Method and system for classifying susceptibility to noise-induced hearing loss based on multi-frequency hearing threshold

Abstract

The present application relates to the technical field of noise-induced hearing loss susceptibility grading, and particularly relates to a noise-induced hearing loss susceptibility grading method and system based on multi-band hearing threshold, which comprises presetting measurement frequency points, generating a measurement frequency point set in order from high frequency to low frequency, and completing M independent measurements for each frequency point; recording a subject response time stamp to generate a confidence value label; collecting an otoacoustic emission signal in a subject ear canal in real time, extracting a time domain signal with the highest confidence of a single frequency point, calculating a spectral feature, and constructing an otoacoustic emission joint parameter. The present application can comprehensively capture the time sequence diffusion law of high frequency to low frequency hair cell damage, significantly improve the recognition accuracy of potential hearing damage, simulate the propagation mechanism of high frequency damage to medium frequency by constructing a physically constrained spectral migration operator, and realize individualized and dynamic damage risk prediction by combining the otoacoustic emission time domain-frequency domain joint parameter and physiological indicators.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of noise-induced hearing loss susceptibility grading, in particular to a noise-induced hearing loss susceptibility grading method and system based on multi-frequency hearing threshold. BACKGROUND

[0002] Noise-induced hearing loss refers to the hearing loss or permanent hearing impairment caused by the damage to cochlear hair cells or auditory nerves due to long-term or short-term exposure to high-intensity noise environment. The sensitivity of different individuals to hearing impairment under the same noise exposure conditions varies significantly, which is mainly due to the differences in cochlear structure, physiological metabolism level, genetic factors and exposure history.

[0003] In order to realize early warning and individualized protection of noise-induced hearing loss, the concept of "noise-induced hearing loss susceptibility grading" has been proposed in recent years. This grading method comprehensively analyzes the multi-frequency hearing threshold, otoacoustic emission characteristics, auditory brainstem response and other hearing-related parameters of the tested person to evaluate the vulnerability of the individual's auditory system under continuous noise action.

[0004] However, the existing noise-induced hearing loss susceptibility grading technology generally only uses the change of hearing threshold in a single frequency band to judge the hearing impairment risk of individuals, ignoring the time sequence diffusion rule of noise-induced hearing loss in different frequency bands. Actual research shows that high-frequency noise first causes the function of high-frequency hair cells at the base of the cochlea to decrease, and then the damage gradually spreads to the medium and low frequency regions. When the tested person is exposed to a steady-state noise environment for a long time, the medium frequency hair cells may have hidden signs of damage such as reduced activity and distorted otoacoustic emission waveform, but the function of high-frequency hair cells is still within the normal range. The existing technology is likely to misjudge the tested person as a low-risk state because it does not monitor the medium frequency band signal or the low frequency band signal. SUMMARY

[0005] To solve the above problems, the present application provides a noise-induced hearing loss susceptibility grading method and system based on multi-frequency hearing threshold.

[0006] The present application adopts the following technical solution: a noise-induced hearing loss susceptibility grading method based on multi-frequency hearing threshold, comprising: presetting measurement frequency points, generating a measurement frequency point set in order from high frequency to low frequency, and completing M independent measurements for each frequency point;

[0007] Recording the subjective response timestamp of the tested person, judging the cooperation degree of the tested person through the reaction time distribution and generating a confidence value label;

[0008] Real-time acquisition of the otoacoustic emission signal in the ear canal of the tested person to form a time-domain signal of sound reflection changing with time, extraction of the time-domain signal with the highest confidence value for a single frequency point, calculation of the spectral characteristics, and construction of the otoacoustic emission joint parameters;

[0009] Synchronously record the auditory brainstem response signal and extract the hidden impairment feature of the auditory brainstem response;

[0010] Collect physiological parameters of the subject, the physiological parameters including heart rate, ear canal temperature and electrode impedance;

[0011] Based on the physiological parameters and the otoacoustic emission time-frequency domain joint parameters, a time sequence coupling framework between frequency bands is established, and a physically constrained spectral transfer operator is used to formalize the impairment propagation mechanism from high frequency to medium frequency, so that the operator can receive the physiological parameters and the otoacoustic emission time-frequency domain joint parameters and output a frequency band transfer probability field;

[0012] The obtained frequency band transfer probability field and the hidden impairment feature are input into a pre-constructed future risk evolution prediction model, and a risk evolution curve of each frequency band in the next one to three months is output;

[0013] Curve features of the frequency band risk evolution curve are extracted, and a single risk score of the frequency band is obtained by weighting and synthesizing with preset weights, and a noise deafness susceptibility classification level is generated.

[0014] As a further description of the above technical solutions, the method for determining the cooperation degree of the subject by the reaction time distribution and generating a confidence value label comprises:

[0015] The subject presses the response button when stimulated according to the prompt;

[0016] The starting time of each stimulation and the key pressing time of the subject are recorded, and the stimulation frequency point, frequency point number Fi and measurement number Fij are recorded; j∈{1,2,…,M};

[0017] The average reaction time, standard deviation and frequency point sequence consistency score of each frequency point are calculated, and the average reaction time, standard deviation and frequency point sequence consistency score are used for weighted summation to generate a comprehensive cooperation degree score;

[0018] The comprehensive cooperation degree score is assigned a confidence value as the measurement result each time.

[0019] As a further description of the above technical solutions, the method for obtaining the frequency point sequence consistency score comprises:

[0020] The measurement frequency point set is arranged from high frequency to low frequency to form a reference sequence, and each frequency point is assigned a position number in the reference sequence, and the position number of each frequency point represents its position in the reference sequence;

[0021] The subjective responses of the subject to each frequency point in the measurement process are sorted according to the response time from early to late, and each frequency point obtains a position number in the actual response sequence, which is used to represent the sequence of the response of the frequency point;

[0022] For each actual response frequency point, the position number in the reference sequence is found, then the difference value of the actual response sequence and the reference sequence is compared, the absolute value is taken, the sequence deviation of all frequency points is statistically integrated as a whole, the average value is obtained, and the difference value is divided by the average value to obtain the sequence consistency score.

[0023] As a further description of the above technical solution: the spectral features include: peak frequency average, harmonic intensity coefficient and average drift rate.

[0024] The extraction method of the spectral features includes:

[0025] The original time domain waveform is band-pass filtered, and the band-pass range covers the target analysis frequency band, and a linear phase or zero phase filter is used to avoid phase distortion.

[0026] The fixed length window containing the complete reflection packet is intercepted with the stimulation trigger point as the time zero point, the preset time window length, the Hanning window or Kaiser window is used for each window, and the overlap rate is set.

[0027] Short-time Fourier transform is performed on each window length to obtain the corresponding time-frequency matrix.

[0028] In the average time-frequency matrix, search for the local energy peak value for each time frame, identify the main frequency peak, connect the main frequency peak in the continuous time frame to form a peak frequency trajectory, and obtain the peak frequency average.

[0029] In each time window, the ratio of the fundamental frequency energy to the harmonic band energy is calculated, and the average value is taken as the harmonic intensity coefficient.

[0030] Based on the obtained peak frequency trajectory, the total drift amount and drift time of the trajectory are calculated to obtain the average drift rate.

[0031] As a further description of the above technical solution: the method for synchronously recording the auditory brainstem response signal includes:

[0032] When measuring, the recording electrode is arranged at the top of the head of the measured person, the reference electrode is arranged at the earlobe, and the ground electrode is arranged at the forehead or the sternum.

[0033] Digital sampling is performed using a high sampling frequency greater than or equal to 20 kHz;

[0034] One recording frame is saved for each stimulation, recorded as a single auditory brainstem response, the multiple responses are added point by point and divided by the number of responses to obtain the average auditory brainstem response waveform.

[0035] As a further description of the above technical solution: the method for extracting the hidden damage features of the auditory brainstem response includes:

[0036] The peak search window is set on the average auditory brainstem response waveform, wherein the wave I window is 1-3 ms, the wave III window is 3-5 ms, and the wave V window is 5-8 ms;

[0037] In each search window, the highest positive peak is found, and the time axis value of the positive peak position is obtained, which is recorded as the corresponding latency in the search window. The latencies of wave I, wave III and wave V are obtained in turn.

[0038] The interpeak latency difference values of wave I-wave III, wave III-wave V and wave I-wave V are obtained.

[0039] As a further description of the above technical solution: the training method of the future risk evolution prediction model comprises:

[0040] Q sets of time series training data are collected in advance, Q is a positive integer greater than 0, the time series training data includes a frequency band migration probability field and a hidden damage feature, and a frequency band risk evolution curve corresponding to the frequency band migration probability field and the hidden damage feature;

[0041] The long short-term memory network is selected as the future risk evolution prediction model, the Adam optimizer is selected, the classification cross-entropy is selected as the loss function, and the constructed time series training data is divided into a training set, a validation set and a test set according to a designed proportion;

[0042] The training set data is input into the model in batches, the prediction probability distribution is obtained through forward propagation, the network weights are updated through back propagation after the loss is calculated, and the weighted F1 score is calculated using the validation set after each round in the training process. If the performance of the validation set does not improve within the preset number of rounds, the training is automatically stopped, and the optimal model at this time is saved to prevent overfitting and save training time;

[0043] On the test set, the weighted F1 score of the model for predicting the risk level in the future 1 month and 3 months needs to reach a preset threshold, and after the model training is completed, the trained model is converted into an ONNX lightweight format and deployed to a cloud or edge computing node.

[0044] As a further description of the above technical solution: the curve feature includes a peak probability, a growth rate and a persistence length, and the curve features are weighted and summed according to a preset weight to obtain a single risk score FXxs of the frequency band.

[0045] As a further description of the above technical solution: the method for generating the noise-induced hearing loss susceptibility classification level comprises:

[0046] The single risk scores of each frequency band are summed and divided by the number of frequency bands to obtain an average risk score FXxs;

[0047] The preset risk score gradient threshold is FX1, FX2, FX3, FX4, wherein FX1<FX2<FX3<FX4;

[0048] When FXxs≤FX1, a low risk level is generated;

[0049] When FX1<FXxs≤FX2, a low-medium risk level is generated;

[0050] When FX2<FXxs≤FX3, a medium risk level is generated;

[0051] When FX3<FXxs≤FX4, a high-medium risk level is generated;

[0052] When FXxs>FX4, a high risk level is generated.

[0053] A multi-band hearing threshold noise-induced hearing loss susceptibility grading system is used to implement the multi-band hearing threshold noise-induced hearing loss susceptibility grading method, and the system comprises:

[0054] A frequency point generation module, which generates a set of measurement frequency points in order from high frequency to low frequency, and each frequency point completes M independent measurements;

[0055] A response evaluation module, which records the measured subjective response timestamp, determines the measured cooperation degree through the reaction time distribution, and generates a confidence value label;

[0056] An otoacoustic emission acquisition module, which acquires the otoacoustic emission signals in the ear canal of the measured person in real time, forms a time-domain signal of acoustic reflection changing with time, extracts the time-domain signal with the highest confidence value of a single frequency point, calculates the spectral feature, and constructs the otoacoustic emission joint parameter;

[0057] A brainstem feature extraction module, which synchronously records the auditory brainstem response signal and extracts the hidden damage feature of the auditory brainstem response;

[0058] A physiological parameter acquisition module, which acquires the physiological parameters of the measured person, and the physiological parameters include heart rate, ear canal temperature, and electrode impedance;

[0059] A coupling modeling module, which establishes a frequency band time sequence coupling framework based on the physiological parameters and the otoacoustic emission time-domain-frequency-domain joint parameters, uses a physically constrained spectral transfer operator to formalize the damage propagation mechanism from high frequency to medium frequency, so that the operator can receive the physiological parameters and the otoacoustic emission time-domain-frequency-domain joint parameters and output a frequency band transfer probability field;

[0060] A risk prediction module, which inputs the obtained frequency band transfer probability field and hidden damage feature into a pre-constructed future risk evolution prediction model, and outputs the risk evolution curve of each frequency band in the next one to three months;

[0061] The grade dividing module extracts curve features of the frequency band risk evolution curve, performs weighted synthesis by preset weights, obtains a single risk score of the frequency band, and generates a noise-induced deafness susceptibility grading level.

[0062] The present application has the following advantages:

[0063] The multi-frequency band hearing threshold noise-induced deafness susceptibility grading method and system provided by the present application can realize early perception and quantitative analysis of noise-induced deafness hidden damage by presetting multi-frequency band measurement points, synchronously collecting otoacoustic emission signals, auditory brainstem responses and physiological parameters, and establishing a time sequence coupling framework between frequency bands, which can capture the time sequence diffusion law of high-frequency to low-frequency hair cell damage, significantly improve the recognition accuracy of potential auditory damage, simulate the propagation mechanism of high-frequency damage to medium frequency by constructing a physically constrained spectral migration operator, and realize individualized and dynamic damage risk prediction by combining otoacoustic emission time domain-frequency domain joint parameters and physiological indexes.

[0064] Further, by inputting the frequency band migration probability field and the hidden damage features into a future risk evolution prediction model, the risk evolution curve of each frequency band in the next one to three months can be output, and a single risk score can be generated by weighted synthesis to realize noise-induced deafness susceptibility grading. This method not only can detect hidden damage signs of medium and low frequency hair cells in advance to avoid misjudgment of traditional technology, but also can provide scientific basis for individualized noise protection and intervention measures, thereby improving the timeliness and accuracy of noise-induced deafness protection. BRIEF DESCRIPTION OF DRAWINGS

[0065] The present application will be further explained in conjunction with the drawings and examples:

[0066] Figure 1 The flowchart of the multi-frequency band hearing threshold noise-induced deafness susceptibility grading method provided for the present application example 1 is shown in the figure;

[0067] Figure 2 The method flowchart for generating confidence labels provided for the present application example 1 is shown in the figure;

[0068] Figure 3 The method flowchart for synchronously recording auditory brainstem response signals provided for the present application example 1 is shown in the figure;

[0069] Figure 4 The flowchart of the multi-frequency band hearing threshold noise-induced deafness susceptibility grading system provided for the present application example 2 is shown in the figure. DETAILED DESCRIPTION

[0070] In order to make the technical means, creative features, purposes and effects of the present application easy to understand, the present application will be further described below in combination with specific drawings. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0071] Embodiment 1

[0072] Please refer to Figures 1-3 A technical solution provided by the embodiment of the present application is: a multi-frequency band hearing threshold noise-induced hearing loss susceptibility grading method;

[0073] It should be noted that noise-induced hearing loss susceptibility grading refers to establishing a quantitative or grading evaluation system for judging the susceptibility of a person to noise-induced hearing loss in different populations according to the susceptibility difference of individuals to hearing impairment caused by noise exposure;

[0074] The method comprises:

[0075] A preset measurement frequency point is generated in order from high frequency to low frequency to form a measurement frequency point set, M (M≥3) independent measurements are completed for each frequency point, a subject response time stamp is recorded, and the cooperation degree of the subject is judged by the reaction time distribution and a confidence value label is generated;

[0076] Optionally, the measurement frequency point set is: 16 kHz, 12.5 kHz, 10 kHz, 8 kHz, 6 kHz, 4 kHz, 3 kHz, 2 kHz, 1 kHz, 750 Hz, 500 Hz, 250 Hz and 125 Hz.

[0077] It should be noted that the measurement frequency point set is used to apply a sound stimulus signal of a set frequency and duration to the ear canal of the subject, and the stimulus signal can be a short pulse signal or a double-frequency combined signal, which is used to induce the outer hair cells of the cochlea to produce a reflective sound wave response.

[0078] Optionally, the stimulus parameters include a pure tone duration of ≥500 ms, a gradual rise / fall window of 10-20 ms, a starting sound level gradually adjusted from 30 dB SPL, and a step size of 5 dB until the threshold.

[0079] The method of judging the cooperation degree of the subject by the reaction time distribution and generating a confidence value label comprises:

[0080] The subject presses the response button when prompted to receive the stimulus;

[0081] The starting time of each stimulus and the key time of the subject are recorded, and the stimulus frequency point, frequency point number F i and measurement number F ij are recorded simultaneously; j∈{1,2,…,M}.

[0082] The average reaction time, the standard deviation and the frequency point sequence consistency score of each frequency point are calculated, the average reaction time, the standard deviation and the frequency point sequence consistency score are weighted and summed to generate a comprehensive cooperation degree score;

[0083] The comprehensive cooperation degree score is taken as a confidence value of each measurement result. It should be noted that the lower the confidence value, the higher the confidence.

[0084] The frequency point sequence consistency score is obtained by the following method:

[0085] The measurement frequency point set is arranged from high frequency to low frequency to form a reference sequence, and each frequency point is assigned a position number in the reference sequence, and the position number of each frequency point represents its position in the reference sequence;

[0086] The subjective responses of the subject to each frequency point during the measurement process are sorted according to the response time from early to late, and each frequency point obtains a position number in the actual response sequence, which is used to represent the order of the response of the frequency point;

[0087] For each actual response frequency point, the position number in the reference sequence is found, and then the difference value between the actual response sequence and the reference sequence is compared, the absolute value is taken, the sequence deviation of all frequency points is statistically analyzed, the average value is obtained, and the difference value is divided by the average value to obtain the sequence consistency score;

[0088] It should be noted that a positive difference indicates that the frequency point responds later than the reference sequence, a negative difference indicates that the frequency point responds earlier than the reference sequence, and a difference of zero indicates complete compliance with the reference sequence.

[0089] The sequence consistency score is used to judge the cooperation degree of the subject during the measurement process, and provides a basis for subsequent confidence label generation.

[0090] During the measurement, the otoacoustic emission signal in the ear canal of the subject is collected in real time through a miniature microphone, the otoacoustic emission signal is amplified and anti-aliasing filtered, and then enters a high-precision sampling channel for collection, and the original waveform information is recorded in a time-domain continuous manner to form a time-domain signal of sound reflection varying with time, the time-domain signal with the highest confidence of a single frequency point is extracted, the spectral feature is calculated, and the otoacoustic emission joint parameter is constructed;

[0091] It should be noted that the otoacoustic emission signal is a weak sound wave reflection signal generated by the inner ear (outer hair cells of the cochlea) after being stimulated by sound.

[0092] In this embodiment, by extracting the time-domain signal with the highest confidence of a single frequency point, calculating the spectral feature, and constructing the otoacoustic emission joint parameter, the time-domain signal with low subject cooperation degree is proposed, and the accuracy of data acquisition is improved.

[0093] The spectral features include: peak frequency average, harmonic intensity coefficient, and average drift rate.

[0094] The extraction method of the spectral features includes:

[0095] The original time-domain waveform is band-pass filtered, and the band-pass range covers the target analysis frequency band, for example, 125 Hz to 20 kHz or as set by the experiment, and the filter uses linear phase or zero phase to avoid phase distortion;

[0096] Taking the stimulation trigger point as the time zero point, a fixed length window containing the complete reflection packet is intercepted, the preset time window length is used, a Hanning window or a Kaiser window is used for each window, and an overlap rate is set; optionally, the overlap degree is 50%-75%;

[0097] A short-time Fourier transform is performed on each window length to obtain a corresponding time-frequency matrix;

[0098] In the average time-frequency matrix, a local energy peak value is searched for each time frame, a main frequency peak is identified, a peak frequency trajectory is formed by connecting the main frequency peaks in the continuous time frames, and a peak frequency average is obtained;

[0099] In each time window, the ratio of the fundamental frequency energy to the harmonic band energy is calculated, and the average value is obtained as the harmonic intensity coefficient;

[0100] Based on the obtained peak frequency trajectory, the total drift amount (end frequency minus start frequency) and drift time of the trajectory are calculated, and the average drift rate is obtained.

[0101] The otoacoustic emission combined parameters include time-domain signals, peak frequency average, harmonic intensity coefficient, and average drift rate.

[0102] In this embodiment, multi-dimensional information such as time delay, wave packet shape change, and spectral features of the otoacoustic emission can be obtained simultaneously in a single test, high-resolution description of the dynamic behavior of the cochlear acoustic reflection is realized, and through joint recording in the time domain and the frequency domain, early nonlinear distortion and waveform delay change in the high-frequency region can be captured, thereby providing a reliable data basis for early identification and classification of noise-induced hearing loss susceptibility.

[0103] The auditory brainstem response signal is recorded synchronously, and the hidden damage features of the auditory brainstem response are extracted;

[0104] The method for recording the auditory brainstem response signal synchronously includes:

[0105] When measuring, a recording electrode is arranged at the top of the head (Cz or Fz) of the measured person, a reference electrode is arranged at the earlobe, and a ground electrode is arranged at the forehead or the sternum;

[0106] Digital sampling is performed using a high sampling frequency greater than or equal to 20 kHz (preferably 40 kHz);

[0107] Each stimulus is recorded as a single auditory brainstem response. Multiple responses are added together and divided by the number of responses to obtain the average auditory brainstem response waveform.

[0108] Methods for extracting latent impairment features of auditory brainstem responses include:

[0109] Peak search windows were set on the average auditory brainstem response waveform, with wave I window being 1–3 ms, wave III window being 3–5 ms, and wave V window being 5–8 ms.

[0110] Within each search window, find the highest amplitude positive peak, obtain the time axis value of the location where the positive peak appears, and record the time axis value as the corresponding latency within the search window. Obtain the latencies of wave I, wave III and wave V in sequence.

[0111] And obtain the interwave latency differences between wave I-wave III, wave III-wave V, and wave I-wave V.

[0112] It should be noted that wave I mainly originates from the peripheral auditory nerve conduction velocity, wave III mainly originates from the cochlear nucleus and superior olivary complex, and wave V mainly originates from the lateral lemniscus to the inferior colliculus. By obtaining the latencies of waves I, III, and V, and by extracting the latencies of waves I, III, and V respectively, the nerve conduction velocities of the peripheral auditory nerve, lower brainstem, and upper brainstem can be quantitatively characterized. Further calculation of the interwave latency differences between waves I and III, III and V, and I and V can be used to identify abnormalities such as auditory neuropathy, brainstem conduction block, and demyelinating lesions, thereby achieving a precise assessment of the functional status of the auditory pathway.

[0113] The physiological parameters of the subjects were collected, including heart rate, ear canal temperature, and electrode impedance.

[0114] Based on physiological parameters and joint time-frequency parameters of otoacoustic emission, a temporal coupling framework between frequency bands is established. A physically constrained spectral transfer operator is used to formalize the damage propagation mechanism from high frequency to mid frequency. Individualized adjustment rules for operator parameters are defined and a noise robust estimator is introduced to suppress environmental interference, enabling the operator to receive physiological parameters and joint time-frequency parameters of otoacoustic emission and output a frequency band transfer probability field.

[0115] Specifically, the inter-band temporal coupling framework uses a physically constrained spectral migration operator to simulate the damage propagation mechanism from high frequency to mid frequency. The operator satisfies directional constraints, local attenuation constraints, and time reversibility terms. The operator parameters are calibrated through population priors and individualized Bayesian updates. The observation noise is estimated by combining confidence and environmental noise, and a robust temporal filter is used to achieve recursive estimation.

[0116] Divide the continuous frequency band into segments B, using time indexing. The evolution of the system state is represented, and the inter-band damage propagation process is modeled as a discrete-time dynamical system:

[0117]

[0118] where, is the risk coefficient at time is the spectral shift operator, is the exogenous input at t, including recent exposure intensity, individual protection behavior, transient event annotation, is the process noise.

[0119] The spectral shift operator is represented by a hybrid representation of linearly coupled kernels plus band-activated nonlinear terms:

[0120] ;

[0121] where, is the shift kernel matrix, characterizes the direct coupling strength between bands, and B represents the number of bands, is the influence matrix of exogenous input, , characterizes the number of exogenous inputs, is the bias term, is the component-wise saturated nonlinear term.

[0122] It should be noted that the spectral shift operator with physical constraints maps the multi-band time series state to the future state; combined with the hierarchical Bayesian individualized parameter updating strategy, the band shift probability field can be robustly estimated in a noisy environment.

[0123] It should be noted that the band shift probability field refers to the spatial (band) -time (prediction time) field representation of the probability distribution of the occurrence of a significant risk increase event for each band in the system within a given prediction time window, and the output includes uncertainty and main driving sources for decision module calling.

[0124] wherein the representation of the band shift probability field is:

[0125] The band set is ; from high frequency to low frequency;

[0126] The prediction time step set is ; T represents the Tth step in the time step plan;

[0127] The band shift probability field is represented as: ; represents the probability of "risk" in the future τth step of band b.

[0128] ​The acquired frequency band migration probability field and latent damage features are input into a pre-constructed future risk evolution prediction model, which outputs the risk evolution curves of each frequency band in the next one to three months.

[0129] The training method for the future risk evolution prediction model includes:

[0130] Q sets of time-series training data were collected in advance, where Q is a positive integer greater than 0. The time-series training data included the frequency band migration probability field and latent damage features, as well as the frequency band risk evolution curves corresponding to the frequency band migration probability field and latent damage features.

[0131] Long Short-Term Memory Network (LSTM) is selected as the future risk evolution prediction model. The Adam optimizer is selected, the initial learning rate is set to 0.001, and a learning rate decay strategy can be set. The classification cross-entropy is used as the loss function. The constructed time-series training data is divided into training set, validation set and test set according to the design ratio (optionally, 7:2:1).

[0132] The training set data is input into the model in batches. The predicted probability distribution is obtained through forward propagation. After calculating the loss, the network weights are updated through backpropagation. During the training process, the weighted F1 score is calculated using the validation set after each round. If the validation set performance does not improve within a preset number of rounds (e.g., 20 rounds), the training is automatically stopped, and the optimal model at this time is saved to prevent overfitting and save training time.

[0133] On the test set, the model's weighted F1 score for predicting risk levels for the next 1 to 3 months must reach preset thresholds (e.g., ≥0.85 and ≥0.80 respectively). After the model training is completed, the trained model is converted into the ONNX lightweight format and deployed to the cloud or edge computing nodes.

[0134] The curve features of the frequency band risk evolution curve are extracted, and weighted synthesis is performed with preset weights to obtain a single risk score for the frequency band, thereby generating a noise-induced hearing loss susceptibility classification level.

[0135] The curve features include peak probability, growth rate, and duration. The curve features are weighted and summed according to preset weights to obtain a single risk score FXxs for the frequency band.

[0136] Methods for generating susceptibility ratings for noise-induced hearing loss include:

[0137] The average risk score FXxs is obtained by summing the individual risk scores FXxs of each frequency band and dividing by the number of frequency bands.

[0138] The preset risk score gradient thresholds are FX1, FX2, FX3, and FX4, where FX1 < FX2 < FX3 < FX4;

[0139] When FXxs≤FX1, a low risk level is generated;

[0140] When FX1<FXxs≤FX2, a low-medium risk level is generated;

[0141] When FX2<FXxs≤FX3, a medium risk level is generated;

[0142] When FX3<FXxs≤FX4, a medium-high risk level is generated;

[0143] When FXxs>FX4, a high risk level is generated.

[0144] It should be noted that the peak probability is that each time point of the risk evolution curve is traversed, and the maximum risk value is found as the peak probability; it represents the possibility that the frequency band has the highest damage risk in the prediction period, and the higher the peak, the more likely that the subject will suffer significant hearing damage in the short term, which is an important reference index for early intervention.

[0145] The difference of the continuous time points of the risk curve is calculated to obtain the risk increment of each time period, and the growth rate represents the rapid rate of damage spread, which describes the speed of risk accumulation from low value to high value;

[0146] The persistence length, i.e., the risk threshold is set, and the length of the time period in which the curve is continuously higher than the threshold is counted.

[0147] It should be noted that the weight coefficient in the above and each preset threshold in the analysis process is set by a person skilled in the art according to the actual situation or obtained by a large amount of data simulation. The size of the weight coefficient is to obtain a specific numerical value by quantifying each parameter, which is convenient for subsequent comparison. The size of the weight coefficient depends on the number of sample data and the preliminary processing coefficient corresponding to each group of sample data set by the person skilled in the art; as long as it does not affect the proportional relationship between the parameters and the quantized numerical value.

[0148] In this embodiment, the present application constructs a noise-induced deafness susceptibility grading method based on multi-frequency hearing threshold. Through multi-frequency point measurement from high frequency to low frequency, ear sound emission signal collection, auditory brainstem response monitoring and physiological parameter recording, the present application realizes the full range and dynamic perception of individual auditory damage. Compared with the existing method which only relies on single frequency hearing threshold, the present application can capture the time sequence rule of high frequency hair cell damage spreading to the middle and low frequency region, effectively identify the hidden damage signs of middle and low frequency hair cells, and thus significantly improve the accuracy and sensitivity of early identification of noise-induced deafness. At the same time, by introducing the subject cooperation degree evaluation and confidence label, the present application can improve the reliability and accuracy of the measurement data and reduce the influence of individual subjective response deviation on the results.

[0149] In addition, the present application establishes a spectrum migration operator based on physical constraints and a frequency band inter-temporal coupling framework, maps the otoacoustic emission joint parameters and physiological parameters into a frequency band migration probability field, and combines a future risk evolution prediction model to predict the frequency band risk evolution curve in the next one to three months. By extracting the curve features and weighting and synthesizing a single risk score, the present application can quantitatively classify the individual noise-induced hearing loss risk and realize dynamic and individualized risk assessment. The method not only provides a scientific basis for early intervention, but also can assist in developing individualized noise protection strategies, improve the timeliness and accuracy of protection measures, and has significant clinical application value and public health significance.

[0150] Embodiment 2

[0151] Please refer to Figure 4 The multi-frequency hearing threshold noise-induced hearing loss susceptibility grading system provided by the embodiment of the present application is used to realize the multi-frequency hearing threshold noise-induced hearing loss susceptibility grading method, and the system comprises:

[0152] A frequency point generation module, which generates a set of measurement frequency points in the order from high frequency to low frequency, and completes M independent measurements for each frequency point;

[0153] A response evaluation module, which records the measured subjective response timestamp, judges the measured cooperation degree through the reaction time distribution, and generates a confidence value label;

[0154] An otoacoustic emission acquisition module, which acquires the otoacoustic emission signals in the ear canal of the measured person in real time, forms a time-domain signal of acoustic reflection changing with time, extracts the time-domain signal with the highest confidence of a single frequency point, calculates the spectral features, and constructs the otoacoustic emission joint parameters;

[0155] A brainstem feature extraction module, which synchronously records the auditory brainstem response signals and extracts the hidden damage features of the auditory brainstem response;

[0156] A physiological parameter acquisition module, which acquires the physiological parameters of the measured person, and the physiological parameters include heart rate, ear canal temperature and electrode impedance;

[0157] A coupling modeling module, which establishes a frequency band inter-temporal coupling framework based on the physiological parameters and the otoacoustic emission time-domain-frequency-domain joint parameters, and formalizes the damage propagation mechanism from high frequency to medium frequency in the form of a physically constrained spectral migration operator, so that the operator can receive the physiological parameters and the otoacoustic emission time-domain-frequency-domain joint parameters and output a frequency band migration probability field;

[0158] A risk prediction module, which inputs the obtained frequency band migration probability field and hidden damage features into a pre-constructed future risk evolution prediction model, and outputs the frequency band risk evolution curve in the next one to three months;

[0159] The grade dividing module extracts curve features of the frequency band risk evolution curve, performs weighted synthesis by preset weights, obtains a single risk score of the frequency band, and generates a noise-induced deafness susceptibility grading level.

[0160] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for classifying susceptibility to noise-induced hearing loss based on multi-band hearing thresholds, characterized in that: include: Preset measurement frequency points, generate a set of measurement frequency points in order from high frequency to low frequency, and complete M independent measurements for each frequency point; Record the timestamps of the subjective responses of the test subjects, determine the degree of cooperation of the test subjects through the reaction time distribution, and generate confidence value labels; Otoacoustic emission signals in the ear canal under test are acquired in real time to form a time-domain signal of sound reflection changing over time. The time-domain signal with the highest confidence at a single frequency point is extracted, the spectral characteristics are calculated, and joint parameters of otoacoustic emission are constructed. Simultaneously record auditory brainstem response signals and extract latent damage characteristics of auditory brainstem responses; The physiological parameters of the subjects were collected, including heart rate, ear canal temperature, and electrode impedance. A temporal coupling framework between frequency bands is established based on physiological parameters and joint time-frequency parameters of otoacoustic emission. A physically constrained spectral migration operator is used to formalize the damage propagation mechanism from high frequency to mid frequency, enabling the operator to receive physiological parameters and joint time-frequency parameters of otoacoustic emission and output a frequency band migration probability field. The acquired frequency band migration probability field and latent damage features are input into a pre-constructed future risk evolution prediction model, which outputs the risk evolution curves of each frequency band in the next one to three months. The curve features of the frequency band risk evolution curve are extracted, and weighted synthesis is performed with preset weights to obtain a single risk score for the frequency band, thereby generating a noise-induced hearing loss susceptibility classification level.

2. The method for classifying susceptibility to noise-induced hearing loss based on multi-band hearing thresholds according to claim 1, characterized in that, The method for determining the compatibility degree of the tested object and generating confidence value labels based on the reaction time distribution includes: The test subject pressed a response button as prompted when stimulated; Record the start time of each stimulus and the time the subject presses the key, and also record the stimulus frequency, frequency number Fi, and number of tests Fij; j∈{1,2,…,M}; Calculate the average response time, standard deviation, and frequency order consistency score for each frequency point. Use the average response time, standard deviation, and frequency order consistency score to perform a weighted summation to generate a comprehensive coordination score. The overall cooperation score is used as the confidence level for each measurement result.

3. The method for classifying susceptibility to noise-induced hearing loss based on multi-band hearing thresholds according to claim 2, characterized in that, The method for obtaining the frequency point order consistency score includes: The set of measurement frequency points is arranged from high frequency to low frequency to form a reference sequence. Each frequency point is assigned a position number in the reference sequence, and the position number of each frequency point represents its position in the reference sequence. During the measurement process, the subjective responses of the test subjects to each frequency point are sorted from earliest to latest according to the response time. Each frequency point obtains a position number in the actual response sequence to indicate the order in which the frequency point is responded. For each actual response frequency point, find its position number in the reference sequence, then compare the difference between the actual response sequence and the reference sequence, take the absolute value, perform overall statistical analysis on the order deviation of all frequency points, obtain the average value, and divide the difference value by the average value to obtain the order consistency score.

4. The method for classifying susceptibility to noise-induced hearing loss based on multi-band hearing thresholds according to claim 1, characterized in that, The spectral characteristics include: peak frequency average value, harmonic intensity coefficient, and average drift rate; The method for extracting the spectral features includes: The original time-domain waveform is bandpass filtered, and the bandpass range covers the target analysis frequency band. The filter uses linear phase or zero phase to avoid phase distortion. Using the stimulus trigger point as the zero point of time, a fixed-length window containing the complete reflection packet is extracted. The time window length is preset, and each window uses a Hanning window or a Kaiser window with an overlap rate set. Perform a short-time Fourier transform on each window length to obtain the corresponding time-frequency matrix; In the average time-frequency matrix, local energy peaks are searched for each time frame to identify the main frequency peak. Trajectories of the main frequency peaks are connected in consecutive time frames to form peak frequency trajectories and obtain the average peak frequency. Within each time window, the ratio of fundamental frequency energy to harmonic band energy is calculated, and the average value is used as the harmonic intensity coefficient. Based on the acquired peak frequency trajectory, the total drift amount and drift time of the trajectory are calculated, and the average drift rate is obtained.

5. The method for classifying susceptibility to noise-induced hearing loss based on multi-band hearing thresholds according to claim 1, characterized in that, The method for synchronously recording auditory brainstem response signals includes: During the measurement, recording electrodes are placed on the top of the subject's head, reference electrodes are placed on the earlobe, and ground electrodes are placed on the forehead or sternum. Digital sampling is performed using a high sampling frequency greater than or equal to 20kHz; Each stimulus is recorded as a single auditory brainstem response. Multiple responses are added together and divided by the number of responses to obtain the average auditory brainstem response waveform.

6. The method for classifying susceptibility to noise-induced hearing loss based on multi-band hearing thresholds according to claim 5, characterized in that, Methods for extracting latent impairment features of auditory brainstem responses include: Peak search windows were set on the average auditory brainstem response waveform, with wave I window being 1–3 ms, wave III window being 3–5 ms, and wave V window being 5–8 ms. Within each search window, find the highest amplitude positive peak, obtain the time axis value of the location where the positive peak appears, and record the time axis value as the corresponding latency within the search window. Obtain the latencies of wave I, wave III and wave V in sequence. Then obtain the interwave latency differences between wave I and wave III, wave III and wave V, and wave I and wave V.

7. The method for classifying susceptibility to noise-induced hearing loss based on multi-band hearing thresholds according to claim 1, characterized in that, The training method for the future risk evolution prediction model includes: Q sets of time-series training data were collected in advance, where Q is a positive integer greater than 0. The time-series training data included the frequency band migration probability field and latent damage features, as well as the frequency band risk evolution curves corresponding to the frequency band migration probability field and latent damage features. Long Short-Term Memory Network was selected as the future risk evolution prediction model. The Adam optimizer was selected and classification cross-entropy was used as the loss function. The constructed time series training data was divided into training set, validation set and test set according to the design ratio. The training set data is input into the model in batches. The predicted probability distribution is obtained through forward propagation. After calculating the loss, the network weights are updated through backpropagation. During the training process, the weighted F1 score is calculated using the validation set after each round. If the validation set performance does not improve within a preset number of rounds, the training is automatically stopped and the optimal model at this time is saved to prevent overfitting and save training time. On the test set, the model's weighted F1 score for predicting risk levels for the next 1 month and 3 months must reach a preset threshold. After the model training is completed, the trained model is converted into the ONNX lightweight format and deployed to the cloud or edge computing nodes.

8. The method for classifying susceptibility to noise-induced hearing loss based on multi-band hearing thresholds according to claim 1, characterized in that, The curve features include peak probability, growth rate, and duration. The curve features are weighted and summed according to preset weights to obtain a single risk score for the frequency band.

9. The method for classifying susceptibility to noise-induced hearing loss based on multi-band hearing thresholds according to claim 8, characterized in that, The method for generating a noise-induced hearing loss susceptibility rating includes: The average risk score FXxs is obtained by summing the individual risk scores of each frequency band and dividing by the number of frequency bands. The preset risk score gradient thresholds are FX1, FX2, FX3, and FX4, where FX1 < FX2 < FX3 < FX4; When FXxs≤FX1, a low-risk level is generated; When FX1 < FXxs ≤ FX2, a low to medium risk level is generated; When FX2 < FXxs ≤ FX3, a medium risk level is generated; When FX3 < FXxs ≤ FX4, a medium-to-high risk level is generated; When FXxs > FX4, a high-risk level is generated.

10. A multi-band hearing threshold noise-induced hearing loss susceptibility classification system, used to implement the multi-band hearing threshold noise-induced hearing loss susceptibility classification method according to any one of claims 1-9, characterized in that, The system includes: The frequency point generation module presets measurement frequencies and generates a set of measurement frequencies in order from high frequency to low frequency. Each frequency point completes M independent measurements. The response assessment module records the timestamps of the subjective responses of the test subjects, determines the degree of cooperation of the test subjects through the reaction time distribution, and generates confidence value labels. The otoacoustic acquisition module acquires otoacoustic emission signals in the ear canal in real time, forms a time-domain signal of sound reflection changing over time, extracts the time-domain signal with the highest confidence at a single frequency point, calculates spectral characteristics, and constructs joint otoacoustic emission parameters; The brainstem feature extraction module synchronously records auditory brainstem response signals and extracts latent damage features of auditory brainstem responses; The physiological parameter acquisition module collects the subject's physiological parameters, including heart rate, ear canal temperature, and electrode impedance. The coupling modeling module establishes a temporal coupling framework between frequency bands based on physiological parameters and joint time-frequency parameters of otoacoustic emission. It formalizes the damage propagation mechanism from high frequency to mid frequency using a physically constrained spectral migration operator, enabling the operator to receive physiological parameters and joint time-frequency parameters of otoacoustic emission and output a frequency band migration probability field. The risk prediction module inputs the acquired frequency band migration probability field and latent damage characteristics into a pre-built future risk evolution prediction model, and outputs the risk evolution curves of each frequency band within the next one to three months. The classification module extracts the curve features of the frequency band risk evolution curve, performs weighted synthesis with preset weights, obtains a single risk score for the frequency band, and generates a noise-induced hearing loss susceptibility classification level.

Images