A task-aware query guidance based chronic obstructive pulmonary disease auxiliary judgment system

Abstract

The present application relates to a kind of task perception query guide-based slow obstructive lung auxiliary judgment system, it relates to medical data auxiliary technical field, the system includes: biological characteristic extraction module is used to extract biological characteristic vector;Audio feature extraction module is used to process the cough audio data of subject to extract audio global feature vector and audio frame level feature sequence;Task perception query generation module is used to fuse biological characteristic vector and audio global feature vector, generates a task perception query vector constrained by lung function index regression task;Cross-modal attention analysis module is used to utilize task perception query vector to carry out time and frequency dimension attention analysis to audio frame level feature sequence, output and pathology related enhanced acoustic characteristics, and according to enhanced acoustic characteristics generate auxiliary judgment chronic obstructive pulmonary disease two classification signals, to solve the technical problem that existing technology is susceptible to acoustic heterogeneity.

Description

Technical Field

[0001] This invention relates to the field of medical data assistance technology, and in particular to a COPD auxiliary diagnosis system based on task-aware query guidance. Background Technology

[0002] Early screening for COPD is crucial for improving patient outcomes. However, pulmonary function tests, which are the gold standard for diagnosis, are difficult to popularize in primary healthcare. This makes automated screening based on cough sounds a promising alternative. However, acoustic heterogeneity caused by individual differences in biomarkers seriously interferes with the model's accurate identification of pathological features.

[0003] Existing studies have attempted to integrate patient biometric data into acoustic diagnostic models, but most of these methods remain at the level of simple feature splicing or shallow fusion, essentially passively combining two types of information. While this can decouple the complex relationship between biometrics and pathological features to some extent, it also introduces noise due to information redundancy. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a COPD auxiliary judgment system based on task-aware query guidance, which solves the technical problem of being susceptible to acoustic heterogeneity in the prior art.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the main technical solutions adopted by the present invention include:

[0008] In a first aspect, embodiments of the present invention provide a COPD auxiliary diagnosis system based on task-aware query guidance, comprising:

[0009] The biometric extraction module is used to generate descriptive text and extract biometric vectors based on the subject's biometric metadata.

[0010] The audio feature extraction module is used to process the subject's cough audio data to extract audio frame-level feature sequences and audio global feature vectors;

[0011] The task-aware query generation module is connected to the biometric feature extraction module and the audio feature extraction module respectively. It is used to fuse the biometric feature vector and the global audio feature vector to generate a task-aware query vector constrained by the lung function index regression task.

[0012] The cross-modal attention analysis module is connected to the audio feature extraction module and the task-aware query generation module, respectively. It is used to perform attention analysis on the audio frame-level feature sequence in terms of time and frequency using the task-aware query vector, output pathologically relevant enhanced acoustic features, and generate a binary classification signal to help judge chronic obstructive pulmonary disease based on the enhanced acoustic features.

[0013] Optionally, the biometric extraction module is also used to combine biometric metadata, including age, gender, height, and weight, into descriptive text according to a structured template, and extract biometric vectors from the descriptive text using a pre-trained text encoder.

[0014] Optionally, the audio feature extraction module is also used to convert cough audio data into a log-Mel spectrogram, segment and expand the log-Mel spectrogram into a patch sequence, process the patch sequence through a pre-trained audio encoder to obtain an audio frame-level feature sequence, and obtain an audio global feature vector based on the audio frame-level feature sequence.

[0015] Optionally, the task-aware query generation module includes:

[0016] The gating unit is used to convert the biometric feature vector into a gating signal through linear projection and nonlinear activation, and to initially suppress individual difference noise in the global audio feature vector by multiplying the gating signal with the global audio feature vector element by element.

[0017] The generation unit, connected to the gating unit, is used to concatenate the global audio feature vector after element-wise multiplication with the biometric feature vector, and then perform a nonlinear transformation through a multilayer perceptron network to generate a task-aware query vector.

[0018] Optionally, the task-aware query generation module further includes a constraint unit connected to the generation unit. The constraint unit constitutes a supervisory mechanism for the generation of task-aware query vectors, and this mechanism is activated only during the system training phase. The constraint unit is also used to input the task-aware query vectors output by the generation unit into a regression prediction head to obtain predicted values ​​for at least one pulmonary function test index, calculate the regression loss between the predicted value and the corresponding true value, and optimize the parameters of the gating unit and the generation unit by backpropagating the regression loss. Through the optimization of the supervisory mechanism, the finally generated task-aware query vector is endowed with clinical semantics that directly characterize the degree of pulmonary function impairment.

[0019] Optionally, the regression prediction head consists of multiple linear layers.

[0020] Optionally, the regression loss is calculated using the mean squared error loss function.

[0021] Optionally, the cross-modal attention analysis module includes:

[0022] The temporal attention branch is used to perform the first cross-attention calculation with the task-aware query vector as the query and the audio frame-level feature sequence as both the key and value. The output of this calculation is then added to the task-aware query vector by residual to obtain the first addition result. The first addition result is then subjected to layer normalization to obtain the temporal dimension enhanced feature.

[0023] The frequency attention branch is used to transpose the audio frame-level feature sequence in the time and feature dimensions to obtain the transposed feature sequence. At the same time, the task-aware query vector is projected through a linear transformation layer to obtain a transformed query vector that matches the dimension of the transposed feature sequence. The second cross-attention calculation is performed with the transformed query vector as the query and the transposed feature sequence as the key and value. The output of this calculation is added to the residual of the transformed query vector to obtain the second sum result. The second sum result is then subjected to layer normalization to obtain the frequency-dimensional enhanced feature.

[0024] The splicing unit is used to splice the time-dimensional enhancement features and the frequency-dimensional enhancement features along the feature dimensions to obtain the enhanced acoustic features.

[0025] In a second aspect, embodiments of the present invention provide an electronic device, including the COPD auxiliary judgment system based on task-aware query guidance as described in any one of the first aspects.

[0026] (III) Beneficial Effects

[0027] The beneficial effects of this invention are:

[0028] This invention proposes a COPD auxiliary judgment system based on task-aware query guidance. Through a task-aware query generation module, biometric vectors and audio features are deeply fused under the implicit constraints of a high-level clinical task (lung function prediction) to generate a task-aware query vector containing clear pathological semantics. This process transforms potentially noisy individual differences into intelligent guidance signals for actively modulating and interpreting audio features, fundamentally suppressing the interference of acoustic heterogeneity on pathological identification. Finally, the system uses this query vector to actively focus on and retrieve the original audio sequence through a cross-modal attention analysis module, accurately enhancing time-frequency patterns related to COPD and filtering irrelevant variations. Ultimately, it outputs a highly reliable risk auxiliary judgment signal, overcoming the information redundancy and noise problems caused by simple feature splicing in traditional methods, and achieving a comprehensive improvement in screening accuracy, robustness, and generalization ability. Attached Figure Description

[0029] Figure 1This illustration shows a schematic diagram of a COPD auxiliary judgment system based on task-aware query guidance provided in an embodiment of this application;

[0030] Figure 2 This illustration shows a specific schematic diagram of a COPD auxiliary judgment system based on task-aware query guidance provided in an embodiment of this application;

[0031] Figure 3 This illustration shows a partial structural diagram of a cross-modal attention analysis module provided in an embodiment of this application. Detailed Implementation

[0032] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] Chronic obstructive pulmonary disease (COPD) is a common, preventable, and treatable chronic respiratory disease characterized by persistent airflow limitation. Without intervention, it can progress to diseases such as pulmonary heart disease and respiratory failure, resulting in extremely high rates of disability and mortality.

[0034] Clinically, a pulmonary function test result with an initial expiratory volume in one second (FEV1 / FVC) <0.7 is considered a necessary condition for diagnosing COPD. However, its application faces significant practical bottlenecks. On the one hand, the availability of pulmonary function testing equipment in primary healthcare institutions is severely insufficient, resulting in a large number of potential patients not being screened in a timely manner. On the other hand, the testing process requires a high degree of patient cooperation, which is often difficult for elderly, frail, or severely ill patients to complete according to standard procedures, thus affecting the accuracy of the diagnostic results.

[0035] In recent years, research on the diagnosis of respiratory diseases based on audio signals such as breath sounds and cough sounds has attracted much attention. Although deep learning has achieved significant success in single-modal audio analysis, the problem of acoustic heterogeneity persists in clinical applications. This means that acoustic signals from different individuals exhibit significant acoustic differences, even under similar pathological conditions. This heterogeneity caused by individual differences severely interferes with the model's accurate identification of core pathological patterns, leading to poor model generalization performance. This problem is particularly prominent when training data is limited, becoming a key bottleneck restricting further improvement in model performance. To mitigate the impact of acoustic heterogeneity, some studies have attempted to integrate patient auxiliary information into acoustic diagnostic models. However, most of these methods remain at the level of simple feature splicing or shallow fusion, essentially passively combining two types of information. While this can decouple the complex relationship between biological and pathological features to some extent, it also introduces noise due to information redundancy.

[0036] Based on this, this invention proposes a COPD auxiliary diagnosis system guided by task-aware query. It generates task-aware query vectors using a task-aware query generation module and innovatively introduces a pulmonary function test index regression auxiliary task to constrain the generation end-to-end. Since pulmonary function testing (PFT) is the gold standard for diagnosing COPD, this constraint mechanism forces the generated task-aware query vector to contain deep semantic information capable of accurately predicting the patient's pulmonary function status. Subsequently, this task-aware query vector carrying clinical prior knowledge is fed into a cross-modal attention analysis module to dynamically guide the model to retrieve, weight, and amplify the acoustic patterns most directly related to pulmonary function impairment from complex raw cough acoustic features. In this way, the model can accurately separate the core acoustic features related to COPD pathology, thereby achieving efficient and accurate screening.

[0037] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0038] Please see Figure 1 , Figure 1 This illustration shows a schematic diagram of a COPD auxiliary judgment system based on task-aware query guidance provided in an embodiment of this application. Specifically, the auxiliary judgment system includes:

[0039] The biometric extraction module 110 is used to generate descriptive text and extract biometric vectors based on the subject's biometric metadata.

[0040] The audio feature extraction module 120 is used to process the subject's cough audio data to extract the global audio feature vector and the audio frame-level feature sequence;

[0041] The task-aware query generation module 130 is connected to the biometric feature extraction module 110 and the audio feature extraction module 120 respectively. It is used to fuse the biometric feature vector and the global audio feature vector to generate a task-aware query vector constrained by the lung function index regression task.

[0042] The cross-modal attention analysis module 140 is connected to the audio feature extraction module 120 and the task-aware query generation module 130, respectively. It is used to perform attention analysis on the audio frame-level feature sequence in terms of time and frequency using the task-aware query vector, output pathologically related enhanced acoustic features, and generate a binary classification signal to assist in the judgment of chronic obstructive pulmonary disease based on the enhanced acoustic features.

[0043] To facilitate understanding of this COPD auxiliary assessment system, specific embodiments are described below.

[0044] Optionally, the biometric extraction module is specifically used to combine biometric metadata, including age, gender, height, and weight, into descriptive text according to a structured template; and to extract biometric vectors from the descriptive text using a pre-trained text encoder.

[0045] For example, audio clips of a subject coughing are collected, and their PFT core indicators and biometrics are recorded simultaneously. The PFT core indicators mainly include forced expiratory volume in one second (FEV1), forced vital capacity (FVC), FEV1 / FVC, predicted FEV1, and predicted FVC, while age, gender, height, and weight are used as biometric metadata. Simultaneously, a COPD diagnostic label given to the subject by a professional physician based on the PFT results and other clinical criteria is obtained, thus constructing a multimodal dataset for COPD screening.

[0046] It should be understood that biometric metadata may include other information besides age, gender, height, and weight, and this application is not limited to this.

[0047] And, as Figure 2 As shown, the biometric extraction module may include a conversion unit and a first encoder unit. The conversion unit can combine biometric metadata, including age, gender, height, and weight, into descriptive text. For example, the descriptive text could be the cough sound of a 58-year-old male patient who is 175cm tall and weighs 70kg. The first encoder unit can extract biometric vectors from the structured descriptive text based on the subject's biometric metadata using a pre-trained GPT-2 text encoder.

[0048] Furthermore, the GPT-2 model originates from the text encoder branch of the pre-trained CLAP model, and the encoded biometric vector F T Represented as:

[0049] F T :[X1,X2,...,X i ,...,X D ]∈R D ;

[0050] In the formula, F T X is a biological feature vector; i R represents the i-th data point, where 1 < i < D; D represents the feature dimension; D This represents a feature vector of dimension D.

[0051] It should be understood that although the above description uses the GPT-2 text encoder as an example, those skilled in the art can replace it with other text encoders according to actual needs, and this application is not limited thereto.

[0052] Optionally, the audio feature extraction module is specifically used to convert cough audio data into a log-Mel spectrogram; segment and expand the log-Mel spectrogram into a patch sequence; process the patch sequence through a pre-trained audio encoder to obtain an audio frame-level feature sequence, and obtain an audio global feature vector based on the audio frame-level feature sequence.

[0053] For example, after acquiring cough audio data, each segment is compressed or stretched to 0.5 seconds to ensure consistent input size.

[0054] And, such as Figure 2 As shown, the audio feature extraction module may include a conversion unit and a second encoder unit. The conversion unit converts cough audio data (e.g., cough sound fragments) into a log-Mel spectrogram. Its core calculations include short-time Fourier transform (STFT) and Mel filtering, and it takes the logarithm of the output energy of all Mel filters to obtain the final log-Mel spectrogram, which is a two-dimensional matrix of size M×N, where M is the number of Mel filters and N is the number of time frames. The second encoder unit divides the log-Mel spectrogram into several patch windows, further splits the patches within each patch window, and expands the patches along each patch window in a left-to-right order to obtain a patch sequence. This patch sequence is then input into the HTS-AT backbone network to obtain the audio frame-level feature sequence F. L Finally, the audio frame-level feature sequence F L After patching and embedding, the global audio feature vector F is obtained. A That is, embedding audio frame-level feature sequences F through patching. L The pathological information is embedded into the audio global feature vector F. A Go to the middle.

[0055] Optionally, the task-aware query generation module includes:

[0056] The gating unit is used to convert the biometric feature vector into a gating signal through linear projection and nonlinear activation, and to initially suppress individual difference noise in the global audio feature vector by multiplying the gating signal with the global audio feature vector element by element.

[0057] The generation unit, connected to the gating unit, is used to concatenate the global audio feature vector after element-wise multiplication with the biometric feature vector, and then perform a nonlinear transformation through a multilayer perceptron network to generate a task-aware query vector.

[0058] Furthermore, the task-aware query generation module also includes a constraint unit connected to the generation unit. The constraint unit constitutes a supervisory mechanism for the generation of task-aware query vectors, and this mechanism is activated only during the system training phase. Specifically, the constraint unit is used to: input the task-aware query vector output by the generation unit into a regression prediction head to obtain a predicted value for at least one pulmonary function test indicator; calculate the regression loss between the predicted value and the corresponding true value; and optimize the parameters of the gating unit and the generation unit by backpropagating the regression loss. Through the optimization of the supervisory mechanism, the finally generated task-aware query vector is endowed with clinical semantics that directly characterize the degree of pulmonary function impairment.

[0059] For example, such as Figure 2 As shown, the gating unit receives the biometric vector F output by the biometric extraction module. T The audio global feature vector F output by the audio feature extraction module A and biological feature vector F T After passing through a linear projection layer and being activated by the Sigmoid function, a set of gated signals between 0 and 1 is generated. This gated signal is then compared with the audio global feature vector F. A Element-wise multiplication is performed to dynamically weight and adjust audio features based on the context of the text description, thereby initially suppressing irrelevant individual difference noise.

[0060] Subsequently, the generation unit combines the gated global audio feature vector with the biometric feature vector F. T Concat concatenation is performed, followed by deep fusion through a multilayer perceptron (MLP) network to ultimately generate the task-aware query vector Q. T .

[0061] Furthermore, to ensure the task-aware query vector Q T To truly characterize a patient's lung function status, rather than simply combining features, the constraint unit introduces a parallel regression task with a task-aware query vector Q. T The generation of this unit is constrained, and it only takes effect during the training phase. Specifically, the generated task-aware query vector Q is... T A dedicated regression head, consisting of multiple linear layers, is fed into the system to predict the actual lung function test indicators, i.e., the generated task-aware query vector Q. T It is fed into a regression head, which predicts a five-dimensional vector V. R Used to fit the true PFT index V PFurthermore, the regression loss uses the mean squared error loss function (MSE Loss) to measure the difference between predicted and true values. During training, this regression loss optimizes the model's parameters through a deep learning backpropagation mechanism, forcing the model to extract deep representations from audio and biological information that can directly predict the degree of lung function impairment. Given that the fundamental basis for the clinical diagnosis of COPD is these quantitative lung function indicators, introducing this auxiliary regression task for direct supervision ensures that the final generated query Q... T It inherently encapsulates high-level clinical semantics that are highly relevant to the pathological state of COPD.

[0062] Optionally, the cross-modal attention analysis module includes:

[0063] The temporal attention branch is used to perform the first cross-attention calculation with the task-aware query vector as the query and the audio frame-level feature sequence as both the key and value. The output of this calculation is then added to the task-aware query vector as a residual, and the result is then subjected to layer normalization to obtain the temporal dimension enhanced features.

[0064] The frequency attention branch is used to transpose the audio frame-level feature sequence in the time and feature dimensions to obtain the transposed feature sequence. At the same time, the task-aware query vector is projected through a linear transformation layer to obtain a transformed query vector that matches the dimension of the transposed feature sequence. The second cross-attention calculation is performed with the transformed query vector as the query and the transposed feature sequence as the key and value. The output of this calculation is added to the residual of the transformed query vector, and the addition result is subjected to layer normalization to obtain the frequency-dimensional enhanced feature.

[0065] The splicing module is used to splice the time-dimensional enhancement features and the frequency-dimensional enhancement features along the feature dimensions to obtain enhanced acoustic features.

[0066] For example, such as Figure 3 As shown, this cross-modal attention analysis module can analyze the audio frame-level feature sequence F from two orthogonal and complementary dimensions: time and frequency. L The analysis was conducted. Considering the complexity of diagnostic clues for pathological cough sounds, which exist not only in their temporal dynamics but also within their spectral characteristics, the acoustic analysis unit employed methods such as... Figure 3 The parallel multi-head cross-attention structure is shown. Specifically,

[0067] In the temporal attention branch, the task-aware query vector Q T As a query, the audio frame-level feature sequence F LThe first cross-attention calculation is performed, using the key and value as the key and value. To prevent the loss of key summary information, which is composed of audio and text features carried by the query itself, during the attention process, the output of this calculation is compared with the task-aware query vector Q. T The residuals are summed to obtain the first sum result. Then, the first sum result is subjected to layer normalization to obtain the time dimension enhancement feature. The main function of this branch is to identify and weight acoustic events along the time axis, so that the model can focus on time patterns with diagnostic significance.

[0068] In the frequency attention branch, the audio frame-level feature sequence F L Transpose the sequence in both the time and feature dimensions to obtain the transposed feature sequence F'. L For example, the audio frame-level feature sequence F L Originally, it was a two-dimensional matrix, with the first dimension representing the time steps (T steps in total) and the second dimension representing the feature channels (D dimensions in total), with a shape represented as [T,D]. The transpose operation swapped it to [D,T], meaning the first dimension became the feature channels and the second dimension became the time steps. And, the task-aware query vector Q... T The transposed feature sequence F' is obtained through a linear layer. L Transform query vector Q' that matches the dimension T and transform the query vector Q' T As a query, the transposed feature sequence F' L The second cross-attention computation is performed, using the key and value as the key and value, respectively. This design enables the model to directly identify and aggregate high-level acoustic feature combinations that are diagnostically significant throughout the entire event, regardless of their specific temporal location. To prevent the loss of key summary information, composed of both audio and textual features, carried by the query itself during the attention process, the output of the second cross-attention computation is combined with the transformed query vector Q'. T The residuals are summed to obtain a second sum, which is then subjected to layer normalization to obtain frequency-dimensional enhanced features. This design ensures that the final output feature vector contains both the acoustically enhanced pathological patterns and fully preserves the global diagnostic context defined by the bimodal information and the regression task.

[0069] Subsequently, the splicing unit concatenates the two obtained features along the feature dimension to form the final enhanced acoustic feature F. I .

[0070] Subsequently, the enhanced acoustic features F can be used I The input is fed into a binary classifier, which outputs binary classification signals for COPD and non-COPD.

[0071] Furthermore, during the model training phase of the auxiliary judgment system, an end-to-end joint optimization strategy is adopted to learn the network parameters of trainable parts such as the task-aware query generation module and the cross-modal attention analysis module as a whole. During training, the AdamW optimizer can be used to optimize the network parameters, with a batch size of 32. The total number of training epochs can be 100, and the learning rate scheduling can employ a strategy that includes warm-up and decay: for example, warming up the learning rate in the first three training epochs, setting it to 0.05, 0.1, and 0.2 respectively; then using step decay, for example, halving the learning rate every 10 epochs until it reaches 0.05 and remains constant, to facilitate model convergence. The core optimization objective is achieved through a custom composite loss function, which is a weighted sum of the binary cross-entropy loss of the main classification task and the mean squared error loss of the auxiliary lung function index regression task. Its general form can be expressed as L... total =(1- )×L BCE + ×L MSE L total Indicates total loss; L BCE L represents the binary cross-entropy loss for the main classification task; MSE This represents the mean squared error loss in the regression task for assisted pulmonary function indicators. It is a hyperparameter used to balance the weights of the two losses, in this specific embodiment It can be set to 0.2. Therefore, through the backpropagation of this composite loss, the lung function regression loss (i.e., L...) MSE ) is used to directly constrain and optimize the generation of task-aware query vectors, making them contain clinical semantics; main classification loss (i.e., L BCE This simultaneously optimizes the classification decisions of the entire system. This training mechanism collectively drives the system to learn a deep fusion representation that can suppress acoustic heterogeneity and accurately focus on pathology-related acoustic patterns.

[0072] The above parameters include batch size, learning rate, training epochs, decay period, and hyperparameters. The specific values ​​can be adjusted according to the actual training data and computing resources. The values ​​given in this embodiment are only one feasible example.

[0073] Therefore, by utilizing the aforementioned technical solution, this embodiment of the application, through the design of a task-aware query generation module and the introduction of PFT index regression as an auxiliary task, successfully transforms static biometric data into a dynamic query containing advanced clinical semantics. This query is then fed into a cross-modal attention analysis module to proactively guide the model to accurately retrieve and amplify the acoustic patterns most relevant to the pathological state from complex and noisy acoustic signals. This successfully overcomes the challenges posed by acoustic heterogeneity and provides a novel paradigm for addressing multimodal fusion and heterogeneity issues in biomedical signal analysis.

[0074] It should be understood that the above-described COPD auxiliary judgment system based on task-aware query guidance is merely an example. Those skilled in the art can make various modifications based on the above system, and the modified solutions also fall within the protection scope of this application.

[0075] In addition, embodiments of the present invention also provide an electronic device, which may include a COPD auxiliary judgment system based on task-aware query guidance.

[0076] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0077] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0078] It should be noted that any reference numerals placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims that enumerate several means, several of these means may be embodied by the same hardware. The use of the terms first, second, third, etc., is merely for convenience of expression and does not indicate any order. These terms can be understood as part of the component names.

[0079] Furthermore, it should be noted that in the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0080] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the claims should be interpreted to include both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0081] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, then this invention should also include these modifications and variations.

Claims

1. A COPD auxiliary diagnosis system based on task-aware query guidance, characterized in that, include: The biometric extraction module is used to generate descriptive text and extract biometric vectors based on the subject's biometric metadata. An audio feature extraction module is used to process the subject's cough audio data to extract audio frame-level feature sequences and audio global feature vectors; The task-aware query generation module is connected to the biometric extraction module and the audio feature extraction module respectively. It is used to fuse the biometric vector and the global audio feature vector to generate a task-aware query vector constrained by the lung function index regression task. The cross-modal attention analysis module is connected to the audio feature extraction module and the task-aware query generation module, respectively. It is used to perform attention analysis on the audio frame-level feature sequence in terms of time and frequency using the task-aware query vector, output pathologically related enhanced acoustic features, and generate a binary classification signal to assist in the judgment of chronic obstructive pulmonary disease based on the enhanced acoustic features. The task-aware query generation module includes: The gating unit is used to convert the biometric vector into a gating signal through linear projection and nonlinear activation, and to initially suppress individual difference noise in the audio global feature vector by multiplying the gating signal with the audio global feature vector element by element. A generation unit, connected to the gating unit, is used to concatenate the audio global feature vector processed by the element-wise multiplication operation with the biological feature vector, and perform a nonlinear transformation through a multilayer perceptron network to generate the task-aware query vector. The task-aware query generation module also includes a constraint unit connected to the generation unit. The constraint unit constitutes a supervisory mechanism for the generation of the task-aware query vector, and this mechanism is activated only during the system training phase. The constraint unit is also used to input the task-aware query vector output by the generation unit into a regression prediction head to obtain a predicted value for at least one pulmonary function test indicator, calculate the regression loss between the predicted value and the corresponding true value, and optimize the parameters of the gating unit and the generation unit by backpropagating the regression loss. Through the optimization of the supervisory mechanism, the finally generated task-aware query vector is endowed with clinical semantics that directly characterizes the degree of pulmonary function impairment. The cross-modal attention analysis module includes: The temporal attention branch is used to perform a first cross-attention calculation with the task-aware query vector as the query and the audio frame-level feature sequence as both the key and the value. The output of the calculation is then added to the task-aware query vector by residual to obtain a first addition result. The first addition result is then subjected to layer normalization to obtain the temporal dimension enhanced feature. The frequency attention branch is used to transpose the audio frame-level feature sequence in the time and feature dimensions to obtain a transposed feature sequence. Simultaneously, the task-aware query vector is projected through a linear transformation layer to obtain a transformed query vector that matches the dimension of the transposed feature sequence. The transformed query vector is used as the query, and the transposed feature sequence is used as the key and value to perform a second cross-attention calculation. The output of this calculation is added to the transformed query vector by residual addition to obtain a second addition result. The second addition result is then subjected to layer normalization to obtain the frequency-dimensional enhanced feature. The splicing unit is used to splice the time-dimensional enhancement feature and the frequency-dimensional enhancement feature along the feature dimension to obtain the enhanced acoustic feature.

2. The COPD auxiliary judgment system according to claim 1, characterized in that, The biometric extraction module is further configured to combine biometric metadata, including age, gender, height, and weight, into descriptive text according to a structured template, and extract the biometric vector from the descriptive text using a pre-trained text encoder.

3. The COPD auxiliary judgment system according to claim 1, characterized in that, The audio feature extraction module is further configured to convert the cough audio data into a log-Mel spectrogram, segment and expand the log-Mel spectrogram into a patch sequence, process the patch sequence through a pre-trained audio encoder to obtain an audio frame-level feature sequence, and obtain an audio global feature vector based on the audio frame-level feature sequence.

4. The COPD auxiliary judgment system according to claim 1, characterized in that, The regression prediction head consists of multiple linear layers.

5. The COPD auxiliary judgment system according to claim 1, characterized in that, The regression loss is calculated using the mean squared error loss function.

6. The auxiliary judgment system according to claim 1, characterized in that, Both the first cross-attention calculation and the second cross-attention calculation are implemented using a multi-head attention mechanism.

7. An electronic device, characterized in that, The COPD auxiliary judgment system based on task-aware query guidance as described in any one of claims 1 to 6.

Images