40 dB HL, masking noise needs to be added to the non-test ear. If a preliminary bone conduction test has been performed, and the test audio frequency i1 of the test ear exceeds the bone conduction hearing threshold of the non-test ear by 40 dB HL, the non-test ear should be masked. The masking noise is the frequency corresponding to the basic air conduction hearing threshold of the non-test ear plus 15 dB EM. EM refers to the effective masking level, and 15 dB EM is the noise level that can just mask a 15 dB HL pure tone. The masking noise should be narrowband noise, and according to GB / T 4854.4 (ISO 384.4-1994), its bandwidth should not be less than 1 / 3 octave and not more than 1 / 2 octave.

[0130] If the test audio i1 is more than 60 dB HL greater than the same frequency in the previous test, this test point should be ignored to protect hearing. Instead, select the point with slightly smaller entropy in I = max(I1-I2) in step 4). If the test conditions are still not met, continue to the next test point.

[0131] Play the selected test audio x to the subject * If masking is required, add masking noise to the non-test ear. Observe the subject's feedback and record whether the subject heard the test audio. * Record it.

[0132] In a preferred embodiment of the present invention, S500 is implemented, x * ,y * Add D, and use the updated dataset D to perform Gaussian process regression to generate a new audiogram. Repeat steps S300 and S400 until the regression result is greater than the set threshold, then stop the loop. Generate the final binaural audiogram based on the regression result.

[0133] The frequency of the new test audio, hearing level data x * and feedback results y * Add the test dataset D, and still based on the Gaussian process model in S100, perform Gaussian process regression on the updated dataset according to the steps in S200 to obtain a new posterior probability distribution map, which corresponds to the audiogram of the test ear. Then, according to the entropy calculation method in S300, select the test point with the maximum uncertainty, determine the test audio frequency and hearing level intensity, and then adjust the test audio and masking strategy according to the steps in S500 to conduct pure tone audiometry tests on the subjects.

[0134] Repeat the above steps to continuously acquire new test audio and new audiogram distributions until a threshold is reached. The threshold is set according to the following rules: 1. The maximum number of hearing test cycles cannot exceed 50. After 50 tests, the test automatically stops and the current audiogram is selected as the final audiogram for the corresponding test ear. 2. When the mean distance between the mean matrix of the current iteration and the mean matrix of the previous iteration is less than 0.1, it can be determined that there is no significant change between the two iterations, and the test stops.

[0135] After the test is stopped, the points with a posterior probability of 0.707 corresponding to each frequency are selected from the final obtained posterior probability distribution map and fitted and connected to generate the corresponding air conduction (bone conduction) pure tone audiometry audiogram of the test ear.

[0136] The above embodiments utilize Bayesian principles to infer from existing dataset results, thereby identifying the test signal with the greatest uncertainty. Each signal inferred from Bayesian principles is used as the hearing test signal, played to the subject, and the test results are added to the existing dataset after feedback is received. Therefore, each round of testing generates new data results and updates the original dataset, ultimately forming an iterative process between the dataset and the test signal—a form of active learning. It is evident that the above embodiments utilize an active learning mechanism, ensuring that each audio test signal is the most meaningful and information-rich signal for the hearing test subject, greatly improving the efficiency of pure-tone audiometry.

[0137] The above embodiments abstract audiogram data into two-dimensional input data of a Gaussian process. Utilizing the interpretability, smoothness, and excellent uncertainty estimation characteristics of the Gaussian process in relation to the physical properties of the data, the hearing data is regressed into a smooth curve, which fits the audiogram well. It is evident that the above embodiments employ a kernel function related to the physical properties of otoacoustics, making the regressed hearing curve highly similar to the human ear's hearing curve. Furthermore, because the data modeled by the Gaussian process has infinite scalability, the regression result data volume is extremely large, far exceeding the characteristic frequency results of traditional audiograms.

[0138] Based on the same inventive concept, in other embodiments of the present invention, a terminal is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, can be used to perform the method described thereon.

[0139] Based on the same inventive concept, in other embodiments of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, can be used to perform the method described thereon.

[0140] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention. The above preferred features can be used in any combination without conflict.

Claims

1. An automatic pure-tone audiometry method based on Bayesian active learning and Gaussian process regression, characterized in that, include: Pure tones of different frequencies at the same hearing level are taken and benchmark tests are performed on the left and right ears respectively. Based on the benchmark tests and Bayesian active learning, a Gaussian process model is established. The dataset obtained from the benchmark test Based on this, Gaussian process regression is performed using the Gaussian process model to obtain the baseline audiograms for the left and right ears. Analyze the difference between the expected response entropy of the left and right ear benchmark audiograms and the average information entropy under a given distribution, and take the maximum value of the difference to maximize the mutual information between the expected response and the Gaussian process regression. The test signal corresponding to the maximum value is then used. As the signal for the next pure tone test; The test signal is determined based on the characteristics of the human ear. Feasibility, adjust the test signal Intensity range and according to the masking rules, add masking to the non-test ear, using the adjusted test signal. Play the corresponding test audio to the subjects, observe their feedback, and obtain new test sets. ; Will , Add to dataset Using the updated dataset Perform the Gaussian process regression to generate a new audiogram, and repeat the acquisition process. and and updating the dataset D The loop continues until the regression result exceeds the set threshold, at which point it stops; the final left and right ear baseline audiograms are generated based on the regression results. Specifically, the difference between the expected response entropy of the left and right ear benchmark audiograms and the average information entropy under a given distribution is analyzed. The maximum value of this difference is taken to maximize the mutual information between the expected response and the Gaussian process regression. The signal corresponding to the maximum value is then... As the signal for the next pure tone test, it includes: Using the Shannon entropy calculation method, we can find the test data point that maximizes the decrease in expected posterior entropy. In the formula, the first part calculates the value of the existing training set. In the case of input point Predicted value The information entropy, the latter part represents the information entropy of a given distribution. Input in the case The difference between the two values ​​represents the difference between the information entropy of a given input sample and the average information entropy of the entire population. Transform the first part of the formula based on the mathematical expectation formula of the composite function of random variables. , Denotes the entropy under the Bernoulli distribution. The probability density function is the standard Gaussian distribution. In the existing Gaussian process model The corresponding probability value, express The step size for discretizing the corresponding normal probability density function is determined by the number of sampling points in the Gaussian distribution. The decision will be made to proceed according to the following: Divided The probability distribution matrix of the Gaussian process regression is given in a small interval. according to I 1. Calculate the information entropy corresponding to conducting a hearing test at this point; Transform the latter half into This formula discretizes the expected entropy and calculates the distribution. For each input Average information entropy under discrete standard Gaussian probability density; Depend on Calculations yielded Corresponding coordinates This coordinate will be used to temporarily determine the audio frequency and hearing level test signal to be used in the next test. .

2. The automatic pure-tone audiometry method based on Bayesian active learning and Gaussian process regression according to claim 1, characterized in that, Pure tones of different frequencies at the same hearing level were selected, and benchmark tests were performed on the left and right ears respectively. Based on the benchmark tests and Bayesian active learning, a Gaussian process model was established, including: Take 40dB HL pure tones at 250Hz, 500Hz, 1000Hz, 2000Hz, 4000Hz, and 8000Hz and perform air conduction pure tone tests on the subjects; or take 20dB HL pure tones at 250Hz, 500Hz, 1000Hz, 2000Hz, 3000Hz, and 4000Hz and perform bone conduction pure tone tests on the subjects. Test the left and right ears separately and record the test results, including the frequency of the audio, the hearing level, and the corresponding audiometry results of whether the audio was heard; Using the test results as the basic input data, a Gaussian process model is established based on Bayesian active learning.

3. The automatic pure-tone audiometry method based on Bayesian active learning and Gaussian process regression according to claim 2, characterized in that, The step of using the test results as basic input data to establish a Gaussian process model based on Bayesian active learning includes: According to Bayes' theorem ,in Given the existing training set, The current input dataset contains two components: signal frequency and hearing level. , Label the audiometry results; Represents the model weight components. For prior probability, Indicates in Input dataset with weights get The conditional likelihood function of the label. Marginal probability; because As a constant, we get ; Assuming the noise follows a Gaussian distribution, the likelihood function is transformed. ,get Conforms to a Gaussian distribution ; Obtained from the properties of Gaussian processes Follows a normal distribution The likelihood function and prior probability Substitute to get ; The posterior probability results conform to a multidimensional Gaussian distribution. Based on the posterior distribution of the weights, the given test point is obtained. Corresponding results Predicted probability distribution Using kernel functions Perform nonlinear transformation on the input data; For the objective function Substitute the objective function into the input of the posterior distribution. The posterior mean and covariance of the Gaussian process after nonlinear transformation are found to be related to the kernel function. Therefore, a Gaussian process under kernel function conditions is established. The model is ,in The mean of this Gaussian process distribution, It is both the kernel function and the covariance matrix of the distribution of this Gaussian process.

4. The automatic pure-tone audiometry method based on Bayesian active learning and Gaussian process regression according to claim 1, characterized in that, The dataset obtained from the benchmark test Based on this, Gaussian process regression is performed using the aforementioned Gaussian process model to obtain the baseline audiograms for the left and right ears, including: The kernel function parameters for the input data of the Gaussian process regression model are designed as follows: For the audio listening level dimension input in the test data, a linear kernel function is established as input. For the audio frequency input in the test data, a radial base was selected. As input to the kernel function along the frequency dimension of the test data, where , Lengthscal is used to control the degree of its fluctuation; Construct a covariance matrix kernel function that is an additive combination of linear kernel functions and radial basis kernel functions. ,in For the frequency components of the input data, For the input data intensity components; Based on the established Gaussian process model Knowing in the function space The mean matrix of this Gaussian process. It is the variance matrix; For the input data training function , and objective function Establishing Relationships ; Using the properties of the multivariate Gaussian conditional distribution, a given predicted value can be calculated. posterior probability distribution ; Using the normal cumulative distribution function As the classification fitting function, the ExpectationPropagation approximation method is used as the likelihood function. Based on the Gaussian process model, Gaussian process classification is performed, generating frequency-based classification functions. The horizontal axis represents hearing level. The vertical axis represents the probability of audiometric feedback. A three-dimensional probability distribution plot with the vertical axis as the coordinate; In the three-dimensional probability distribution graph, the point on each horizontal axis whose feedback probability is closest to the set value is selected. The vertical coordinate of the point is extracted, and the point is plotted as a line. The line is then projected onto a graph with a frequency. The horizontal axis represents hearing level. On the plane with ordinate as the vertical axis, the image corresponding to this plane is the corresponding audiogram for this test.

5. The automatic pure-tone audiometry method based on Bayesian active learning and Gaussian process regression according to claim 1, characterized in that, The signal is judged based on the characteristics of the human ear. Feasibility, adjust the signal Intensity range and according to the masking rules, add masking to non-test ears, using the adjusted... Play the corresponding test audio to the subjects, observe their feedback, and obtain new test sets. ,include: If a basic bone conduction hearing threshold test is not performed, and the middle ear is not understood, and only air conduction audio is performed, the generated test audio... When the dB HL is reached, masking noise needs to be added to the non-test ear; If a preliminary bone conduction test has already been performed, the test audio for the test ear... If the bone conduction hearing threshold of the non-test ear exceeds 40 dB HL, the non-test ear should be masked. The noise level is the frequency corresponding to the baseline air conduction hearing threshold of the non-tested ear plus 15 dB EM. If testing audio If the value is more than 60 dB HL greater than the same frequency in the previous test, ignore this test point and select... If the test conditions are still not met, proceed to the next test point with slightly lower entropy. Play the selected test audio to the subject. If masking is required, masking noise should be added to the non-test ear. Observe the feedback from the subjects and record whether the subjects heard the test audio. Record it.

6. The automatic pure-tone audiometry method based on Bayesian active learning and Gaussian process regression according to claim 1, characterized in that, The , join in Using the updated dataset Gaussian process regression is used to generate new audiograms, and the data is repeatedly acquired. and and updating the dataset D The loop continues until the regression result exceeds the set threshold, at which point it stops. The final binaural hearing map is generated based on the regression results, including: The frequency and hearing level data of the new test audio. and feedback results Add to test dataset Based on the Gaussian process model, Gaussian process regression is performed with the updated dataset as input to obtain a new posterior probability distribution map, which is the audiogram of the corresponding test ear. Then, the test point with the maximum uncertainty is selected according to the entropy calculation method. The test point is selected to determine the test audio frequency and hearing level intensity. The test audio and masking strategy are adjusted, and pure tone hearing tests are performed on the subjects. The above steps are repeated to continuously obtain new test audio and new audiogram distributions until the threshold is reached. After the test is stopped, the points corresponding to the posterior probability of each frequency in the final obtained posterior probability distribution map are selected and fitted and connected to generate the pure tone audiometry chart of the air conduction or bone conduction of the corresponding test ear.

7. The automatic pure-tone audiometry method based on Bayesian active learning and Gaussian process regression according to claim 6, characterized in that, The setting of the threshold includes: The maximum number of hearing test cycles cannot exceed the set number of cycles. That is, after the set number of hearing tests, the test will automatically stop and the current audiogram will be selected as the final audiogram for the corresponding ear being tested. If the mean Euclidean distance between the mean matrix of the current iteration and the mean matrix of the previous iteration is less than a set value, it can be determined that there is no significant change between the two iterations, and the test can be stopped.

8. A terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it can be used to perform the method of any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program can be used to perform the method of any one of claims 1-7.

Images