A method and system for auxiliary identification of respiratory lung sounds for clinical care

Abstract

The present application relates to the technical field of respiratory lung sound recognition, and particularly relates to a respiratory lung sound auxiliary recognition method and system for clinical nursing. The present application provides a respiratory lung sound auxiliary recognition method and system for clinical nursing, which solves the technical problems of data scarcity, complex noise interference and insufficient cross-device generalization capability in the lung sound signal through generative data enhancement and cross-modal migration, combines self-supervised contrast learning and causal feature discovery, realizes multi-source signal feature fusion and noise robustness improvement, dynamically adapts to real-time noise environment, breaks through the limitations of traditional methods which rely on labeled data and are easily affected by mixed noise in feature extraction, and constructs a full-link closed-loop system from synthetic data generation, cross-modal alignment to causal driving decision, effectively improving the accuracy of clinical lung sound analysis and clinical applicability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of respiratory lung sound recognition, and particularly relates to a respiratory lung sound auxiliary recognition method and system for clinical nursing. BACKGROUND

[0002] The lung sound signal generally refers to the respiratory sound collected by a stethoscope, such as normal respiratory sound, dry and wet rales and the like. Changes in these sounds can reflect lung diseases such as pneumonia, asthma, COPD and the like, so a method for identifying these features should involve signal processing and pattern recognition.

[0003] In clinical lung sound analysis, existing methods face core challenges such as data scarcity, complex noise interference and insufficient cross-device generalization ability. Traditional models relying on labeled data are difficult to cover rare disease characteristics and are easily confused by environmental noise. Although single generative data augmentation technology can expand data diversity, it may introduce non-physiological features. Cross-modal transfer can supplement multi-source information, but lacks deep analysis of causal relationships. Self-supervised contrast learning can mine the internal patterns of unlabeled data, but cannot effectively eliminate the influence of mixed noise. Causal feature discovery can improve feature interpretability, but is limited by insufficient labeled data and fuzzy intervention variable definition. SUMMARY

[0004] The present application provides a respiratory lung sound auxiliary recognition method and system for clinical nursing, which solves the technical problems of data scarcity, complex noise interference and insufficient cross-device generalization ability in lung sound signals through generative data augmentation and cross-modal transfer. Combined with self-supervised contrast learning and causal feature discovery, it realizes multi-source signal feature fusion and noise robustness improvement, dynamically adapts to real-time noise environment, breaks through the limitations of traditional methods relying on labeled data and feature extraction being easily affected by mixed noise, and constructs a full-link closed-loop system from synthetic data generation, cross-modal alignment to causal driven decision-making, effectively improving the accuracy and clinical applicability of clinical lung sound analysis.

[0005] To achieve the above purpose, the embodiments of the present application disclose the following technical solutions:

[0006] On the one hand, the present application discloses a respiratory lung sound auxiliary recognition method for clinical nursing, comprising the following steps: acquiring original lung sound signals, respiratory signals, CT image features, historical low-noise period distribution features and patient anatomical features, generating synthetic lung sound data through a conditional generative adversarial network, covering rare disease characteristics and noise combination scenarios.

[0007] In the embodiment of the scheme, the synthetic lung sound data comprises: the conditional generative adversarial network is inputted with disease types, noise types, and patient anatomical characteristics as conditions; the synthetic lung sound data verifies its physiological rationality through an anatomical constraint loss function, and the anatomical constraint loss function is constructed based on a mapping relationship between patient anatomical characteristics and lung sound signal frequency domain characteristics.

[0008] The synthetic lung sound data, the original lung sound signal, the respiratory signal, and the CT image feature are inputted into a cross-modal transfer module to construct a multi-source feature space.

[0009] In the embodiment of the scheme, the construction of the multi-source feature space comprises: aligning the time sequence characteristics of the lung sound signal and the respiratory signal; aligning the lung sound signal and the CT image feature; and the multi-source feature space comprises a time domain lung sound waveform, a frequency domain mel spectrum, a respiratory signal time sequence characteristic, and a CT image lesion region mask.

[0010] Self-supervised contrastive learning pre-training is performed in the multi-source feature space to generate a robust feature vector.

[0011] In the embodiment of the scheme, the loss function of the self-supervised contrastive learning pre-training comprises: an anatomical consistency loss function, which constrains the consistency of the generated lung sound data and the patient anatomical characteristics; and a noise robustness contrast loss function, which enhances the robustness of the model to environmental noise by introducing a noise sample weight coefficient.

[0012] A causal relationship between lung sound features and diseases and noise is constructed based on a causal graph model to screen a causal feature subset.

[0013] In the embodiment of the scheme, the screening of the causal feature subset comprises: constructing a causal graph comprising lung sound features, disease labels, and noise types; applying a virtual intervention to a feature vector, calculating an average causal effect of the feature and the disease, and removing non-causal features with an effect value lower than a preset threshold.

[0014] In the embodiment of the scheme, the virtual intervention comprises: defining an intervention variable as a key frequency band energy value in the lung sound feature; calculating a potential result difference after the intervention, and screening features with a causal effect significance higher than a statistical threshold.

[0015] According to historical low-noise period distribution characteristics and a real-time noise influence coefficient, model parameters of the causal feature subset are dynamically adjusted to output a final lung sound analysis result.

[0016] In the embodiment of the scheme, the loss dynamically adjusts the model parameters, which comprises: calculating a model weight decay factor according to the real-time noise influence coefficient; and updating the model parameters through online learning to optimize the weighted balance of the classification loss and the noise loss in the objective function.

[0017] In another aspect, the present scheme discloses a respiratory lung sound auxiliary recognition system for clinical care, comprising: a generative data enhancement module: an integrated conditional generative adversarial network generates synthetic lung sound data conforming to anatomical constraints;

[0018] A cross-modal migration module: aligns the multi-source features of lung sound, respiratory signal and CT image;

[0019] A self-supervised contrast learning module: a noise-robust contrast loss function is used to pretrain the model;

[0020] A causal inference module: a causal graph model is used to filter a causal feature subset;

[0021] A dynamic adaptation module: dynamically adjusts the model parameters according to the real-time noise influence coefficient;

[0022] The output of the generative data enhancement module is processed by the cross-modal migration module, and then input into the self-supervised contrast learning module and the causal inference module, and finally the analysis result is output by the dynamic adaptation module.

[0023] In an embodiment of the present scheme, the generative data enhancement module comprises:

[0024] An anatomical constraint loss calculation unit verifies the physiological reasonableness of the generated data through the mapping relationship between the patient's anatomical features and the frequency domain features of the lung sound signal;

[0025] A noise superposition unit superimposes monitor alarm sound, personnel conversation sound and muscle contraction noise in the synthetic lung sound data, and the superimposition process is based on a time domain masking mechanism to avoid timing conflicts.

[0026] The present scheme effectively solves the dual problems of data scarcity and noise interference in traditional methods by combining generative adversarial networks with causal inference, proposes a spatio-temporal attention-driven cross-modal alignment framework, breaks through the difficulty of feature fusion of multi-source medical signals, and designs a noise dynamic adaptation mechanism to realize model self-optimization in a clinical complex environment. In addition to improving rare disease recognition ability, enhancing noise robustness, and improving multi-modal fusion accuracy, the self-adaptive learning based on historical data reduces the need for manual annotation, is suitable for real-time bedside monitoring scenarios, and has significant technical innovation and clinical practical value. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 The method flowchart of the first embodiment of the present scheme;

[0028] Figure 2 The overall block diagram of the second embodiment system of the present scheme;

[0029] Figure 3 The interaction diagram of the generative data enhancement module of the second embodiment system of the present scheme;

[0030] Figure 4 Interaction diagram of the cross-modal transfer module of the embodiment two system of the present application;

[0031] Figure 5 Interaction diagram of the self-supervised contrastive learning module of the embodiment two system of the present application;

[0032] Figure 6 Interaction diagram of the causal inference module of the embodiment two system of the present application;

[0033] Figure 7 Interaction diagram of the dynamic adaptation module of the embodiment two system of the present application. DETAILED DESCRIPTION

[0034] Reference will now be made in detail to the present application. While the application will be described in conjunction with these specific embodiments, it will be understood that they are not intended to limit the application to these specific embodiments. On the contrary, the application is intended to cover alternatives, modifications, and equivalents, which can be included within the spirit and scope of the application as defined by the claims. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. The present application can be practiced without some or all of these specific details. In other instances, well known process operations have not been described in detail in order not to unnecessarily obscure the present application.

[0035] The singular forms "a," "an," and "the" include plural referents unless the context clearly dictates otherwise. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0036] SUMMARY OF THE APPLICATION: In the prior art, clinical lung sound analysis relies on models of labeled data, which is difficult to cover rare disease characteristics and is easily confused by environmental noise, while single generative data augmentation technology can expand data diversity, but may introduce non-physiological characteristics, cross-modal transfer can supplement multi-source information but lacks deep analysis of causal relationships, while self-supervised contrastive learning can mine the internal patterns of unlabeled data but cannot effectively eliminate the influence of mixed noise, and causal feature discovery can improve feature interpretability but is limited by insufficient labeled data and fuzzy intervention variable definition.

[0037] To solve the above problems, the scheme uses conditional generative adversarial network combined with anatomical constraints to generate synthetic data covering rare disease scenarios, integrates the timing and spatial features of lung sound, respiratory signal and CT image through a cross-modal alignment module, pre-trains robust representations using noise-robust contrastive learning, filters out mixed noise by screening key features with a causal graph model, and finally adjusts the model parameters of the causal feature subset in real time to adapt to changes in environmental noise through a dynamic parameter adjustment module, forming a full-link solution from data generation to decision optimization, thereby breaking the dependence of traditional methods on labeled data and improving the clinical applicability in complex scenarios.

[0038] Embodiment one

[0039] A respiratory lung sound auxiliary identification method for clinical care, comprising the following steps:

[0040] Obtain the original lung sound signal, respiratory signal, CT image feature, historical low-noise period distribution feature and patient anatomy feature, generate synthetic lung sound data covering rare disease characteristics and noise combination scenarios through conditional generative adversarial network; input the synthetic lung sound data, the original lung sound signal, the respiratory signal and the CT image feature into the cross-modal transfer module to construct a multi-source feature space; perform self-supervised contrastive learning pre-training in the multi-source feature space to generate a robust feature vector; construct the causal relationship between lung sound features and diseases, noise based on a causal graph model, and filter a causal feature subset; dynamically adjust the model parameters of the causal feature subset according to the historical low-noise period distribution feature and the real-time noise influence coefficient, and output the final lung sound analysis result.

[0041] In the present embodiment, the original lung sound signal and the respiratory signal are synchronously collected by a multi-channel biosensor array, combined with the bronchial wall thickness and lung parenchyma density extracted from high-resolution CT images, and integrated with the historical low-noise period distribution characteristics recorded in the long-term electronic health records of the patient. A synthetic lung sound dataset containing rare pathological rales, crackles and environmental noise interference combinations is constructed using a conditional generative adversarial network, breaking through the sample imbalance limitation in real data collection. The generated synthetic lung sound data and the original multi-modal physiological signals are input into a cross-modal migration module, the time-frequency domain alignment algorithm is used to fuse different sampling rate feature streams, the graph convolutional neural network is used to construct a multi-source feature space containing time convolution layers and modal interaction attention mechanisms, and the semantic space alignment of lung sound waveform features, respiratory mechanics parameters, image texture features and anatomical structure parameters is realized. Self-supervised contrastive learning pre-training is performed in the multi-source feature space based on negative sample sampling, the disease-specific feature vector with noise robustness is generated by comparing the deep feature representation of normal and abnormal pathological patterns, and the fault tolerance of the model to baseline drift, motion artifacts and other interference is enhanced. Based on the structural causal model, a causal graph is constructed between the lung sound features and the types of respiratory diseases and environmental noise sources, and a causal discovery algorithm is used to select a strong causal feature subset that directly affects disease diagnosis, eliminating false associations caused by confounding factors. According to the circadian rhythm pattern extracted from the historical low-noise period distribution characteristics and the real-time noise influence coefficient, the model weight parameters of the causal feature subset are dynamically adjusted, the real-time monitoring data is fused through an online incremental learning mechanism, and the final lung sound analysis results containing disease risk probability, noise interference confidence and anatomical structure abnormality index are output, realizing an end-to-end closed-loop diagnosis link from original signal collection to clinical decision support.

[0042] The application further proposes a method for generating synthetic lung sound data, which comprises: the conditional generative adversarial network is inputted with disease type, noise type and patient anatomy feature as conditions; the synthetic lung sound data is verified for its physiological rationality by an anatomy constraint loss function, and the anatomy constraint loss function is constructed based on the mapping relationship between patient anatomy feature and lung sound signal frequency domain feature. In the embodiment, a multi-modal input fusion architecture is constructed by the conditional generative adversarial network, the disease type parameter is converted into a pathological feature encoding vector, the noise type parameter is mapped into a time-frequency domain interference mode, the patient anatomy feature is generated into a multi-scale anatomy mask corresponding to lung lobe partition through a three-dimensional space interpolation algorithm, and the dynamic coupling of pathological features, noise mode and anatomy structure is realized in the generator network through a cross-channel attention mechanism; the mapping supervision mechanism of generated data and real physiological features is constructed by using the anatomy constraint loss function, the fundamental frequency, harmonic distribution and energy attenuation features of the lung sound signal are extracted through frequency domain power spectrum analysis, the frequency domain-anatomy joint loss function is constructed by combining the lung volume parameter, bronchial bifurcation angle and alveolar elastic coefficient in the patient anatomy feature, the generator output is optimized by using the adversarial training to improve the matching degree of the frequency energy concentration area of the synthetic lung sound and the anatomy feature, and a dynamic noise injection module is introduced to simulate the time-varying superposition characteristics of environmental noise such as chest wall friction and electronic device interference, the propagation attenuation consistency of the synthetic lung sound under the bronchial tree structure is verified through the multi-scale feature fusion module in the anatomy constraint loss function, and it is ensured that the frequency spectrum distortion features of the generated pathological rhonchi in the bronchiole and alveolar region meet the real physiological conduction rules.

[0043] The application further proposes a method for constructing a multi-source feature space, which comprises: aligning the time sequence features of the lung sound signal and the respiratory signal; aligning the lung sound signal and the CT image feature; and the multi-source feature space comprises time domain lung sound waveform, frequency domain mel spectrum, respiratory signal time sequence feature and CT image lesion area mask.

[0044] In the embodiment, the time sequence characteristics of the lung sound signal and the respiratory signal are aligned by a dynamic time warping algorithm, the non-stationary lung sound waveform is mapped to a time-frequency domain by a short-time Fourier transform and a phase synchronization mechanism is established with the respiratory flow rate fluctuation curve, cross-modal time sequence feature fusion is realized in a dual-flow graph attention network; a three-dimensional convolutional neural network is used to extract a lesion region mask of the CT image, the lung sound frequency domain mel spectrum and the image texture features are spatially aligned by an anatomic constraint cross-scale feature pyramid, and an adversarial domain adaptation module is used to eliminate the distribution difference between different imaging modalities; the short-time energy envelope of the time domain lung sound waveform, the critical band energy distribution of the frequency domain mel spectrum, the flow rate-volume loop features of the respiratory signal time sequence, and the morphological parameters of the CT image lesion region mask are integrated in the multi-source feature space, a time sequence feature memory module is constructed by a gated recurrent unit, a multi-head self-attention mechanism is used to establish a hidden space association between the lung sound pathological features and the anatomical structure abnormalities in the cross-modal transfer module, the feature decoupling ability of the multi-source feature space is optimized by adversarial training, and cross-modal collaborative representation of physiological signals and image features is realized.

[0045] The application further proposes a loss function of self-supervised contrastive learning pre-training, including: an anatomic consistency loss function, which constrains the consistency of generated lung sound data and patient anatomic features; a noise robustness contrast loss function, which enhances the robustness of the model to environmental noise by introducing a noise sample weight coefficient.

[0046] In the embodiment, the loss function of self-supervised contrastive learning pre-training includes: an anatomic consistency loss function, which constrains the cross-modal mapping between the generated lung sound frequency domain features and the patient anatomic features, and uses adversarial training to optimize the output of the generator to greatly improve the matching degree of the bronchial resonance peak distribution of the synthesized lung sound with the airway morphological parameters in the three-dimensional CT image, and introduces a dynamic noise injection module to simulate the time-varying superposition characteristics of environmental noise such as chest wall friction and electronic device interference; a noise robustness contrast loss function, which constructs positive and negative sample pairs containing feature representations under normal breathing mode and noise pollution mode, uses a ternary loss function to enhance the robustness of the model to environmental noise, and aligns the pathological feature distribution under different noise intensities in the feature space; a noise robustness contrast loss function, which dynamically adjusts the contrast learning target by introducing a noise sample weight coefficient, gradually increases the noise sample weight proportion by using a curriculum learning strategy, constructs a noise-aware feature decoupling mechanism in the cross-modal transfer module, optimizes the feature decoupling ability of the multi-source feature space by adversarial training, and realizes cross-modal collaborative representation of physiological signals and image features.

[0047] The application further proposes that the noise robustness contrast loss function is defined as:

[0048] ;

[0049] in, This represents the feature vector of the current sample, i.e., the lung sound signal feature representation extracted by the model; Indicates and Feature vectors of the same category refer to the feature representations of similar breathing patterns or the same disease type in the same patient. Indicates and Feature vectors of different categories refer to the characteristic representations of different disease types or normal breathing patterns; Temperature is a hyperparameter used to control the discriminative power of feature distribution, affecting the degree of clustering of positive and negative samples in the feature space. represents the feature similarity calculation function, using cosine similarity or other distance metrics; N represents the number of negative samples, the number of different class samples used for comparison in contrastive learning; k represents the index variable, which iterates through all negative samples from 1 to N in the summation. The noise weighting coefficient dynamically adjusts the influence of noisy samples in the loss function. Dynamically adjusted via real-time noise power spectral density; The noise sample features are the feature representations extracted from environmental noise or interference signals, and the feature of the k-th noise sample is represented.

[0050] In this embodiment, a noise-robust contrastive loss function is proposed. This function constructs a contrastive learning framework containing positive and negative sample pairs to align the distribution of pathological features between normal breathing patterns and noise-polluted patterns in the feature space. The noise weight coefficient λ is dynamically adjusted based on the real-time noise power spectral density, enabling the model to adaptively enhance its ability to suppress high-frequency noise interference during training. This loss function incorporates noise sample features. The joint optimization mechanism with anatomical constraints superimposes a similarity penalty term for noise samples into the denominator and combines the temperature parameter τ to control the discriminative power of the feature distribution. This allows the model to capture disease-specific features while suppressing the propagation of environmental noise in the feature space through dynamically adjusted noise weight coefficient λ, thereby enhancing the model's robustness to complex noises such as monitor alarms and conversations.

[0051] This application further proposes a method for screening a subset of causal features, including: constructing a causal graph containing lung sound features, disease labels, and noise types; applying virtual intervention to the feature vectors; calculating the average causal effect between the features and the disease; and removing non-causal features with effect values ​​below a preset threshold.

[0052] In this embodiment, a dynamic causal graph model of lung sound time-frequency features, disease progression status and noise interference types is constructed, a structural equation model is used to describe the nonlinear causal relationship between the features, the interference of confounding factors is adjusted through the backdoor criterion, and the propagation path of feature intervention is simulated under the Bayesian network framework; a virtual intervention operation is applied to the feature vector, the causal effect strength of the feature on disease classification is calculated through do-calculus, the sample distribution after feature disturbance is generated by counterfactual reasoning, and the marginal contribution of each feature to the prediction result is quantified by Shapley value; a feature subset is dynamically selected based on the causal effect threshold, redundant features with a causal effect lower than a preset threshold are removed through permutation importance evaluation, the feature selection process is optimized using a regularization path algorithm, a minimal causal feature closure is constructed using a Markov blanket algorithm, and the selected lung sound feature subset is independent of the noise interference source and retains the disease-specific causal association path. Through a dynamic threshold adjustment mechanism, the feature distribution difference of different patient groups is adapted, and diagnostic features with causal power are accurately extracted from high-dimensional multi-modal data.

[0053] The application further proposes a virtual intervention method, which comprises: defining the intervention variable as the key frequency band energy value in the lung sound feature; calculating the potential result difference after intervention, and screening the features with a causal effect significance higher than a statistical threshold.

[0054] In this embodiment, the key frequency band energy value in the lung sound frequency domain feature is defined as the intervention variable through a structural causal model, the potential result distribution difference after intervention is calculated using do-calculus, the causal effect strength of the feature on disease classification is quantified using a non-parametric statistical test, the statistical significance threshold is dynamically adjusted through a sliding window, and a lung sound feature subset with stable causal effect is screened out by combining multiple hypothesis testing correction.

[0055] The application further proposes a method for dynamically adjusting model parameters, which comprises: calculating a model weight attenuation factor according to a real-time noise influence coefficient; and updating the model parameters through online learning to optimize the weighted balance of the classification loss and the noise loss in the objective function.

[0056] In this embodiment, the noise influence coefficient is estimated in real time through a Kalman filter, the model weight attenuation factor is calculated by combining an exponential weighted moving average, the weight coefficient of the classification loss and the noise loss is dynamically adjusted using an LSTM network, the momentum optimization algorithm is introduced when updating the model parameters to accelerate convergence, the balance coefficient of the classification error and the noise robustness index in the loss function is dynamically adjusted through reinforcement learning, and the time sequence dependence of parameter updating is optimized by combining an adaptive learning rate mechanism.

[0057] Embodiment two

[0058] A respiratory lung sound auxiliary recognition system for clinical nursing, comprising:

[0059] The generative data augmentation module is a conditional generative adversarial network that generates synthetic lung sound data that meets anatomical constraints.

[0060] The cross-modal transfer module aligns multi-source features of lung sound, respiratory signal, and CT image.

[0061] The self-supervised contrastive learning module uses a noise-robust contrastive loss function to pre-train the model.

[0062] The causal inference module filters a causal feature subset based on a causal graph model.

[0063] The dynamic adaptation module dynamically adjusts model parameters according to real-time noise influence coefficients.

[0064] The output of the generative data augmentation module is processed by the cross-modal transfer module, and then input into the self-supervised contrastive learning module and the causal inference module, and finally output by the dynamic adaptation module. The generative data augmentation module includes:

[0065] The anatomical constraint loss calculation unit verifies the physiological reasonableness of the generated data through the mapping relationship between the patient's anatomical features and the frequency domain features of the lung sound signal.

[0066] The noise superposition unit superimposes monitor alarm sound, human conversation sound, and muscle contraction noise in the synthetic lung sound data. The superimposition process is based on a time domain masking mechanism to avoid timing conflicts.

[0067] The cross-modal transfer module aligns the timing features of lung sound and respiratory signal through dynamic time warping algorithm, uses short-time Fourier transform to realize phase synchronization of lung sound frequency domain features and respiratory flow rate curve, and completes cross-modal feature fusion in a double-flow graph attention network. The self-supervised contrastive learning module constructs an anatomical consistency loss function to constrain the matching degree of the generated data and the patient's anatomical features, and introduces a noise robustness contrastive loss function to enhance the model's ability to distinguish monitor alarm sound and electronic interference noise. The causal inference module establishes a dynamic causal graph model containing lung sound time-frequency features, disease labels, and noise types, describes the nonlinear causal relationship between features through a structural equation model, and uses Do-calculus to calculate the potential result difference after feature intervention. The dynamic adaptation module uses Kalman filter to estimate the noise power spectral density in real time, combines with the exponential weighted moving average to calculate the model weight decay factor, dynamically adjusts the weight coefficient of classification loss and noise robustness loss through LSTM network, and introduces a momentum optimization algorithm to accelerate convergence when updating parameters online.

[0068] It should be pointed out finally that the above examples are only used to illustrate the technical solutions of the present application but not to limit it; although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or some technical features can be replaced by equivalent ones without departing from the spirit of the technical solutions of the present application, and all of them should be covered in the technical solution range claimed by the present application.

Claims

1. A method for assisting in the recognition of respiratory lung sounds in clinical nursing, characterized in that, Includes the following steps: Step S1: Obtain raw lung sound signals, respiratory signals, CT image features, historical low-noise period distribution features, and patient anatomical features. Generate synthetic lung sound data through a conditional generative adversarial network to cover rare disease features and noise combination scenarios. Step S2: Input the synthesized lung sound data, the original lung sound signal, respiratory signal, and CT image features into the cross-modal transfer module to construct a multi-source feature space; Step S3: Perform self-supervised contrastive learning pre-training in the multi-source feature space to generate robust feature vectors; Step S4: Construct the causal relationship between lung sound features and disease and noise based on the causal graph model, and filter the subset of causal features; Step S5: Based on the historical low-noise period distribution characteristics and the real-time noise impact coefficient, dynamically adjust the model parameters of the causal feature subset and output the final lung sound analysis results; The loss function for self-supervised contrastive learning pre-training in step S3 includes: Anatomical consistency loss function constrains the consistency between generated lung sound data and patient anatomical features; The noise robustness contrast loss function enhances the model's robustness to environmental noise by introducing noise sample weight coefficients. The noise robustness contrast loss function is defined as follows: ； in, This represents the feature vector of the current sample; Indicates and Feature vectors of the same category; Indicates and Feature vectors of different categories; For temperature parameters; This represents the feature similarity calculation function; N represents the number of negative samples; Noise weighting coefficient; The noise sample features are represented by k; k represents the index variable; the noise weight coefficient is... Dynamic adjustment is achieved through real-time noise power spectral density.

2. The method for assisting in the recognition of respiratory sounds in clinical nursing according to claim 1, characterized in that, In step S1, the method for generating synthetic lung sound data includes: The conditional generative adversarial network takes disease type, noise type, and patient anatomical features as conditional inputs. The synthetic lung sound data were validated for physiological rationality using an anatomically constrained loss function, which was constructed based on the mapping relationship between the patient's anatomical features and the frequency domain features of the lung sound signal.

3. The method for assisting in the recognition of respiratory sounds in clinical nursing according to claim 1, characterized in that, In step S2, the method for constructing the multi-source feature space includes: aligning the temporal features of lung sound signals and respiratory signals; Align lung sound signals with CT image features; The multi-source feature space includes time-domain lung sound waveforms, frequency-domain Mel spectrum, respiratory signal temporal features, and CT image lesion area masks.

4. The method for assisting in the recognition of respiratory sounds in clinical nursing according to claim 1, characterized in that, The method for selecting a subset of causal features in step S4 includes: Construct a causal graph that includes lung sound features, disease labels, and noise types; Virtual intervention is applied to the feature vector to calculate the average causal effect between the feature and the disease, and non-causal features with effect values ​​below a preset threshold are eliminated.

5. The method for assisting in the recognition of respiratory sounds in clinical nursing according to claim 4, characterized in that, The virtual intervention is achieved through the following steps: The intervention variable is defined as the energy value of the key frequency band in lung sound characteristics; Calculate the potential difference in outcomes after intervention and screen for features where the causal effect is significant above the statistical threshold.

6. The method for assisting in the recognition of respiratory lung sounds in clinical nursing according to claim 1, characterized in that, The method for dynamically adjusting model parameters in step S5 includes: Calculate the model weight attenuation factor based on the real-time noise impact coefficient; By updating model parameters through online learning, the weighted balance between classification loss and noise loss in the objective function is optimized.

7. A respiratory sound recognition system for clinical nursing, used to perform the respiratory sound recognition method for clinical nursing as described in any one of claims 1-6, characterized in that, include: Generative data augmentation module: Integrates conditional generative adversarial networks to generate synthetic lung sound data that conforms to anatomical constraints; Cross-modal transfer module: Aligns multi-source features of lung sounds, respiratory signals, and CT images; Self-supervised contrastive learning module: The model is pre-trained using a noise-robust contrastive loss function; Causal inference module: Filters causal feature subsets based on causal graph model; Dynamic adaptation module: dynamically adjusts model parameters based on real-time noise impact coefficient; The output of the generative data augmentation module is processed by the cross-modal transfer module and then input into the self-supervised contrastive learning module and the causal inference module. Finally, the dynamic adaptation module outputs the analysis results.

8. The respiratory sound recognition system for clinical nursing according to claim 7, characterized in that, The generative data augmentation module includes: The anatomical constraint loss calculation unit verifies the physiological rationality of the generated data by mapping the patient's anatomical features to the frequency domain features of lung sound signals. The noise superposition unit superimposes environmental noise and physiological noise into the synthetic lung sound data. The superposition process is based on a temporal masking mechanism to avoid temporal conflicts.

Images