Vein thrombosis risk assessment method and system based on large language model and deep learning

Abstract

The invention discloses a venous thrombosis risk assessment method and system based on a large language model and deep learning, and the method and system can comprehensively capture the clinical information of a patient through the comprehensive utilization of multi-dimensional and multi-modal clinical diagnosis and treatment data, and improve the accuracy of risk assessment. An automatic feature extraction and selection technology is realized by applying a large language model, manual intervention is reduced, and the comprehensiveness and accuracy of feature extraction are improved. The introduction of the deep learning model enhances the prediction performance of the model, supports real-time data processing and risk prediction, and realizes the automation and real-time performance of risk assessment. Besides, the model framework supports continuous learning and updating, can adapt to new clinical data and research progress, and keeps advancement and effectiveness of the model, thereby providing high-quality hospitalized patient VTE risk assessment services.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and in particular, relates to a venous thrombosis risk assessment method based on a large language model and deep learning, and a venous thrombosis risk assessment system based on a large language model and deep learning according to this method. Background Art

[0002] In recent years, with the advancement of medical technology and the deepening of disease research, venous thromboembolism (VTE) has received widespread attention as a common complication. VTE includes deep vein thrombosis (DVT) and pulmonary thromboembolism (PTE), which are particularly common in hospitalized patients. Traditional VTE risk assessment methods rely on the experience of clinicians and known risk factors such as age, medical history, and surgical conditions. However, these methods are often subjective and it is difficult to comprehensively and accurately predict patients' risks.

[0003] The rapid development of big language models and deep learning technologies has brought new opportunities to the medical field. The text parsing capabilities of big language models can extract key clinical information from a large amount of unstructured data such as electronic medical records and medical records, while deep learning technology can achieve large-scale, multi-dimensional medical data processing and prediction by training complex neural network models. The combination of these technologies makes it possible to build an intelligent risk assessment system based on big data.

[0004] At present, there are some VTE risk assessment methods based on machine learning and statistical models. These methods usually use simple machine learning algorithms such as regression analysis and decision trees, and require a lot of feature engineering processing. Although the application of existing methods has improved the accuracy of VTE risk assessment to a certain extent, there are also some shortcomings. First, these methods mainly rely on structured data, while ignoring a large amount of potentially valuable unstructured text data, such as doctors' diagnosis records, nursing records, medical images, physiological signals, etc. Secondly, the complexity and expression ability of traditional machine learning models are limited, and it is difficult to capture deep relationships in high-dimensional data. In addition, the feature engineering process is complex and relies on expert knowledge, which is difficult to automate. And the model's expression ability is limited, making it difficult to process complex clinical data.

[0005] The following is a brief description of the existing methods.

[0006] (1) Caprini scoring system

[0007] First, the patient's basic information and clinical data are collected, such as age, surgical history, cancer history, heart disease history, etc. Based on the collected information, each risk factor is assigned a corresponding score according to the Caprini score table. The scores of all risk factors are added together to obtain the patient's total score. Based on the total score, the patient is divided into low, medium, high, and very high risk groups.

[0008] Problems: The Caprini scoring system mainly relies on structured data and ignores key information in unstructured data. The scoring system is relatively simple in design and cannot capture the deep relationships in complex clinical data, which may limit the accuracy of risk assessment results. In addition, the manual scoring process is cumbersome and prone to human errors.

[0009] (2) Padua scoring system

[0010] First, the patient's clinical characteristics are assessed, including active cancer, previous history of VTE, reduced mobility, hematological abnormalities, etc. According to the risk factor of each characteristic, a corresponding score is given. The scores of each characteristic are added together to calculate the total score. Based on the total score, the patient is divided into a low-risk or high-risk group.

[0011] Existing problems: The Padua scoring system also relies mainly on structured data and fails to fully utilize the valuable information in unstructured data. Its scoring criteria are relatively fixed and cannot dynamically adapt to the individual differences of patients and changes in clinical diagnosis and treatment. In addition, the scoring process needs to rely on manual operation, which is time-consuming and highly subjective, affecting the efficiency and consistency of the evaluation.

[0012] (3) IMPROVE risk model

[0013] First, the patient's clinical data was collected, including age, cancer, previous VTE history, mobility, D-dimer level, etc. Then, each risk factor was scored according to the IMPROVE scoring criteria. The total score of all risk factors was calculated to obtain the patient's risk score. According to the risk score, the patient was divided into different risk groups.

[0014] Existing problems: The IMPROVE risk model also relies on structured data and cannot effectively utilize important information in unstructured data. The feature variables used in the model are relatively limited and may not fully reflect the overall risk of the patient. The fixedness of the scoring system and the reliance on manual operation also exist, resulting in a certain degree of subjectivity and uncertainty in the evaluation results.

[0015] (4) Khorana Scoring System

[0016] By collecting specific clinical data of cancer patients, such as tumor type, white blood cell count, hemoglobin level, platelet count, etc., each risk factor is scored according to the Khorana scoring criteria. The total score of all risk factors is calculated to obtain the patient's risk score. Based on the risk score, the patient is divided into low, medium, and high risk groups.

[0017] Problems: The Khorana scoring system focuses on cancer patients and has a relatively narrow scope of application, which cannot cover all hospitalized patients. The scoring system also relies on structured data and cannot fully utilize key information in unstructured data. The scoring process relies on manual operation, which is highly subjective and inefficient, affecting the convenience and accuracy of clinical application.

[0018] In summary, existing VTE risk assessment methods mainly rely on structured data and fixed scoring criteria, and cannot fully utilize important information in unstructured data, resulting in limited comprehensiveness and accuracy of risk assessment. In addition, most of these methods require manual operation, which is time-consuming, subjective, and error-prone, affecting the efficiency and consistency of clinical decision-making. Therefore, there is an urgent need for an innovative method that can comprehensively utilize structured and unstructured data, automated processing, and real-time evaluation to improve the accuracy and convenience of VTE risk assessment. Summary of the invention

[0019] In order to overcome the defects of the prior art, the technical problem to be solved by the present invention is to provide a venous thrombosis risk assessment method based on a large language model and deep learning, which can comprehensively capture the patient's clinical information, improve the accuracy of risk assessment, enhance the comprehensiveness and accuracy of feature extraction, realize the automation and real-time nature of risk assessment, adapt to new clinical data and research progress, maintain the advancement and effectiveness of the model, and thus provide high-quality VTE risk assessment services for hospitalized patients.

[0020] The technical solution of the present invention is: the method for assessing venous thrombosis risk based on a large language model and deep learning comprises the following steps:

[0021] (1) Obtain the basic information of the patient to be evaluated and obtain the patient's medical treatment data based on the patient's unique identity;

[0022] (2) Classify the acquired diagnosis and treatment data according to the content format into three categories: structured data, semi-structured data, and unstructured data;

[0023] (3) Perform data standardization and preprocessing for data clarity on structured data; apply large language models to extract text features from semi-structured data and convert them into feature vectors; apply deep learning algorithms to remove noise and enhance images from unstructured data;

[0024] (4) Apply feature engineering to extract and select features from the three categories of data content, and use the selected features for feature fusion to form multimodal features;

[0025] (5) Feature linking of multimodal features to form a joint feature vector;

[0026] (6) Use a multi-layer neural network to construct and train a comprehensive model of the joint feature vector, and use the trained model to predict the risk of VTE in patients in real time;

[0027] (7) The prediction results are fed back to clinicians and a detailed risk report is generated.

[0028] The present invention can comprehensively capture the patient's clinical information and improve the accuracy of risk assessment by comprehensively utilizing multi-dimensional and multi-modal clinical diagnosis and treatment data. Automated feature extraction and selection technology reduces manual intervention and improves the comprehensiveness and accuracy of feature extraction. The introduction of deep learning models enhances the predictive performance of the model, supports real-time data processing and risk prediction, and realizes the automation and real-time nature of risk assessment. In addition, the model framework of the present invention supports continuous learning and updating, can adapt to new clinical data and research progress, maintain the advancement and effectiveness of the model, and thus provide high-quality VTE risk assessment services for hospitalized patients.

[0029] A venous thrombosis risk assessment system based on a large language model and deep learning is also provided, the system comprising:

[0030] An information acquisition module, which acquires the basic information of the patient to be evaluated and obtains the patient's previous medical treatment data based on the patient's unique identity identifier;

[0031] The classification module classifies the acquired diagnosis and treatment data according to the content format into three categories: structured data, semi-structured data and unstructured data;

[0032] A preprocessing module that performs data standardization and data clarification preprocessing on structured data;

[0033] Apply large language models to semi-structured data to extract text features and convert them into feature vectors;

[0034] Apply deep learning algorithms to remove noise and enhance images of unstructured data;

[0035] The multimodal module applies feature engineering to extract and select features from the three categories of data content, and then uses the selected features for feature fusion to form multimodal features;

[0036] A feature linking module that links multimodal features to form a joint feature vector;

[0037] The training module uses a multi-layer neural network to build and train a comprehensive model for the joint feature vector, and uses the trained model to make real-time predictions on the patient's risk of VTE. The feedback module feeds back the prediction results to clinicians and generates a detailed risk report. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a flow chart of the method for assessing venous thrombosis risk based on a large language model and deep learning according to the present invention. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0040] In order to make the description of the present disclosure more detailed and complete, the following provides an illustrative description of the implementation and specific embodiments of the present invention; however, this is not the only form of implementing or using the specific embodiments of the present invention. The implementation covers the features of multiple specific embodiments and the method steps and sequences used to construct and operate these specific embodiments. However, other specific embodiments can also be used to achieve the same or equivalent functions and step sequences.

[0041] like Figure 1 As shown, this venous thrombosis risk assessment method based on a large language model and deep learning includes the following steps:

[0042] (1) Obtain the basic information of the patient to be evaluated and obtain the patient's medical treatment data based on the patient's unique identity;

[0043] (2) Classify the acquired diagnosis and treatment data according to the content format into three categories: structured data, semi-structured data, and unstructured data;

[0044] (3) Perform data standardization and preprocessing for data clarity on structured data; apply large language models to extract text features from semi-structured data and convert them into feature vectors; apply deep learning algorithms to remove noise and enhance images from unstructured data;

[0045] (4) Apply feature engineering to extract and select features from the three categories of data content, and use the selected features for feature fusion to form multimodal features;

[0046] (5) Feature linking of multimodal features to form a joint feature vector;

[0047] (6) Use a multi-layer neural network to construct and train a comprehensive model of the joint feature vector, and use the trained model to predict the risk of VTE in patients in real time;

[0048] (7) The prediction results are fed back to clinicians and a detailed risk report is generated.

[0049] The present invention can comprehensively capture the patient's clinical information and improve the accuracy of risk assessment by comprehensively utilizing multi-dimensional and multi-modal clinical diagnosis and treatment data. Automated feature extraction and selection technology reduces manual intervention and improves the comprehensiveness and accuracy of feature extraction. The introduction of deep learning models enhances the predictive performance of the model, supports real-time data processing and risk prediction, and realizes the automation and real-time nature of risk assessment. In addition, the model framework of the present invention supports continuous learning and updating, can adapt to new clinical data and research progress, maintain the advancement and effectiveness of the model, and thus provide high-quality VTE risk assessment services for hospitalized patients.

[0050] Preferably, in step (1), the patient's unique identity identifier is an ID number or a medical insurance card number, and the medical treatment data includes the patient's basic information, medical advice information, test indicators, examination results, medical images, electrocardiogram data, electroencephalogram data, electromyogram data, medical records, past medical history, current medical history, family history, nursing records, and surgical records.

[0051] Preferably, in step (2), structured data includes two-dimensional table data such as basic information of patients and laboratory test results, and semi-structured data includes large text data such as medical records, past medical history, current medical history, family history, nursing records, surgical records, and examination reports. Unstructured data includes medical data such as medical images, ultrasound images, electrocardiogram data, and electromyogram data.

[0052] Preferably, in step (3), the following steps are performed on the structured data:

[0053] (3.1.1) Apply big data processing technology to clean, standardize and encode structured data;

[0054] (3.1.2) Clean, standardize and encode all clinical diagnosis and treatment data;

[0055] (3.1.3) Use multiple imputation to fill missing values ​​in combination with other patient information

[0056]

[0057] For missing data in clinical diagnosis and treatment records, represents the estimate of the i-th interpolation value in the clinical diagnosis and treatment data, which refers to the missing data point X when generating the interpolation data. iFilling value; μ X represents the expected value of variable X; Z i It is a random variable that conforms to the standard normal distribution and is automatically generated by the algorithm. It is used to introduce randomness in the generation of missing values ​​and ensure the diversity of interpolation values. X is the standard deviation of variable X, which indicates the distribution range of data points in clinical diagnosis and treatment data relative to the mean;

[0058] (3.1.4) Combine clinical knowledge and use the upper and lower limit truncation method to handle abnormal data exceeding a certain threshold;

[0059] (3.1.5) Achieve data normalization and standardization to form a standardized clinical diagnosis and treatment data set;

[0060] (3.1.6) Each clinical test parameter is normalized using its clinical normal range.

[0061] Transform the eigenvalues ​​according to their position in the normal range;

[0062] (3.1.7) Check the outliers in the data and use Box-Cox transformation and inverse transformation to process the existing outliers.

[0063] Preferably, in step (3), the following steps are performed on the semi-structured data:

[0064] (3.2.1) Segment and embed text data, convert text into vector representation, extract text features through a pre-trained large language model, and input these features into the deep learning model together with the features of structured data;

[0065] (3.2.2) Apply word segmentation algorithms, combine clinical terminology and knowledge base to analyze medical text data

[0066] Structural word segmentation

[0067]

[0068] The clinical text data is segmented, t is the specific segmentation to be extracted, d represents the clinical text currently being processed, D represents the set of all clinical text documents in the currently processed inpatient record, N represents the total number of documents in the clinical text document set D, and f t,d represents the number of times the word to be extracted t appears in the clinical text d currently being processed, t' represents other words that appear in the clinical text d currently being processed, and f t',d Represents the total number of all word segments in the currently processed clinical text d;

[0069] (3.2.3) Embed medical text data;

[0070] (3.2.4) Introduce a clinical context-aware attention mechanism, which enhances the model's ability to perceive important clinical information by calculating the importance of each word or phrase in the context;

[0071] (3.2.5) Use large language models to segment and embed text data, converting text into

[0072] is a vector representation;

[0073] (3.2.6) Text features are extracted through a pre-trained large language model and converted into numerical structured feature vectors. These feature vectors will be input into the deep learning model together with the features of the structured data.

[0074] Preferably, in step (3), the following steps are performed on the unstructured data:

[0075] (3.3.1) Image enhancement through generative adversarial networks, improving image quality through adversarial training of the generator and the discriminator;

[0076] (3.3.2) Use deep learning models to segment medical images and extract regions of interest;

[0078] (3.3.3) Use convolutional neural networks to extract high-level features from the segmented images.

[0079] These features reflect the shape and texture of the image;

[0080] (3.3.4) Introducing perceptual loss and structural similarity loss to retain the high quality of the original medical image

[0081] Frequency details and structural information;

[0082] (3.3.5) Calculate the mean and variance of the input image batches, and improve the stability and speed of model training by standardizing the input images of each batch;

[0083] (3.3.6) Scale and translate the standardized image to convert the data back to an intuitive and physically meaningful range so that the model can understand and utilize its features when processing the image;

[0085] (3.3.7) Combine the attention mechanism to process electrophysiological data of electrocardiogram and electromyography to improve the ability to capture long-term dependencies

[0086]

[0087] Where Q is the query matrix, which represents the information that needs to be processed at present. K is the key matrix, which represents all the information to be processed. T represents the transposed matrix of the key matrix K, dk is the characteristic dimension of the key vector, V is the value matrix, which is a set of information corresponding to the key matrix K, containing specific information related to each key in the key matrix K, and softmax() is a normalization function used to convert the query matrix Q and the key matrix K and the characteristic dimension d of the key vector into k The calculated values ​​are converted into probability distributions.

[0088] Preferably, the step (4) comprises:

[0089] (4.1) Feature the structured data, text extraction data, and image analysis data separately

[0090] Extraction, based on dynamic time warping of features to form multimodal features;

[0091] (4.2) The extracted image features and text features are fused to construct a multimodal model, and the attention mechanism is used to weight the fusion of features of different modalities to obtain a comprehensive patient

[0092] information;

[0093] (4.3) Add feature importance analysis, calculate the importance of different features, and identify key risks

[0094] Risk factors;

[0095]

[0096] Feature Importance(x j ) represents the jth node x j The feature importance of N t is the number of samples of node t, ΔImpurity(t) is the purity change after the node splits,

[0097] Importance(f j ) represents the jth feature f j importance, T represents the set of all nodes in the decision tree, t represents a node in the decision tree, I(v(t)) is an indicator function, when node t uses feature f j The value is 1 when splitting, otherwise it is 0. v(t) represents the feature used by node t, Δi(t) represents the information gain of node t, and the impurity before and after the node split is calculated.

[0098] The degree of measurement difference is calculated;

[0099] (4.4) Combined with radial basis function kernel to process nonlinear features

[0100]

[0101] K(x,y) is the input vector, which represents the kernel function value between x and y. γ is an adjustable parameter that controls the width of the kernel and affects the complexity of the model. ∥xy∥ 2 represents the square of the Euclidean distance between x and y;

[0102] (4.5) Multimodal features are connected through features to form a joint feature vector;

[0103] (4.6) Training the joint feature vector through an integrated learning model;

[0104] (4.7) Use multiple machine learning models as base learners, train each base learner, and obtain different feature representations and classification results;

[0105] (4.8) Voting or weighted averaging the classification results of all base learners to obtain the final classification result;

[0106] (4.9) The classification results of the base learner are used as new features and fused with the original features;

[0107] (4.10) Perform linear discriminant analysis on the fused features to obtain the optimized feature representation.

[0108] Preferably, the step (6) comprises:

[0109] (6.1) Apply convolutional neural network to model comprehensive features;

[0110] (6.2) Use cross-validation and hyperparameter tuning methods to train the model;

[0111] (6.3) Use the trained model to predict VTE risk in real time for hospitalized patients.

[0112] Preferably, the step (7) comprises:

[0113] (7.1) Combine the generative capabilities of the large language model to generate natural language explanations to help doctors understand the model's diagnostic process and results. This can be achieved by generating dialogues or reports, and the VTE risk prediction results can be fed back to clinicians through the hospital information system.

[0114] (7.2) Generate a VTE risk assessment report for the patient based on the patient's VTE risk profile, including risk score, key feature explanations, and recommended clinical decision support;

[0115] (7.3) Doctors carry out targeted VTE risk prevention and treatment for patients based on the patient's VTE risk assessment report.

[0116] A venous thrombosis risk assessment system based on a large language model and deep learning is also provided, the system comprising:

[0117] An information acquisition module, which acquires the basic information of the patient to be evaluated and obtains the patient's previous medical treatment data based on the patient's unique identity identifier;

[0118] The classification module classifies the acquired diagnosis and treatment data according to the content format into three categories: structured data, semi-structured data and unstructured data;

[0119] A preprocessing module, which performs data standardization and data clarification preprocessing on structured data;

[0120] Apply large language models to semi-structured data to extract text features and convert them into feature vectors;

[0121] Apply deep learning algorithms to remove noise and enhance images of unstructured data;

[0122] The multimodal module applies feature engineering to extract and select features from the three categories of data content, and then uses the selected features for feature fusion to form multimodal features;

[0123] A feature linking module that links multimodal features to form a joint feature vector;

[0124] A training module, which uses a multi-layer neural network to build and train a comprehensive model for the joint feature vector, and uses the trained model to make real-time predictions of the risk of VTE in patients;

[0125] A feedback module that feeds back the prediction results to clinicians and generates a detailed risk report.

[0126] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the protection scope of the technical solution of the present invention.

Claims

1. A venous thrombosis risk assessment method based on a large language model and deep learning, characterized by: It includes the following steps: (1) Obtain the basic information of the patient to be evaluated and obtain the patient's medical treatment data based on the patient's unique identity; (2) Classify the acquired diagnosis and treatment data according to the content format into three categories: structured data, semi-structured data, and unstructured data; (3) Perform data standardization and preprocessing for data clarity on structured data; apply large language models to extract text features from semi-structured data and convert them into feature vectors; apply deep learning algorithms to remove noise and enhance images from unstructured data; (4) Apply feature engineering to extract and select features from the three categories of data content, and use the selected features for feature fusion to form multimodal features; (5) Feature linking of multimodal features to form a joint feature vector; (6) Use a multi-layer neural network to construct and train a comprehensive model of the joint feature vector, and use the trained model to predict the risk of VTE in patients in real time; (7) The prediction results are fed back to clinicians and a detailed risk report is generated.

2. The method for assessing venous thrombosis risk based on a large language model and deep learning according to claim 1, characterized in that: In step (1), the patient's unique identity identifier is the ID number or the medical insurance card number, and the medical treatment data includes the patient's basic information, medical advice information, test indicators, examination results, medical images, electrocardiogram data, electroencephalogram data, electromyogram data, medical history, past medical history, current medical history, family history, nursing records, and surgical records.

3. The method for assessing venous thrombosis risk based on a large language model and deep learning according to claim 2, characterized in that: In step (2), structured data includes basic patient information and laboratory test results; semi-structured data includes medical records, past medical history, current medical history, family history, nursing records, surgical records, and examination reports; and unstructured data includes medical images, ultrasound images, electrocardiogram data, and electromyogram data.

4. The method for assessing venous thrombosis risk based on a large language model and deep learning according to claim 3, characterized in that: In step (3), the following steps are performed on the structured data: (3.1.1) Apply big data processing technology to clean, standardize and encode structured data; (3.1.2) Clean, standardize and encode all clinical diagnosis and treatment data; (3.1.3) Use multiple imputation to fill missing values ​​in combination with other patient information For missing data in clinical diagnosis and treatment records, represents the estimate of the i-th interpolation value in the clinical diagnosis and treatment data, which refers to the missing data point X when generating the interpolation data. i Filling value; μ X represents the expected value of variable X; Z i It is a random variable that conforms to the standard normal distribution and is automatically generated by the algorithm. It is used to introduce randomness in the generation of missing values ​​and ensure the diversity of interpolation values. X is the standard deviation of variable X, which indicates the distribution range of data points in clinical diagnosis and treatment data relative to the mean; (3.1.4) Combine clinical knowledge and use the upper and lower limit truncation method to handle abnormal data exceeding a certain threshold; (3.1.5) Achieve data normalization and standardization to form a standardized clinical diagnosis and treatment data set; (3.1.6) For each clinical test parameter, normalize it using its clinical normal range and transform the characteristic value according to its position within the normal range; (3.1.7) Check the outliers in the data and use Box-Cox transformation and inverse transformation to process the existing outliers.

5. The method for assessing venous thrombosis risk based on a large language model and deep learning according to claim 4, characterized in that: In step (3), the following steps are performed on the semi-structured data: (3.2.1) Segment and embed text data, convert text into vector representation, extract text features through a pre-trained large language model, and input these features into the deep learning model together with the features of structured data; (3.2.2) Apply word segmentation algorithms and combine clinical terminology lists and knowledge bases to perform structural word segmentation on medical text data The clinical text data is segmented, t is the specific segmentation to be extracted, d represents the clinical text currently being processed, D represents the set of all clinical text documents in the currently processed inpatient record, N represents the total number of documents in the clinical text document set D, and f t,d represents the number of times the word to be extracted t appears in the clinical text d currently being processed, t' represents other words that appear in the clinical text d currently being processed, and f t',d Represents the total number of all word segments in the currently processed clinical text d; (3.2.3) Embed medical text data; (3.2.4) Introduce a clinical context-aware attention mechanism, which enhances the model's ability to perceive important clinical information by calculating the importance of each word or phrase in the context; (3.2.5) Use a large language model to segment and embed text data and convert the text into a vector representation; (3.2.6) Text features are extracted through a pre-trained large language model and converted into numerical structured feature vectors. These feature vectors will be input into the deep learning model together with the features of the structured data.

6. The method for assessing venous thrombosis risk based on a large language model and deep learning according to claim 5, characterized in that: In step (3), the following steps are performed on the unstructured data: (3.3.1) Image enhancement through generative adversarial networks, improving image quality through adversarial training of the generator and the discriminator; (3.3.2) Use deep learning models to segment medical images and extract regions of interest; (3.3.3) Use convolutional neural networks to extract high-level features from the segmented images. These features reflect the morphology and texture of the image. (3.3.4) Introducing perceptual loss and structural similarity loss to retain the high-frequency details and structural information of the original medical image; (3.3.5) Calculate the mean and variance of the input image batches, and improve the stability and speed of model training by standardizing the input images of each batch; (3.3.6) Scale and translate the standardized image to convert the data back to an intuitive and physically meaningful range so that the model can understand and utilize its features when processing the image; (3.3.7) Combine the attention mechanism to process electrophysiological data of electrocardiogram and electromyography to improve the ability to capture long-term dependencies Q is the query matrix, which represents the information that needs to be processed currently, K is the key matrix, which represents all the information to be processed, and K T represents the transposed matrix of the key matrix K, d k is the characteristic dimension of the key vector, V is the value matrix, which is a set of information corresponding to the key matrix K, containing specific information related to each key in the key matrix K, and softmax() is a normalization function used to convert the query matrix Q and the key matrix K and the characteristic dimension d of the key vector into k The calculated values ​​are converted into probability distributions.

7. The method for assessing venous thrombosis risk based on a large language model and deep learning according to claim 6, characterized in that: The step (4) comprises: (4.1) Extract features from structured data, text extraction data, and image analysis data respectively, and form multimodal features based on dynamic time warping of features; (4.2) The extracted image features and text features are fused to construct a multimodal model, and the attention mechanism is used to weight the fusion of features of different modalities to obtain comprehensive patient information; (4.3) Add feature importance analysis to calculate the importance of different features and identify key risk factors; Feature Importance(x j ) represents the jth node x j The feature importance of N t is the number of samples of node t, ΔImpurity(t) is the purity change after the node splits, Importance(f j ) represents the jth feature f j importance, T represents the set of all nodes in the decision tree, t represents a node in the decision tree, I(v(t)) is an indicator function, when node t uses feature f j The value is 1 when splitting, otherwise it is 0. v(t) represents the features used by node t, and Δi(t) represents the information gain of node t, which is calculated by the difference in impurity metrics before and after the node split; (4.4) Combined with radial basis function kernel to process nonlinear features K(x,y) is the input vector, which represents the kernel function value between x and y. γ is an adjustable parameter that controls the width of the kernel and affects the complexity of the model. ∥xy∥ 2 represents the square of the Euclidean distance between x and y; (4.5) Multimodal features are connected through features to form a joint feature vector; (4.6) Training the joint feature vector through an integrated learning model; (4.7) Use multiple machine learning models as base learners, train each base learner, and obtain different feature representations and classification results; (4.8) Voting or weighted averaging the classification results of all base learners to obtain the final classification result; (4.9) The classification results of the base learner are used as new features and fused with the original features; (4.10) Perform linear discriminant analysis on the fused features to obtain the optimized feature representation.

8. The method for assessing venous thrombosis risk based on a large language model and deep learning according to claim 7, characterized in that: The step (6) comprises: (6.1) Apply convolutional neural network to model comprehensive features; (6.2) Use cross-validation and hyperparameter tuning methods to train the model; (6.3) Use the trained model to predict VTE risk in real time for hospitalized patients.

9. The method for assessing venous thrombosis risk based on a large language model and deep learning according to claim 7, characterized in that: The step (7) comprises: (7.1) Combine the generative capabilities of the large language model to generate natural language explanations to help doctors understand the model's diagnostic process and results. This can be achieved by generating dialogues or reports, and the VTE risk prediction results can be fed back to clinicians through the hospital information system. (7.2) Generate a VTE risk assessment report for the patient based on the patient's VTE risk profile, including risk score, key feature explanations, and recommended clinical decision support; (7.3) Doctors carry out targeted VTE risk prevention and treatment for patients based on the patient's VTE risk assessment report.

10. Venous thrombosis risk assessment system based on large language model and deep learning, characterized by: It includes: An information acquisition module, which acquires the basic information of the patient to be evaluated and obtains the patient's previous medical treatment data based on the patient's unique identity identifier; The classification module classifies the acquired diagnosis and treatment data according to the content format into three categories: structured data, semi-structured data and unstructured data; A preprocessing module, which performs data standardization and data clarification preprocessing on structured data; Apply large language models to semi-structured data to extract text features and convert them into feature vectors; Apply deep learning algorithms to remove noise and enhance images of unstructured data; The multimodal module applies feature engineering to extract and select features from the three categories of data content, and then uses the selected features for feature fusion to form multimodal features; The feature linking module links the multimodal features to form a joint feature vector. The training module uses a multi-layer neural network to build and train a comprehensive model for the joint feature vector, and uses the trained model to make real-time predictions on the risk of VTE in patients. The feedback module feeds back the prediction results to clinicians and generates a detailed risk report.

Images