Intelligent diagnosis and treatment integrated system for respiratory diseases

Abstract

The invention discloses an intelligent diagnosis and treatment integrated system for respiratory diseases, and the system comprises a data collection module which introduces a hospital information system to record clinical data, medical images and physiological signals of a to-be-detected object; the edge calculation module is used for carrying out filtering processing on the physiological signal and carrying out noise reduction and lung field segmentation processing on the medical image; the multi-modal fusion diagnosis module is used for generating disease prediction types, severity and treatment suggestions according to the clinical data, the medical images and the physiological signals, and introducing an analysis module to generate a diagnosis report; the application service module is used for providing diagnosis and treatment feedback for the doctor end, the patient end and the management end and supporting the model optimization module to dynamically optimize the multi-modal fusion diagnosis module; according to the invention, multi-modal data is deeply fused through the multi-modal fusion diagnosis module, so that the limitation of single data is overcome; therefore, the method and the device play a role in improving the diagnosis accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a respiratory disease intelligent diagnosis and treatment integrated system. BACKGROUND

[0002] Traditional respiratory disease diagnosis and treatment relies on the experience of doctors, and is prone to misdiagnosis and missed diagnosis due to the similarity of symptoms between diseases (such as cough and dyspnea). With the rapid development of artificial intelligence and medical Internet of Things technology, intelligent diagnosis and treatment has become an important development direction of modern medicine. However, the existing respiratory disease intelligent diagnosis and treatment system has the following limitations:

[0003] 1. Lack of diagnosis and treatment accuracy: Chinese patent application No. CN201711097135.4 provides a respiratory disease intelligent diagnosis and treatment system and control method, including: a server, the server includes one or more processors configured to execute computer program modules, characterized in that: the computer program modules include: a diagnosis publishing system module, an asthma management scheme execution system module, and a user device access authentication system module; the present application seamlessly connects the respiratory disease diagnosis / treatment scheme and the treatment scheme execution, realizes the positive feedback and circulation of (doctor) diagnosis→treatment (patient self-management) and treatment (patient self-management)→(doctor) diagnosis in asthma disease treatment, improves the control and cure efficiency, and avoids the disease recurrence or even aggravation caused by ineffective asthma self-management; however, it only focuses on asthma management closed loop, does not integrate image and physiological signal, and often misdiagnosis occurs;

[0004] 2. Lack of dynamic optimization: the existing diagnosis and treatment system relies on hardware modular design and lacks deep learning driven multi-modal analysis, which cannot continuously optimize the model according to the doctor's correction opinion. SUMMARY

[0005] The technical problems solved by the present application are: 1. Lack of diagnosis and treatment accuracy: single data source cannot fully capture the complex characteristics of respiratory diseases, leading to misdiagnosis and missed diagnosis; 2. Lack of dynamic optimization: the system relies on hardware modular design and lacks deep learning driven multi-modal analysis, which cannot continuously optimize the model according to the doctor's correction opinion.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] A respiratory disease intelligent diagnosis and treatment integrated system, comprising:

[0008] A data acquisition module introduces the clinical data, medical images and physiological signals of the object to be measured recorded by the hospital information system;

[0009] An edge computing module filters the physiological signals and performs denoising and lung field segmentation processing on the medical images;

[0010] A multi-modal fusion diagnosis module generates disease prediction types, severity and treatment suggestions according to the clinical data, medical images and physiological signals, and introduces an analysis module to generate a diagnosis report;

[0011] An application service module is configured to provide diagnosis and treatment feedback to a doctor end, a patient end and a management end, and support a model optimization module to dynamically optimize the multi-modal fusion diagnosis module.

[0012] As a preferred scheme of the intelligent diagnosis and treatment integrated system for respiratory diseases, the data acquisition layer comprises:

[0013] A medical record acquisition unit is configured to acquire clinical data of a patient; the clinical data comprises medical history text data, physical examination numerical values and laboratory reports;

[0014] An image acquisition unit is configured to acquire medical images of the patient; the medical images comprise chest X-ray images and CT scan images;

[0015] A wearable device unit is configured to acquire physiological signals of the patient; the physiological signals comprise respiratory frequency and respiratory waveform.

[0016] As a preferred scheme of the intelligent diagnosis and treatment integrated system for respiratory diseases, the edge computing layer comprises:

[0017] A physiological signal processing unit is configured to introduce a band-pass filter to filter the physiological signals;

[0018] An image preprocessing unit is configured to construct a lung field segmentation model, and perform noise reduction and lung field segmentation processing on the medical images based on a computer vision library.

[0019] As a preferred scheme of the intelligent diagnosis and treatment integrated system for respiratory diseases, the multi-modal fusion diagnosis module comprises:

[0020] A data set construction unit is configured to construct a multi-source data set according to the clinical data, medical images and physiological signals;

[0021] A data processing unit comprises a clinical data processing subunit, a medical image processing subunit and a physiological signal processing subunit; wherein,

[0022] The clinical data processing subunit is configured to perform standardization processing on the medical history text data to generate text feature quantities, and perform normalization processing on the physical examination numerical values to generate numerical feature quantities; the clinical data processing subunit constructs a fusion model, and extracts clinical features according to the text feature quantities and the numerical feature quantities;

[0023] The medical image processing subunit extracts multi-scale features from the medical image based on a 3D convolution network model, constructs a feature aggregation layer, and generates image features according to the multi-scale features;

[0024] The physiological signal processing subunit constructs a continuous wavelet transform model, processes the respiratory frequency to obtain a time-frequency matrix, constructs a time series convolution network model, extracts respiratory fluctuation feature quantities according to the respiratory waveform, and generates time series features according to the time-frequency matrix and the respiratory fluctuation feature quantities.

[0025] The data fusion unit constructs a multi-modal fusion model, fuses the clinical features, image features and time series features to generate fusion features, constructs a diagnosis model, and generates disease prediction types, severity and treatment suggestions according to the fusion features.

[0026] As a preferred scheme of the respiratory disease intelligent diagnosis and treatment integrated system, the multi-modal fusion model is constructed based on an early fusion model and a late fusion stage model; wherein,

[0027] The early fusion model adopts a cross-modal Transformer model, uses a multi-head cross-attention mechanism, and generates correlation weights according to the clinical features and image features.

[0028] The late fusion model adopts a graph neural network model, constructs a modal relationship graph, determines the weight between nodes through a dynamic correlation coefficient, and outputs fusion features based on a graph convolution layer.

[0029] As a preferred scheme of the respiratory disease intelligent diagnosis and treatment integrated system, the diagnosis model is constructed based on a multi-task output model and an analysis model; wherein,

[0030] The multi-task output model synchronously outputs disease prediction types, severity and treatment suggestions.

[0031] The analysis model integrates a SHAP analyzer to calculate the contribution of each modal feature to the diagnosis result, and integrates a report generator to generate a diagnosis report and write the diagnosis report into a hospital information system.

[0032] As a preferred scheme of the respiratory disease intelligent diagnosis and treatment integrated system, the model optimization module comprises:

[0033] The reinforcement model unit records the changes of classification confidence of prediction types by collecting doctor correction data and patient monitoring data, generates training samples, classifies difficult samples that are difficult for the multi-task output model to stably predict, and triggers parameter update of the multi-task output model.

[0034] As a preferred scheme of the intelligent diagnosis and treatment integrated system for respiratory diseases, the model optimization module further comprises:

[0035] The meta-learning framework unit optimizes the initial parameters of the model on a few-shot task based on the MAML framework, and improves the capturing ability of new disease characteristics.

[0036] As a preferred scheme of the intelligent diagnosis and treatment integrated system for respiratory diseases, the application service module comprises:

[0037] The doctor end is used for displaying the diagnosis result and is integrated into a hospital information system; and the correction data is generated according to the contribution of each modality feature to the diagnosis data, and is input into the model optimization module;

[0038] The patient end is internally provided with a symptom sign association engine, and actively reminds the patient based on an abnormal physiological signal detection algorithm.

[0039] The management end is used for epidemic trend early warning, and dynamically adjusts the early warning threshold based on historical baseline.

[0040] An intelligent diagnosis and treatment method for respiratory diseases is applied to the above-mentioned intelligent diagnosis and treatment integrated system for respiratory diseases, and comprises the following steps:

[0041] S1. The clinical data, medical images and physiological signals of the patient are acquired through the data acquisition module;

[0042] S2. The edge computing module performs filtering processing on the physiological signals, and performs noise reduction and lung field segmentation processing on the medical images;

[0043] S3. The multi-modal fusion diagnosis module extracts fusion features, and generates disease prediction types, severity and treatment suggestions according to the fusion features;

[0044] S4. The diagnosis and treatment feedback is output through the application service module, and the parameters of the multi-modal fusion diagnosis module are dynamically optimized based on the model optimization module.

[0045] The beneficial effects of the present application are as follows:

[0046] Firstly, the present application provides an intelligent diagnosis and treatment integrated system for respiratory diseases, which deeply fuses multi-modal data through a multi-modal fusion diagnosis module, and overcomes the limitations of single data. At the same time, the cross-modal attention mechanism establishes the association between clinical data and medical images, and the graph neural network model dynamically fuses multi-source features, so that the model can accurately distinguish similar symptom diseases in complex cases. Therefore, the present application plays a role in improving the accuracy of diagnosis.

[0047] Secondly, the application provides a respiratory disease intelligent diagnosis and treatment integrated system. Through an edge computing module, physiological signals and medical images are locally preprocessed to reduce cloud transmission load. Meanwhile, a band-pass filter and a lung field segmentation model are used to optimize calculation efficiency to meet the second-level response demand in an emergency scene, especially in a resource-limited environment of primary medical institutions. Therefore, the application plays a role in guaranteeing real-time response.

[0048] Thirdly, the application provides a respiratory disease intelligent diagnosis and treatment integrated system. Through a closed-loop incremental learning mechanism of a reinforcement model unit, doctor's correction opinions are converted into training samples to dynamically optimize model parameters. Meanwhile, through a MAML framework, an meta-learning framework unit quickly optimizes initial model parameters based on a few sample tasks to adapt the system to new diseases and solve the bottleneck of traditional models relying on big data training. The system can be continuously iterated with clinical practice to maintain long-term effectiveness. Therefore, the application plays a role in dynamically evolving the system. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 A flowchart of a respiratory disease intelligent diagnosis and treatment integrated system provided by an embodiment of the application is shown in the figure.

[0050] Figure 2 A flowchart of a diagnosis model in a respiratory disease intelligent diagnosis and treatment integrated system provided by an embodiment of the application is shown in the figure.

[0051] Figure 3 A deployment flowchart of a respiratory disease intelligent diagnosis and treatment integrated system provided by an embodiment of the application is shown in the figure.

[0052] Figure 4 A structural diagram of a respiratory disease intelligent diagnosis and treatment integrated system provided by an embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0053] In order to make the above-mentioned purposes, features and advantages of the application more obvious and easy to understand, the specific embodiments of the application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments.

[0054] Respiratory disease diagnosis and treatment rely on doctor's experience, and misdiagnosis and missed diagnosis are easily caused due to the similarity of symptoms between diseases (such as cough and dyspnea). With the rapid development of artificial intelligence and medical Internet of Things technology, intelligent diagnosis and treatment has become an important development direction of modern medicine. However, the existing respiratory disease intelligent diagnosis and treatment system only focuses on asthma management closed loop, does not integrate images and physiological signals, and often causes misdiagnosis. Meanwhile, it excessively relies on hardware modular design and lacks deep learning driven multi-modal analysis, and cannot continuously optimize the model according to doctor's correction opinions.

[0055] To solve the above problems, with reference to Figure 1 To solve the above problems, with reference to

[0056] Specifically, the data acquisition layer includes: a medical record acquisition unit that acquires clinical data of a patient; the clinical data includes medical history text data, physical examination values, and laboratory reports; an image acquisition unit that acquires medical images of the patient; the medical images include chest X-ray images and CT scan images; and a wearable device unit that acquires physiological signals of the patient; the physiological signals include respiratory rate and respiratory waveform.

[0057] The acquisition of clinical data covers the entire cycle of diagnosis and treatment records of outpatients and inpatients in the respiratory department. The text data is divided into structured fields and unstructured fields, the structured fields include smoking history, occupational exposure history, and medication records. The unstructured field text is subjected to entity recognition using a clinical natural language processing tool (CLAMP), and descriptive signs such as "barrel chest" are mapped to SNOMED CT concept IDs. At the same time, the data acquisition of medical images needs to comply with the ACR technical standard, the CT scan layer thickness is not greater than 1mm, and the contrast-enhanced phase contains dual-phase data of arterial phase and delay phase. In addition, the physiological signal data acquisition device needs to pass FDA Class II certification; the respiratory waveform sampling rate is not less than 200Hz, and the body position change event marker is recorded synchronously. Synchronously, patient report data is generated, which is collected through a REDCap spreadsheet and includes weekly scores of the mMRC dyspnea scale and an acute exacerbation event diary.

[0058] Further, the edge computing layer provided by the present application includes: a physiological signal processing unit that introduces a band-pass filter to filter the physiological signals; an image preprocessing unit that constructs a lung field segmentation model and performs noise reduction and lung field segmentation processing on the medical images based on a computer vision library.

[0059] Specifically, medical images need to be annotated, and the specific annotation work is performed by a team of respiratory specialists. Each case is independently annotated, and the inter-group consistency is evaluated by Kappa coefficient. Disagreement cases are resolved through multidisciplinary consultation. At the same time, the disease label adopts ICD-11 extended coding, including phenotype typing (such as Th2-high subtype of asthma). Among them, the image annotation includes pixel-level segmentation label (ITK-SNAP tool draws the emphysema region) and sign description label (such as "tree bud sign", "mosaic perfusion"). At this time, the data enhancement strategy adopts a conditional diffusion model to generate synthetic CT images for small sample diseases, and controls the latent space vector to ensure the anatomical rationality of the lesion.

[0060] Specifically, the physiological signal needs to be preprocessed. Among them, the time series signal data is expanded by wavelet interpolation through a Wavenet variational autoencoder, and the frequency domain feature power spectrum distribution remains unchanged.

[0061] In the respiratory disease intelligent diagnosis and treatment integrated system provided by the application, the multi-modal fusion diagnosis module comprises a data set construction unit which constructs a multi-source data set according to clinical data, medical images and physiological signals. Among them, the data processing unit comprises a clinical data processing subunit, a medical image processing subunit and a physiological signal processing subunit.

[0062] Regarding the preprocessing of clinical data, the clinical data processing subunit proposed in the application standardizes the medical history text data to generate text feature quantities, and normalizes the physical examination numerical values to generate numerical feature quantities; the clinical data processing subunit constructs a fusion model, and extracts clinical features according to the text feature quantities and the numerical feature quantities.

[0063] Specifically, for the mixed structured text and numerical data in the clinical data, medical terminology standardization and statistical normalization double-channel processing are adopted. The medical history text data is mapped to standard concept codes through the UMLS medical ontology library, eliminating the description difference; the physical examination data is normalized through quantile normalization to eliminate dimensional interference. Specifically, the original clinical text data needs to be input into the system and the numerical matrix . The term standardization is performed through the UMLS Metathesaurus to generate a concept unique identifier vector , wherein represents the CUI code of the i th clinical entity in the UMLS.

[0064] At this time, the numerical features are processed by quantile normalization to obtain the parameters :

[0065] ;

[0066] Among them, is the inverse function of the standard normal distribution, is the sample size.

[0067] Final features are obtained by cross-modal attention mechanism fusion, and clinical features are obtained

[0068]

[0069] Among them, the attention head output dimension is set to 512 dimensions aligned with the image feature space.

[0070] Regarding the preprocessing of medical images, the medical image processing subunit proposed in the application extracts multi-scale features from medical images based on a 3D convolutional network model, and constructs a feature aggregation layer to generate image features according to the multi-scale features.

[0071] Specifically, based on the data of the registered three-dimensional CT volume (CT scan image), a deep supervision 3D convolutional architecture is used to realize multi-scale lesion feature extraction. The lung field segmentation model incorporates lung field segmentation prior knowledge in four decoding stages based on a 3D convolutional network model, and the geometric consistency of the feature map and the true lesion area is constrained by the Dice loss. Among them, the feature aggregation layer uses generalized mean pooling (GeM) to adaptively adjust the local feature weight, highlighting the contribution degree of ground glass shadow and bronchiectasis key signs. The finally output image feature not only retains the spatial structure information but also has lesion specificity, and completes the lung field segmentation processing.

[0072] First, the registered CT volume data (CT scan image) is input , and multi-scale features are extracted by a 3D convolution kernel ( ) :

[0073] ;

[0074] Then, a deep supervision strategy is used to inject lesion area prior knowledge in four decoder stages. Among them, the four decoder stages inject anatomical prior knowledge of lung lesions into the feature learning process through a deep supervision strategy, establish spatial correspondence between abstract features extracted by the convolutional network and the true lesion area labeled by doctors at multiple scales, and guide the network to focus on the imaging findings with clinical diagnostic significance through a Dice loss function. Gradual refinement from low-level visual features to high-level semantic features is realized, and the output image feature is ensured to retain complete spatial context information:

[0075] ;

[0076] Among them, ​​to focus on imaging findings with clinical diagnostic significance, to the down-sampled real segmentation mask, to the hierarchical weight, to the Dice similarity coefficient loss function.

[0077] Specifically, based on multi-scale features, the feature projection uses a GeM pooling layer, and then the image features :

[0078] ;

[0079] wherein the local feature aggregation strength is controlled by a learnable parameter , is the pooling region (e.g., the entire feature map or a sub-region), is is each spatial position in

[0080] Regarding the preprocessing of the physiological signal, the physiological signal processing subunit proposed in the present application constructs a continuous wavelet transform model to process the respiratory frequency, obtains a time-frequency matrix, constructs a time series convolution network model, and extracts respiratory fluctuation feature quantities according to the respiratory waveform; and generates time series features according to the time-frequency matrix and the respiratory fluctuation feature quantities.

[0081] Specifically, the diagnosis and treatment system proposed in the present application constructs a time-frequency joint analysis pipeline in view of the non-stationary characteristics of the respiratory waveform. The continuous wavelet transform model converts the time-domain signal into a time-frequency matrix based on continuous wavelet transform, and the Morlet wavelet basis function parameter is set according to the optimization of respiratory physiological characteristics. The time series convolution network model is based on a multi-scale time series convolution network (TCN), which captures short-term fluctuations and long-period patterns of respiratory rhythm through differentiated dilated convolution.

[0082] First, the preprocessed respiratory waveform is input to construct a time-frequency joint representation :

[0083] ;

[0084] wherein, is the preprocessed respiratory waveform, the center frequency of the Morlet wavelet basis function is set to 1.5 times the respiratory fundamental frequency, is the time position.

[0085] Next, the time series modeling adopts a multi-scale TCN architecture:

[0086] ;

[0087] wherein, is an expanded convolution layer, is a time step of the feature representation.

[0088] Then, the expansion factor is set according to an exponential growth rule to cover the complete respiratory cycle. Wherein, the feature dimension reduction is realized by differentiable pooling to obtain the converted feature representation of the time step :

[0089] ;

[0090] wherein, the projection matrix matches the dimension of the feature space of other modalities, is a soft pooling function.

[0091] Finally, the time sequence feature is obtained.

[0092] In the intelligent diagnosis and treatment integrated system for respiratory diseases proposed by the present application, as shown in Figure 2 , a data fusion unit is constructed to build a multi-modal fusion model, to fuse clinical features, image features and time sequence features to generate fusion features, to build a diagnosis model, and to generate disease prediction types, severity and treatment suggestions according to the fusion features.

[0093] In the intelligent diagnosis and treatment integrated system for respiratory diseases proposed by the present application, the multi-modal fusion model is constructed based on an early fusion model and a late fusion stage model; wherein, the early fusion model adopts a cross-modal Transformer model, uses a multi-head cross-attention mechanism, and generates correlation weights according to clinical features and image features; the late fusion model adopts a graph neural network model, constructs a modal relationship graph, determines the weight between nodes through a dynamic correlation coefficient, and outputs fusion features based on a graph convolution layer.

[0094] Specifically, first, input clinical features , image features and physiological signal features , and adopt a two-stage hybrid fusion strategy.

[0095] Next, the early fusion model maps heterogeneous features to a unified space through cross-modal embedding alignment, uses a multi-head cross-attention mechanism to calculate the correlation weight between clinical text and image regions, and the specific model is as follows:

[0096] ;

[0097] wherein, / is a trainable projection matrix,​ is a feature dimension.

[0098] Then, the late fusion model constructs a modal relation graph GG=(V,E), a node vi∈V represents a single modal feature, and an edge eij∈E is weighted by a dynamic correlation coefficient, and the specific model is as follows:

[0099]

[0100] wherein, is a prior inter-modal correlation, λ is a modal prior correlation strength (such as the physiological coupling of lung sound and blood oxygen), is an indicator function, is a node feature and is a standard deviation, is a node feature and is a covariance, is an edge weight between node features and .

[0101] Finally, the fusion feature is output through a graph convolution layer :

[0102]

[0103] wherein, is a graph convolution weight matrix, denotes a neighbor node set.And the multi-modal fusion model is realized by PyTorch Geometric and supports dynamic graph structure learning.

[0104] In the intelligent diagnosis and treatment integrated system for respiratory diseases provided by the application, a diagnosis model is constructed based on a multi-task output model and an analysis model; wherein the multi-task output model synchronously outputs disease prediction types, severity and treatment suggestions; the analysis model integrates a SHAP analyzer to calculate the contribution of each modal feature to the diagnosis result; and a report generator is integrated to generate a diagnosis report and write the diagnosis report into a hospital information system.

[0105] Specifically, the diagnosis model adopts a multi-task learning architecture to synchronously output disease type prediction, severity grading and treatment suggestions. Among them, the classification task of disease type prediction is constructed based on a hierarchical label system, which first distinguishes between infectious and non-infectious diseases, and then refines to specific pathogen types. The evaluation module of severity grading combines the quantitative standards defined in the clinical guidelines (such as respiratory rate and oxygenation index) to output a continuous risk score. The generator of treatment suggestions matches the medication plan and nursing points that meet the current patient characteristics from the structured knowledge base through retrieval enhancement technology.​

[0106] Specific multi-task learning architecture, first build multi-task output head, disease type prediction classification task uses hierarchical label encoding (K is the number of disease categories), severity regression task output . The treatment recommendation generation module retrieves key passages from the clinical guideline text library based on the fused features through the pointer network:

[0107] ;

[0108] Wherein, is the alignment matrix, is the passage selection probability distribution output by the pointer network, is the normalized exponential function.

[0109] The analysis model integrates the SHAP analyzer through the explainability module to calculate the contribution of the modal features :

[0110] ;

[0111] Wherein, is the set of all modalities, is the model prediction function, is any subset of modalities not containing the target modality, is the new subset after adding the target modality to the subset . The report generator writes the key evidence of the diagnosis into a JSON format diagnosis for the EMR system (hospital information system, intelligent diagnosis and treatment system) to call.

[0112] In the respiratory disease intelligent diagnosis and treatment integrated system proposed by the application, the model optimization module comprises: a reinforcement model unit, which records the classification confidence change of the predicted type by collecting doctor correction data and patient monitoring data, and generates training samples, classifies difficult samples that are difficult for the multi-task output model to predict stably, and triggers parameter update of the multi-task output model. It also includes: a meta-learning framework unit that optimizes the initial parameters of the model on a few-shot task based on the MAML framework to improve the ability to capture new disease characteristics. Specifically, there are two optimization processes:

[0113] First, focus on difficult samples:

[0114] ​To solve the class imbalance problem in the respiratory disease data, a dynamic sample re-weighting mechanism is designed to construct a reinforced model unit. The reinforced model unit monitors the changes in the classification confidence of the samples in real time during the training process and automatically identifies difficult samples that are difficult for the model to predict stably. For these samples, the model learns the key discriminative features by adaptively adjusting the loss function weights.

[0115] Specifically, the sample difficulty coefficient is defined as The cross-entropy of the model output probability and the true label is dynamically weighted :

[0116] ;

[0117] wherein, is a smoothing factor, is the prediction value at the tthiteration, is the cross-entropy loss function.

[0118] Improved Focal Loss function is:

[0119] ;

[0120] wherein is the prediction probability of class c, β is the difficulty scaling coefficient, and the difficult sample detector is The top 20% of samples trigger the feature enhancement pipeline, is the natural logarithm function.

[0121] Second, use the meta-learning framework for optimization:

[0122] To improve the adaptability of the model to new respiratory diseases, the meta-learning framework unit in the diagnosis and treatment system provided in the present application uses the meta-learning paradigm to construct the model training framework. First, in the internal optimization loop, the meta-learning paradigm model learns to extract cross-disease discriminative features by quickly fine-tuning on a few-shot task of each disease. The external loop optimizes the initial parameters of the model to enable it to quickly capture new disease feature patterns from limited samples. At the same time, the MAML framework is used to perform second-order optimization on the support set and the query set .

[0123] Specifically, the inner loop quickly adapts to the new disease sub-task:

[0124] ;

[0125] wherein, is the adapted parameter, is the initial parameter of the model (meta parameter), For the inner loop learning rate, For parameters The model prediction function, For initial parameters Find the gradient. This is the loss function.

[0126] Specifically, the outer loop updates the meta-parameters:

[0127] ;

[0128] in, The learning rate for the outer loop. For the adapted task-specific model, For loss function, For the first A query set for each task.

[0129] Next, the task sampler constructs an N-way K-shot task (such as the identification of 5 types of respiratory viruses, 3 cases for each type) from the epidemiological database, uses the Higher library to implement the higher-order differential of the gradient, and accurately calculates the second derivative of the model parameters with respect to the loss of the few-sample task in the inner loop of meta-learning, so that the model can quickly adapt to new disease features with a small number of training samples.

[0130] The integrated intelligent diagnosis and treatment system for respiratory diseases proposed in this invention includes the following application service modules: a doctor's end, which displays diagnostic results and is integrated into the hospital information system; and generates corrected data based on the contribution of each modality feature to the diagnostic data and inputs it into the model optimization module; a patient's end, which has a built-in symptom and sign association engine and actively reminds patients based on an abnormal physiological signal detection algorithm; and a management end, which is used for epidemic trend early warning and dynamically adjusts the early warning threshold based on historical baselines.

[0131] Specifically, the doctor's workstation integrates a model interpretation component, which uses quantitative methods to evaluate the contribution of each modal feature to the diagnostic results and displays it as a heatmap overlaid on the original image. The patient's application has a built-in symptom-sign correlation engine that automatically triggers an abnormal respiratory waveform monitoring algorithm when specific symptoms are reported. The management system deploys an epidemiological early warning module that uses statistical methods to detect disease clustering trends, and the early warning threshold is dynamically adjusted based on historical baselines.

[0132] like Figure 3 and 4 As shown, the deployment and implementation process of the intelligent diagnosis and treatment system involved in this application includes:

[0133] First, edge deployment: In the local environment of medical institutions, a lightweight inference engine is deployed to receive real-time data streams from the hospital information system, including continuously monitored physiological parameters, periodically collected medical images, and clinician-entered clinical observation records. At the same time, to ensure that the system response speed meets the real-time requirements of clinical practice, model quantization and pruning techniques are used to optimize computational efficiency, and the core discriminative ability of the multi-modal fusion module is preserved. The input interface is compatible with the medical industry standard data protocol, enabling seamless integration with existing medical equipment. The output module generates a structured diagnostic recommendation document, supporting deep integration with the electronic medical record system to ensure the coherence of diagnostic information in the workflow.

[0134] Second, closed-loop feedback: First, a continuous model optimization mechanism is established to automatically trigger model iteration and update by collecting the difference between actual clinical diagnosis and system recommendations. Further, when the doctor modifies the system's diagnostic conclusion, the difference analysis module extracts key correction basis (such as new laboratory test results or imaging features) and generates labeled training samples accordingly. Moreover, the model update process uses a flexible parameter adjustment strategy to absorb new knowledge while avoiding damage to existing diagnostic capabilities. At the same time, the system also integrates a clinical warning function that automatically triggers a hierarchical alarm mechanism when the patient's condition meets predefined critical criteria, ensuring that abnormal situations are promptly addressed. The entire closed-loop process is strictly controlled through version control and effectiveness evaluation to ensure the reliability and transparency of system evolution.

[0135] The application also provides a respiratory disease intelligent diagnosis and treatment method, which is applied to the above-mentioned integrated system for respiratory disease intelligent diagnosis and treatment, and includes the following steps:

[0136] S1. Obtain the clinical data, medical images, and physiological signals of the patient through the data acquisition module;

[0137] S2. The edge computing module filters the physiological signals and performs noise reduction and lung field segmentation on the medical images;

[0138] S3. The multi-modal fusion diagnosis module extracts fusion features and generates disease prediction types, severity, and treatment recommendations based on the fusion features;

[0139] S4. Output the diagnosis and treatment feedback through the application service module, and dynamically optimize the parameters of the multi-modal fusion diagnosis module based on the model optimization module.

[0140] The application also provides experimental data for the diagnosis system:

[0141] This experiment is based on the real clinical environment of the respiratory department of a certain third-grade hospital. In the past six months, multi-modal diagnosis and treatment data of respiratory disease patients have been continuously collected, including:

[0142] (1) Clinical data: medical history records (structured text), symptom description (unstructured text), laboratory examination results (blood routine, inflammation indicators, etc.);

[0143] (2) Image data: chest X-ray (DICOM format, 512x512 resolution), CT (slice thickness≤1mm);

[0144] (3) Physiological signals: respiratory rate (Hz), blood oxygen saturation (SpO2), lung sound waveform (sampling rate 100Hz).

[0145] After data collection, the training set and the test set are divided in the ratio of 8:2, and the following two tasks are verified:

[0146] 1. Multimodal disease classification task: the classification performance of the model on respiratory diseases (asthma, COPD, pneumonia, interstitial lung disease) is evaluated by accuracy (Accuracy), F1-Score (macro average) and AUC-ROC, and the experimental results of this task are shown in Table 1.

[0147] Table 1 Experimental results of the multimodal disease classification task

[0148] Disease type Accuracy (%) F1-Score AUC-ROC Asthma 92.4 0.91 0.97 COPD 88.7 0.86 0.93 Pneumonia 94.2 0.93 0.98 Interstitial lung disease 85.6 0.82 0.91

[0149] 2. Abnormal physiological signal detection task: for abnormal events (such as apnea, hypoxemia, abnormal lung sound) in respiratory rate, blood oxygen saturation (SpO2) and lung sound signals, the recall (Recall) and precision (Precision) are used to evaluate the detection ability of the model, and the experimental results of this task are shown in Table 2.

[0150] Table 2 Experimental results of the abnormal physiological signal detection task

[0151] Abnormality type Recall (%) Precision (%) Apnea 95.2 93.8 Hypoxemia (SpO2 < 90%) 89.6 88.3 Abnormal lung sounds (wheeze / wet rales) 91.4 90.1

[0152] The experimental results show that the multimodal fusion model of the method exhibits excellent performance in the disease classification task, and the abnormal physiological signal detection has precision. It is confirmed that the present application has practical value in intelligent diagnosis and treatment of respiratory system diseases, and can effectively assist doctors in comprehensive judgment.

[0153] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, a system or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (or computer- readable storage media) having computer-usable program code embodied in the medium. The medium can be any available medium or combination thereof that is accessible by a general purpose or special purpose computer. By way of example, such computer-usable storage media can include a volatile memory, a non-volatile memory, a hard disk drive, a solid state drive, a magnetic diskette, an optical disk (e.g., a compact disk (CD) or a digital versatile disk (DVD)), or a floppy disk, all of which are tangible of computer- readable media. The computer-usable program code, when executed, can enable a general purpose computer, special purpose computer, or other programmable data processing apparatus to perform a method, process, or system specified in flowcharts of the figures. The computer program product can be deployed in a distributed system and can be stored on or transmitted between one or more computer-usable storage media of the distributed system. The computer-usable program code can be implemented in any of a variety of programming languages, including but not limited to C, C++, Java, JavaScript, or Visual Basic, among others. Figure 1 one or more flows and / or blocks Figure 1 one or more flows and / or blocks

[0154] It should be noted that the above-mentioned embodiments are only used to illustrate but not to limit the technical solutions of the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalent replaced without departing from the spirit and scope of the technical solutions of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A respiratory disease intelligent diagnosis and treatment integrated system, characterized in that, The application comprises: a data acquisition module that introduces a hospital information system to record clinical data, medical images, and physiological signals of a test subject; an edge computing module that filters the physiological signals and performs denoising and lung field segmentation on the medical images; a multi-modal fusion diagnosis module that generates disease prediction types, severity, and treatment recommendations based on the clinical data, medical images, and physiological signals, and introduces an analysis module to generate a diagnosis report; an application service module that provides diagnosis feedback to a doctor terminal, a patient terminal, and a management terminal, and supports a model optimization module to dynamically optimize the multi-modal fusion diagnosis module.

2. The intelligent diagnosis and treatment integrated system for respiratory diseases according to claim 1, characterized in that, The data acquisition layer comprises: a medical record acquisition unit that acquires clinical data of a patient; the clinical data includes medical history text data, physical examination values, and laboratory reports; an image acquisition unit that acquires medical images of a patient; the medical images include chest X-ray images and CT scan images; a wearable device unit that acquires physiological signals of a patient; the physiological signals include respiratory rate and respiratory waveform. 3.The intelligent diagnosis and treatment integrated system for respiratory diseases of claim 1, wherein, The edge computing layer comprises: a physiological signal processing unit that introduces a band-pass filter to filter the physiological signals; an image preprocessing unit that constructs a lung field segmentation model and performs denoising and lung field segmentation on the medical images based on a computer vision library.

4. The intelligent diagnosis and treatment integrated system for respiratory diseases according to claim 1, characterized in that, The multi-modal fusion diagnosis module comprises: a data set construction unit that constructs a multi-source data set based on the clinical data, medical images, and physiological signals; a data processing unit that includes a clinical data processing subunit, a medical image processing subunit, and a physiological signal processing subunit; wherein, the clinical data processing subunit performs standardization processing on the medical history text data to generate text feature quantities and normalization processing on the physical examination values to generate numerical feature quantities; the clinical data processing subunit constructs a fusion model and extracts clinical features based on the text feature quantities and numerical feature quantities; the medical image processing subunit extracts multi-scale features from the medical images based on a 3D convolutional network model and constructs a feature aggregation layer to generate image features based on the multi-scale features; the physiological signal processing subunit constructs a continuous wavelet transform model to process the respiratory rate and obtain a time-frequency matrix, constructs a time series convolutional network model to extract respiratory fluctuation feature quantities based on the respiratory waveform, and generates time series features based on the time-frequency matrix and respiratory fluctuation feature quantities; a data fusion unit that constructs a multi-modal fusion model to fuse the clinical features, image features, and time series features to generate fusion features, constructs a diagnosis model, and generates disease prediction types, severity, and treatment recommendations based on the fusion features.

5. The intelligent diagnosis and treatment integrated system for respiratory diseases according to claim 4, characterized in that, The multi-modal fusion model is constructed based on an early fusion model and a late fusion stage model; wherein, the early fusion model uses a cross-modal Transformer model and uses a multi-head cross-attention mechanism to generate correlation weights based on the clinical features and image features; the late fusion model uses a graph neural network model to construct a modal relationship graph, determines the weights between nodes through a dynamic correlation coefficient, and outputs fusion features based on a graph convolution layer.

6. The intelligent diagnosis and treatment integrated system for respiratory diseases according to claim 4, characterized in that, The diagnosis model is constructed based on a multi-task output model and an analysis model; wherein, A multi-task output model outputs disease prediction type, severity and treatment suggestion simultaneously; An analysis model integrates a SHAP analyzer to calculate the contribution of each modality feature to the diagnosis result, and a report generator to generate a diagnosis report and write the diagnosis report into a hospital information system.

7. The intelligent diagnosis and treatment integrated system for respiratory diseases according to claim 4, characterized in that, The model optimization module comprises: A reinforcement model unit records the change of classification confidence of prediction type by collecting doctor correction data and patient monitoring data, generates training samples, classifies difficult samples that are difficult for the multi-task output model to predict stably, and triggers parameter update of the multi-task output model.

8. The intelligent diagnosis and treatment integrated system for respiratory diseases according to claim 1, wherein, The model optimization module further comprises: An meta-learning framework unit optimizes the initial parameters of the model on few-shot tasks based on the MAML framework, and improves the ability to capture the characteristics of new diseases.

9. The intelligent diagnosis and treatment integrated system for respiratory diseases according to claim 1, wherein, The application service module comprises: A doctor end for displaying diagnosis results and integrated into a hospital information system, and generating correction data according to the contribution of each modality feature to the diagnosis data and inputting the correction data into the model optimization module; A patient end with a symptom and sign correlation engine, which actively reminds the patient based on an abnormal physiological signal detection algorithm; A management end for epidemic trend early warning, which dynamically adjusts the early warning threshold based on historical baseline.

10. The intelligent diagnosis and treatment method of respiratory diseases is applied to the integrated system for intelligent diagnosis and treatment of respiratory diseases in any one of claims 1-9, characterized in that, The method comprises the following steps: S1. Obtain the clinical data, medical images and physiological signals of the patient through the data acquisition module; S2. The edge computing module filters the physiological signals, and denoises and lung field segments the medical images; S3. The multi-modality fusion diagnosis module extracts fusion features, and generates disease prediction type, severity and treatment suggestion according to the fusion features; S4. Output diagnosis and treatment feedback through the application service module, and dynamically optimize the parameters of the multi-modality fusion diagnosis module based on the model optimization module.

Images