Acute pulmonary embolism auxiliary diagnosis decision-making system based on large model

Abstract

The invention relates to the technical field of medical information processing, and discloses an acute pulmonary embolism auxiliary diagnosis decision-making system based on a large model, and the system comprises a data collection and preprocessing module which is used for receiving multi-source clinical data and preprocessing the multi-source clinical data; the large model reasoning module is used for inputting the preprocessed multi-source clinical data into a large model fused with a multi-modal causal attention and identification adversarial learning mechanism, and outputting a reasoning result; the dynamic learning module is used for accessing desensitized case data of a hospital electronic medical record system, continuously learning new cases through federal learning, and optimizing the recognition capability of a large model pair; the diagnosis suggestion generation module is used for generating structured diagnosis suggestions according to the reasoning result output by the large model; the result output and interaction module is used for displaying diagnosis suggestions to doctors and updating reasoning results in real time by a large model when the doctors supplement data; the diagnosis efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical information processing, in particular to an acute pulmonary embolism auxiliary diagnosis decision system based on a large model. BACKGROUND

[0003] Acute pulmonary embolism is an acute respiratory and circulatory disorder syndrome caused by endogenous or exogenous emboli blocking the pulmonary artery, with the characteristics of acute onset, high mortality, and atypical symptoms, and missed diagnosis or misdiagnosis may lead to serious consequences (such as sudden death). Currently, clinical diagnosis mainly relies on doctors to comprehensively judge in combination with clinical symptoms, laboratory tests (such as D-dimer), imaging data (such as CT pulmonary angiography CTPA), and classic scoring (such as Wells score and Geneva score), but there are the following problems: 1. Low data integration efficiency: diagnosis relies on multi-source data (structured data such as test values, unstructured data such as medical record texts, and imaging reports), and manual integration is prone to omissions or errors, especially in emergency scenarios, which takes a long time; 2. Difficulty in identifying atypical cases: about 30% of acute pulmonary embolism patients show atypical symptoms (such as only fatigue and syncope), and the traditional scoring system has insufficient sensitivity in identifying such cases, which is prone to missed diagnosis; 3. High complexity of differential diagnosis: the symptoms of acute pulmonary embolism (chest pain, dyspnea, etc.) are highly overlapped with acute myocardial infarction, aortic dissection, and pneumonia, and inexperienced doctors at the grassroots level are prone to confusion; 4. Poor dynamic case adaptation: existing auxiliary systems are mostly based on fixed rules or traditional machine learning models, which are difficult to quickly integrate new clinical guidelines and rare case data, and have insufficient diagnostic adaptability for special populations (such as pregnant women, the elderly, and patients with multiple underlying diseases). SUMMARY

[0005] The purpose of the present application is to solve the above problems, and an acute pulmonary embolism auxiliary diagnosis decision system based on a large model is designed.

[0006] The present application provides an acute pulmonary embolism auxiliary diagnosis decision system based on a large model, which comprises: a data acquisition and preprocessing module for receiving multi-source clinical data and preprocessing the multi-source clinical data, wherein the multi-source clinical data includes structured data, unstructured data, and imaging data; a large model inference module for inputting the preprocessed multi-source clinical data into a large model that integrates multi-modal causal attention and discriminative adversarial learning mechanism, and outputting an inference result; A dynamic learning module is configured to access desensitized case data of a hospital electronic medical record system, continuously learn new cases through federated learning, and optimize the recognition ability of the large model; A diagnosis suggestion generation module is configured to generate a structured diagnosis suggestion according to an inference result output by the large model, wherein the structured diagnosis suggestion includes a preliminary diagnosis tendency, a next examination recommendation, and a differential diagnosis priority; A result output and interaction module is configured to display the diagnosis suggestion to a doctor, and the large model updates the inference result in real time when the doctor supplements data.

[0007] Optionally, in the first implementation manner of the present application, the structured data at least includes age, gender, vital signs, D-dimer value and electrocardiogram index, the unstructured data at least includes electronic medical record and image report text, and the image data at least includes CTPA original image and echocardiogram dynamic image.

[0008] Optionally, in the second implementation manner of the present application, the data acquisition and preprocessing module comprises: A preliminary processing submodule is configured to first identify and process missing values in the structured data by using an interpolation method, and then remove irrelevant symbols in the unstructured data by using a regular expression, to complete preliminary processing; A category classification submodule is configured to convert the structured data after preliminary processing into a unified format by field mapping, perform word segmentation processing on the unstructured data after preliminary processing based on a word segmentation model trained by BioBERT, extract entity information, and classify the entity information according to clinical categories; An image processing submodule is configured to remove noise of the image data by using a Gaussian filter, enhance the contrast between embolus foci and surrounding tissues by using adaptive histogram equalization, and crop the processed image to retain an effective field of view containing a pulmonary artery region; An association and integration submodule is configured to associate and integrate data of the same patient by using a unique patient identification index, to obtain the preprocessed multi-source clinical data.

[0009] Optionally, in the third implementation manner of the present application, the image processing submodule comprises: A 3*3 filter kernel is selected, the filter kernel is slid at each pixel position of the image, a weighted average value of all pixels in the current pixel and its neighborhood is calculated, the original pixel value is replaced by the average value, and noise removal of the whole image is completed; The image after noise removal is divided into non-overlapping rectangular sub-regions, a gray level histogram of each sub-region is calculated, and the histogram of each sub-region is truncated, and the pixel gray level values exceeding the threshold are uniformly distributed to other gray levels; According to the processed sub-region histogram, the cumulative distribution function of each pixel is calculated, the original pixel gray value is mapped to a new gray range, local contrast enhancement is realized, and a bilinear interpolation smoothing process is used at the junction of the sub-regions; The U-Net architecture is used for pixel-level segmentation of the enhanced image to generate a binary mask of the pulmonary artery region, the boundary coordinates of the pulmonary artery region are determined according to the mask, and the effective field of view image obtained by cutting is scaled to 512*512 pixels, and the gray value of the scaled image is normalized.

[0010] Optionally, in the fourth implementation manner of the present application, the large model inference module comprises: An input sub-module is configured to input the preprocessed structured data, text data, and image data into the encoder of the corresponding modality, respectively; A fusion sub-module is configured to calculate the causal attention weight between the features of different modalities, obtain the final attention weight through weighted calculation, and fuse the multi-modal features; A first calculation sub-module is configured to input the fused multi-modal feature vector into a fully connected layer to calculate the pulmonary embolism suspected probability, and output the patient risk level through a multi-classifier; A comparison sub-module is configured to input the fused multi-modal feature vector into a discrimination sub-network, compare the features with the features of similar diseases, output a feature difference score, and finally generate a differentiated feature list; A summary sub-module is configured to summarize the pulmonary embolism suspected probability, the patient risk level, and the differentiated feature list, and output the inference result.

[0011] Optionally, in the fifth implementation manner of the present application, the input sub-module comprises: The structured data is processed through a feature selection branch and a dynamic weighting branch to generate a structured feature vector; The unstructured data is processed through a medical word segmentation tool to extract entities including symptoms and medical history, combined with time sequence encoding and semantic mapping, to generate a text semantic vector; The image data is processed through a convolutional network to extract spatial features of embolism foci, and combined with clinical indicators to generate an image feature vector.

[0012] Optionally, in the sixth implementation manner of the present application, the input sub-module comprises: The structured data is processed through a feature selection branch and a dynamic weighting branch to generate a structured feature vector; The unstructured data is processed through a medical word segmentation tool to extract entities including symptoms and medical history, combined with time sequence encoding and semantic mapping, to generate a text semantic vector; ​The image data extracts spatial features of the embolus through a convolutional network, and is associated with clinical indicators to generate an image feature vector.

[0013] Optionally, in the seventh implementation manner of the present application, the diagnostic suggestion generation module comprises: A determination sub-module is configured to receive and parse the inference result output by the large model, and determine a diagnostic expression according to the pulmonary embolism suspected probability and the patient risk level; A screening sub-module is configured to match a corresponding examination path according to the patient risk level, screen out 3-4 diseases with the highest probability according to the differential feature list, determine a priority list of differential diagnosis according to the probability from high to low, and finally form a complete structured diagnostic suggestion.

[0014] Optionally, in the eighth implementation manner of the present application, the method for implementing the acute pulmonary embolism auxiliary diagnosis decision system based on a large model comprises the following steps: Receiving multi-source clinical data, and pre-processing the multi-source clinical data, wherein the multi-source clinical data comprises structured data, unstructured data and image data; Inputting the pre-processed multi-source clinical data into a large model integrating multi-modal causal attention and differential adversarial learning mechanism, and outputting an inference result; Accessing desensitized case data of a hospital electronic medical record system, continuously learning new cases through federated learning, and optimizing the recognition ability of the large model; Generating a structured diagnostic suggestion according to the inference result output by the large model, wherein the structured diagnostic suggestion comprises a preliminary diagnosis tendency, a next examination recommendation and a differential diagnosis priority; Displaying the diagnostic suggestion to a doctor, and updating the inference result of the large model in real time when the doctor supplements data.

[0015] Optionally, in the ninth implementation manner of the present application, the method for implementing the acute pulmonary embolism auxiliary diagnosis decision system based on a large model comprises the following steps: First, an interpolation method is used to identify and process missing values in the structured data, and then a regular expression is used to remove irrelevant symbols in the unstructured data, to complete preliminary processing; The structured data after preliminary processing is converted into a unified format through field mapping, the unstructured data after preliminary processing is subjected to word segmentation processing based on a word segmentation model trained by BioBERT, entity information is extracted, and the entity information is classified according to clinical categories; Gaussian filtering is used to remove noise of the image data, adaptive histogram equalization is used to enhance the contrast between the embolus and the surrounding tissue, the processed image is cropped to retain an effective field of view containing the pulmonary artery region; The same patient data is associated and integrated to obtain preprocessed multi-source clinical data by taking the patient unique identification index.

[0016] In the technical solution provided by the application, multi-source clinical data is received and preprocessed, the preprocessed multi-source clinical data is input into a large model with a fusion multi-modal causal attention and discriminative adversarial learning mechanism, and a reasoning result is output; desensitized case data of a hospital electronic medical record system is accessed, new cases are continuously learned through federated learning, and the identification ability of the large model is optimized; structured diagnosis suggestions are generated according to the reasoning result output by the large model; the diagnosis suggestions are displayed to doctors, and the large model updates the reasoning result in real time when the doctors supplement data; the application improves the diagnosis efficiency, completes multi-source data integration and reasoning within 10-15 minutes, shortens the emergency diagnosis time, and reduces the risk of death; improves the diagnosis accuracy, improves the identification sensitivity of the large model to atypical symptoms, and reduces missed diagnosis; enhances the differential diagnosis ability, generates a differentiated list through feature comparison, helps primary doctors to exclude similar diseases, and reduces the misdiagnosis rate; has strong dynamic adaptability, continuously learns new cases and guidelines, adapts to special populations and rare cases, prolongs the system life cycle, reduces unreasonable costs, and reduces the over-examination and over-diagnosis of acute pulmonary embolism patients. BRIEF DESCRIPTION OF DRAWINGS

[0018] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are for purposes of illustration only and are not considered a limitation of the present application.

[0019] Figure 1 A structural schematic diagram of an acute pulmonary embolism auxiliary diagnosis and decision system based on a large model provided by an embodiment of the application is shown in the figure. Figure 2 A structural schematic diagram of a data acquisition and preprocessing module provided by an embodiment of the application is shown in the figure. Figure 3 A structural schematic diagram of a large model reasoning module provided by an embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0021] The terms "first", "second", "third", "fourth" and the like in the description and in the claims of the present application, and above-described drawings, if any, are used to distinguish between similar objects and are not necessarily used to describe a particular sequential or chronological order. It is to be understood that the use of the terms so-termed, where appropriate, can be interchanged with each other to allow the embodiments described herein to be implemented in an order other than that illustrated or described herein. Furthermore, the terms "comprise" or "have" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, device, product or apparatus comprising a list of steps or units does not necessarily limit to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to such processes, methods, products or apparatuses.

[0022] For the sake of understanding, the specific flow of the embodiments of the present application is described below, please refer to Figure 1 The acute pulmonary embolism auxiliary diagnosis and decision system based on a large model provided by the embodiments of the present application comprises: A data acquisition and preprocessing module is configured to receive multi-source clinical data and preprocess the multi-source clinical data, wherein the multi-source clinical data comprises structured data, unstructured data and image data; A large model inference module is configured to input the preprocessed multi-source clinical data into a large model with a multi-modal causal attention and discriminative adversarial learning mechanism, and output an inference result; A dynamic learning module is configured to access desensitization case data of a hospital electronic medical record system, continuously learn new cases through federated learning, and optimize the identification ability of the large model; A diagnosis suggestion generation module is configured to generate a structured diagnosis suggestion according to the inference result output by the large model, wherein the structured diagnosis suggestion comprises a preliminary diagnosis tendency, a next step examination recommendation and a differential diagnosis priority; A result output and interaction module is configured to display the diagnosis suggestion to a doctor, and the large model updates the inference result in real time when the doctor supplements data.

[0023] In the embodiments, the structured data at least comprises age, gender, vital signs, D-dimer value and electrocardiogram index, the unstructured data at least comprises electronic medical record and image report text, and the image data at least comprises CTPA original image and echocardiogram dynamic image.

[0024] In the embodiments, please refer to Figure 2 The data acquisition and preprocessing module comprises: A preliminary processing submodule is configured to first identify and process missing values in the structured data by using an interpolation method, and then remove irrelevant symbols in the unstructured data by using a regular expression, to complete preliminary processing; The category classification submodule is configured to convert the preliminarily processed structured data into a unified format through field mapping, perform word segmentation processing on the preliminarily processed unstructured data based on a word segmentation model trained by BioBERT, extract entity information, and classify the entity information according to clinical categories; The image processing submodule is configured to remove noise of image data through Gaussian filtering, enhance the contrast between embolus foci and surrounding tissues by using adaptive histogram equalization, and crop the processed image to retain an effective field of view containing a pulmonary artery region; The association integration submodule is configured to associate and integrate data of the same patient by using a unique patient identifier index, to obtain preprocessed multi-source clinical data.

[0025] In this embodiment, the image processing submodule includes: A 3*3 filter kernel is selected, the filter kernel is slid at each pixel position of the image, a weighted average value of all pixels in the current pixel and its neighborhood is calculated, the average value is used to replace the original pixel value, and noise removal of the whole image is completed; The denoised image is divided into non-overlapping rectangular sub-regions, a gray level histogram of each sub-region is calculated, and the pixel gray level values exceeding the threshold are uniformly distributed to other gray levels through truncation processing of the histogram of each sub-region; According to the processed sub-region histogram, a cumulative distribution function of each pixel is calculated, the original pixel gray level value is mapped to a new gray level range, local contrast enhancement is realized, and a bilinear interpolation smoothing processing is performed at the junction of the sub-regions; The U-Net architecture is used to perform pixel-level segmentation on the enhanced image, to generate a binary mask of the pulmonary artery region, to determine the boundary coordinates of the pulmonary artery region according to the mask, and to crop according to the boundary coordinates; The effective field of view image obtained by cropping is scaled to 512*512 pixels, and the scaled image is subjected to gray value normalization processing.

[0026] In this embodiment, referring to Figure 3 The large model inference module includes: The input submodule is configured to input the preprocessed structured data, text data and image data into the corresponding modality encoder, respectively; The fusion submodule is configured to calculate the causal attention weight between the features of each modality, obtain the final attention weight through weighted calculation, and perform fusion of the multi-modal features; The first calculation submodule is configured to input the fused multi-modal feature vector into a full connection layer to calculate the pulmonary embolism suspected probability, and output the patient risk level through a multi-classifier; The comparison submodule is configured to input the fused multi-modal feature vector into a discrimination subnetwork, compare the feature with the features of similar diseases, output a feature difference score, and finally generate a differentiated feature list. a summary submodule configured to summarize the pulmonary embolism suspected probability, the patient risk level, and the differential feature list, and output a reasoning result.

[0027] In this embodiment, the input submodule includes: The structured data is converted into a structured feature vector through the feature selection branch and the dynamic weighting branch; The unstructured data is converted into a text semantic vector by extracting entities including symptoms and medical history through a medical word segmentation tool, combining time sequence coding and semantic mapping; The image data is converted into an image feature vector by extracting spatial features of embolism foci through a convolutional network and associating the features with clinical indicators.

[0028] In this embodiment, the dynamic learning module includes: a second calculation submodule configured to input the desensitized case data into the large model, calculate the deviation between the output of the large model and the actual diagnosis result, and calculate the gradient change of the parameters of each layer of the model according to the deviation through back propagation; an encryption submodule configured to encrypt the gradient values by using a homomorphic encryption algorithm of federated learning, add differential privacy noise, upload the encrypted gradient information to a federated learning central server, and perform differential privacy processing on the gradient values; an application submodule configured to apply the global gradient update quantity aggregated by the central server to the current large model to generate new large model parameters.

[0029] In this embodiment, the diagnosis suggestion generation module includes: a determination submodule configured to receive and analyze the reasoning result output by the large model, and determine a diagnosis expression according to the pulmonary embolism suspected probability and the patient risk level; a screening submodule configured to match a corresponding examination path according to the patient risk level, screen out the 3-4 diseases with the highest probability according to the differential feature list, determine a priority list of differential diagnosis according to the probability from high to low, and finally form a complete structured diagnosis suggestion.

[0030] In this embodiment, the desensitized case data of the hospital electronic medical record system is accessed through a standardized interface such as a medical data transmission protocol conforming to HIPAA and GDPR, and the data needs to be pre-processed to remove patient personal identification information such as name, ID number, and medical record number, and only keep clinical features such as symptoms, test results, image features, and diagnosis conclusions; the accessed data is verified for compliance: checking the desensitization integrity to ensure that there is no implicit identity information, such as the case where the birth date and gender can be uniquely identified, and the data format consistency such as the naming of structured fields and the uniformity of units, and after verification, the data is stored in a local encrypted database, and the data access time and source department are marked; Each participating hospital performs preprocessing on the newly accessed desensitized case data locally: structured data supplements missing values with median filling of the same batch data, and abnormal values are corrected based on the clinical reasonable range truncation; text data extracts entities through medical word segmentation tools and standardizes terms; image features call preset feature extraction modules and generate feature vectors consistent with image processing logic during model training; after preprocessing, through the feature mapping table provided by the federated learning coordination layer, the local features of each hospital are aligned to the unified feature space, such as mapping A hospital's D-dimer and B hospital's D-dimer quantitative to the same feature field; Each hospital uses the locally aligned new case data to perform parameter update calculation based on the current local model copy without uploading the original data: input the new case features into the model, calculate the deviation of the model output from the actual diagnosis result, and calculate the gradient change of each layer parameter of the model according to the deviation. Only the change amount is calculated, and the complete training process is not involved; for special samples in new cases such as rare PE subtypes and complex cases with multiple underlying diseases, higher calculation weight is given through sample weight mechanism, such as 1.0 for regular cases and 1.5 for special cases, to strengthen the model's learning of such samples; Each hospital encrypts the calculated parameter gradient: uses a homomorphic encryption algorithm such as Paillier encryption to encrypt the gradient value, ensuring that the central server cannot decrypt the original gradient information; at the same time, differential privacy noise such as Laplace noise is added, and the noise intensity is dynamically adjusted according to the data volume. The smaller the data volume, the higher the noise intensity, further preventing the original data from being inferred through the gradient; the encrypted gradient information is attached with the hospital identifier and data quality label such as case annotation accuracy score, and uploaded to the federated learning central coordination server; After the central server receives the encrypted gradient uploaded by each hospital, it first verifies the data integrity, checks whether the gradient dimension and encryption format meet the specifications, and then calculates the global gradient based on the preset aggregation strategy: according to the data volume proportion weight and data quality label quality score weight of each hospital, the aggregation weight is allocated, such as data volume proportion 60% and quality score proportion 40%, and the encrypted gradient is weighted and summed to obtain the update amount of the global model parameter. The aggregation process is completed in an encrypted state without decrypting the gradient; The central server applies the aggregated global gradient update amount to the current global model to generate new global model parameters; the new parameters are signed and authenticated by adding a timestamp and a server digital signature to ensure that the parameter transmission process cannot be tampered with, and then distributed to each participating hospital; After each hospital receives the new global model parameters, it first verifies the signature for legality and confirms that the parameters have not been tampered with; then it replaces the old parameters of the local model copy with the new parameters to complete the local model update; the local reserved verification set of desensitized cases that did not participate in parameter update is used to verify the updated model: the diagnostic accuracy, sensitivity, and other indicators of the model on the verification set are calculated, if the indicators decrease by more than a preset threshold such as 5% compared to before the update, the old model is temporarily retained and an exception is fed back to the central server, otherwise the model update is confirmed to be effective; A fixed update cycle such as once a week is set, and the following is repeatedly performed: new desensitized case data is continuously accessed, and a closed loop of local calculation-encrypted upload-global aggregation-parameter distribution is realized through the federated learning mechanism, so that the model continuously absorbs new case information from each hospital without sharing the original data, and gradually optimizes the recognition ability of various cases, especially rare and complex cases.

[0031] In this embodiment, key information is extracted from the large model output, including acute pulmonary embolism suspected probability, risk stratification results, feature contribution degrees such as CTPA image features accounting for 45%, D-dimer value accounting for 20%, differential probability of similar diseases such as acute myocardial infarction 10%, aortic dissection 5%, and uncertainty quantification results such as confidence 92%; these information is classified and stored according to categories to form a structured reasoning result data set, providing a basis for subsequent suggestion generation; According to the acute pulmonary embolism suspected probability and the risk stratification result, a core diagnosis expression is determined: if the suspected probability is ≥70%, it is expressed as highly suspected acute pulmonary embolism; if the suspected probability is between 30%-70%, it is expressed as moderately suspected acute pulmonary embolism; if the suspected probability is <30%, it is expressed as low suspected acute pulmonary embolism; at the same time, 2-3 most key supporting evidences are extracted according to the feature contribution degree, such as CTPA showing right pulmonary artery main stem embolism + Wells score 6 points, which are attached to the diagnosis tendency to form a complete preliminary diagnosis tendency content; According to the risk stratification result, the corresponding examination path is matched: high-risk patients are preferentially recommended for emergency diagnosis examinations such as CT pulmonary angiography CTPA; medium-risk patients are recommended for stratified progressive examinations such as first checking lower limb venous ultrasound + D-dimer retest, and then deciding whether to perform CTPA according to the results; low-risk patients are recommended for basic screening such as D-dimer + electrocardiogram recheck; at the same time, the existing examination completion information is obtained from the electronic medical record system, the completed items are excluded, and the recommended examination items are ensured to be complementary and necessary to form specific suggestions for the next examination; According to the differential probability of similar diseases, the top 3-4 diseases with the highest probability are screened out, such as acute myocardial infarction, aortic dissection, and severe pneumonia, and the priority is determined from high to low according to the probability; for each disease, the differential features output by the large model are extracted, such as acute myocardial infarction: ST segment arch upward, significant increase in troponin, as a differential basis attached to the corresponding disease, and the core difference between the disease and acute pulmonary embolism is clearly distinguished, forming a priority list for differential diagnosis; The preliminary diagnosis tendency, next step examination recommendation, and differential diagnosis priority are combined in a fixed format: the preliminary diagnosis tendency is the first part, the next step examination recommendation is the second part, and the differential diagnosis priority is the third part, each part adopts a structured form of title + content; At the same time, if the confidence in the uncertainty quantization result is less than 70%, a prompt is added at the end of the suggestion: the current diagnosis confidence is low, it is recommended to combine multi-disciplinary consultation for further clarification, and finally a complete structured diagnosis suggestion is formed.

[0032] In this embodiment, the generated structured diagnosis suggestion including preliminary diagnosis tendency, next step examination recommendation, and differential diagnosis priority is converted into a clinically readable interface format: the content of each part is displayed in a hierarchical list, and key information such as high suspicion and CTPA examination is highlighted by bold or color marking; the reasoning basis such as feature contribution degree is attached to the corresponding suggestion below in the form of a folding panel, and the doctor can click to expand and view the details; At the same time, data supplement entry such as text input box and drop-down selection menu is reserved in the interface to support doctors to add or correct patient information such as supplementing past thrombosis history, family history, etc. Real-time monitoring of the doctor's operation in the data supplement entry, when the doctor inputs new information such as the patient has a history of lower extremity varicose veins or corrects the existing data such as correcting the heart rate of 90 times / min to 110 times / min, the system immediately captures the input content; Analyze the input data: structured data such as numerical values and options are directly extracted into fields and values; unstructured text such as symptom description is extracted into entities by medical word segmentation tool, such as lower extremity varicose veins classified as risk factors, and matched with the system's pre-set feature library to determine its corresponding feature type such as thrombosis-related risk factors; The supplemented data after analysis is preprocessed in the same way as the initial data: structured data is standardized such as unit conversion and format unification; text entities are converted into corresponding feature vectors and marked as new features; the preprocessed supplemented data is integrated into the original feature set, and the feature matrix is updated: if it is a new feature such as no family history in the original data, the feature matrix dimension is extended and the corresponding value is filled; if it is a corrected data such as heart rate value update, replace the value at the corresponding position in the original feature matrix, and record the data update time and type (new / modified) at the same time; The incremental inference interface of the large model is called, the updated feature matrix is input into the model, and real-time calculation is triggered: the model only recalculates the feature path affected by the supplemented data, such as after supplementing the varicose vein history of lower limbs, only recalculating the attention weight and inference branch related to the risk factors of thrombosis, instead of full-model re-inference; during the calculation process, the weight and inference result of other unchanged features are kept unchanged, and only the associated parameters of the newly added or corrected features are adjusted; the varicose vein history of lower limbs will increase the causal coefficient of deep vein thrombosis->PE; The new results obtained according to the incremental inference, such as the suspected probability from 85% to 92%, and the risk stratification from moderate risk to high risk, are used to reconstruct the contents of each part according to the structured diagnosis suggestion generation rule: update the expression of the preliminary diagnosis tendency and the supporting evidence; adjust the next examination recommendation according to the new risk stratification, such as changing the lower limb venous ultrasound from moderate risk to high risk emergency CTPA; recalculate and sort the differential probability of similar diseases; at the same time, mark the changed part in the updated suggestion, such as using the symbol Δ to mark the adjusted content, so as to facilitate the doctor to identify the difference; Replace the original display content in the interface with the reconstructed diagnosis suggestion, and keep the format consistency; if there is data correction, display the modification record in the corresponding position, such as heart rate: original 90 times / min->present 110 times / min; at the same time, keep the backtracking entrance of the historical suggestion version, and the doctor can compare the suggestions before and after the update to ensure that the whole updating process is traceable.

[0033] The basic principles, main features and advantages of the present application are shown and described above. Those skilled in the art should understand that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only preferred examples of the present application, and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A large model-based acute pulmonary embolism auxiliary diagnosis decision system, characterized in that, The system comprises: a data acquisition and preprocessing module for receiving multi-source clinical data and preprocessing the multi-source clinical data, wherein the multi-source clinical data comprises structured data, unstructured data and image data; a large model inference module for inputting the preprocessed multi-source clinical data into a large model with a fusion multi-modal causal attention and discriminative adversarial learning mechanism, and outputting inference results; a dynamic learning module for accessing desensitized case data of a hospital electronic medical record system, continuously learning new cases through federated learning, and optimizing the identification ability of the large model; a diagnosis suggestion generation module for generating structured diagnosis suggestions according to the inference results output by the large model, wherein the structured diagnosis suggestions comprise preliminary diagnosis tendency, next step examination recommendation and differential diagnosis priority; a result output and interaction module for displaying the diagnosis suggestions to doctors, and updating the inference results in real time when the doctors supplement data.

2. The large model-based acute pulmonary embolism auxiliary diagnosis decision system of claim 1, wherein, The structured data at least comprises age, gender, vital signs, D-dimer value and electrocardiogram indicators, the unstructured data at least comprises electronic medical records and image report texts, and the image data at least comprises CTPA original images and echocardiogram dynamic images.

3. The large model-based acute pulmonary embolism auxiliary diagnosis decision system of claim 1, wherein, The data acquisition and preprocessing module comprises: a preliminary processing submodule for identifying and processing missing values in the structured data by using an interpolation method first, and then removing irrelevant symbols in the unstructured data by using a regular expression to complete preliminary processing; a category classification submodule for converting the preliminarily processed structured data into a unified format through field mapping, performing word segmentation processing on the preliminarily processed unstructured data based on a word segmentation model trained by BioBERT, extracting entity information, and classifying the entity information according to clinical categories; an image processing submodule for removing image data noise by using Gaussian filtering, enhancing the contrast between embolus foci and surrounding tissues by using adaptive histogram equalization, and performing cropping on the processed images to retain the effective field of view containing the pulmonary artery region; an association and integration submodule for associating and integrating the data of the same patient by using a patient unique identification index to obtain the preprocessed multi-source clinical data.

4. The large model-based acute pulmonary embolism auxiliary diagnosis decision system of claim 3, wherein, The image processing submodule comprises: a 3x3 filter kernel is selected, the filter kernel is slid at each pixel position of the image, the weighted average value of all pixels in the current pixel and its neighborhood is calculated, the average value is used to replace the original pixel value, and the noise removal of the whole image is completed; the denoised image is divided into non-overlapping rectangular sub-regions, the gray level histogram of each sub-region is calculated, and the pixel gray level values exceeding the threshold are uniformly distributed to other gray levels through truncation processing of the histogram of each sub-region; according to the processed sub-region histogram, the cumulative distribution function of each pixel is calculated, the original pixel gray level value is mapped to a new gray level range, local contrast enhancement is realized, and a bilinear interpolation smoothing process is performed at the sub-region boundary; a U-Net architecture is used to perform pixel-level segmentation on the enhanced image to generate a binary mask of the pulmonary artery region, the boundary coordinates of the pulmonary artery region are determined according to the mask, and cropping is performed according to the boundary coordinates; The effective field of view image obtained by cutting is scaled to 512*512 pixels, and the scaled image is subjected to gray value normalization processing.

5. The large model based acute pulmonary embolism aided diagnosis decision system of claim 1, wherein, The large model inference module comprises: An input sub-module configured to input the preprocessed structured data, text data, and image data into the corresponding modal encoder; A fusion sub-module configured to calculate the causal attention weight between the features of each modality, obtain the final attention weight through weighted calculation, and fuse the multi-modal features; A first calculation sub-module configured to input the fused multi-modal feature vector into a fully connected layer to calculate the pulmonary embolism suspected probability, and output the patient risk level through a multi-classifier; A comparison sub-module configured to input the fused multi-modal feature vector into a discrimination sub-network, compare it with the features of similar diseases, output a feature difference score, and finally generate a differential feature list; A summary sub-module configured to summarize the pulmonary embolism suspected probability, patient risk level, and differential feature list, and output the inference result.

6. The large model based acute pulmonary embolism aided diagnosis decision system of claim 5, wherein, The input sub-module comprises: The structured data is processed through a feature selection branch and a dynamic weighting branch to generate a structured feature vector; The unstructured data is processed through a medical word segmentation tool to extract entities including symptoms and medical history, combined with time encoding and semantic mapping, to generate a text semantic vector; The image data is processed through a convolutional network to extract spatial features of embolic lesions and associate them with clinical indicators to generate an image feature vector.

7. The large model based acute pulmonary embolism aided diagnosis decision system of claim 1, wherein, The dynamic learning module comprises: A second calculation sub-module configured to input the desensitized case data into the large model, calculate the deviation of the large model output from the actual diagnosis result, and calculate the gradient change of the parameters of each layer of the model according to the deviation; An encryption sub-module configured to encrypt the gradient value using a homomorphic encryption algorithm of federated learning, add differential privacy noise, and upload the encrypted gradient information to a federated learning central server; An application sub-module configured to apply the global gradient update quantity aggregated by the central server to the current large model to generate new large model parameters.

8. The large model based acute pulmonary embolism aided diagnosis decision system of claim 1, wherein, The diagnosis suggestion generation module comprises: A determination sub-module configured to receive and parse the inference result output by the large model, and determine the diagnosis expression according to the pulmonary embolism suspected probability and patient risk level; A screening sub-module configured to match the corresponding examination path according to the patient risk level, screen out the top 3-4 diseases according to the differential feature list, determine the priority list of differential diagnosis according to the probability from high to low, and finally form a complete structured diagnosis suggestion.

9. The method of implementing the large model-based acute pulmonary embolism auxiliary diagnosis decision system according to claim 1, characterized in that, The method comprises the following steps: Receiving multi-source clinical data and preprocessing the multi-source clinical data, wherein the multi-source clinical data comprises structured data, unstructured data, and image data; Inputting the preprocessed multi-source clinical data into a large model with a fusion multi-modal causal attention and discriminative adversarial learning mechanism to output an inference result; Accessing desensitized case data of a hospital electronic medical record system, continuously learning new cases through federated learning, and optimizing the recognition ability of the large model; Generating a structured diagnosis suggestion according to the inference result output by the large model, wherein the structured diagnosis suggestion comprises a preliminary diagnosis tendency, a next examination recommendation, and a differential diagnosis priority; Displaying the diagnosis suggestion to the doctor, and updating the inference result in real time when the doctor supplements the data.

10. The method of implementing the large model-based acute pulmonary embolism auxiliary diagnosis decision system according to claim 1, characterized in that, The method comprises the following steps: First, the missing values in the structured data are identified and processed by using an interpolation method, and then irrelevant symbols in the unstructured data are removed by using a regular expression to complete preliminary processing; The structured data after preliminary processing is converted into a unified format through field mapping, the unstructured data after preliminary processing is subjected to word segmentation processing based on a word segmentation model trained by BioBERT, entity information is extracted, and the entity information is classified according to clinical categories; Gaussian filtering is used to remove noise of the image data, adaptive histogram equalization is used to enhance the contrast between the embolus and the surrounding tissue, the processed image is cropped, and the effective field of view containing the pulmonary artery region is retained; The same patient data is associated and integrated by using a patient unique identification index, and the preprocessed multi-source clinical data is obtained.

Images