AI-based child respiratory system diagnosis auxiliary method and system

Abstract

The invention provides an AI-based child respiratory system diagnosis assistance method and system, and relates to the technical field of diagnosis assistance, and the method comprises the steps: collecting multiple layers of breathing sound signals through a bionic sensing array, building an airflow track model, carrying out the signal enhancement, and generating an enhanced breathing sound signal. Generating three-dimensional breathing sound propagation data by using the acoustic reconstruction network, evaluating a breathing function in combination with airflow dynamics and acoustic parameters, tracking abnormal breathing sound and generating a characteristic distribution diagram; symptom evolution analysis is carried out through a medical knowledge inference engine, real-time monitoring data is output, accurate auxiliary diagnosis is achieved, and the diagnosis efficiency and accuracy of children respiratory system diseases are improved.

Description

Technical Field

[0001] The present invention relates to the field of diagnostic assistance technologies, and particularly to an AI-based diagnostic assistance method and system for children's respiratory systems. Background Art

[0002] Diseases of children's respiratory systems are one of the most common diseases in pediatrics, and their diagnosis mainly relies on means such as auscultation and imaging examinations. The traditional auscultation method diagnoses by collecting breath sounds through a stethoscope, and doctors need to analyze and judge the characteristics of breath sounds based on rich clinical experience. With the development of artificial intelligence technology, respiratory system assisted diagnostic methods based on digital sound collection and intelligent analysis have gradually emerged, collecting breath sound signals through a sensor array and realizing automatic localization and classification of lesions by combining machine learning algorithms.

[0003] However, the existing breath sound diagnostic technologies have the following problems. Traditional single-point sound collection cannot accurately reflect the spatial distribution characteristics of respiratory airflow, and the collected signals are easily interfered by body surface conduction and environmental noise; the existing signal processing methods lack consideration of the anatomical structure characteristics of children's respiratory tracts and are difficult to accurately restore the propagation characteristics of sounds at different anatomical levels; the current diagnostic models mainly classify based on statistical features and fail to make full use of medical expertise for pathological mechanism analysis, and the diagnostic results lack interpretability; existing methods generally ignore the dynamic evolution process of disease development and cannot continuously monitor and warn of the disease condition.

[0004] In summary, there is an urgent need for a diagnostic method for children's respiratory systems based on multi-layer acoustic collection and medical knowledge reasoning, which realizes hierarchical collection of breath sounds through a bionic sensor array, combines aerodynamic analysis to establish a sound field propagation model; designs feature extraction and diagnostic algorithms based on anatomical structure constraints to improve the accuracy and interpretability of diagnosis; realizes continuous monitoring and warning of the disease condition through medical knowledge reasoning and dynamic feedback optimization, providing a new technical solution for the intelligent diagnosis of children's respiratory system diseases. The present invention can solve the problems in the prior art. Summary of the Invention

[0005] An embodiment of the present invention provides an AI-based diagnostic assistance method and system for children's respiratory systems, which can solve the problems in the prior art.

[0006] In a first aspect of an embodiment of the present invention, There is provided an AI-based diagnostic assistance method for children's respiratory systems, including: Based on the anatomical structure of children's respiratory systems, multi-layer respiratory sound signals are collected hierarchically through a bionic sensing array, and the respiratory airflow direction vector is obtained. An airflow trajectory model is established based on the respiratory airflow direction vector to drive an adaptive signal enhancer to adjust signal acquisition parameters, denoise and optimize the multi-layer respiratory sound signals, and generate enhanced respiratory sound signals; the enhanced respiratory sound signals are input into an acoustic reconstruction network, and three-dimensional respiratory sound propagation data is generated through a spatial sound field mapping algorithm. The three-dimensional respiratory sound propagation data is dynamically segmented to generate a respiratory feature dataset containing airflow dynamics parameters and hierarchical acoustic parameters; The respiratory feature dataset is input into an anatomy-guided diagnosis network. A respiratory function evaluation model is established based on physiological structure constraints. The propagation path of abnormal respiratory sounds is traced by combining airflow dynamics parameters and hierarchical acoustic parameters in the respiratory function evaluation model through a feature propagation algorithm to generate an abnormal feature distribution map; a lesion area localization matrix is calculated based on the abnormal feature distribution map. The lesion area localization matrix determines abnormal parameters through multi-scale feature aggregation; an acoustic feature combination is extracted from the lesion area localization matrix to generate a pathological state feature map and determine a preliminary diagnosis result; The preliminary diagnosis result is input into a medical knowledge inference engine. A symptom evolution model is constructed through a pathological mechanism analysis module. The preliminary diagnosis result is recursively state-estimated based on the symptom evolution model to generate disease development prediction data; a diagnostic optimization feedback loop is established to correct the disease development prediction data through symptom association to obtain a corrected diagnosis result. A disease risk warning model is constructed based on the corrected diagnosis result. The disease risk warning model generates real-time monitoring data through a dynamic threshold algorithm. The real-time monitoring data is combined with the corrected diagnosis result to output the final auxiliary diagnosis result.

[0007] In an alternative embodiment, Based on the anatomical structure of children's respiratory systems, multi-layer respiratory sound signals are collected hierarchically through a bionic sensing array, and the respiratory airflow direction vector is obtained. An airflow trajectory model is established based on the respiratory airflow direction vector to drive an adaptive signal enhancer to adjust signal acquisition parameters, and denoising and optimizing the multi-layer respiratory sound signals to generate enhanced respiratory sound signals includes: According to the anatomical hierarchical structure of children's respiratory systems, a bionic sensing array is set on the anterior chest wall, back, and trachea regions; multi-layer respiratory sound signals are collected hierarchically through the bionic sensing array; a mapping relationship matrix between the bionic sensing array and the anatomical hierarchy is established based on the distance parameter, sound wave incident angle parameter, and tissue attenuation parameter between the sensing unit and the anatomical hierarchy; Input the multi-layer breath sound signal into the cross-correlation analysis module of adjacent sensing units to obtain a correlation function based on the signal time sequence relationship. By detecting the peak position of the correlation function, obtain the acoustic wave propagation time delay between the adjacent sensing units. Combine the acoustic wave propagation time delay with the relative position relationship of the adjacent sensing units in three-dimensional space to construct a direction vector characterizing the movement characteristics of the respiratory airflow; Fit the basic airflow movement characteristics in the respiratory airflow direction vector through a polynomial model to obtain a basic airflow component. Describe the vortex airflow movement characteristics in the respiratory airflow direction vector through an attenuation period model to obtain a vortex airflow component; Input the polynomial parameters of the basic airflow component and the periodic attenuation parameters of the vortex airflow component into the least squares model, and iteratively adjust the polynomial parameters and the periodic attenuation parameters according to the real-time measurement value of the respiratory airflow direction vector to establish a dynamic model of the airflow movement trajectory; Calculate the modulation parameter corresponding to signal enhancement according to the dynamic model of the airflow movement trajectory. Perform a multiplication operation on the modulation parameter, the adaptive filter coefficient, and the multi-layer breath sound signal to generate a primary enhanced signal; Based on the error between the primary enhanced signal and the desired signal, update the adaptive filter coefficient in combination with the modulation parameter, and perform noise reduction optimization processing on the multi-layer breath sound signal through the updated adaptive filter coefficient, and finally output an enhanced breath sound signal.

[0008] In an alternative embodiment, Input the enhanced breath sound signal into an acoustic reconstruction network, and generate three-dimensional breath sound propagation data through a spatial sound field mapping algorithm, including: The acoustic reconstruction network includes an encoder and a decoder. The encoder includes three three-dimensional convolutional layers to extract time-domain features and spatial features. The decoder reconstructs the feature map through a deconvolution structure; Through the acoustic reconstruction network, obtain the time-frequency features and spatial features of the enhanced breath sound signal; Based on the time-frequency features and the spatial features, establish an acoustic wave propagation control equation including sound pressure and wave number using the Helmholtz equation. Discretize the acoustic wave propagation control equation through the boundary element method to obtain a sound field transfer model. Determine the Green's function matrix of the sound field transfer model by calculating the acoustic wave propagation path between the sound source position and the measurement point. Based on the acoustic wave radiation energy at the sound source position, determine the sound source intensity vector of the sound field transfer model; Construct an initial sound source intensity distribution based on the time-frequency characteristics, input the initial sound source intensity distribution into the sound field transfer model to obtain the calculated sound pressure, compare the calculated sound pressure with the measured sound pressure to obtain an error vector; determine the iteration step factor through the line search method, multiply the transpose matrix of the Green's function matrix by the error vector to obtain a gradient vector, multiply the gradient vector by the iteration step factor to obtain an update amount, and perform iterative update on the sound source intensity distribution based on the update amount. When the two-norm of the error vector is less than the preset threshold, obtain the optimal sound source intensity distribution; Determine the grid node division in the three-dimensional space based on the spatial characteristics, calculate the sound wave propagation path between the sound source position and the grid nodes to obtain the grid propagation Green's function, perform a spatial domain convolution operation on the optimal sound source intensity distribution and the grid propagation Green's function to obtain the complex sound pressure value of the grid nodes, and calculate the sound pressure amplitude distribution, phase distribution, and sound intensity distribution based on the complex sound pressure value to generate three-dimensional breath sound propagation data.

[0009] In an alternative embodiment, Perform dynamic segmentation processing on the three-dimensional breath sound propagation data to generate a breath feature dataset including aerodynamic parameters and stratified acoustic parameters, including: Divide the three-dimensional breath sound propagation data into an inhalation phase, an exhalation phase, and a transition phase according to the change in the airflow direction; in the inhalation phase, the exhalation phase, and the transition phase, calculate the velocity potential based on the three-dimensional breath sound propagation data, calculate the airflow velocity based on the gradient of the velocity potential, calculate the vorticity based on the curl of the airflow velocity, and calculate the Reynolds stress based on the pulsating component of the airflow velocity to obtain aerodynamic parameters; Perform time-frequency analysis on the acoustical signals collected in layers to obtain the spectral energy distribution, calculate the sound pressure amplitude ratio between adjacent anatomical layers based on the sound pressure amplitude distribution to obtain the inter-layer sound pressure ratio, and calculate the sound propagation attenuation coefficient based on the sound pressure amplitude distribution to obtain stratified acoustic parameters; Form a feature vector by arranging the aerodynamic parameters and the stratified acoustic parameters in time series, perform covariance matrix decomposition and eigenvalue sorting on the feature vector, and select the feature vector corresponding to the eigenvalue with a cumulative contribution rate greater than the preset threshold to form a breath feature dataset.

[0010] In an alternative embodiment, Input the breath feature dataset into an anatomy-guided diagnosis network, establish a respiratory function evaluation model based on physiological structure constraints, and generate an abnormal feature distribution map by tracking the propagation path of abnormal breath sounds in the respiratory function evaluation model by combining aerodynamic parameters and stratified acoustic parameters through a feature propagation algorithm, including: Input the respiratory feature dataset into the anatomy-guided diagnosis network, construct the three-dimensional bronchial tree network topology based on the chest CT images of children, extract the positional relationship of bronchial branches and the lumen connectivity features, calculate the anatomical position correlation matrix, and establish a respiratory function assessment model; In the respiratory function assessment model, extract the aerodynamic parameters and hierarchical acoustic parameters in the respiratory feature dataset, and fuse them with the anatomical position correlation matrix to generate a fused feature map; Construct a feature propagation network based on the anatomical position correlation matrix, set anatomical region nodes in the feature propagation network, allocate the fused feature map to each anatomical region node, calculate the node feature state, and calculate the propagation coefficient between adjacent anatomical region nodes based on the node feature state, where the propagation coefficient decays as the anatomical distance increases, and optimize the node feature state by iteratively updating the propagation coefficient; Compare the optimized node feature state with the preset normal breathing reference features, calculate the degree of abnormality of each anatomical region node, and combine the preset anatomical structure importance to weight the degree of abnormality to generate a spatial abnormality distribution, perform continuous processing and extreme value analysis on the spatial abnormality distribution to determine the position of the abnormal propagation center, and generate an abnormal feature distribution map based on the position of the abnormal propagation center.

[0011] In an alternative embodiment, Calculate a lesion area localization matrix based on the abnormal feature distribution map, where the lesion area localization matrix determines abnormal parameters through multi-scale feature aggregation; extract an acoustic feature combination from the lesion area localization matrix to generate a pathological state feature map, and the preliminary diagnosis results include: Perform multi-scale decomposition on the abnormal feature distribution map, extract features using Gaussian convolution kernels of different scales to generate a multi-scale feature atlas, calculate the local significance value of each feature map in the multi-scale feature atlas, determine the feature fusion weight based on the local significance value, and perform weighted combination on the multi-scale feature atlas according to the feature fusion weight to generate a lesion area localization matrix; Calculate the regional average abnormal intensity, regional boundary gradient, regional shape complexity, and regional tissue contrast in the lesion area localization matrix to obtain an abnormal parameter set, perform regional segmentation on the lesion area localization matrix based on the abnormal parameter set to determine candidate lesion areas; perform feature clustering on the candidate lesion areas, and merge the areas whose feature similarity meets the preset similarity threshold to obtain an optimized lesion area localization matrix; construct a lesion area distribution feature according to the distribution positions and inter-region connection relationships of each region in the optimized lesion area localization matrix; Extract the respiratory audio spectral envelope features, energy distribution features, and sound conduction features from the optimized lesion area localization matrix to construct an acoustic feature combination; calculate the deviation degree between the acoustic feature combination and the normal reference features to obtain an acoustic feature deviation value; fuse the acoustic feature combination, the acoustic feature deviation value, and the abnormal parameter set to generate a pathological state feature vector; calculate the distribution probability of the pathological state feature vector by the kernel density estimation method to generate a pathological state feature map; Establish a diagnostic decision function based on the pathological state feature map. The diagnostic decision function performs a comprehensive evaluation by combining the lesion area distribution features and the acoustic feature deviation value; output the lesion localization result and the abnormal type judgment through the diagnostic decision function to determine the preliminary diagnosis result.

[0012] In an alternative embodiment, Input the preliminary diagnosis result into a medical knowledge inference engine, construct a symptom evolution model through a pathological mechanism analysis module, perform recursive state estimation on the preliminary diagnosis result based on the symptom evolution model to generate disease development prediction data; establish a diagnostic optimization feedback loop, correct the disease development prediction data through symptom association to obtain a corrected diagnosis result, construct a disease risk warning model based on the corrected diagnosis result. The disease risk warning model generates real-time monitoring data through a dynamic threshold algorithm, combines the real-time monitoring data with the corrected diagnosis result, and outputs the final auxiliary diagnosis result including: Convert the preliminary diagnosis result into an input feature vector, map the input feature vector to a preset medical knowledge space through a knowledge graph mapping function to generate a knowledge node set; calculate the inter-node correlation weights of the knowledge node set using a multi-head attention mechanism to form an attention weight matrix; rank the importance of the knowledge nodes based on the attention weight matrix, and extract a subset of pathological mechanism knowledge related to the current case; Input the subset of pathological mechanism knowledge into the symptom evolution model to construct a disease state vector including anatomical structure indicators, functional indicators, and symptom indicators; use the treatment response index, immune function index, stress level index, and concurrent risk index as time-varying influencing factors, and combine them with the disease state vector to construct a state transition probability matrix; perform recursive state estimation on the preliminary diagnosis result based on the state transition probability matrix, predict the state distribution at the next moment, and perform Bayesian update on the predicted state in combination with real-time observation data to generate disease development prediction data; Build a symptom association network, using clinical symptoms as network nodes, calculating the temporal dependence intensity and co-occurrence probability between symptom nodes as edge weights to form a symptom association matrix; use the symptom association matrix to weight-optimize the disease development prediction data to generate a corrected diagnosis result; train a disease risk warning model based on the corrected diagnosis result, and transform clinical indicators, imaging features, and functional scores into a risk assessment space through a feature mapping function; Implement dynamic threshold monitoring for the monitoring indicators in the risk assessment space, including: calculating the moving average and standard deviation of the indicators, extracting the periodic pattern of temporal features, evaluating the variation characteristics of feature weights over time, updating the risk scoring function, and generating real-time monitoring data; perform deep feature fusion on the real-time monitoring data and the corrected diagnosis result, and construct a comprehensive evaluation indicator using an attention weight mechanism; calculate the decision function value based on the comprehensive evaluation indicator, map the fused features to the diagnosis result space, and output the final auxiliary diagnosis result including the probability distribution of abnormal types, severity score, and development trend.

[0013] In the second aspect of the embodiments of the present invention, Provide an AI-based diagnostic assistance system for children's respiratory system, including: A first unit for hierarchically collecting multi-layer respiratory sound signals through a bionic sensing array based on the anatomical structure of the children's respiratory system, obtaining the respiratory airflow direction vector, establishing an airflow trajectory model based on the respiratory airflow direction vector, driving an adaptive signal enhancer to adjust signal acquisition parameters, and performing noise reduction optimization on the multi-layer respiratory sound signals to generate enhanced respiratory sound signals; input the enhanced respiratory sound signals into an acoustic reconstruction network, generate three-dimensional respiratory sound propagation data through a spatial sound field mapping algorithm, and perform dynamic segmentation processing on the three-dimensional respiratory sound propagation data to generate a respiratory feature dataset containing airflow dynamics parameters and hierarchical acoustic parameters; A second unit for inputting the respiratory feature dataset into an anatomy-guided diagnosis network, establishing a respiratory function evaluation model based on physiological structure constraints, tracking the propagation path of abnormal respiratory sounds by combining airflow dynamics parameters and hierarchical acoustic parameters in the respiratory function evaluation model through a feature propagation algorithm, and generating an abnormal feature distribution map; calculating a lesion area localization matrix based on the abnormal feature distribution map, where the lesion area localization matrix determines abnormal parameters through multi-scale feature aggregation; extracting an acoustic feature combination from the lesion area localization matrix to generate a pathological state feature map and determining a preliminary diagnosis result; A third unit is configured to input the preliminary diagnosis result into a medical knowledge inference engine, construct a symptom evolution model through a pathological mechanism analysis module, perform recursive state estimation on the preliminary diagnosis result based on the symptom evolution model to generate disease development prediction data; establish a diagnostic optimization feedback loop, correct the disease development prediction data through symptom association to obtain a corrected diagnosis result, construct a disease risk warning model based on the corrected diagnosis result, and the disease risk warning model generates real-time monitoring data through a dynamic threshold algorithm, and combines the real-time monitoring data with the corrected diagnosis result to output a final auxiliary diagnosis result.

[0014] In a third aspect of the embodiments of the present invention, a kind of electronic device is provided, including: a processor; a memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0015] In a fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0016] In the embodiments of the present invention, through the multi-layer respiratory sound signal acquisition based on the anatomical structure of the children's respiratory system and the establishment of the airflow trajectory model, the quality of the respiratory sound signal can be effectively improved, the clarity and accuracy of the signal can be enhanced, so as to provide more reliable data support for the diagnosis of the children's respiratory system; by using the acoustic reconstruction network and the dynamic segmentation processing technology, high-precision three-dimensional respiratory sound propagation data can be generated, and combined with the airflow dynamics parameters and acoustic parameters, the propagation path of abnormal respiratory sounds can be comprehensively analyzed to help doctors better understand and locate the lesions of the respiratory system; through the combination of the medical knowledge inference engine and the symptom evolution model, the recursive state estimation of the preliminary diagnosis result and the disease development prediction can be realized, and then a disease risk warning model can be established to provide real-time monitoring data, and the ability of early detection and intervention of children's respiratory diseases can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a schematic flowchart of the AI-based children's respiratory system diagnosis assistance method according to the embodiments of the present invention; Figure 2 is a schematic structural diagram of the AI-based children's respiratory system diagnosis assistance system according to the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part rather than all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0019] The following uses specific embodiments to elaborate on the technical solutions of the present invention in detail. These several specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0020] Figure 1 The following is a schematic flowchart of the AI-based child respiratory system diagnosis assistance method according to the embodiments of the present invention. As Figure 1 shown, the method includes: S101. Based on the anatomical structure of the child respiratory system, use a bionic sensing array to hierarchically collect multi-layer respiratory sound signals, obtain the respiratory airflow direction vector, establish an airflow trajectory model based on the respiratory airflow direction vector, drive an adaptive signal enhancer to adjust the signal acquisition parameters, perform noise reduction and optimization on the multi-layer respiratory sound signals, and generate enhanced respiratory sound signals; input the enhanced respiratory sound signals into an acoustic reconstruction network, generate three-dimensional respiratory sound propagation data through a spatial sound field mapping algorithm, perform dynamic segmentation processing on the three-dimensional respiratory sound propagation data, and generate a respiratory feature dataset containing airflow dynamics parameters and hierarchical acoustic parameters; In this embodiment, based on the unique anatomical structure of the child respiratory system, a bionic sensing array is used to achieve hierarchical collection of multi-level respiratory sound signals, making the signal collection more refined, capable of capturing minute changes during the child's breathing process, and improving the overall detection accuracy; by obtaining the respiratory airflow direction vector and establishing an airflow trajectory model, the movement direction and trajectory of the respiratory airflow can be accurately described, providing a reliable kinematic basis for subsequent signal processing, thus helping to better understand and evaluate respiratory dynamics; driving an adaptive signal enhancer to dynamically adjust the acquisition parameters according to the collected signals, effectively performing noise reduction processing, generating enhanced respiratory sound signals, significantly improving the signal-to-noise ratio and overall quality of the signals, and laying a solid foundation for subsequent analysis and processing; using an acoustic reconstruction network and a spatial sound field mapping algorithm to generate three-dimensional respiratory sound propagation data and perform dynamic segmentation processing on it, thereby extracting a rich respiratory feature dataset containing airflow dynamics parameters and hierarchical acoustic parameters. This method can not only comprehensively reflect the spatio-temporal dynamic changes of the breathing process but also reveal subtle pathological changes.

[0021] S102. Input the respiratory feature dataset into the anatomy-guided diagnosis network, establish a respiratory function assessment model based on physiological structure constraints, use the feature propagation algorithm to combine aerodynamic parameters and hierarchical acoustic parameters in the respiratory function assessment model to track the propagation path of abnormal breath sounds, and generate an abnormal feature distribution map; calculate a lesion area localization matrix based on the abnormal feature distribution map, where the lesion area localization matrix determines abnormal parameters through multi-scale feature aggregation; extract an acoustic feature combination from the lesion area localization matrix, generate a pathological state feature map, and determine a preliminary diagnosis result; In this embodiment, by inputting the respiratory feature dataset into the anatomy-guided diagnosis network and using physiological structure constraints to establish a respiratory function assessment model, it is possible to effectively combine aerodynamic parameters and hierarchical acoustic parameters, accurately track the propagation path of abnormal breath sounds, and the generated abnormal feature distribution map provides an intuitive basis for subsequent diagnosis; using the feature propagation algorithm and multi-scale feature aggregation technology, calculate a lesion area localization matrix from the abnormal feature distribution map, making the identification of abnormal parameters more accurate, improving the localization accuracy of the lesion area, and helping to refine the diagnosis result; by extracting an acoustic feature combination from the lesion area localization matrix and generating a pathological state feature map, it provides quantitative and structured abnormal feature information for clinical practice, further determines the preliminary diagnosis result, and thus enhances the objectivity and accuracy of the diagnosis.

[0022] S103. Input the preliminary diagnosis result into the medical knowledge inference engine, construct a symptom evolution model through the pathological mechanism analysis module, perform recursive state estimation on the preliminary diagnosis result based on the symptom evolution model, and generate disease development prediction data; establish a diagnostic optimization feedback loop, correct the disease development prediction data through symptom association, obtain a corrected diagnosis result, construct a disease risk warning model based on the corrected diagnosis result, where the disease risk warning model generates real-time monitoring data through a dynamic threshold algorithm, and combine the real-time monitoring data with the corrected diagnosis result to output the final auxiliary diagnosis result.

[0023] In this embodiment, by inputting the preliminary diagnosis result into the medical knowledge inference engine and using the pathological mechanism analysis module to construct a symptom evolution model, it realizes recursive state estimation of symptom changes, thereby generating reliable disease development prediction data, providing a scientific basis for early intervention and personalized treatment; establish a diagnostic optimization feedback loop, use symptom association to correct the prediction data, effectively compensate for errors in the preliminary diagnosis, obtain a more accurate corrected diagnosis result, and improve the accuracy and reliability of the overall diagnosis; construct a disease risk warning model based on the corrected diagnosis result, generate real-time monitoring data with the help of the dynamic threshold algorithm, realize dynamic monitoring and timely warning of disease risks, and enhance the prevention and control ability of the disease deterioration trend.

[0024] In an alternative embodiment, based on the anatomical structure of a child's respiratory system, multi-layer respiratory sound signals are collected through a bionic sensing array in layers, and the respiratory airflow direction vector is obtained. An airflow trajectory model is established based on the respiratory airflow direction vector to drive an adaptive signal enhancer to adjust signal acquisition parameters, and noise reduction optimization is performed on the multi-layer respiratory sound signals to generate enhanced respiratory sound signals, including: According to the anatomical hierarchical structure of a child's respiratory system, a bionic sensing array is arranged on the anterior chest wall, back, and trachea regions; multi-layer respiratory sound signals are collected through the bionic sensing array in layers; a mapping relationship matrix between the bionic sensing array and the anatomical hierarchy is established based on the distance parameter between the sensing unit and the anatomical hierarchy, the acoustic wave incident angle parameter, and the tissue attenuation parameter; The multi-layer respiratory sound signals are input into the cross-correlation analysis module of adjacent sensing units to obtain a correlation function based on the signal time sequence relationship. By detecting the peak position of the correlation function, the acoustic wave propagation time delay between the adjacent sensing units is obtained. The acoustic wave propagation time delay is combined with the relative position relationship of the adjacent sensing units in three-dimensional space to construct a direction vector representing the movement characteristics of the respiratory airflow; The basic airflow movement characteristics in the respiratory airflow direction vector are fitted through a polynomial model to obtain a basic airflow component, and the vortex airflow movement characteristics in the respiratory airflow direction vector are described through an attenuation period model to obtain a vortex airflow component; the polynomial parameters of the basic airflow component and the periodic attenuation parameters of the vortex airflow component are input into the least squares model, and the polynomial parameters and the periodic attenuation parameters are iteratively adjusted according to the real-time measurement value of the respiratory airflow direction vector to establish a dynamic model of the airflow movement trajectory; The modulation parameters corresponding to signal enhancement are calculated according to the dynamic model of the airflow movement trajectory. The modulation parameters are multiplied by the adaptive filter coefficients and the multi-layer respiratory sound signals to generate a primary enhanced signal; based on the error between the primary enhanced signal and the desired signal, the adaptive filter coefficients are updated in combination with the modulation parameters, and the multi-layer respiratory sound signals are subjected to noise reduction optimization processing through the updated adaptive filter coefficients, and finally enhanced respiratory sound signals are output.

[0025] In a specific implementation manner, a layout scheme of the bionic sensing array is designed based on the anatomical structure characteristics of a child's respiratory system. 8 sensing units are arranged on each side of the midline of the sternum, corresponding to the intercostal spaces; 12 sensing units are symmetrically arranged on both sides of the spinal column in the back, covering the main regions of the lungs; 4 sensing units are arranged in the front region of the trachea. Each sensing unit consists of a piezoelectric ceramic sensor with a diameter of 8 mm and a supporting signal conditioning circuit, and the sampling frequency is set to 4096 Hz and the resolution is 24 bits.

[0026] Establish the spatial mapping relationship between the sensing array and the anatomical hierarchy. Obtain the anatomical stratification data of the tissue under the sensing unit through high-resolution CT images, and measure the thickness of each layer of tissue. Taking the skin surface as the reference plane, the distances of the superficial layer, middle layer, and deep layer are set to 0-20 mm, 20-40 mm, and 40-60 mm respectively. Combining the propagation speeds of sound waves in different tissues, calculate the time delay of sound waves propagating from each anatomical hierarchy to the sensing unit, and establish a spatial mapping matrix based on time delay.

[0027] Determine the direction of respiratory airflow through the cross-correlation analysis of signals from adjacent sensing units. Select the respiratory sound signals of adjacent sensing unit pairs, and calculate the cross-correlation function using a 32-ms analysis window. Determine the signal propagation time delay through peak detection, with a time delay resolution of 0.25 ms. Combining the three-dimensional coordinate information of the sensing units, convert the time delay into the sound wave propagation direction, and construct a three-dimensional direction vector characterizing the airflow movement characteristics.

[0028] Perform eigen-decomposition and dynamic modeling on the respiratory airflow direction vector. Use a third-order polynomial model to fit the basic laminar flow movement characteristics, and set the initial values of the polynomial coefficients to [1.0, -0.5, 0.2, -0.05]. Use a decaying sine function to describe the eddy current movement characteristics, with an initial frequency set to 2 Hz and a decay coefficient of 0.8. Update the model parameters through least squares iterative optimization, with the number of iterations set to 100 and the convergence threshold set to 0.001. Finally, establish a dynamic prediction model for the airflow movement trajectory.

[0029] Realize adaptive signal enhancement based on the airflow trajectory model. Calculate the signal modulation parameters, with the amplitude modulation depth range of 0.6-1.4 and the phase modulation range of ±30 degrees. Process the modulation parameters, the 64th-order FIR filter coefficients, and the original signal to generate a primary enhanced signal. Iteratively update the filter coefficients through the mean square error criterion, with the learning rate set to 0.01 and the number of iterations to 200. After multi-level optimization, output the final enhanced respiratory sound signal, with the signal-to-noise ratio increased by 15-20 dB.

[0030] In clinical applications, this method can effectively extract the respiratory sound characteristics of children aged 6 months to 12 years. For typical cases, such as children with pneumonia, it can accurately capture the spatial distribution characteristics of local moist rales. The moist rales are mainly distributed in the chest wall area corresponding to the lesion, with the sound intensity 10-15 dB higher than that of normal respiratory sounds and the frequency range between 100-400 Hz. Through dynamic tracking analysis, the changing trend of the gradual weakening of the moist rale intensity and the gradual shrinking of the range during the treatment process can be monitored, providing an objective basis for evaluating the clinical treatment effect.

[0031] In this embodiment, by establishing a mapping relationship matrix between the sensing unit and the anatomical hierarchy of the child's respiratory system, each sensor can be accurately corresponding to the corresponding anatomical site, realizing the fixed-point acquisition of multi-layer breath sound signals, and improving the positioning accuracy and reliability of data acquisition; by using the cross-correlation analysis between adjacent sensing units, combining the acoustic wave propagation time delay and the three-dimensional spatial position relationship, a direction vector characterizing the movement characteristics of the respiratory airflow is effectively constructed to ensure that the dynamic change characteristics of the airflow during the breathing process can be accurately captured; using the polynomial model to fit the basic airflow movement characteristics and describing the vortex airflow movement characteristics through the attenuation period model, realizing the effective decomposition and accurate description of the complex components of the airflow movement, providing a solid data basis for the subsequent construction of the dynamic model; combining the least squares method model to iteratively adjust the polynomial parameters and the periodic attenuation parameters, establishing a dynamic model of the airflow movement trajectory, so as to realize the real-time tracking and accurate modeling of the respiratory airflow movement state, and enhancing the adaptability of the system to dynamic changes; calculating the signal modulation parameters according to the dynamic model of the airflow movement trajectory, multiplying them with the adaptive filter coefficients and the multi-layer breath sound signals to generate a primary enhanced signal. Subsequently, the filter coefficients are updated through error feedback to further optimize the noise reduction of the breath sound signal, significantly improving the clarity and signal-to-noise ratio of the output signal.

[0032] In an alternative embodiment, the enhanced breath sound signal is input into an acoustic reconstruction network, and the three-dimensional breath sound propagation data is generated through a spatial sound field mapping algorithm, including: The acoustic reconstruction network includes an encoder and a decoder. The encoder includes three layers of three-dimensional convolutional layers to extract time-domain features and spatial features, and the decoder reconstructs the feature map through a deconvolution structure; through the acoustic reconstruction network, the time-frequency features and spatial features of the enhanced breath sound signal are obtained; Based on the time-frequency features and the spatial features, an acoustic wave propagation control equation including sound pressure and wave number is established using the Helmholtz equation, and the acoustic wave propagation control equation is discretized by the boundary element method to obtain a sound field transfer model. By calculating the acoustic wave propagation path between the sound source position and the measurement point, the Green's function matrix of the sound field transfer model is determined, and based on the acoustic wave radiation energy at the sound source position, the sound source intensity vector of the sound field transfer model is determined; Based on the time-frequency features, an initial sound source intensity distribution is constructed, and the initial sound source intensity distribution is input into the sound field transfer model to obtain the calculated sound pressure. The calculated sound pressure is compared with the measured sound pressure to obtain an error vector; the iteration step factor is determined by the line search method, the gradient vector is obtained by multiplying the transpose matrix of the Green's function matrix with the error vector, the update amount is obtained by multiplying the gradient vector by the iteration step factor, and the sound source intensity distribution is iteratively updated based on the update amount. When the two-norm of the error vector is less than the preset threshold, the optimal sound source intensity distribution is obtained; Based on the spatial features, determine the grid node division in the three-dimensional space, calculate the acoustic wave propagation path between the sound source position and the grid nodes to obtain the grid propagation Green's function, perform a spatial domain convolution operation on the optimal sound source intensity distribution and the grid propagation Green's function to obtain the complex sound pressure value of the grid nodes, calculate the sound pressure amplitude distribution, phase distribution, and sound intensity distribution based on the complex sound pressure value, and generate three-dimensional breath sound propagation data.

[0033] The Helmholtz equation specifically refers to a mathematical expression method for describing steady-state wave phenomena. It reflects the spatial distribution characteristics of the wave field when the wave propagates in the medium. This equation is usually used to analyze the propagation behavior of acoustic waves, light waves, etc. at a certain frequency, and can reveal how the wave is distributed and attenuated in space.

[0034] The boundary element method specifically refers to a numerical solution technique mainly used to handle boundary value problems. The characteristic of this method is that only the boundary of the problem needs to be discretized, rather than meshing the entire region. By converting the original problem into a form that only depends on the boundary integral, the boundary element method shows unique advantages in solving problems in infinite or semi-infinite regions, and can improve the calculation efficiency and accuracy.

[0035] The Green's function specifically refers to a function that describes the response characteristics of a system, used to reflect the response of a system to a local excitation or a point source. Using the Green's function, a complex excitation problem can be transformed into an integral form, so as to obtain the response of the system at any position. This concept plays an important role in establishing the transfer model of physical systems, especially in the fields of acoustics and electromagnetic fields, helping to accurately describe the propagation process from the sound source to the measurement point.

[0036] In a specific embodiment, construct an acoustic reconstruction network architecture. The encoder adopts a three-layer three-dimensional convolution structure. The size of the convolution kernel in the first layer is 3×3×3, and 64 feature maps are output; the size of the convolution kernel in the second layer is 3×3×3, and 128 feature maps are output; the size of the convolution kernel in the third layer is 3×3×3, and 256 feature maps are output. After each layer of convolution, a batch normalization layer and a ReLU activation function are connected. The decoder adopts a corresponding transposed convolution structure to restore the feature dimension layer by layer. Through this network, extract the time-frequency features of the enhanced breath sound signal, including the spectrum information in the range of 0 - 2000 Hz, and the spatial features reflecting the propagation characteristics of the acoustic wave in the three-dimensional space.

[0037] Based on the extracted time-frequency features and spatial features, a sound field propagation model is established. First, a grid division is constructed in three-dimensional space with a grid spacing of 5 mm, covering a spatial range of approximately 300×200×150 mm in the front and back regions of the chest cavity. At each grid node position, the relationship between sound pressure and wave number is calculated to establish the control equation for sound wave propagation. The boundary element method is used to discretize the continuous equation into matrix form, where the transfer matrix reflects the sound wave propagation characteristics between the sound source position and the measurement point.

[0038] For the divided grid space, the propagation paths from the sound source to each measurement point are calculated to construct the Green's function matrix. This matrix describes the propagation attenuation characteristics of sound waves in space, considering factors such as geometric divergence and medium absorption. Based on the measured sound pressure amplitude, the radiation energy intensity of the sound source is determined. The initial sound source intensity distribution adopts a Gaussian distribution form, with the peak set to 1.5 times the measured sound pressure and the standard deviation to 20 mm.

[0039] The initial sound source intensity distribution is input into the sound field transfer model to calculate the theoretical sound pressure values at each measurement point. The calculated values are compared with the measured sound pressure to obtain the error vector. The optimal iteration step size is determined through line search, with the step size range being 0.1 - 0.5. The gradient direction is obtained by multiplying the transpose of the Green's function matrix with the error vector, and the sound source intensity distribution is updated in combination with the iteration step size. The error threshold is set to 1% of the initial error, and the optimal sound source intensity distribution is obtained when the error is less than the threshold.

[0040] Based on the optimal sound source intensity distribution, the three-dimensional sound field is reconstructed. The propagation paths from the sound source to the spatial grid nodes are calculated to construct the grid propagation Green's function. This function includes an amplitude attenuation term and a phase delay term, reflecting the energy loss and phase change of sound waves during spatial propagation. The complex sound pressure values at the grid nodes are obtained through spatial domain convolution operation, and further the sound pressure amplitude distribution, phase distribution, and sound intensity distribution are calculated.

[0041] In clinical verification, this method can accurately reconstruct the three-dimensional propagation characteristics of children's breath sounds. For normal breath sounds, the sound pressure amplitude is approximately 60 - 70 dB on the chest wall surface, with an inward attenuation rate of 6 - 8 dB / cm. Pathological breath sounds such as wheezing can reach a sound pressure of 80 - 85 dB at the lesion site, and a local sound intensity enhancement area is formed around it, with a diameter of approximately 30 - 40 mm. The sound source position can be accurately located through three-dimensional sound field reconstruction, providing a basis for subsequent lesion localization.

[0042] In this embodiment, the time-domain and spatial features of the signal can be extracted simultaneously, realizing the all-round and multi-angle information reconstruction of the enhanced breath sound signal, and improving the ability to capture the details of the breath signal; using the physical model method to construct the acoustic wave propagation control equation and adopting the numerical discretization technology helps to accurately establish the sound field transfer model, thus more realistically reflecting the propagation behavior of acoustic waves in space; by calculating the propagation path between the sound source and the measurement point and constructing the response matrix, the accurate quantification of the radiation energy of the sound source is realized, which helps to improve the accuracy of the acoustic reconstruction result; adopting the iterative optimization method to continuously correct the sound source intensity distribution can effectively reduce the calculation error, making the finally reconstructed sound field data more accurate; by combining the optimal sound source intensity distribution with the propagation characteristics in the three-dimensional space, the comprehensive analysis of sound pressure, phase and sound intensity is realized, thus generating high-quality three-dimensional breath sound propagation data, providing solid data support for subsequent diagnosis and analysis.

[0043] In an alternative embodiment, the dynamic segmentation process is performed on the three-dimensional breath sound propagation data to generate a breath feature data set including aerodynamic parameters and stratified acoustic parameters, including: The three-dimensional breath sound propagation data is divided into an inhalation phase, an exhalation phase and a transition phase according to the change of the air flow direction; in the inhalation phase, the exhalation phase and the transition phase, the velocity potential is calculated based on the three-dimensional breath sound propagation data, the air flow velocity is calculated based on the gradient of the velocity potential, the vorticity is calculated based on the curl of the air flow velocity, and the Reynolds stress is calculated based on the pulsating component of the air flow velocity to obtain the aerodynamic parameters; The time-frequency analysis is performed on the acoustical signals collected in layers to obtain the spectral energy distribution. The sound pressure amplitude ratio between adjacent anatomical layers is calculated based on the sound pressure amplitude distribution to obtain the inter-layer sound pressure ratio, and the sound propagation attenuation coefficient is calculated based on the sound pressure amplitude distribution to obtain the stratified acoustic parameters; The aerodynamic parameters and the stratified acoustic parameters are composed into a feature vector according to the time sequence, the covariance matrix decomposition and eigenvalue sorting are performed on the feature vector, and the feature vector corresponding to the eigenvalue with the cumulative contribution rate greater than the preset threshold is selected to constitute the breath feature data set.

[0044] The velocity potential specifically refers to a scalar function describing the fluid motion. Under the condition of irrotational flow, the velocity field of the fluid can be expressed as the gradient of this scalar function. It can simplify the complex flow field into a single function, which helps to quantitatively analyze the motion characteristics of the fluid.

[0045] The Reynolds stress specifically refers to the stress component caused by the pulsating part of the fluid velocity in turbulent flow. It reflects the role of momentum transfer in turbulence and provides key parameters for understanding and predicting the turbulent behavior by describing the interaction between pulsating velocities.

[0046] In a specific embodiment, the respiratory phases are divided based on three-dimensional breath sound propagation data. By detecting the variation characteristics of the airflow direction, the respiratory cycle is divided into an inhalation phase, an exhalation phase, and a transition phase. The characteristics of the inhalation phase are that the airflow flows from the outside to the inside, with a duration of about 1.2 - 1.5 seconds; the characteristics of the exhalation phase are that the airflow flows from the inside to the outside, with a duration of about 1.0 - 1.2 seconds; the transition phase is between inhalation and exhalation, with a duration of about 0.1 - 0.2 seconds, and it is manifested as a process of airflow direction conversion.

[0047] The airflow dynamics characteristic parameters are extracted in each respiratory phase. First, the velocity potential distribution is calculated, with a grid division of 32×32×16 and a grid spacing of 5 mm. Based on the spatial gradient of the velocity potential, the airflow velocity field is obtained. During normal breathing, the peak airflow velocity is about 2 - 3 m / s at the bronchi and decreases to 0.5 - 1 m / s at the bronchioles. The vorticity distribution is obtained by calculating the curl of the airflow velocity field to reflect the rotational characteristics of the airflow. The vorticity intensity at the bronchial bifurcation is about 20 - 30 Hz. The Reynolds stress tensor is calculated based on the pulsating component of the airflow velocity to characterize the turbulence intensity. During normal breathing, the peak Reynolds stress is about 0.2 - 0.4 Pa.

[0048] Multi-scale time-frequency analysis is performed on the acoustical signals collected in layers. Using a Hanning window with a window length of 512 points and an overlap rate of 50%, the short-time Fourier transform is calculated to obtain the time-frequency spectrum. The main energy distribution of normal breath sounds is in the range of 100 - 600 Hz, and pathological breath sounds can extend to 2000 Hz. The sound pressure amplitude ratio between adjacent anatomical levels is calculated to obtain the inter-layer sound pressure ratio. During normal breathing, the sound pressure ratio between the deep layer and the middle layer is about 0.6 - 0.8, and the sound pressure ratio between the middle layer and the shallow layer is about 0.7 - 0.9. Based on the attenuation law of the sound pressure amplitude during spatial propagation, the sound propagation attenuation coefficient is calculated. The attenuation coefficient in normal tissues is about 0.7 - 0.9 dB / cm.

[0049] The time-series structure of the respiratory feature dataset is constructed. The airflow dynamics parameters and the layered acoustic parameters are combined into a feature vector according to the sampling time series. Each feature vector contains parameters such as the velocity field, vorticity field, Reynolds stress, spectral energy, inter-layer sound pressure ratio, and attenuation coefficient. The covariance matrix is calculated for the feature vector sequence, and eigenvalue decomposition is performed to obtain eigenvalues and eigenvectors. The eigenvalues are arranged in descending order, and the eigenvectors corresponding to the eigenvalues with a cumulative contribution rate greater than 85% are selected to form the reduced-dimensional respiratory feature dataset.

[0050] In clinical applications, this feature extraction method can effectively distinguish normal and abnormal breathing patterns. When normal children breathe, the aerodynamic parameters show regular changes and the interlayer sound pressure ratio is stable; while for children with pneumonia, features such as reduced airflow velocity, increased vorticity, and elevated Reynolds stress will appear in the lesion area, and the interlayer sound pressure ratio changes significantly. Through the time series analysis of the breathing feature dataset, the dynamic change process of these abnormal features can be traced, providing a quantitative basis for clinical diagnosis. Under pathological conditions, the airflow velocity can be reduced by 30%-50%, the vorticity intensity increases by 1.5-2 times, the Reynolds stress increases by 2-3 times, and the variation degree of the interlayer sound pressure ratio increases by 40%-60%.

[0051] In this embodiment, the three-dimensional breathing sound propagation data is divided into the inhalation phase, exhalation phase, and transition phase according to the change of the airflow direction, which helps to accurately distinguish different stages in the breathing process and facilitates subsequent refined analysis; in each breathing stage, by calculating the velocity potential and further obtaining the airflow velocity, vorticity, and Reynolds stress, various parameters reflecting the aerodynamic characteristics can be systematically obtained, thereby revealing the subtle changes in the airflow movement during the breathing process; by performing time-frequency analysis on the acoustical signals collected in layers and calculating the interlayer sound pressure ratio and propagation attenuation coefficient, layered acoustical parameters can be obtained, which provides an intuitive basis for reflecting the sound wave transmission characteristics between different anatomical levels; the aerodynamic parameters and layered acoustical parameters are composed into a feature vector according to the time series, and the main features are extracted by using covariance matrix decomposition and eigenvalue sorting to form a breathing feature dataset. This multi-dimensional and multi-scale feature extraction method improves the data representation ability and provides a more comprehensive and accurate basis for further breathing function evaluation and disease diagnosis.

[0052] In an alternative embodiment, the breathing feature dataset is input into an anatomy-guided diagnosis network, a breathing function evaluation model is established based on physiological structure constraints, and the propagation path of abnormal breath sounds is traced by combining the aerodynamic parameters and layered acoustical parameters in the breathing function evaluation model through a feature propagation algorithm. The generated abnormal feature distribution map includes: The breathing feature dataset is input into an anatomy-guided diagnosis network, a three-dimensional bronchial tree network topology structure is constructed based on the chest CT images of children, the positional relationship of bronchial branches and the lumen connection characteristics are extracted, the anatomical position correlation matrix is calculated, and a breathing function evaluation model is established; In the breathing function evaluation model, the aerodynamic parameters and layered acoustical parameters in the breathing feature dataset are extracted and fused with the anatomical position correlation matrix to generate a fused feature map; Construct a feature propagation network based on the anatomical position correlation matrix. Set anatomical region nodes in the feature propagation network, allocate the fused feature map to each of the anatomical region nodes, calculate the node feature states, and calculate the propagation coefficients between adjacent anatomical region nodes based on the node feature states, where the propagation coefficients decay as the anatomical distance increases. Optimize the node feature states by iteratively updating the propagation coefficients; Compare the optimized node feature states with the preset normal breathing reference features, calculate the abnormality degree of each anatomical region node, and combine the preset anatomical structure importance to weight the abnormality degree to generate a spatial abnormality distribution. Continuously process the spatial abnormality distribution and perform extreme value analysis to determine the position of the abnormality propagation center, and generate an abnormal feature distribution map based on the position of the abnormality propagation center.

[0053] In a specific embodiment, construct an anatomy-guided diagnosis model based on children's chest CT images. Collect chest CT images with a resolution of 0.5 mm, and extract the bronchial tree structure through image segmentation. Starting from the tracheal inlet, trace the bronchial branches layer by layer, and record the spatial coordinates, diameter sizes, and bifurcation angles of each level of bronchi. The diameter range of the main bronchi is 8-10 mm, the diameter of the secondary bronchi is 5-7 mm, and the diameter of the tertiary bronchi is 3-4 mm. Construct a network topology structure based on the bronchial branch relationship, calculate the spatial distance and connectivity between adjacent anatomical positions, and generate an anatomical position correlation matrix.

[0054] Construct a respiratory function assessment model. Normalize the aerodynamic parameters in the respiratory feature dataset, including the air velocity field, vorticity field, and Reynolds stress field. At the same time, process the stratified acoustic parameters, including the spectral energy distribution, interlayer sound pressure ratio, and sound propagation attenuation coefficient. Perform a tensor fusion operation on these feature parameters and the anatomical position correlation matrix to generate a fused feature map reflecting the local anatomical features and functional states.

[0055] Set anatomical region nodes in the feature propagation network. The number of nodes corresponds to the number of bronchial branch levels, usually including 50-80 nodes. Allocate the fused feature map to each node according to the anatomical correspondence relationship. The initial node feature states include local airflow features and acoustic features. Calculate the propagation coefficients between adjacent nodes. The propagation coefficients decay exponentially as the anatomical distance increases, and the decay coefficient is about 0.3 / cm. Set the number of iterative updates to 50, and update the propagation coefficients and node feature states in each iteration until the feature distribution tends to be stable.

[0056] Establish a normal breathing reference feature library. Collect the breathing data of 100 healthy children and establish a reference range according to age groups. The standard deviation range of air velocity is 0.2 - 0.4 m / s, the standard deviation range of vorticity is 5 - 8 Hz, and the coefficient of variation of the interlayer sound pressure ratio is less than 15%. Compare the measured data with the reference features and calculate the deviation degree of each node feature. At the same time, introduce the importance weight of anatomical structures, with the weight of the main bronchus being 1.0, the weight of the secondary bronchus being 0.8, and the weight of the tertiary bronchus being 0.6, and perform weighted processing on the abnormal degree.

[0057] Perform post - processing on the weighted spatial abnormal distribution. Use cubic spline interpolation to achieve the continuous processing of the abnormal distribution, and improve the spatial resolution to 2 mm. Determine the position of the abnormal propagation center through local extreme value analysis, and set the abnormal determination threshold to 2 times the standard deviation of the reference range. Based on the abnormal center, calculate the spatial diffusion range of the abnormal features and generate a three - dimensional abnormal feature distribution map.

[0058] In clinical verification, this method can effectively identify and locate respiratory system abnormalities. For children with bronchitis, characteristic regions with a 40 - 60% reduction in local air velocity, a 1.5 - 2.5 - fold increase in vorticity, and a 30 - 50% abnormal increase in sound pressure ratio can be detected. The abnormal feature distribution map clearly shows the location and scope of the lesion, has good consistency with the clinical diagnosis results, and the spatial positioning error is less than 5 mm. By continuously monitoring the dynamic changes of the abnormal feature distribution, an objective evaluation of the treatment effect can be achieved.

[0059] In this embodiment, by constructing a three - dimensional bronchial tree network topology based on children's chest CT images, the positional relationship of bronchial branches and the lumen connection characteristics can be accurately extracted, thereby constructing a correlation matrix reflecting the anatomical structure, providing a solid structural basis for respiratory function evaluation; integrating aerodynamic parameters and stratified acoustic parameters and integrating them with the anatomical position correlation matrix helps to comprehensively characterize respiratory features from multiple perspectives and improve the description accuracy of the respiratory function state; the constructed feature propagation network realizes the effective propagation and information aggregation of features in the anatomical structure by distributing fused feature maps on nodes in different anatomical regions and calculating the propagation coefficient that decays with anatomical distance between nodes, optimizing the node feature state; comparing the optimized node feature state with the preset normal breathing reference features and weighting them in combination with the importance of the anatomical structure can generate an intuitive spatial abnormal distribution, and determine the position of the abnormal propagation center through continuous processing and extreme value analysis, achieving precise positioning of abnormal signals.

[0060] In an alternative embodiment, a lesion area localization matrix is calculated based on the abnormal feature distribution map. The abnormal parameters are determined through multi-scale feature aggregation in the lesion area localization matrix. Extracting an acoustic feature combination from the lesion area localization matrix to generate a pathological state feature map, and determining the preliminary diagnosis result includes: Perform multi-scale decomposition on the abnormal feature distribution map, extract features using Gaussian convolution kernels of different scales to generate a multi-scale feature map set, calculate the local significance value of each feature map in the multi-scale feature map set, determine the feature fusion weight based on the local significance value, and perform weighted combination on the multi-scale feature map set according to the feature fusion weight to generate a lesion area localization matrix; Calculate the regional average abnormal intensity, regional boundary gradient, regional shape complexity, and regional tissue contrast in the lesion area localization matrix to obtain an abnormal parameter set. Perform regional segmentation on the lesion area localization matrix based on the abnormal parameter set to determine candidate lesion areas. Perform feature clustering on the candidate lesion areas, and merge the areas with feature similarity meeting the preset similarity threshold to obtain an optimized lesion area localization matrix. Construct the lesion area distribution feature according to the distribution positions and inter-regional connection relationships of the regions in the optimized lesion area localization matrix; Extract the respiratory audio spectral envelope feature, energy distribution feature, and sound conduction feature from the optimized lesion area localization matrix to construct an acoustic feature combination. Calculate the deviation degree between the acoustic feature combination and the normal reference feature to obtain an acoustic feature deviation value. Perform feature fusion on the acoustic feature combination, the acoustic feature deviation value, and the abnormal parameter set to generate a pathological state feature vector. Calculate the distribution probability of the pathological state feature vector through the kernel density estimation method to generate a pathological state feature map; Establish a diagnostic decision function based on the pathological state feature map. The diagnostic decision function performs comprehensive evaluation by combining the lesion area distribution feature and the acoustic feature deviation value. Output the lesion localization result and abnormal type judgment through the diagnostic decision function to determine the preliminary diagnosis result.

[0061] In a specific implementation manner, multi-scale feature extraction is performed on the abnormal feature distribution map. Four Gaussian convolution kernels with different scales are set, and the standard deviations are 2, 4, 8, and 16 millimeters respectively, to generate a multi-scale feature map set. Calculate the local significance value of each feature map, use a 9×9 pixel sliding window to calculate the contrast between the pixel values within the window and the surrounding area. Set the feature fusion weight based on the significance value, and assign a larger weight to the feature map with a higher significance value. The weight range is 0.1 - 0.4. Combine the multi-scale feature maps according to the weight to generate a lesion area localization matrix.

[0062] Extract abnormal parameters from the lesion area location matrix. Calculate the average abnormal intensity of the area, with the normal reference value range being 0.2 - 0.4; calculate the boundary gradient of the area, using the Sobel operator, and set the gradient threshold to 0.3; calculate the shape complexity of the area, based on the perimeter - area ratio, and set the complexity threshold to 1.5; calculate the tissue contrast of the area, based on the gray - level co - occurrence matrix, and set the contrast threshold to 0.4. Combine these parameters into an abnormal parameter set for area segmentation. Use the region growing algorithm for segmentation, with the growth threshold being 1.5 times the local average value, to obtain the candidate lesion area.

[0063] Perform feature clustering optimization on the candidate lesion area. Calculate the feature similarity between regions, including abnormal intensity similarity, shape similarity, and position similarity. Set the similarity threshold to 0.85, and merge adjacent regions that meet the threshold. Based on the spatial position relationship between regions, establish a minimum spanning tree to determine the region connection relationship. Obtain the optimized lesion area location matrix and construct a topological structure reflecting the spatial distribution characteristics of the lesions.

[0064] Extract acoustic features from the optimized lesion area location matrix. Analyze the respiratory audio spectrum envelope, calculate the main frequency range, spectrum skewness, and harmonic ratio; analyze the energy distribution characteristics, calculate the energy proportion of each frequency band and the energy aggregation degree; analyze the sound conduction characteristics, calculate the conduction delay and attenuation characteristics. Construct an acoustic feature combination containing these features. At the same time, calculate the standardized deviation from the normal reference features to obtain the acoustic feature deviation value.

[0065] Perform feature fusion on the acoustic feature combination, deviation value, and abnormal parameter set. Use a 128 - dimensional feature vector to represent the pathological state, including 64 - dimensional acoustic features, 32 - dimensional morphological features, and 32 - dimensional position features. Calculate the distribution probability of the feature vector through kernel density estimation, with the bandwidth parameter set to 0.1, to generate the pathological state feature map.

[0066] Establish a diagnostic decision function based on the feature map. The decision function comprehensively considers the spatial consistency of the lesion distribution characteristics, the significance of the acoustic feature deviation, and clinical prior knowledge. Set the diagnostic confidence threshold to 0.8. When the probability value in the feature map exceeds the threshold, output the corresponding lesion location result and abnormal type judgment.

[0067] In clinical applications, this method has good recognition effects on common pediatric respiratory diseases. For pneumonia patients, a local abnormal intensity increase of 2 - 3 times, a 50 - 70% increase in the boundary gradient, and a significant enhancement of the spectrum energy in the range of 200 - 400 Hz can be detected. The lesion location accuracy reaches over 90%, and the abnormal type judgment accuracy exceeds 85%. For bronchitis patients, the wheezing sound characteristics in the breath sound can be identified, with a frequency range of 400 - 600 Hz and a duration greater than 200 milliseconds, and the anatomical location of the affected bronchus can be accurately located.

[0068] In this embodiment, by performing multi-scale decomposition and using Gaussian convolution kernels of different scales to extract features, it is possible to comprehensively capture the detailed information of the lesion area at each scale, improving the resolution ability for lesion features; calculating the local saliency value of each feature map and determining the feature fusion weight accordingly, realizing the adaptive weighted combination of multi-scale features, making the localization of the lesion area more accurate and robust; the generated lesion area localization matrix can quantitatively describe the characteristics of the abnormal area by analyzing the regional average abnormal intensity, boundary gradient, shape complexity, and tissue contrast, providing a basis for the fine segmentation and subsequent merging optimization of the candidate lesion area; by merging the candidate areas through feature clustering, the lesion area localization matrix is further optimized, making the overall localization result more coherent and accurate, and helping to eliminate false detections and redundant areas; integrating various acoustic information such as the respiratory audio spectrum envelope, energy distribution, and sound conduction characteristics with the abnormal parameters to construct a comprehensive acoustic feature combination, providing multi-dimensional data support for judging the abnormal type of the lesion; using kernel density estimation to calculate the distribution probability of the pathological state feature vector and establishing a diagnostic decision function, which can comprehensively consider the lesion area distribution and acoustic deviation to achieve accurate judgment of the lesion location and abnormal type, thereby outputting a preliminary diagnosis result and providing an effective auxiliary diagnosis basis for clinical practice.

[0069] In an alternative embodiment, the preliminary diagnosis result is input into a medical knowledge inference engine. A symptom evolution model is constructed through a pathological mechanism analysis module, and recursive state estimation is performed on the preliminary diagnosis result based on the symptom evolution model to generate disease development prediction data; a diagnostic optimization feedback loop is established, and the disease development prediction data is corrected through symptom association to obtain a corrected diagnosis result. Based on the corrected diagnosis result, a disease risk warning model is constructed. The disease risk warning model generates real-time monitoring data through a dynamic threshold algorithm, and combines the real-time monitoring data with the corrected diagnosis result to output the final auxiliary diagnosis result, including: Convert the preliminary diagnosis result into an input feature vector, map the input feature vector to a preset medical knowledge space through a knowledge graph mapping function to generate a knowledge node set; use a multi-head attention mechanism to calculate the inter-node association weights for the knowledge node set to form an attention weight matrix; based on the attention weight matrix, rank the importance of the knowledge nodes and extract a subset of pathological mechanism knowledge related to the current case; Input the subset of pathological mechanism knowledge into the symptom evolution model to construct a disease state vector including anatomical structure indicators, functional indicators, and symptom indicators; use the treatment response index, immune function index, stress level index, and concurrent risk index as time-varying influencing factors, and combine them with the disease state vector to construct a state transition probability matrix; perform recursive state estimation on the preliminary diagnosis result based on the state transition probability matrix, predict the state distribution at the next moment, and perform Bayesian update on the predicted state in combination with real-time observation data to generate disease development prediction data; Establish a symptom association network, use clinical symptoms as network nodes, calculate the temporal dependence intensity and co-occurrence probability between symptom nodes as edge weights to form a symptom association matrix; use the symptom association matrix to perform weighted optimization on the disease development prediction data to generate a corrected diagnosis result; train a disease risk early warning model based on the corrected diagnosis result, and transform clinical indicators, imaging features, and functional scores into the risk assessment space through a feature mapping function; Implement dynamic threshold monitoring on the monitoring indicators in the risk assessment space, including: calculating the moving average and standard deviation of the indicators, extracting the temporal feature periodic pattern, evaluating the change characteristics of the feature weights over time, updating the risk scoring function, and generating real-time monitoring data; perform deep feature fusion on the real-time monitoring data and the corrected diagnosis result, and use an attention weight mechanism to construct a comprehensive evaluation indicator; calculate the decision function value based on the comprehensive evaluation indicator, map the fused features to the diagnosis result space, and output the final auxiliary diagnosis result including the probability distribution of abnormal types, severity score, and development trend.

[0070] The time-varying influencing factors specifically refer to the key variables that have a dynamic impact during the evolution of the disease state over time. These factors will continuously adjust with the changes in the patient's physiological state, treatment process, or external environment, and have an important impact on the development trend of the disease, the risk of disease deterioration, and the treatment effect. In this solution, the time-varying influencing factors include the treatment response index, immune function index, stress level index, and concurrent risk index, which jointly act on the disease state vector to construct a more accurate state transition probability matrix, thereby improving the accuracy of disease prediction and diagnosis.

[0071] In a specific implementation, convert the preliminary diagnosis result into a 256-dimensional feature vector, including information such as lesion location, morphological features, acoustic features, and clinical manifestations. Project the feature vector into the medical knowledge space through a pre-trained knowledge graph mapping model, which contains 10,000 disease knowledge nodes. Use an 8-head attention mechanism to calculate the association intensity between nodes, with the attention head dimension being 32, to obtain the node-to-node attention weight matrix. Sort the knowledge nodes based on the weight matrix, and select the nodes with a weight sum exceeding 0.8 to form a subset of pathological mechanism knowledge.

[0072] Construct a symptom evolution model to analyze the disease development pattern. Combine anatomical structure indicators (airway diameter, wall thickness), functional indicators (airflow velocity, ventilation volume), and symptom indicators (cough frequency, abnormal breath sound) into a disease state vector. Introduce time-varying influencing factors: the treatment response index ranges from 0 to 1, the immune function index ranges from 0 to 100, the stress level index ranges from 0 to 10, and the concurrent risk index ranges from 0 to 5. Combine these factors to construct a state transition probability matrix with a dimension of 128×128.

[0073] Perform state estimation and prediction on the preliminary diagnosis results. Set the prediction time window to 24 hours and the time step to 1 hour. Predict the state distribution at the next moment based on the state transition probability matrix, and use the observed data such as real-time collected vital signs and test results for Bayesian update to generate disease development prediction data. The prediction accuracy reaches 90% within 6 hours, 85% within 12 hours, and 75% within 24 hours.

[0074] Establish a symptom association network to optimize the prediction results. Take 50 common respiratory symptoms as network nodes, calculate the temporal dependence strength and co-occurrence probability between symptoms based on 3000 clinical case data, and construct a symptom association matrix. Use the association matrix to weight and optimize the prediction data to generate a corrected diagnosis result. The optimized prediction accuracy is improved by 5 - 10 percentage points.

[0075] Train a disease risk warning model. Convert clinical indicators (blood routine, CRP, procalcitonin), imaging features (CT density value, lesion volume), and functional scores (dyspnea score, cough score) to the risk assessment space through a feature mapping function. The dimension of the mapping space is 64, and the feature mapping adopts a three-layer neural network structure.

[0076] Implement a dynamic threshold monitoring mechanism. Use a 24-hour sliding window to calculate the moving average and standard deviation of the monitoring indicators, with a window step size of 1 hour. Extract the periodic pattern of the indicator changes and evaluate the dynamic changes of the feature weights over time. Update the risk scoring function with a scoring range of 0 - 100. Generate real-time monitoring data, and the monitoring frequency can be adaptively adjusted according to the risk level. The monitoring interval is shortened to 5 minutes in the high-risk state.

[0077] Construct a comprehensive evaluation system. Perform deep feature fusion on the real-time monitoring data and the corrected diagnosis results. The fusion network adopts a residual structure and contains 6 convolutional layers. Calculate the feature weights through a bidirectional attention mechanism to generate a 128-dimensional comprehensive evaluation index. Calculate the decision function value based on the evaluation index, with the function value range of 0 - 1. Map the fused features to the diagnosis result space and output the final auxiliary diagnosis result.

[0078] In clinical applications, this system can achieve intelligent auxiliary diagnosis of children's respiratory diseases. For typical cases such as pneumonia, the system can give the probability distribution of abnormal types (bacterial 0.85, viral 0.12, other 0.03), severity scores (mild 0 - 30, moderate 31 - 70, severe 71 - 100), and predictions of the development trend in the next 24 hours. The results of clinical verification show that the diagnostic coincidence rate of the system reaches 92%, the early warning accuracy rate reaches 88%, and the accuracy rate of predicting the disease progression reaches 85%. The system can promptly detect changes in the condition and provide an objective basis for clinical decision-making.

[0079] In this embodiment, the initial diagnosis result is converted into a set of knowledge nodes in the medical knowledge space through a knowledge graph mapping function, and the multi-head attention mechanism is used to calculate the association weights between nodes, thereby accurately extracting the pathological mechanisms related to the current case and improving the medical interpretability and rationality of the diagnosis; a disease state vector is constructed by combining anatomical structure indicators, functional indicators, and symptom indicators, and dynamic influencing factors such as treatment response index and immune function index are incorporated, which can more comprehensively simulate the evolution process of the disease state and improve the accuracy of disease development prediction; a state transition probability matrix is used for recursive state estimation, and Bayesian update is combined with real-time observation data to dynamically adjust the disease prediction data, making the prediction result more in line with the individual condition changes of the patient; by constructing a symptom association network, calculating the temporal dependence and co-occurrence probability between symptoms, the weighted optimization of disease development prediction data is realized, thereby improving the accuracy of the corrected diagnosis result and the clinical guiding value; a disease risk early warning model is trained, and the clinical indicators, imaging features, and functional scores are converted into the risk assessment space by using a feature mapping function, making the disease risk assessment more systematic and data-driven; a dynamic threshold monitoring method is adopted to analyze the temporal characteristics and change trends of the monitoring indicators, and the risk scoring function is adaptively adjusted to ensure the real-time and sensitivity of the risk assessment; the comprehensive evaluation index is calculated through deep feature fusion and attention weight mechanism, and the probability distribution of abnormal types, severity score, and development trend are output based on the decision function, providing a more valuable final auxiliary diagnosis result for doctors and optimizing the clinical decision support.

[0080] Figure 2 FIG. is a schematic structural diagram of the AI-based children's respiratory diagnosis assistance system according to an embodiment of the present invention, as Figure 2 shown, the system includes: The first unit is configured to collect multi-layer breath sound signals through a bionic sensing array based on the anatomical structure of a child's respiratory system, obtain the respiratory airflow direction vector, establish an airflow trajectory model based on the respiratory airflow direction vector, drive an adaptive signal enhancer to adjust signal acquisition parameters, denoise and optimize the multi-layer breath sound signals, and generate enhanced breath sound signals; input the enhanced breath sound signals into an acoustic reconstruction network, generate three-dimensional breath sound propagation data through a spatial sound field mapping algorithm, perform dynamic segmentation processing on the three-dimensional breath sound propagation data, and generate a breath feature dataset including airflow dynamics parameters and layered acoustic parameters; The second unit is configured to input the breath feature dataset into an anatomy-guided diagnosis network, establish a respiratory function evaluation model based on physiological structure constraints, trace the propagation path of abnormal breath sounds by combining airflow dynamics parameters and layered acoustic parameters in the respiratory function evaluation model through a feature propagation algorithm, and generate an abnormal feature distribution map; calculate a lesion area localization matrix based on the abnormal feature distribution map, and the lesion area localization matrix determines abnormal parameters through multi-scale feature aggregation; extract an acoustic feature combination from the lesion area localization matrix, generate a pathological state feature map, and determine a preliminary diagnosis result; The third unit is configured to input the preliminary diagnosis result into a medical knowledge inference engine, construct a symptom evolution model through a pathological mechanism analysis module, perform recursive state estimation on the preliminary diagnosis result based on the symptom evolution model, and generate disease development prediction data; establish a diagnosis optimization feedback loop, correct the disease development prediction data through symptom association to obtain a corrected diagnosis result, construct a disease risk warning model based on the corrected diagnosis result, and the disease risk warning model generates real-time monitoring data through a dynamic threshold algorithm, combine the real-time monitoring data with the corrected diagnosis result, and output a final auxiliary diagnosis result.

[0081] In the third aspect of the embodiments of the present invention, a kind of electronic device is provided, including: a processor; a memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0082] In the fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0083] The present invention can be a method, a device, a system and / or a computer program product. The computer program product can include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-based method for assisting in the diagnosis of children's respiratory systems, characterized in that, Comprising: Based on the anatomical structure of the children's respiratory system, multi-layer respiratory sound signals are collected layer by layer through a bionic sensing array, and the respiratory airflow direction vector is obtained. An airflow trajectory model is established based on the respiratory airflow direction vector to drive an adaptive signal enhancer to adjust signal acquisition parameters, denoise and optimize the multi-layer respiratory sound signals, and generate enhanced respiratory sound signals; The enhanced respiratory sound signals are input into an acoustic reconstruction network, and three-dimensional respiratory sound propagation data is generated through a spatial sound field mapping algorithm. The three-dimensional respiratory sound propagation data is dynamically segmented to generate a respiratory feature dataset containing airflow dynamics parameters and layered acoustic parameters; The respiratory feature dataset is input into an anatomy-guided diagnosis network. A respiratory function evaluation model is established based on physiological structure constraints. The propagation path of abnormal respiratory sounds is traced by combining airflow dynamics parameters and layered acoustic parameters in the respiratory function evaluation model through a feature propagation algorithm to generate an abnormal feature distribution map; A lesion area localization matrix is calculated based on the abnormal feature distribution map. The lesion area localization matrix determines abnormal parameters through multi-scale feature aggregation; An acoustic feature combination is extracted from the lesion area localization matrix to generate a pathological state feature map and determine a preliminary diagnosis result; The preliminary diagnosis result is input into a medical knowledge inference engine. A symptom evolution model is constructed through a pathological mechanism analysis module. Recursive state estimation is performed on the preliminary diagnosis result based on the symptom evolution model to generate disease development prediction data; A diagnostic optimization feedback loop is established. The disease development prediction data is corrected through symptom association to obtain a corrected diagnosis result. A disease risk warning model is constructed based on the corrected diagnosis result. The disease risk warning model generates real-time monitoring data through a dynamic threshold algorithm. The real-time monitoring data is combined with the corrected diagnosis result to output a final auxiliary diagnosis result.

2. The method according to claim 1, wherein Based on the anatomical structure of the children's respiratory system, multi-layer respiratory sound signals are collected layer by layer through a bionic sensing array, and the respiratory airflow direction vector is obtained. An airflow trajectory model is established based on the respiratory airflow direction vector to drive an adaptive signal enhancer to adjust signal acquisition parameters, denoise and optimize the multi-layer respiratory sound signals, and generate enhanced respiratory sound signals, including: According to the anatomical hierarchical structure of the children's respiratory system, a bionic sensing array is set on the anterior chest wall, back and trachea regions; Multi-layer respiratory sound signals are collected layer by layer through the bionic sensing array; A mapping relationship matrix between the bionic sensing array and the anatomical hierarchy is established based on the distance parameter, sound wave incident angle parameter and tissue attenuation parameter between the sensing unit and the anatomical hierarchy; The multi-layer respiratory sound signals are input into a cross-correlation analysis module of adjacent sensing units to obtain a correlation function based on the signal time series relationship. By detecting the peak position of the correlation function, the sound wave propagation time delay between the adjacent sensing units is obtained. The sound wave propagation time delay is combined with the relative position relationship between the adjacent sensing units in three-dimensional space to construct a direction vector characterizing the respiratory airflow movement characteristics; Fitting the basic airflow movement characteristics in the respiratory airflow direction vector through a polynomial model to obtain a basic airflow component, and describing the vortex airflow movement characteristics in the respiratory airflow direction vector through an attenuation period model to obtain a vortex airflow component; inputting the polynomial parameters of the basic airflow component and the periodic attenuation parameters of the vortex airflow component into a least squares model, and iteratively adjusting the polynomial parameters and the periodic attenuation parameters according to the real-time measured value of the respiratory airflow direction vector to establish a dynamic model of the airflow movement trajectory; Calculating modulation parameters corresponding to signal enhancement according to the dynamic model of the airflow movement trajectory, performing a multiplication operation on the modulation parameters, the adaptive filter coefficients, and the multi-layer respiratory sound signals to generate a primary enhanced signal; updating the adaptive filter coefficients based on the error between the primary enhanced signal and the desired signal in combination with the modulation parameters, and performing noise reduction optimization processing on the multi-layer respiratory sound signals through the updated adaptive filter coefficients, and finally outputting an enhanced respiratory sound signal.

3. The method according to claim 1, wherein Inputting the enhanced respiratory sound signal into an acoustic reconstruction network, and generating three-dimensional respiratory sound propagation data through a spatial sound field mapping algorithm, including: The acoustic reconstruction network includes an encoder and a decoder. The encoder includes three layers of three-dimensional convolutional layers to extract time-domain features and spatial features, and the decoder reconstructs the feature map through a deconvolution structure; through the acoustic reconstruction network, the time-frequency features and spatial features of the enhanced respiratory sound signal are obtained; Based on the time-frequency features and the spatial features, using the Helmholtz equation to establish a sound wave propagation control equation including sound pressure and wave number, discretizing the sound wave propagation control equation through the boundary element method to obtain a sound field transfer model, calculating the sound wave propagation path between the sound source position and the measurement point, determining the Green's function matrix of the sound field transfer model, and determining the sound source intensity vector of the sound field transfer model based on the sound wave radiation energy at the sound source position; Constructing an initial sound source intensity distribution based on the time-frequency features, inputting the initial sound source intensity distribution into the sound field transfer model to obtain a calculated sound pressure, comparing the calculated sound pressure with the measured sound pressure to obtain an error vector; determining an iterative step factor through a line search method, multiplying the transpose matrix of the Green's function matrix by the error vector to obtain a gradient vector, multiplying the gradient vector by the iterative step factor to obtain an update amount, and iteratively updating the sound source intensity distribution based on the update amount. When the two-norm of the error vector is less than a preset threshold, an optimal sound source intensity distribution is obtained; Determining the grid node division in the three-dimensional space based on the spatial features, calculating the sound wave propagation path between the sound source position and the grid nodes to obtain the grid propagation Green's function, performing a spatial domain convolution operation on the optimal sound source intensity distribution and the grid propagation Green's function to obtain the complex sound pressure value of the grid nodes, and calculating the sound pressure amplitude distribution, phase distribution, and sound intensity distribution based on the complex sound pressure value to generate three-dimensional respiratory sound propagation data.

4. The method according to claim 3, wherein Performing dynamic segmentation processing on the three-dimensional respiratory sound propagation data to generate a respiratory feature dataset including aerodynamic parameters and stratified acoustic parameters includes: Dividing the three-dimensional respiratory sound propagation data into an inhalation phase, an exhalation phase, and a transition phase according to the change in the airflow direction; in the inhalation phase, the exhalation phase, and the transition phase, calculating the velocity potential based on the three-dimensional respiratory sound propagation data, calculating the airflow velocity based on the gradient of the velocity potential, calculating the vorticity based on the curl of the airflow velocity, and calculating the Reynolds stress based on the pulsating component of the airflow velocity to obtain aerodynamic parameters; Performing time-frequency analysis on the acoustical signals collected in layers to obtain the spectral energy distribution, calculating the sound pressure amplitude ratio between adjacent anatomical levels based on the sound pressure amplitude distribution to obtain the inter-layer sound pressure ratio, and calculating the sound propagation attenuation coefficient based on the sound pressure amplitude distribution to obtain stratified acoustic parameters; Combining the aerodynamic parameters and the stratified acoustic parameters into a feature vector according to time sequence, performing covariance matrix decomposition and eigenvalue sorting on the feature vector, and selecting the feature vectors corresponding to the eigenvalues with a cumulative contribution rate greater than a preset threshold to form a respiratory feature dataset.

5. The method according to claim 1, wherein Inputting the respiratory feature dataset into an anatomy-guided diagnosis network, establishing a respiratory function evaluation model based on physiological structure constraints, and tracing the propagation path of abnormal breath sounds by combining aerodynamic parameters and stratified acoustic parameters in the respiratory function evaluation model through a feature propagation algorithm to generate an abnormal feature distribution map, including: Inputting the respiratory feature dataset into an anatomy-guided diagnosis network, constructing a three-dimensional bronchial tree network topology structure based on children's thoracic CT images, extracting the positional relationship of bronchial branches and the lumen connection characteristics, calculating the anatomical position correlation matrix, and establishing a respiratory function evaluation model; In the respiratory function evaluation model, extracting the aerodynamic parameters and stratified acoustic parameters in the respiratory feature dataset, and fusing them with the anatomical position correlation matrix to generate a fused feature map; Constructing a feature propagation network based on the anatomical position correlation matrix, setting anatomical region nodes in the feature propagation network, allocating the fused feature map to each anatomical region node, calculating the node feature state, calculating the propagation coefficient between adjacent anatomical region nodes based on the node feature state, where the propagation coefficient decays with the increase of the anatomical distance, and optimizing the node feature state by iteratively updating the propagation coefficient; Comparing the optimized node feature state with a preset normal breathing reference feature, calculating the degree of abnormality of each anatomical region node, and weighting the degree of abnormality in combination with the preset anatomical structure importance to generate a spatial abnormality distribution, performing continuous processing and extreme value analysis on the spatial abnormality distribution to determine the position of the abnormal propagation center, and generating an abnormal feature distribution map based on the position of the abnormal propagation center.

6. The method according to claim 1, characterized in that Calculating a lesion area localization matrix based on the abnormal feature distribution map, where the lesion area localization matrix determines abnormal parameters through multi-scale feature aggregation; extracting an acoustic feature combination from the lesion area localization matrix to generate a pathological state feature map to determine a preliminary diagnosis result, including: Perform multi-scale decomposition on the abnormal feature distribution map, extract features using Gaussian convolution kernels of different scales to generate a multi-scale feature map set, calculate the local significance value of each feature map in the multi-scale feature map set, determine the feature fusion weight based on the local significance value, and perform weighted combination on the multi-scale feature map set according to the feature fusion weight to generate a lesion area localization matrix; Calculate the regional average abnormal intensity, regional boundary gradient, regional shape complexity, and regional tissue contrast in the lesion area localization matrix to obtain an abnormal parameter set. Based on the abnormal parameter set, perform regional segmentation on the lesion area localization matrix to determine candidate lesion areas; perform feature clustering on the candidate lesion areas, and merge areas with feature similarity meeting a preset similarity threshold to obtain an optimized lesion area localization matrix; construct a lesion area distribution feature according to the distribution positions and inter-regional connection relationships of each region in the optimized lesion area localization matrix; Extract the respiratory audio spectral envelope feature, energy distribution feature, and sound conduction feature from the optimized lesion area localization matrix to construct an acoustic feature combination; calculate the deviation degree between the acoustic feature combination and the normal reference feature to obtain an acoustic feature deviation value; perform feature fusion on the acoustic feature combination, the acoustic feature deviation value, and the abnormal parameter set to generate a pathological state feature vector; calculate the distribution probability of the pathological state feature vector by the kernel density estimation method to generate a pathological state feature map; Establish a diagnostic decision function based on the pathological state feature map. The diagnostic decision function performs a comprehensive evaluation by combining the lesion area distribution feature and the acoustic feature deviation value; output the lesion localization result and abnormal type judgment through the diagnostic decision function to determine the preliminary diagnosis result.

7. The method according to claim 1, characterized in that Input the preliminary diagnosis result into a medical knowledge inference engine, construct a symptom evolution model through a pathological mechanism analysis module, perform recursive state estimation on the preliminary diagnosis result based on the symptom evolution model to generate disease development prediction data; establish a diagnostic optimization feedback loop, correct the disease development prediction data through symptom association to obtain a corrected diagnosis result, construct a disease risk warning model based on the corrected diagnosis result, and the disease risk warning model generates real-time monitoring data through a dynamic threshold algorithm. Combine the real-time monitoring data with the corrected diagnosis result to output the final auxiliary diagnosis result, including: Convert the preliminary diagnosis result into an input feature vector, map the input feature vector to a preset medical knowledge space through a knowledge graph mapping function to generate a knowledge node set; calculate the inter-node association weight of the knowledge node set using a multi-head attention mechanism to form an attention weight matrix; rank the importance of the knowledge nodes based on the attention weight matrix, and extract a subset of pathological mechanism knowledge related to the current case; Input the subset of pathological mechanism knowledge into the symptom evolution model to construct a disease state vector including anatomical structure indicators, functional indicators, and symptom indicators; use the treatment response index, immune function index, stress level index, and concurrent risk index as time-varying influencing factors, and combine them with the disease state vector to construct a state transition probability matrix; perform recursive state estimation on the preliminary diagnosis result based on the state transition probability matrix, predict the state distribution at the next moment, and perform Bayesian update on the predicted state in combination with real-time observation data to generate disease development prediction data; Establish a symptom association network, use clinical symptoms as network nodes, calculate the temporal dependence strength and co-occurrence probability between symptom nodes as edge weights to form a symptom association matrix; use the symptom association matrix to perform weighted optimization on the disease development prediction data to generate a corrected diagnosis result; train a disease risk warning model based on the corrected diagnosis result, and transform clinical indicators, imaging features, and functional scores into the risk assessment space through a feature mapping function; Perform dynamic threshold monitoring on the monitoring indicators in the risk assessment space, including: calculating the moving average and standard deviation of the indicators, extracting the temporal feature periodic pattern, evaluating the change characteristics of the feature weights over time, updating the risk scoring function to generate real-time monitoring data; perform deep feature fusion on the real-time monitoring data and the corrected diagnosis result, and use an attention weight mechanism to construct a comprehensive evaluation indicator; calculate the decision function value based on the comprehensive evaluation indicator, map the fused features to the diagnosis result space, and output the final auxiliary diagnosis result including the probability distribution of abnormal types, severity score, and development trend.

8. An AI-based diagnostic assistance system for children's respiratory system, which is used to implement the method described in any one of the preceding claims 1-7, characterized in that, Including: The first unit is used to hierarchically collect multi-layer breath sound signals through a bionic sensing array based on the anatomical structure of the children's respiratory system, obtain the respiratory airflow direction vector, establish an airflow trajectory model based on the respiratory airflow direction vector, drive an adaptive signal enhancer to adjust the signal acquisition parameters, perform noise reduction optimization on the multi-layer breath sound signals to generate enhanced breath sound signals; input the enhanced breath sound signals into an acoustic reconstruction network, generate three-dimensional breath sound propagation data through a spatial sound field mapping algorithm, perform dynamic segmentation processing on the three-dimensional breath sound propagation data to generate a breath feature dataset including airflow dynamics parameters and hierarchical acoustic parameters; The second unit is used to input the breath feature dataset into an anatomy-guided diagnosis network, establish a respiratory function evaluation model based on physiological structure constraints, track the propagation path of abnormal breath sounds in the respiratory function evaluation model by combining airflow dynamics parameters and hierarchical acoustic parameters through a feature propagation algorithm to generate an abnormal feature distribution map; calculate a lesion area localization matrix based on the abnormal feature distribution map, and the lesion area localization matrix determines abnormal parameters through multi-scale feature aggregation; extract an acoustic feature combination from the lesion area localization matrix to generate a pathological state feature mapping and determine a preliminary diagnosis result; A third unit is configured to input the preliminary diagnosis result into a medical knowledge inference engine, construct a symptom evolution model through a pathological mechanism analysis module, perform recursive state estimation on the preliminary diagnosis result based on the symptom evolution model, and generate disease development prediction data; establish a diagnostic optimization feedback loop, correct the disease development prediction data through symptom association to obtain a corrected diagnosis result, construct a disease risk warning model based on the corrected diagnosis result, generate real-time monitoring data through a dynamic threshold algorithm by the disease risk warning model, and combine the real-time monitoring data with the corrected diagnosis result to output a final auxiliary diagnosis result.

9. An electronic device, characterized in that, It includes: a processor; a memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

Images