Auditory brainstem response automatic classification method and device based on time-frequency cross attention network, equipment and storage medium

Abstract

The application discloses an auditory brainstem response automatic classification method and device based on a time-frequency cross attention network, equipment and a storage medium, relates to the technical field of hearing monitoring, and through double-branch parallel processing of ABR signals: a one-dimensional convolution is adopted in a time domain branch to extract local waveform form features, and a short-time Fourier transform is adopted in a frequency domain branch to obtain a time-frequency spectrum feature; a feature linear modulation mechanism is introduced into both branches, a stimulus intensity is taken as condition information to adaptively scale and offset intermediate features; complementary fusion of time-frequency features is realized through a bidirectional cross attention mechanism, time domain features can query frequency domain information, and frequency domain features can query time domain information; global average pooling, stimulus intensity late fusion and a multilayer perception machine are combined to make a binary classification decision. In the automatic classification, multi-modal information and physiological priori can be fully utilized without manual feature engineering, and the accuracy, objectivity and efficiency of hearing screening and diagnosis are significantly improved.

Description

Technical Field

[0001] This application relates to the field of hearing monitoring technology, and in particular to an automatic classification method for auditory brainstem responses based on time-frequency cross-attention networks. Background Technology

[0002] The Auditory Brainstem Response (ABR) is an objective electrophysiological test used to assess the functional status of the auditory conduction pathway from the cochlea and auditory nerve to the brainstem nuclei. It is widely used in newborn hearing screening, post-cochlear implantation rehabilitation monitoring, and as an auxiliary diagnostic tool for various central and peripheral auditory system diseases. The ABR waveform typically contains multiple characteristic peaks, with wave I, wave III, and wave V being the primary basis for clinical interpretation. These peaks generally appear sequentially within milliseconds after auditory stimulation, reflecting neural activity at different levels of the auditory pathway.

[0003] Current clinical ABR interpretation primarily relies on manual methods, where professional technicians identify and judge waveform morphology, peak latency, and amplitude characteristics, then compare the results with a norm database. Because ABR waveforms are susceptible to noise, electromyographic interference, and individual differences, interpretations often vary significantly between interpreters, especially in cases of low signal-to-noise ratios or atypical waveforms. Furthermore, with increased newborn screening coverage and the rise in cochlear implant surgeries, the large-scale ABR testing data generated by medical institutions places higher demands on the efficiency and consistency of traditional manual interpretation methods. This manual process is not only time-consuming but also struggles to ensure consistent interpretation standards across institutions and operators.

[0004] To improve interpretation efficiency and result stability, some studies have proposed using machine learning methods based on manual feature extraction to construct classification models by analyzing indicators such as peak latency and amplitude ratio. However, such methods are highly dependent on feature engineering, requiring extensive domain knowledge for feature selection. Furthermore, manually designed features are insufficient to fully describe the complex variation patterns of ABR signals, resulting in limited generalization ability when facing different populations, different devices, or different acquisition conditions.

[0005] In recent years, deep learning technology has been introduced into the task of automatic ABR analysis. Some studies utilize models such as convolutional neural networks, recurrent neural networks, or Transformers to attempt to automatically learn feature representations from raw signals and improve feature extraction capabilities. However, most existing methods only model single data representations in the time or frequency domains, lacking comprehensive utilization of complementary information between different representation domains. In such models, while time-domain networks can capture waveform morphology features, their ability to express frequency component distribution is limited; frequency-domain representations struggle to maintain high temporal resolution simultaneously. Furthermore, ABR waveforms exhibit a clear physiological pattern of shortening latency and increasing amplitude with changes in stimulus intensity, but existing deep learning models rarely incorporate stimulus intensity information as an explicit condition for feature learning, failing to fully reflect this physiological prior knowledge. On the other hand, a few methods have attempted to fuse time-domain and frequency-domain features, but most employ shallow fusion mechanisms such as simple feature concatenation or weighted merging, lacking deep interaction between cross-domain features and failing to fully exploit the synergistic effects of multi-domain information.

[0006] In summary, existing automated ABR analysis techniques still have limitations in feature representation, utilization of prior information, and multi-domain fusion mechanisms, making it difficult to simultaneously meet the clinical requirements of high accuracy, robustness, and physiological interpretability. Therefore, it is necessary to propose an automated ABR analysis method that can comprehensively utilize time-frequency multi-domain information, explicitly introduce physiological priors on stimulus intensity, and achieve deep fusion, in order to improve interpretation efficiency and recognition accuracy, and meet the needs of large-scale clinical screening and diagnosis scenarios.

[0007] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0008] The main purpose of this application is to provide an automatic classification method for auditory brainstem responses based on a time-frequency cross-attention network, which aims to solve the technical problem of low accuracy in judging hearing status in the prior art.

[0009] To achieve the above objectives, this application provides an automatic classification method for auditory brainstem responses based on a time-frequency cross-attention network, the method comprising:

[0010] The original ABR waveform is preprocessed by Z-score standardization on a sample-by-sample basis, and the stimulus intensity is kept at the original dB SPL value without normalization, resulting in a standardized ABR signal sequence.

[0011] The standardized ABR signal sequence is processed by extracting local temporal morphological features through a one-dimensional convolutional layer, and then subjected to stimulus intensity conditional modulation through a feature linear modulation module. Position embedding is added to obtain a stimulus conditional temporal feature representation.

[0012] The standardized ABR signal sequence is subjected to short-time Fourier transform to extract frequency domain spectral features, and stimulus intensity conditional modulation is performed through the feature linear modulation module. Position embedding is added to obtain the stimulus conditional frequency domain feature representation.

[0013] The stimulus-conditional temporal feature representation and the stimulus-conditional frequency domain feature representation are input into a bidirectional cross-attention module for deep fusion to obtain a joint feature representation;

[0014] The joint feature representation is subjected to global average pooling, concatenated with the stimulus intensity, and then output as an ABR automatic classification result through a classifier.

[0015] In one embodiment, the steps of extracting temporal local morphological features from the standardized ABR signal sequence through a one-dimensional convolutional layer, performing stimulus intensity conditional modulation through a feature linear modulation module, and adding position embedding to obtain a stimulus-conditional temporal feature representation include:

[0016] The standardized ABR signal sequence is convolved by a one-dimensional convolutional layer, and the convolutional output is batch normalized and GELU activated to obtain temporal convolutional features.

[0017] The stimulus intensity value is input into the temporal feature linear modulation network, which contains two layers of multilayer perceptron. The first layer maps the input to the intermediate dimension and applies ReLU activation, and the second layer outputs scaling parameters and offset parameters.

[0018] The temporal convolutional features are modulated by element-wise affine transformation;

[0019] By adding learnable location embeddings, we obtain a temporal feature representation of stimulus-conditionalization.

[0020] In one embodiment, the steps of extracting frequency domain spectral features by performing short-time Fourier transform on the standardized ABR signal sequence, performing stimulus intensity conditional modulation through a feature linear modulation module, and adding position embedding to obtain a stimulus-conditional frequency domain feature representation include:

[0021] The standardized ABR signal sequence is subjected to short-time Fourier transform, the amplitude spectrum is calculated using the Hanning window function, and logarithmic compression is performed.

[0022] The logarithmic amplitude spectrum is projected onto the feature space through a fully connected layer, and layer normalization is performed to obtain the frequency domain features.

[0023] The stimulus intensity is input into a frequency domain linear modulation network, which has the same structure as the time domain linear modulation network but with independent parameters, and outputs frequency domain modulation parameters.

[0024] The frequency domain features are conditionalized by element-level modulation;

[0025] By adding learnable positional embeddings, we obtain a frequency domain feature representation of stimulus-conditionalization.

[0026] In one embodiment, the step of inputting the stimulus-conditioned temporal feature representation and the stimulus-conditioned frequency domain feature into a bidirectional cross-attention module for deep fusion to obtain a joint feature representation includes:

[0027] The temporal feature representation and the frequency domain feature representation of the stimulus conditionation are respectively subjected to layer normalization;

[0028] Normalized time-domain features are used as queries, and frequency-domain features are used as keys and values. Cross-attention is calculated through a multi-head attention layer, and time-domain fusion features are obtained through residual connections and a feedforward network.

[0029] Normalized frequency domain features are used as queries, and time domain features are used as keys and values. Cross attention is calculated through a multi-head attention layer, and frequency domain fusion features are obtained through residual connections and feedforward networks.

[0030] The output time-domain fusion features and frequency-domain fusion features are used as the joint feature representation after bidirectional fusion.

[0031] In one embodiment, the step of performing global average pooling on the joint feature representation, concatenating it with the stimulus intensity, and then outputting the ABR automatic classification result through a classifier includes:

[0032] Global average pooling is performed on the time-domain fusion features and frequency-domain fusion features respectively to obtain feature vectors of fixed dimensions;

[0033] The time-domain feature vector, frequency-domain feature vector, and original stimulus intensity value are concatenated to form a unified representation vector;

[0034] The unified representation vector is input into a two-layer multilayer perceptron classifier. The first layer includes fully connected transformation, GELU activation and Dropout, and the second layer outputs a single logit.

[0035] The probability value is obtained by applying the Sigmoid activation function to the logit, and binary classification is performed using a threshold of 0.5 to output the result of normal hearing or abnormal hearing.

[0036] In one embodiment, the feature linear modulation network includes a multilayer perceptron that learns the mapping relationship from stimulus intensity to modulation parameters through layer-by-layer nonlinear transformation; the element-level affine transformation modulates the features through scaling and offset operations.

[0037] In one embodiment, the short-time Fourier transform uses a window function for frame segmentation, and logarithmic compression is used to reduce the dynamic range; the frequency domain characteristic linear modulation network and the time domain modulation network have completely independent parameters.

[0038] In one embodiment, the multi-head attention layer employs a scaled dot product attention mechanism, and the feedforward network adopts a bottleneck structure, including linear transformations and activation functions.

[0039] In one embodiment, the step of performing sample-by-sample Z-score normalization preprocessing on the original ABR waveform, while keeping the stimulus intensity at its original dB SPL value without normalization, to obtain a normalized ABR signal sequence includes:

[0040] Each ABR signal sample is subjected to sample-by-sample Z-score standardization, and the sample mean and standard deviation are calculated and normalized; however, the stimulus intensity is kept in its original dimension and is not normalized, resulting in a standardized ABR signal sequence.

[0041] In one embodiment, model training employs a cross-entropy loss function with class weights, where class weights are inversely proportional to class frequencies; an adaptive optimizer is used for parameter updates; and early stopping and learning rate decay strategies are implemented to improve training efficiency and model performance.

[0042] Furthermore, to achieve the above objectives, this application also proposes an automatic classification device for auditory brainstem responses based on a time-frequency cross-attention network, wherein the automatic classification device for auditory brainstem responses based on a time-frequency cross-attention network includes:

[0043] The signal processing module is used to perform sample-by-sample Z-score normalization preprocessing on the raw ABR waveform, keeping the stimulus intensity at its original dB SPL value without normalization, to obtain a normalized ABR signal sequence.

[0044] The temporal modulation module is used to extract temporal local morphological features from the standardized ABR signal sequence through a one-dimensional convolutional layer, perform stimulus intensity conditional modulation through the feature linear modulation module, and add position embedding to obtain a stimulus conditional temporal feature representation.

[0045] The frequency domain modulation module is used to extract frequency domain spectral features by performing short-time Fourier transform on the standardized ABR signal sequence, perform stimulus intensity conditional modulation through the feature linear modulation module, and add position embedding to obtain a stimulus conditional frequency domain feature representation.

[0046] The feature union module is used to input the stimulus-conditioned temporal feature representation and the stimulus-conditioned frequency domain feature into the bidirectional cross-attention module for deep fusion to obtain a joint feature representation;

[0047] The automatic classification result is used to perform global average pooling on the joint feature representation, and after being concatenated with the stimulus intensity, it is output as an ABR automatic classification result through a classifier.

[0048] Furthermore, to achieve the above objectives, this application also proposes an automatic classification device for auditory brainstem responses based on a time-frequency cross-attention network. The automatic classification device for auditory brainstem responses based on a time-frequency cross-attention network includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is configured to implement the steps of the automatic classification method for auditory brainstem responses based on a time-frequency cross-attention network as described above.

[0049] In addition, to achieve the above objectives, the present invention also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the automatic classification method for auditory brainstem responses based on time-frequency cross-attention networks as described above.

[0050] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the automatic classification method for auditory brainstem responses based on time-frequency cross-attention networks as described above.

[0051] This application provides an automatic classification method for auditory brainstem response (ABR) based on a time-frequency cross-attention network. It employs a two-branch parallel processing approach for ABR signals: the time-domain branch uses one-dimensional convolution to extract local waveform morphological features, while the frequency-domain branch obtains time-spectrum features through short-time Fourier transform. Both branches introduce a feature linear modulation mechanism, using stimulus intensity as conditional information to adaptively scale and shift intermediate features. A bidirectional cross-attention mechanism enables complementary fusion of time-frequency features, allowing time-domain features to query frequency-domain information, and vice versa. Finally, global average pooling, late-stage stimulus intensity fusion, and multilayer perceptron are combined for binary classification decisions. This invention provides high-precision automatic classification on clinical ABR data, fully utilizing multimodal information and physiological priors without manual feature engineering, significantly improving the accuracy, objectivity, and efficiency of hearing screening and diagnosis. Attached Figure Description

[0052] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0053] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a flowchart of the ABR automatic classification method in Embodiment 1 of the automatic classification method for auditory brainstem response based on time-frequency cross-attention network of this application;

[0055] Figure 2 This is a detailed schematic diagram of the TFCANet network architecture, which is an embodiment of the automatic classification method for auditory brainstem responses based on time-frequency cross-attention network of this application.

[0056] Figure 3 This is a performance comparison chart on a test set of an embodiment of the automatic classification method for auditory brainstem responses based on time-frequency cross-attention networks according to this application.

[0057] Figure 4 This is a confusion matrix result diagram of an embodiment of the automatic classification method for auditory brainstem responses based on time-frequency cross-attention networks in this application;

[0058] Figure 5 This is a comparison of ablation experiment results for an embodiment of the automatic classification method for auditory brainstem responses based on time-frequency cross-attention networks in this application.

[0059] Figure 6 This is a schematic diagram of the module structure of the automatic classification device for auditory brainstem response based on time-frequency cross-attention network according to an embodiment of this application;

[0060] Figure 7 This is a schematic diagram of the device structure of the hardware operating environment involved in the automatic classification method of auditory brainstem response based on time-frequency cross-attention network in the embodiments of this application.

[0061] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0062] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0063] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0064] The main solution of this application embodiment is: to perform sample-by-sample Z-score normalization preprocessing on the original ABR waveform, and to keep the original dB SPL value of the stimulus intensity without normalization, so as to obtain a normalized ABR signal sequence;

[0065] The standardized ABR signal sequence is processed by extracting local temporal morphological features through a one-dimensional convolutional layer, and then subjected to stimulus intensity conditional modulation through a feature linear modulation module. Position embedding is added to obtain a stimulus conditional temporal feature representation.

[0066] The standardized ABR signal sequence is subjected to short-time Fourier transform to extract frequency domain spectral features, and stimulus intensity conditional modulation is performed through the feature linear modulation module. Position embedding is added to obtain the stimulus conditional frequency domain feature representation.

[0067] The temporal and frequency domain features of the stimulus-conditioning are input into a bidirectional cross-attention module for deep fusion to obtain a joint feature representation.

[0068] The joint feature representation is subjected to global average pooling, concatenated with the stimulus intensity value, and then output as an ABR automatic classification result through a classifier.

[0069] Currently, the Auditory Brainstem Response (ABR) is an objective electrophysiological test that can be used to assess the functional status of the auditory conduction pathway from the cochlea and auditory nerve to the brainstem nuclei. It is widely used in newborn hearing screening, post-cochlear implantation rehabilitation monitoring, and the auxiliary diagnosis of various central and peripheral auditory system diseases. The ABR waveform typically contains multiple characteristic peaks, among which wave I, wave III, and wave V are the main criteria for clinical interpretation. These peaks generally appear sequentially within milliseconds after sound stimulation, reflecting neural activity at different levels of the auditory pathway.

[0070] Current clinical ABR interpretation primarily relies on manual methods, where professional technicians identify and judge waveform morphology, peak latency, and amplitude characteristics, then compare the results with a norm database. Because ABR waveforms are susceptible to noise, electromyographic interference, and individual differences, interpretations often vary significantly between interpreters, especially in cases of low signal-to-noise ratios or atypical waveforms. Furthermore, with increased newborn screening coverage and the rise in cochlear implant surgeries, the large-scale ABR testing data generated by medical institutions places higher demands on the efficiency and consistency of traditional manual interpretation methods. This manual process is not only time-consuming but also struggles to ensure consistent interpretation standards across institutions and operators.

[0071] To improve interpretation efficiency and result stability, some studies have proposed using machine learning methods based on manual feature extraction to construct classification models by analyzing indicators such as peak latency and amplitude ratio. However, such methods are highly dependent on feature engineering, requiring extensive domain knowledge for feature selection. Furthermore, manually designed features are insufficient to fully describe the complex variation patterns of ABR signals, resulting in limited generalization ability when facing different populations, different devices, or different acquisition conditions.

[0072] In recent years, deep learning technology has been introduced into the task of automatic ABR analysis. Some studies utilize models such as convolutional neural networks, recurrent neural networks, or Transformers to attempt to automatically learn feature representations from raw signals and improve feature extraction capabilities. However, most existing methods only model single data representations in the time or frequency domains, lacking comprehensive utilization of complementary information between different representation domains. In such models, while time-domain networks can capture waveform morphology features, their ability to express frequency component distribution is limited; frequency-domain representations struggle to maintain high temporal resolution simultaneously. Furthermore, ABR waveforms exhibit a clear physiological pattern of shortening latency and increasing amplitude with changes in stimulus intensity, but existing deep learning models rarely incorporate stimulus intensity information as an explicit condition for feature learning, failing to fully reflect this physiological prior knowledge. On the other hand, a few methods have attempted to fuse time-domain and frequency-domain features, but most employ shallow fusion mechanisms such as simple feature concatenation or weighted merging, lacking deep interaction between cross-domain features and failing to fully exploit the synergistic effects of multi-domain information.

[0073] In summary, existing automated ABR analysis techniques still have limitations in feature representation, utilization of prior information, and multi-domain fusion mechanisms, making it difficult to simultaneously meet the clinical requirements of high accuracy, robustness, and physiological interpretability. Therefore, it is necessary to propose an automated ABR analysis method that can comprehensively utilize time-frequency multi-domain information, explicitly introduce physiological priors on stimulus intensity, and achieve deep fusion, in order to improve interpretation efficiency and recognition accuracy, and meet the needs of large-scale clinical screening and diagnosis scenarios.

[0074] This application provides a solution that performs parallel processing of ABR signals in two branches: the time-domain branch uses one-dimensional convolution to extract local waveform morphological features, and the frequency-domain branch obtains time-spectrum features through short-time Fourier transform; both branches introduce a feature linear modulation mechanism, using stimulus intensity as conditional information to adaptively scale and shift intermediate features; a bidirectional cross-attention mechanism achieves complementary fusion of time-frequency features, enabling time-domain features to query frequency-domain information, and vice versa; finally, global average pooling, late stimulus intensity fusion, and multilayer perceptron are combined for binary classification decision-making. This invention provides high-precision automatic classification on clinical ABR data, fully utilizing multimodal information and physiological priors without manual feature engineering, significantly improving the accuracy, objectivity, and efficiency of hearing screening and diagnosis.

[0075] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions, such as an automatic auditory brainstem response classification device based on a time-frequency cross-attention network, etc. This embodiment does not specifically limit it. The following uses an automatic auditory brainstem response classification device based on a time-frequency cross-attention network as an example to describe this embodiment and the following embodiments.

[0076] All actions involving the acquisition of signals, information, or data in this application are carried out in accordance with the relevant data protection laws and policies of the country where the application is located, and with the authorization of the owner of the relevant device.

[0077] This application provides an automatic classification method for auditory brainstem responses based on a time-frequency cross-attention network, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the automatic classification method for auditory brainstem responses based on time-frequency cross-attention networks according to this application.

[0078] In this embodiment, the automatic classification method for auditory brainstem responses based on time-frequency cross-attention networks includes steps S10-S50:

[0079] Step S10: Perform sample-by-sample Z-score normalization preprocessing on the original ABR waveform, keeping the stimulus intensity at its original dB SPL value without normalization, to obtain a normalized ABR signal sequence.

[0080] In this study, 758 pediatric subjects (511 males and 247 females), aged 0 to 14 years (mean age 2.45 ± 2.34 years), were included. Clicking sounds were used as stimulation signals to examine both ears of each subject, with stimulation intensities ranging from 10 to 100 dB SPL in 10 dB increments, resulting in a total of 6685 ABR recordings. The time window for each ABR recording was -13.675 ms to 11.900 ms relative to the stimulus onset (defined as 0 ms), with a sampling frequency of 40 kHz and a total of 1024 sampling points.

[0081] The data was divided into two categories: normal hearing (4436 records, 66.4%) and abnormal hearing (2249 records, 33.6%), exhibiting an approximately 2:1 class imbalance. The data was stored in CSV format, with each file corresponding to a category label identified by the numbers in the filename. Each column in the CSV file represents an ABR record, with the first row showing the stimulus intensity value (dB SPL), followed by subsequent rows of ABR waveform sampling points. The data loading process involved parsing the filenames using regular expressions to extract the category labels, reading the CSV content column by column, and separating the stimulus intensity from the waveform sequence.

[0082] In this embodiment, for each ABR signal sample, the 1st to 477th sampling points after the stimulus intensity value are extracted as the effective waveform sequence. The sample mean μx and standard deviation σx of this sequence are calculated using the formula... Perform sample-by-sample normalization, where This is the numerical stability constant. This sample-by-sample standardization eliminates amplitude scale differences among different subjects and under different recording conditions, allowing the model to focus on the relative morphological characteristics of the waveform. The stimulus intensity value d retains the original dB SPL dimensions without normalization to preserve the physiological significance carried by its absolute magnitude.

[0083] The dataset was stratified and split: 10% of the total data was first selected as the test set, and another 10% of the remaining data was selected as the validation set, resulting in a training set of approximately 81%. All splits were performed using a fixed random seed (RANDOM_STATE=42) to ensure repeatability.

[0084] Step S20: Extract temporal local morphological features from the standardized ABR signal sequence through a one-dimensional convolutional layer, perform stimulus intensity conditional modulation through a feature linear modulation module, and add position embedding to obtain a stimulus conditional temporal feature representation.

[0085] like Figure 2 As shown, the temporal branch first performs a one-dimensional convolution operation on the standardized ABR signal sequence (shape (B, T, 1), where B is the batch size and T=477 is the sequence length). The convolution kernel size is set to 7, the stride is 1, the padding is "same", and the number of output channels D=128, resulting in the temporal convolution feature Ft with shape (B, T, D). The convolution kernel size of 7 can cover the local neighborhood of the ABR peak, and by sliding 128 learnable filters in the time dimension, local morphological patterns are automatically extracted.

[0086] Batch normalization is performed on Ft, calculating the mean and variance of the current batch in each feature channel. After standardization, learnable scaling and offset parameters are applied. Then, the GELU activation function is applied, defined as follows: ,in It provides a continuously differentiable nonlinear transformation for the cumulative distribution function of the standard normal distribution.

[0087] The stimulus intensity value d (of shape (B, 1)) is input into a temporal feature linear modulation network. This network employs a two-layer multilayer perceptron structure: the first layer is a fully connected layer that maps the input from 1 dimension to D / 2 = 64 dimensions, applying the ReLU activation function; the second layer contains two parallel fully connected branches that map the 64-dimensional intermediate representation to D = 128 dimensions respectively, outputting scaling parameters. and offset parameters The network learns a complex mapping relationship from stimulus intensity to feature modulation parameters.

[0088] Will and (The shapes are all (B, D)) Expanded to (B, 1, D), using the formula Feature modulation is performed, where ⊙ denotes element-wise multiplication. This operation applies an affine transformation to each element of the feature tensor, but the transformation parameters are dynamically determined by the stimulus intensity. In the formula... Using the residual form, when When the modulation operation approaches zero, it degenerates into an identity mapping.

[0089] Add a learnable position embedding matrix of shape (T, D) Initialized using a random normal distribution, it is optimized along with other parameters during training. The location embedding is added via a broadcast mechanism. The final stimulus-conditioned temporal features are obtained. Location embedding assigns a unique learnable vector to each time step in the sequence, enabling the model to distinguish between early and late components and learn the importance weights for different time windows.

[0090] Step S30: Perform short-time Fourier transform on the standardized ABR signal sequence to extract frequency domain spectral features, perform stimulus intensity conditional modulation through the feature linear modulation module, and add position embedding to obtain the stimulus conditional frequency domain feature representation.

[0091] The frequency domain branch first removes the channel dimension of the normalized ABR signal, transforming its shape from (B, T, 1) to (B, T). A short-time Fourier transform (STFT) is then performed on this tensor using a Hanning window function, with a frame length of 64, a frame shift of 16, an FFT length of 64, and zero padding at the end. For an input signal of length T = 477, the STFT processing yields a complex spectrum of shape (B, M, F), where M is the time frame number (approximately 27) and F is the frequency bin number (33).

[0092] Calculate the amplitude |S(t,f)| of the complex spectrum, and then perform logarithmic compression: Slog(t,f) = log(1 + |S(t,f)|). Use log1p to avoid the numerical problems of log(0), while providing a larger relative gain for small amplitudes and compressing large amplitudes to reduce the dynamic range.

[0093] The logarithmic magnitude spectrum is projected from F=33 dimensions to a D=128 dimension feature space through a fully connected layer to obtain the frequency domain features. The shape is (B, M, D). For Layer normalization is performed by calculating the mean and variance of each sample in the feature dimension and applying learnable scaling and offset parameters for normalization.

[0094] The stimulus intensity value d is input into a frequency domain linear modulation network. This network structure is the same as the time-domain modulation network, containing two fully connected layers. The first layer projects to D / 2 = 64 dimensions and is ReLU activated; the second layer outputs a D = 128-dimensional signal. and However, the parameters are completely independent. Parameter independence allows the time-domain and frequency-domain branches to learn stimulus intensity-dependent modulation patterns specific to their respective domains.

[0095] Will and Expanded to (B, 1, D), using the formula The frequency domain features are modulated. A learnable location embedding matrix Pf of shape (64, D) is added (maximum length 64 is a preset fixed value to adapt to different input lengths) to obtain the stimulus-conditional frequency domain features. .

[0096] Step S40: Input the stimulus-conditioned temporal feature representation and the stimulus-conditioned frequency domain feature into the bidirectional cross-attention module for deep fusion to obtain a joint feature representation;

[0097] The bidirectional cross-attention module first focuses on temporal features. and frequency domain features Perform layer normalization separately to obtain and Layer normalization calculates the mean and variance of each sample along the feature dimension, and applies learnable scaling and offset parameters.

[0098] Cross-attention from time domain to frequency domain: As a query , As a key Sum Cross-attention is calculated using a multi-head attention layer. The number of multi-head attention layers is set to 4, and the dimension of each head is [dimension not specified]. Attention calculation uses a scaled dot product:

[0099]

[0100] in A scaling factor of approximately 5.66 is used to prevent excessively large dot product values ​​from causing softmax saturation. The multi-head mechanism will... , , The system is segmented into four heads using a learnable linear projection. Each head independently computes its attention, and the output is then concatenated.

[0101] The attention output is then processed via Dropout (10% dropout rate) and compared with... Perform residual connection to obtain .right Layer normalization is performed before inputting into the feedforward network. The feedforward network employs a bottleneck structure: the first fully connected layer expands the features from D=128 dimensions to 2D=256 dimensions, applying GELU activation and Dropout; the second fully connected layer projects back to 128 dimensions. The output of the feedforward network is then processed through Dropout and compared with... Perform residual connections to obtain temporal fusion features. .

[0102] Cross-attention from the frequency domain to the time domain: performed symmetrically, As a query , As a key Sum Cross-attention is calculated using multi-head attention. Repeated residual connections, layer normalization, and feedforward network processing yield frequency domain fusion features. .

[0103] Bidirectional cross-attention enables time-domain features to selectively focus on relevant frequency-domain information, and frequency-domain features to selectively focus on relevant time-domain information, achieving deep interaction and complementary fusion of features from the two domains.

[0104] Step S50: Perform global average pooling on the joint feature representation, concatenate it with the stimulus intensity, and output the ABR automatic classification result through a classifier.

[0105] Temporal fusion characteristics (shape(B, T, D)) is subjected to global average pooling along the time dimension to obtain a D=128-dimensional vector. Frequency domain fusion features (shape(B, M, D)) is subjected to global average pooling along the time dimension to obtain a D=128-dimensional vector. Global average pooling compresses a variable-length sequence into a fixed-dimensional vector, and obtains the aggregated representation of the entire sequence by averaging over all time steps.

[0106] Will , and the original stimulus intensity scalar value (Shape(B, 1)) is concatenated along the feature dimension to form a unified representation vector z of (2D+1)=257 dimensions. This late fusion strategy enables the classifier to comprehensively consider temporal fusion features, frequency fusion features, and the stimulus intensity itself.

[0107] Dropout (with a dropout rate of 10%) is applied to z, then projected to D=128 dimensions through the first fully connected layer, and GELU activation is applied. Dropout is applied again, and a single scalar logit is output through the second fully connected layer. The sigmoid activation function is applied to the logit to obtain a probability value p in the range [0,1]. A binary classification decision is made using a threshold of 0.5: when p ≥ 0.5, it is predicted as category 1 (abnormal hearing); otherwise, it is predicted as category 0 (normal hearing).

[0108] The model training uses a binary cross-entropy loss function with class weights. The class weights are calculated as follows: ,in Let be the number of samples of category c in the training set. In this embodiment, the weight of the abnormal category is approximately twice that of the normal category.

[0109] The AdamW optimizer is used for parameter updates, with the initial learning rate set to... The weight decay coefficient is Momentum parameters , The numerical stability constant is The batch size is set to 16, and the maximum training duration is 50 epochs.

[0110] An early stopping strategy is used to monitor validation set accuracy. Training is stopped and the optimal weights are restored when validation accuracy shows no improvement for 8 consecutive epochs. A learning rate decay strategy is used to monitor validation set loss. The learning rate is multiplied by 0.5 when validation loss shows no improvement for 4 consecutive epochs, with a minimum learning rate limit of 10⁻⁵.

[0111] like Figure 3As shown, the experimental results demonstrate that the TFCANet model of this invention achieves 91.33% accuracy, 90.43% precision, 90.07% recall, and 90.25% F1 score on the test set. Compared to baseline methods, TFCANet outperforms Transformer (89.24% accuracy) by 2.09 percentage points, DNN (88.64% accuracy) by 2.69 percentage points, and CNN-1D (86.55% accuracy) by 4.78 percentage points.

[0112] like Figure 4 As shown, the confusion matrix results indicate that the model achieves 93.9% specificity (417 / 444 true negatives) and 86.2% sensitivity (194 / 225 true positives) on the test set. The error distribution is 27 false positives and 31 false negatives, indicating that the model does not exhibit significant class bias when handling a 2:1 imbalanced dataset. The slightly higher false negative rate warrants attention in clinical applications, but the 86.2% sensitivity significantly surpasses the baseline method.

[0113] like Figure 5 As shown, a systematic ablation experiment was conducted to verify the effectiveness of each component. The experimental results show:

[0114] The accuracy rate of the time-domain branch alone is 82.06%, indicating that the local morphological features in the time domain have some discriminative power but are insufficient to achieve high-precision classification. The accuracy rate of the frequency-domain branch alone is 88.64%, which is significantly better than the time-domain branch by 6.58 percentage points, indicating that the global spectral energy features in the frequency domain have stronger discriminative power for ABR classification.

[0115] After removing the bidirectional cross-attention mechanism (replacing it with simple concatenation), the accuracy dropped to 88.64%, which is comparable to the performance when used alone in the frequency domain, indicating that simple concatenation cannot effectively fuse time-frequency features. This demonstrates the crucial role of the bidirectional cross-attention mechanism in achieving deep cross-domain feature interaction, contributing 2.69 percentage points.

[0116] Removing the FiLM stimulus intensity modulation mechanism resulted in an accuracy of 90.58%, a decrease of only 0.75 percentage points. Although the contribution of the FiLM mechanism is relatively small, its introduction allows the model to explicitly utilize the physiological prior of stimulus intensity, improving the model's interpretability. Ablation experiments validated the necessity of each component of the model, demonstrating that the complete TFCANet architecture achieves optimal performance.

[0117] This embodiment achieves parallel feature extraction in the time and frequency domains through a dual-branch architecture, explicitly models stimulus intensity as conditional information using the FiLM mechanism, and realizes deep cross-domain feature fusion through bidirectional cross-attention. Experimental results show that this method achieves excellent classification performance on the clinical ABR dataset, providing an effective technical solution for automated ABR analysis.

[0118] This embodiment provides an automatic classification method for auditory brainstem response (ABR) based on a time-frequency cross-attention network. It employs a two-branch parallel processing approach for ABR signals: the time-domain branch uses one-dimensional convolution to extract local waveform morphological features, while the frequency-domain branch obtains time-spectrum features through short-time Fourier transform. Both branches introduce a feature linear modulation mechanism, using stimulus intensity as conditional information to adaptively scale and shift intermediate features. A bidirectional cross-attention mechanism enables complementary fusion of time-frequency features, allowing time-domain features to query frequency-domain information, and vice versa. Finally, global average pooling, late-stage stimulus intensity fusion, and multilayer perceptron are combined for binary classification decisions. This invention provides high-precision automatic classification on clinical ABR data, fully utilizing multimodal information and physiological priors without manual feature engineering, significantly improving the accuracy, objectivity, and efficiency of hearing screening and diagnosis.

[0119] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the automatic classification method of auditory brainstem response based on time-frequency cross-attention network of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0120] This application also provides an automatic auditory brainstem response classification device based on a time-frequency cross-attention network. Please refer to [link / reference]. Figure 6 An automatic classification device for auditory brainstem responses based on a time-frequency cross-attention network includes:

[0121] The signal processing module 10 is used to perform sample-by-sample Z-score normalization preprocessing on the raw ABR waveform, keeping the stimulus intensity at its original dB SPL value without normalization, to obtain a normalized ABR signal sequence.

[0122] The temporal modulation module 20 is used to extract temporal local morphological features from the standardized ABR signal sequence through a one-dimensional convolutional layer, perform stimulus intensity conditional modulation through the feature linear modulation module, and add position embedding to obtain a stimulus conditional temporal feature representation.

[0123] The frequency domain modulation module 30 is used to extract frequency domain spectrum features by performing short-time Fourier transform on the standardized ABR signal sequence, perform stimulus intensity conditional modulation through the feature linear modulation module, and add position embedding to obtain a stimulus conditional frequency domain feature representation.

[0124] The feature joint module 40 is used to input the stimulus-conditioned temporal feature representation and the stimulus-conditioned frequency domain feature into the bidirectional cross-attention module for deep fusion to obtain a joint feature representation;

[0125] The automatic classification result 50 is used to perform global average pooling on the joint feature representation, and after being concatenated with the stimulus intensity, it is output as an ABR automatic classification result through a classifier.

[0126] In one feasible implementation, the time-domain modulation module 20 is further configured to perform a convolution operation on the standardized ABR signal sequence through a one-dimensional convolutional layer, and perform batch normalization and GELU activation on the convolution output to obtain time-domain convolution features.

[0127] The stimulus intensity value is input into the temporal feature linear modulation network, which contains two layers of multilayer perceptron. The first layer maps the input to the intermediate dimension and applies ReLU activation, and the second layer outputs scaling parameters and offset parameters.

[0128] The temporal convolutional features are modulated by element-wise affine transformation;

[0129] By adding learnable location embeddings, we obtain a temporal feature representation of stimulus-conditionalization.

[0130] In one feasible implementation, the frequency domain modulation module 30 is further configured to perform a short-time Fourier transform on the standardized ABR signal sequence, calculate the amplitude spectrum using the Hanning window function, and perform logarithmic compression.

[0131] The logarithmic amplitude spectrum is projected onto the feature space through a fully connected layer, and layer normalization is performed to obtain the frequency domain features.

[0132] The stimulus intensity is input into a frequency domain linear modulation network, which has the same structure as the time domain linear modulation network but with independent parameters, and outputs frequency domain modulation parameters.

[0133] The frequency domain features are conditionalized by element-level modulation;

[0134] By adding learnable positional embeddings, we obtain a frequency domain feature representation of stimulus-conditionalization.

[0135] In one feasible implementation, the feature joint module 40 is further configured to perform layer normalization on the temporal feature representation of the stimulus conditionation and the frequency domain feature of the stimulus conditionation, respectively.

[0136] Normalized time-domain features are used as queries, and frequency-domain features are used as keys and values. Cross-attention is calculated through a multi-head attention layer, and time-domain fusion features are obtained through residual connections and a feedforward network.

[0137] Normalized frequency domain features are used as queries, and time domain features are used as keys and values. Cross attention is calculated through a multi-head attention layer, and frequency domain fusion features are obtained through residual connections and feedforward networks.

[0138] The output time-domain fusion features and frequency-domain fusion features are used as the joint feature representation after bidirectional fusion.

[0139] In one feasible implementation, the automatic classification result 50 is further used to perform global average pooling on the time-domain fusion feature and the frequency-domain fusion feature respectively to obtain a feature vector of fixed dimension.

[0140] The time-domain feature vector, frequency-domain feature vector, and original stimulus intensity value are concatenated to form a unified representation vector;

[0141] The unified representation vector is input into a two-layer multilayer perceptron classifier. The first layer includes fully connected transformation, GELU activation and Dropout, and the second layer outputs a single logit.

[0142] The probability value is obtained by applying the Sigmoid activation function to the logit, and binary classification is performed using a threshold of 0.5 to output the result of normal hearing or abnormal hearing.

[0143] In one feasible implementation, the time-domain modulation module 20 is further configured to include a multilayer perceptron in the feature linear modulation network, which learns the mapping relationship from stimulus intensity to modulation parameters through layer-by-layer nonlinear transformation; the element-level affine transformation modulates the features through scaling and offset operations.

[0144] In one feasible implementation, the frequency domain modulation module 30 is further used to perform frame division processing using a window function for the short-time Fourier transform, and logarithmic compression is used to reduce the dynamic range; the parameters of the frequency domain characteristic linear modulation network and the time domain modulation network are completely independent.

[0145] This application provides an automatic auditory brainstem response classification device based on a time-frequency cross-attention network. The automatic auditory brainstem response classification device based on a time-frequency cross-attention network includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the automatic auditory brainstem response classification method based on a time-frequency cross-attention network in the first embodiment described above.

[0146] The following is for reference. Figure 7This document illustrates a structural schematic diagram of an automatic auditory brainstem response classification device suitable for implementing embodiments of this application based on a time-frequency cross-attention network. The automatic auditory brainstem response classification device based on a time-frequency cross-attention network in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 7 The illustrated automatic classification device for auditory brainstem responses based on time-frequency cross-attention networks is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0147] like Figure 7 As shown, the automatic auditory brainstem response classification device based on a time-frequency cross-attention network may include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in ROM (Read Only Memory) 1002 or a program loaded from storage device 1003 into RAM (Random Access Memory) 1004. The RAM 1004 also stores various programs and data required for the operation of the automatic auditory brainstem response classification device based on the time-frequency cross-attention network. The processing device 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, LCDs (Liquid Crystal Displays), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the auditory brainstem response autoclassification device based on time-frequency cross-attention networks to wirelessly or wiredly communicate with other devices to exchange data. Although the figure shows an auditory brainstem response autoclassification device based on time-frequency cross-attention networks with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0148] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0149] The automatic auditory brainstem response classification device based on time-frequency cross-attention networks provided in this application employs the automatic auditory brainstem response classification method based on time-frequency cross-attention networks described in the above embodiments, and can solve the technical problem of automatic auditory brainstem response classification based on time-frequency cross-attention networks. Compared with the prior art, the beneficial effects of the automatic auditory brainstem response classification device based on time-frequency cross-attention networks provided in this application are the same as the beneficial effects of the automatic auditory brainstem response classification method based on time-frequency cross-attention networks provided in the above embodiments, and other technical features in this automatic auditory brainstem response classification device based on time-frequency cross-attention networks are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0150] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0151] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0152] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the automatic classification method for auditory brainstem responses based on time-frequency cross-attention networks in the above embodiments.

[0153] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable Read Only Memory or Flash Memory), optical fibers, CD-ROM (CD-Read Only Memory), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0154] The aforementioned computer-readable storage medium may be included in an automatic auditory brainstem response classification device based on a time-frequency cross-attention network; or it may exist independently and not be assembled into an automatic auditory brainstem response classification device based on a time-frequency cross-attention network.

[0155] The aforementioned computer-readable storage medium carries one or more programs. When the aforementioned one or more programs are executed by the auditory brainstem response automatic classification device based on time-frequency cross-attention network, the auditory brainstem response automatic classification device based on time-frequency cross-attention network performs sample-by-sample Z-score normalization preprocessing on the raw ABR waveform, keeps the stimulus intensity at the original dB SPL value without normalization, and obtains a normalized ABR signal sequence.

[0156] The standardized ABR signal sequence is processed by extracting local temporal morphological features through a one-dimensional convolutional layer, and then subjected to stimulus intensity conditional modulation through a feature linear modulation module. Position embedding is added to obtain a stimulus conditional temporal feature representation.

[0157] The standardized ABR signal sequence is subjected to short-time Fourier transform to extract frequency domain spectral features, and stimulus intensity conditional modulation is performed through the feature linear modulation module. Position embedding is added to obtain the stimulus conditional frequency domain feature representation.

[0158] The stimulus-conditional temporal feature representation and the stimulus-conditional frequency domain feature representation are input into a bidirectional cross-attention module for deep fusion to obtain a joint feature representation;

[0159] The joint feature representation is subjected to global average pooling, concatenated with the stimulus intensity, and then output as an ABR automatic classification result through a classifier.

[0160] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including LAN (Local Area Network) or WAN (Wide Area Network)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0161] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0162] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0163] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described automatic classification method for auditory brainstem responses based on time-frequency cross-attention networks, thereby solving the technical problem of automatic classification of auditory brainstem responses based on time-frequency cross-attention networks. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the automatic classification method for auditory brainstem responses based on time-frequency cross-attention networks provided in the above embodiments, and will not be repeated here.

[0164] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described automatic classification method for auditory brainstem responses based on a time-frequency cross-attention network.

[0165] The computer program product provided in this application can solve the technical problem of automatic classification of auditory brainstem responses based on time-frequency cross-attention networks. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the automatic classification method of auditory brainstem responses based on time-frequency cross-attention networks provided in the above embodiments, and will not be repeated here.

[0166] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. An automatic classification method for auditory brainstem responses based on a time-frequency cross-attention network, characterized in that, The automatic classification method for auditory brainstem responses based on time-frequency cross-attention networks includes: The original ABR waveform is preprocessed by Z-score standardization on a sample-by-sample basis, and the stimulus intensity is kept at the original dB SPL value without normalization, resulting in a standardized ABR signal sequence. The standardized ABR signal sequence is processed by extracting local temporal morphological features through a one-dimensional convolutional layer, and then subjected to stimulus intensity conditional modulation through a feature linear modulation module. Position embedding is added to obtain a stimulus conditional temporal feature representation. The standardized ABR signal sequence is subjected to short-time Fourier transform to extract frequency domain spectral features, and stimulus intensity conditional modulation is performed through the feature linear modulation module. Position embedding is added to obtain the stimulus conditional frequency domain feature representation. The stimulus-conditional temporal feature representation and the stimulus-conditional frequency domain feature representation are input into a bidirectional cross-attention module for deep fusion to obtain a joint feature representation; The joint feature representation is subjected to global average pooling, concatenated with the stimulus intensity, and then the ABR automatic classification result is output through a classifier. The steps of extracting temporal local morphological features from the standardized ABR signal sequence through a one-dimensional convolutional layer, performing stimulus intensity conditional modulation through a feature linear modulation module, and adding position embedding to obtain a stimulus-conditional temporal feature representation include: The standardized ABR signal sequence is convolved by a one-dimensional convolutional layer, and the convolutional output is batch normalized and GELU activated to obtain temporal convolutional features. The stimulus intensity value is input into the temporal feature linear modulation network, which contains two layers of multilayer perceptron. The first layer maps the input to the intermediate dimension and applies ReLU activation, and the second layer outputs scaling parameters and offset parameters. The temporal convolutional features are modulated by element-wise affine transformation; By adding learnable location embeddings, we obtain a temporal feature representation of stimulus-conditionalization. The steps of extracting frequency domain spectral features by performing short-time Fourier transform on the standardized ABR signal sequence, performing stimulus intensity conditional modulation through a feature linear modulation module, and adding position embedding to obtain a stimulus-conditional frequency domain feature representation include: The standardized ABR signal sequence is subjected to short-time Fourier transform, the amplitude spectrum is calculated using the Hanning window function, and logarithmic compression is performed. The logarithmic amplitude spectrum is projected onto the feature space through a fully connected layer, and layer normalization is performed to obtain the frequency domain features. The stimulus intensity is input into a frequency domain linear modulation network, which has the same structure as the time domain linear modulation network but with independent parameters, and outputs frequency domain modulation parameters. The frequency domain features are conditionalized by element-level modulation; By adding learnable positional embeddings, we obtain a frequency domain feature representation of stimulus-conditionalization.

2. The method as described in claim 1, characterized in that, The step of inputting the stimulus-conditioned temporal feature representation and the stimulus-conditioned frequency domain feature into a bidirectional cross-attention module for deep fusion to obtain a joint feature representation includes: The temporal feature representation and the frequency domain feature representation of the stimulus conditionation are respectively subjected to layer normalization; Normalized time-domain features are used as queries, and frequency-domain features are used as keys and values. Cross-attention is calculated through a multi-head attention layer, and time-domain fusion features are obtained through residual connections and a feedforward network. Normalized frequency domain features are used as queries, and time domain features are used as keys and values. Cross attention is calculated through a multi-head attention layer, and frequency domain fusion features are obtained through residual connections and feedforward networks. The output time-domain fusion features and frequency-domain fusion features are used as the joint feature representation after bidirectional fusion.

3. The method as described in claim 1, characterized in that, The step of performing global average pooling on the joint feature representation, concatenating it with the stimulus intensity, and then outputting the ABR automatic classification result through a classifier includes: Global average pooling is performed on the time-domain fusion features and the frequency-domain fusion features respectively to obtain feature vectors of fixed dimensions; The time-domain feature vector, frequency-domain feature vector, and original stimulus intensity value are concatenated to form a unified representation vector; The unified representation vector is input into a two-layer multilayer perceptron classifier. The first layer includes fully connected transformation, GELU activation and Dropout, and the second layer outputs a single logit. The probability value is obtained by applying the Sigmoid activation function to the logit, and binary classification is performed using a threshold of 0.5 to output the result of normal hearing or abnormal hearing.

4. The method as described in claim 1, characterized in that, The feature linear modulation network includes a multilayer perceptron, which learns the mapping relationship from stimulus intensity to modulation parameters through layer-by-layer nonlinear transformation; the element-level affine transformation modulates the features through scaling and offset operations.

5. The method as described in claim 1, characterized in that, The short-time Fourier transform uses a window function for frame division, and logarithmic compression is used to reduce the dynamic range; the frequency domain characteristic linear modulation network and the time domain modulation network have completely independent parameters.

6. An automatic classification device for auditory brainstem responses based on a time-frequency cross-attention network, characterized in that, The automatic auditory brainstem response classification device based on a time-frequency cross-attention network includes: The signal processing module is used to perform sample-by-sample Z-score normalization preprocessing on the raw ABR waveform, keeping the stimulus intensity at its original dB SPL value without normalization, to obtain a normalized ABR signal sequence. The temporal modulation module is used to extract temporal local morphological features from the standardized ABR signal sequence through a one-dimensional convolutional layer, perform stimulus intensity conditional modulation through the feature linear modulation module, and add position embedding to obtain a stimulus conditional temporal feature representation. The frequency domain modulation module is used to extract frequency domain spectral features by performing short-time Fourier transform on the standardized ABR signal sequence, perform stimulus intensity conditional modulation through the feature linear modulation module, and add position embedding to obtain a stimulus conditional frequency domain feature representation. The feature union module is used to input the stimulus-conditioned temporal feature representation and the stimulus-conditioned frequency domain feature into the bidirectional cross-attention module for deep fusion to obtain a joint feature representation; The automatic classification result is used to perform global average pooling on the joint feature representation, and after being concatenated with the stimulus intensity, it is output as the ABR automatic classification result through a classifier. The steps of extracting temporal local morphological features from the standardized ABR signal sequence through a one-dimensional convolutional layer, performing stimulus intensity conditional modulation through a feature linear modulation module, and adding position embedding to obtain a stimulus-conditional temporal feature representation include: The standardized ABR signal sequence is convolved by a one-dimensional convolutional layer, and the convolutional output is batch normalized and GELU activated to obtain temporal convolutional features. The stimulus intensity value is input into the temporal feature linear modulation network, which contains two layers of multilayer perceptron. The first layer maps the input to the intermediate dimension and applies ReLU activation, and the second layer outputs scaling parameters and offset parameters. The temporal convolutional features are modulated by element-wise affine transformation; By adding learnable location embeddings, we obtain a temporal feature representation of stimulus-conditionalization. The steps of extracting frequency domain spectral features by performing short-time Fourier transform on the standardized ABR signal sequence, performing stimulus intensity conditional modulation through a feature linear modulation module, and adding position embedding to obtain a stimulus-conditional frequency domain feature representation include: The standardized ABR signal sequence is subjected to a short-time Fourier transform, the amplitude spectrum is calculated using the Hanning window function, and logarithmic compression is performed. The logarithmic amplitude spectrum is projected onto the feature space through a fully connected layer, and layer normalization is performed to obtain the frequency domain features. The stimulus intensity is input into a frequency domain linear modulation network, which has the same structure as the time domain linear modulation network but with independent parameters, and outputs frequency domain modulation parameters. The frequency domain features are conditionalized by element-level modulation; By adding learnable positional embeddings, we obtain a frequency domain feature representation of stimulus-conditionalization.

7. An automatic classification device for auditory brainstem responses based on a time-frequency cross-attention network, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the automatic classification method for auditory brainstem responses based on time-frequency cross-attention networks as described in any one of claims 1 to 5.

8. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the automatic classification method for auditory brainstem responses based on time-frequency cross-attention networks as described in any one of claims 1 to 5.

Images