Cardiovascular disease medical decision support method based on large model

Abstract

The invention relates to the technical field of cardiovascular disease treatment, and discloses a cardiovascular disease medical decision support method based on a large model, and the method comprises the steps: obtaining multi-modal cardiovascular disease medical data, constructing a multi-source heterogeneous data set, carrying out the preprocessing and feature fusion, and constructing a combined representation model based on the pre-training large model and the fusion of traditional and western medicine knowledge. And designing a hierarchical space-time diagram convolutional network to generate a dynamic physiological state code of the patient, and constructing a multi-target optimization decision tree to generate a treatment scheme containing western medicines, traditional Chinese medicine prescriptions and non-drug intervention. A federal learning framework is adopted to cooperatively train a local model, data security is guaranteed, a global model is optimized, and a curative effect attribution module is further arranged to correct decision deviation. According to the method, multi-modal data and traditional Chinese and western medicine knowledge are integrated, personalized treatment scheme generation is achieved, data safety is guaranteed, and the cardiovascular disease diagnosis and treatment level is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of cardiovascular disease treatment, and particularly to a medical decision-making support method for cardiovascular diseases based on a large model. Background Art

[0002] Cardiovascular diseases, as major diseases that seriously threaten human health globally, are characterized by high incidence, high mortality, and high disability rates. According to statistics from the World Health Organization, the number of people who die from cardiovascular diseases each year accounts for nearly one-third of the total global deaths, bringing a heavy economic burden to society and families. In the diagnosis and treatment process of cardiovascular diseases, accurate decision-making is crucial for improving the prognosis of patients, but there are still many challenges at present.

[0003] From the data level, the data related to cardiovascular diseases presents the characteristics of multimodality and heterogeneity. Electronic health records contain text data such as patient basic information, medical history, and diagnosis results; medical imaging data such as echocardiograms and coronary angiography images reflect the heart structure and function in the form of images; traditional Chinese medicine diagnosis and treatment texts record unique diagnosis and syndrome differentiation and treatment information of traditional Chinese medicine; biomarker detection results and real-time monitoring data from wearable devices provide quantitative data of physiological indicators. These data sources are extensive and in various formats. How to effectively integrate and utilize these multi-source heterogeneous data and mine valuable information from them has become a key problem in improving the diagnosis and treatment level of cardiovascular diseases. Traditional methods often can only process one or several types of data in isolation and are difficult to fully utilize the comprehensive advantages of multimodal data.

[0004] In terms of diagnosis and treatment decision-making, Western medicine mainly relies on clinical guidelines and evidence-based medicine evidence. However, the individual differences among different patients are relatively large, and clinical guidelines are difficult to fully cover all situations, resulting in limitations in actual applications. Traditional Chinese medicine emphasizes individualized treatment through syndrome differentiation and treatment, but the subjectivity of traditional Chinese medicine diagnosis and treatment is relatively strong, lacking objective quantitative standards, and the inheritance and application of traditional Chinese medicine knowledge rely heavily on doctors' experience. In addition, the selection of treatment plans for cardiovascular diseases needs to comprehensively consider various factors such as treatment effects, side effects, and medical costs. At present, there is a lack of an effective method that can comprehensively weigh these factors and provide accurate and personalized treatment suggestions for doctors.

[0005] In terms of the distribution of medical resources, high-quality medical resources are concentrated in big cities and large hospitals, and the diagnosis and treatment levels of primary medical institutions are relatively low. When facing complex cardiovascular diseases, primary doctors are difficult to make accurate diagnoses and reasonable treatment decisions due to lack of experience and professional knowledge. At the same time, patients' medical data are scattered in various medical institutions, and the phenomenon of data islands is serious, making it impossible to achieve data sharing and collaborative utilization, which restricts the overall development of medical technology and the fairness of medical services.

[0006] In terms of technical means, most of the existing medical decision support systems are based on simple rules or statistical models and are difficult to handle complex medical knowledge and large-scale medical data. With the development of artificial intelligence technology, although some machine learning and deep learning models have been applied to the diagnosis and treatment of cardiovascular diseases, these models still have deficiencies in aspects such as knowledge integration, interpretability, and privacy protection. For example, deep learning models are usually black-box models, and it is difficult for doctors to understand the decision-making process and basis of the models, resulting in trust barriers in practical applications; during the data sharing and model training processes, patients' private data is easily leaked, leading to security risks. Summary of the Invention

[0007] The purpose of the present invention is to provide a medical decision support method for cardiovascular diseases based on a large model to solve the problems proposed in the above background technology.

[0008] To achieve the above purpose, the present invention provides the following technical solution: A medical decision support method for cardiovascular diseases based on a large model, the method includes: Step 1: Obtain multi-modal medical data for cardiovascular diseases, including electronic health records, medical imaging data, traditional Chinese medicine diagnosis and treatment texts, biomarker detection results, and real-time monitoring data from wearable devices, and construct a multi-source heterogeneous data set; Step 2: Preprocess the multi-source heterogeneous data, use an adaptive noise-robust alignment algorithm to fill in missing values, use a deep adversarial generation network to generate synthetic data to balance the class distribution, and achieve feature alignment and fusion between different data modalities through a cross-modal attention mechanism; Step 3: Based on a pre-trained large language model, fuse the traditional Chinese medicine syndrome diagnosis knowledge and the western medicine pathology knowledge graph, construct a combined Chinese and western medicine embedding representation layer, and use dynamic knowledge distillation technology to inject domain expert experience into the model parameters to form a multi-modal combined representation model for cardiovascular diseases; Step 4: Design a hierarchical spatio-temporal graph convolutional network to extract local features from time-series physiological signals, combine global spatio-temporal dependence modeling, and generate a patient's dynamic physiological state encoding; Step 5: Construct a multi-objective optimization decision tree, integrate reinforcement learning strategies and fuzzy logic reasoning, and generate a treatment recommendation plan according to the patient's individual characteristics, including western medicine recommendations, traditional Chinese medicine formula compatibility, and non-drug intervention strategies; Step 6: Use a federated learning framework to collaboratively train the local models of decentralized medical institutions, protect patient data security through differential privacy, and use a model parameter aggregation algorithm to continuously optimize and iterate the global model.

[0009] Preferably, in step 1, the parsing of the traditional Chinese medicine diagnosis and treatment text uses a BERT-based traditional Chinese medicine term entity recognition model, combines a bidirectional long short-term memory network and a conditional random field to extract syndrome elements, and performs semantic enhancement through a traditional Chinese medicine classic prescription knowledge base.

[0010] Preferably, in step 2, the cross-modal attention mechanism adopts a multi-head self-attention structure to allocate weights to the feature vectors of different modal data, screens features with high cross-modal relevance through a gating mechanism, and optimizes the inter-modal consistency using a contrastive loss function.

[0011] Preferably, in step 3, the dynamic knowledge distillation technology includes a teacher-student architecture. The teacher model is a symbolic reasoning system that integrates traditional Chinese and Western medicine expert rules, and the student model is a pre-trained large model based on Transformer. The soft label distribution is adjusted through an adaptive temperature coefficient to achieve knowledge transfer and model lightweighting.

[0012] Preferably, in step 4, the hierarchical spatio-temporal graph convolutional network includes a local spatio-temporal module and a global relationship module. The local module uses dilated convolution to capture high-frequency physiological signal features, and the global module models the interaction relationship between organ systems through a graph attention network.

[0013] Preferably, in step 5, the multi-objective optimization decision tree takes the patient survival rate, minimization of treatment side effects, and medical cost control as optimization objectives, uses a non-dominated sorting genetic algorithm to generate a Pareto optimal solution set, and screens the final solution through fuzzy comprehensive evaluation.

[0014] Preferably, the membership function of the fuzzy comprehensive evaluation is dynamically adjusted according to clinical guidelines, and the matching degree between the treatment plan and the patient's individual characteristics is quantified by combining grey relational analysis.

[0015] Preferably, in step 6, the local model training of the federated learning adopts an adaptive gradient clipping strategy to constrain the parameter update amplitude to prevent privacy leakage, and aggregates heterogeneous distributed model parameters through an elastic averaging algorithm.

[0016] Preferably, the elastic averaging algorithm introduces a dynamic weight allocation mechanism to calculate the aggregation weight according to the data distribution difference and contribution degree of the local model, and reduces the network load through sparse communication.

[0017] Preferably, the method further includes: using a counterfactual neural network to simulate potential outcomes under different treatment plans, and correcting decision biases through causal effect estimation.

[0018] Compared with the prior art, the beneficial effects of the present invention are: The present invention obtains electronic health records, medical imaging data, traditional Chinese medicine (TCM) diagnosis and treatment texts, biomarker detection results, and real-time monitoring data from wearable devices, constructs a multi-source heterogeneous dataset, and performs preprocessing and feature fusion through an adaptive noise robust alignment algorithm, a deep adversarial generation network, and a cross-modal attention mechanism. This not only improves the integrity and quality of the data, but also fully explores the potential connections between different modal data, providing more comprehensive and accurate information for subsequent analysis. For example, by combining medical imaging data with electronic health records, the type and severity of cardiovascular diseases can be more accurately judged; fusing TCM diagnosis and treatment texts with biomarker detection results helps to comprehensively evaluate the patient's physical condition from both the overall and microscopic levels, providing richer basis for personalized treatment.

[0019] Based on a pre-trained large language model, it fuses TCM syndrome diagnosis knowledge and Western medicine pathology knowledge graphs to construct a combined Chinese and Western medicine embedding representation layer. This fusion method gives full play to the respective advantages of Chinese and Western medicine. It not only utilizes the in-depth research on the pathological mechanisms of diseases and precise diagnostic methods in Western medicine, but also combines the holistic concept and syndrome differentiation and treatment characteristics of TCM, providing a more comprehensive and systematic knowledge system for the diagnosis and treatment of cardiovascular diseases. For example, during the diagnosis process, the examination indicators in Western medicine can provide an objective basis for syndrome differentiation in TCM, while the syndrome diagnosis in TCM can grasp the patient's constitution and the development trend of the disease as a whole, assisting Western medicine in formulating a more reasonable treatment plan. Through this combination of Chinese and Western medicine, the accuracy of diagnosis and the effectiveness of treatment can be improved, providing better medical services for patients.

[0020] Construct a multi-objective optimization decision tree, integrating reinforcement learning strategies and fuzzy logic reasoning, with the optimization objectives of maximizing the patient's survival rate, minimizing treatment side effects, and controlling medical costs. Generate a treatment recommendation plan according to the patient's individual characteristics, including Western medicine recommendations, TCM formula compatibility, and non-drug intervention strategies. This method can comprehensively consider the individual differences of patients and multiple treatment objectives, avoiding the limitations of single-objective decision-making. For example, for patients of different ages, genders, physical conditions, and economic conditions, the system can select the treatment plan with the least side effects and the lowest medical costs according to their specific situations while ensuring the treatment effect. At the same time, through the recommendation of non-drug intervention strategies, such as diet adjustment and exercise guidance, it helps to improve the patient's quality of life and promote the recovery of the disease.

[0021] Design a hierarchical spatio-temporal graph convolutional network to perform local feature extraction and global spatio-temporal dependence modeling on time-series physiological signals, and generate patient dynamic physiological state encodings. This network can more accurately capture the changing patterns of physiological signals and the interaction relationships between organ systems, timely detect abnormal changes in the patient's physiological state, and provide a basis for early diagnosis and treatment. For example, by analyzing physiological signals such as heart rate and blood pressure monitored in real time by wearable devices, it is possible to timely detect abnormal fluctuations in the patient's cardiovascular function, early warning of the risk of cardiovascular diseases, enabling doctors to take corresponding intervention measures and reduce the harm of the diseases.

[0022] Adopt a federated learning framework to co-train local models of decentralized medical institutions, protect patient data security through differential privacy, and use a model parameter aggregation algorithm to achieve continuous optimization and iteration of the global model. In this way, on the premise of protecting patient privacy, the sharing and collaborative utilization of medical data are realized, fully integrating the medical resources and data advantages of each medical institution, and improving the generalization ability and accuracy of the model. For example, primary medical institutions can use federated learning technology to leverage the rich data and advanced models of large hospitals to improve their own diagnosis and treatment levels, while not involving the transmission and sharing of patient privacy data, ensuring the information security of patients. In addition, by injecting domain expert experience into model parameters through dynamic knowledge distillation technology, knowledge transfer and model lightweighting are achieved, improving the running efficiency and interpretability of the model, making it easier for doctors to understand and trust the decision-making results of the model.

[0023] Based on a causal inference-based treatment effect attribution module, use a counterfactual neural network to simulate potential outcomes under different treatment plans, and correct decision-making biases through causal effect estimation. This helps doctors more accurately evaluate the effectiveness of treatment plans, understand the impact mechanisms of different treatment measures on the patient's condition, and thus timely adjust treatment plans to improve the precision and effectiveness of treatment. For example, during the treatment process, if it is found that the actual effect of a certain treatment plan does not match the expectation, the treatment effect attribution module can analyze the reasons, determine whether it is a problem with the treatment plan itself or other factors, provide a reference for subsequent treatment decisions, avoid blind treatment, and reduce the pain of patients and the waste of medical resources. Description of the Drawings

[0024] Figure 1 It is the working principle diagram of the cardiovascular disease medical decision-making support method described in the present invention; Figure 2 It is the working flow chart of traditional Chinese medicine diagnosis and treatment text parsing; Figure 3 It is the working flow chart of the hierarchical spatio-temporal graph convolutional network. Detailed Implementation Modes

[0025] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] See also Figures 1-3 The present invention provides a cardiovascular disease medical decision support method based on a large model, and its overall implementation scheme is as follows: Step 1: Obtain multimodal cardiovascular disease medical data, including electronic health records, medical imaging data, TCM diagnosis and treatment texts, biomarker test results, and real-time monitoring data from wearable devices, to construct a multi-source heterogeneous data set. Electronic health records cover basic patient information, past medical history, diagnostic records, etc.; medical imaging data such as cardiac ultrasound and coronary angiography images are used to observe heart structure and function; TCM diagnosis and treatment texts include TCM diagnosis, syndrome differentiation and treatment, etc.; biomarker test results reflect changes in biological indicators related to cardiovascular disease; real-time monitoring data from wearable devices provide dynamic physiological information such as heart rate and blood pressure. Integrate these data from different sources and formats to construct a multi-source heterogeneous data set.

[0027] Step 2: Preprocess multi-source heterogeneous data. Adopt adaptive noise robust alignment algorithm to fill missing values. This algorithm can adaptively determine the filling strategy according to the distribution characteristics and noise conditions of the data, effectively improving the integrity of the data. Use deep adversarial generative network to generate synthetic data to balance the category distribution. By generating synthetic data similar to real data, the problem of too little data in some categories in the data set is solved. The feature alignment and fusion between different data modalities are realized through the cross-modal attention mechanism, so that data of different modalities can complement each other and provide more comprehensive information for subsequent analysis.

[0028] Step 3: Based on the pre-trained large language model, integrate the knowledge of TCM syndrome diagnosis and the knowledge graph of Western medicine pathology to build a joint embedding representation layer of TCM and Western medicine. By integrating the knowledge of TCM and Western medicine, the knowledge reserve of the model is enriched. Dynamic knowledge distillation technology is used to inject the experience of domain experts into the model parameters to form a multimodal joint representation model of cardiovascular disease, so that the model can better simulate the diagnostic thinking of experts.

[0029] Step 4: Design a hierarchical spatiotemporal graph convolutional network to extract local features of temporal physiological signals, and combine them with global spatiotemporal dependency modeling to generate a dynamic physiological state encoding of the patient. Detailed information of physiological signals is obtained through local feature extraction, and the correlation of physiological signals in time and space is considered through global spatiotemporal dependency modeling, so as to more accurately describe the patient's dynamic physiological state.

[0030] Step 5: Construct a multi-objective optimization decision tree, integrate reinforcement learning strategies and fuzzy logic reasoning, and generate a treatment recommendation plan according to the individual characteristics of patients, including western medicine recommendations, traditional Chinese medicine formula compatibility, and non-drug intervention strategies. Considering multiple optimization objectives and comprehensively applying different methods to provide personalized treatment plans for patients.

[0031] Step 6: Use the federated learning framework to co-train the local models of decentralized medical institutions, protect patient data security through differential privacy, and use the model parameter aggregation algorithm to achieve continuous optimization and iteration of the global model. On the premise of protecting patient data privacy, integrate the data of each medical institution and continuously optimize the model performance.

[0032] The following further illustrates the implementation of the present invention in combination with Embodiments 1 to 5.

[0033] Embodiment 1: When constructing a multi-source heterogeneous dataset in Step 1, it involves the parsing of traditional Chinese medicine diagnosis and treatment texts. The parsing of traditional Chinese medicine diagnosis and treatment texts uses a BERT-based traditional Chinese medicine term entity recognition model. The BERT model has strong language understanding ability and can perform in-depth semantic understanding on the words in traditional Chinese medicine diagnosis and treatment texts. First, perform word segmentation on the traditional Chinese medicine diagnosis and treatment texts to split the texts into individual word units. Then input these word-segmented texts into the BERT-based traditional Chinese medicine term entity recognition model, and the model will label each word and identify the traditional Chinese medicine term entities therein, such as symptoms, disease names, formula names, etc.

[0034] Extract syndrome elements by combining bidirectional long short-term memory network (Bi-LSTM) and conditional random field (CRF). Bi-LSTM can learn the forward and backward information of the text simultaneously and better capture the long-distance dependencies in the text. Input the text features processed by the BERT model into Bi-LSTM, and Bi-LSTM further learns and processes these features to extract more representative features. The conditional random field (CRF) is used to constrain and optimize the output results of Bi-LSTM, considering the context relationship between labels to improve the accuracy of syndrome element extraction. For example, when extracting symptoms such as "palpitation, shortness of breath, and fatigue", CRF can accurately judge which syndrome elements they belong to according to the logical relationship between these symptoms.

[0035] Semantic enhancement is performed through the traditional Chinese medicine classic formula knowledge base. The extracted traditional Chinese medicine term entities and syndrome elements are compared and associated with the traditional Chinese medicine classic formula knowledge base. If relevant formula information is found in the knowledge base for the current text, this information can be used to supplement and strengthen the semantics of the text. For example, when symptoms such as "palpitation, insomnia, and dreaminess" are identified and judged as the syndrome of heart blood deficiency, and it is found in the knowledge base that Guipi Decoction is applicable to this syndrome, the understanding of this text can be further enriched, knowing the possible treatment directions, and also providing a reference for subsequent compatibility of traditional Chinese medicine formulas.

[0036] Example 2: When achieving feature alignment and fusion between different data modalities in step 2, a cross-modal attention mechanism is adopted. The cross-modal attention mechanism adopts a multi-head self-attention structure. Assume there are data in different modalities, such as the medical image data modality and the electronic health record data modality etc. For the data of each modality, first, its feature vectors are linearly transformed to obtain multiple different projection vectors. Taking the medical image data as an example, assume its feature vector is , and through the linear transformation the projection vector is obtained, and through the linear transformation the projection vector is obtained, and through the linear transformation the projection vector is obtained.

[0037] Then, the attention scores are calculated. The formula is , where is the dimension of K. Here, the attention scores represent the degree of association between different feature vectors. Through the multi-head self-attention structure, the attention scores can be calculated from multiple different perspectives to obtain multiple different attention representations, and then these representations are concatenated to obtain a richer feature representation.

[0038] Features with high cross-modal relevance are screened through a gating mechanism. The gating mechanism can screen features according to the calculated attention scores. Set a threshold . When the attention score is greater than , it is considered that this feature has high cross-modal relevance and this feature is retained; when the attention score is less than , this feature is discarded. In this way, some noise features with low cross-modal relevance can be removed to improve the quality of feature fusion.

[0039] The contrastive loss function is used to optimize the inter-modal consistency. The definition of the contrastive loss function is:

[0040] where N is the number of samples, denotes the label of the sample pair (if they are of the same class then , otherwise ), d is the distance metric function, is the feature representation of the sample x by the model, is the sample of the same class as is the sample of a different class from, and m is a boundary value. By minimizing the contrastive loss function, the feature representations of similar samples within the same modality become closer, and the feature representations of similar samples between different modalities also become closer, thus optimizing the consistency between modalities.

[0041] Example 3: When constructing the multi-modal joint representation model for cardiovascular diseases in step 3, the dynamic knowledge distillation technique is adopted. The dynamic knowledge distillation technique includes a teacher-student architecture, and the teacher model is a symbolic reasoning system that integrates Chinese and Western medical expert rules. This symbolic reasoning system integrates the diagnostic experience and rules of experts in the fields of Chinese and Western medicine. For example, in Western medicine, based on indicators such as the patient's blood pressure, blood lipids, electrocardiogram, etc., combined with the diagnostic criteria summarized by experts, the type and severity of cardiovascular diseases are judged; in Chinese medicine, based on the principle of syndrome differentiation and treatment in Chinese medicine, the patient's symptoms, tongue manifestations, pulse conditions, etc. are analyzed to obtain the diagnosis of Chinese medicine syndromes.

[0042] The student model is a pre-trained large model based on Transformer. During the training process, the soft label distribution is adjusted through the adaptive temperature coefficient T. Let the output of the teacher model be , and the output of the student model be , then the soft label distribution is adjusted in the following way: , . By adjusting the temperature coefficient T, the smoothness of the soft label can be controlled. When T is large, the soft label distribution is smoother and the knowledge transferred is more ambiguous, but it can prevent the student model from overfitting the output of the teacher model; when T is small, the soft label distribution is more concentrated and the knowledge transferred is more precise, but it may cause the student model to overfit.

[0043] During the training process, by minimizing the loss function between the output of the student model and the soft label, knowledge transfer and model lightweighting are achieved. The loss function can be defined as: , where KL is the KL divergence, which is used to measure the difference between two probability distributions. By continuously optimizing this loss function, the student model gradually learns the knowledge in the teacher model. At the same time, since the student model can adjust and optimize the parameters during the learning process, the model is lightweighted and the running efficiency of the model is improved.

[0044] Example 4: In step 4, a hierarchical spatio-temporal graph convolutional network is designed, which includes a local spatio-temporal module and a global relationship module.

[0045] The local spatio-temporal module uses dilated convolution to capture high-frequency physiological signal features. Suppose the input temporal physiological signal is , and the dilated convolution operation can be expressed as: , where is the output of the dilated convolution, is the convolution kernel weight, M is the convolution kernel size, r is the dilation rate, is the input signal. By setting different dilation rates, the receptive field of the convolution kernel can be expanded without increasing the number of parameters, so as to capture physiological signal features of different frequencies. For example, for heart rate signals, by appropriately setting the dilation rate, short-term fluctuations (high-frequency features) of the heart rate can be captured, and these high-frequency features may reflect the physiological changes of the patient in a short period of time, such as the heart rate change after exercise, etc.

[0046] The global relationship module models the interaction relationships between organ systems through a graph attention network. First, the physiological signals of different organ systems are used as the nodes of the graph. For example, the physiological signal of the heart is used as one node, and the physiological signal of the blood vessels is used as another node. Then, the attention weights between the nodes are calculated. By calculating the attention weights, the interaction intensity between different organ systems can be determined, so as to better model the mutual relationship between organ systems. For example, the contraction and relaxation of the heart will affect the blood pressure in the blood vessels, and this relationship can be captured through the graph attention network, providing a basis for generating a more accurate encoding of the patient's dynamic physiological state.

[0047] Example 5: This embodiment focuses on the implementation method of the optimization objective of the multi-objective optimization decision tree and the local model training and parameter aggregation technologies under the federated learning framework, improving the scientificity of treatment plan generation and the security and efficiency of model training.

[0048] In step 5, a multi-objective optimization decision tree is constructed with the patient survival rate, minimization of treatment side effects, and medical cost control as the optimization objectives. The non-dominated sorting genetic algorithm is used to generate the Pareto optimal solution set. The non-dominated sorting genetic algorithm first sorts the individuals in the initial population according to the non-dominated relationship of multi-objective optimization, divides the individuals not dominated by other individuals into the first level, and then continues to find the non-dominated individuals in the remaining individuals and divides them into the second level, and so on. During the genetic operation process, the population is continuously evolved through operations such as selection, crossover, and mutation. The selection operation uses the tournament selection method, randomly selects multiple individuals from the population for comparison, and selects the individuals with higher fitness to enter the next generation. The crossover operation generates new individuals by exchanging part of the genes of two individuals, and the mutation operation randomly changes the genes of the individuals. Through continuous iteration, the Pareto optimal solution set is generated, and this solution set contains multiple treatment plans that achieve a balance between different objectives.

[0049] The final plan is screened through fuzzy comprehensive evaluation. The membership function of fuzzy comprehensive evaluation is dynamically adjusted according to clinical guidelines. For example, for the evaluation index of treatment side effects, according to the regulations on the severity and occurrence probability of different side effects in clinical guidelines, the parameters of the membership function are determined. The matching degree between the treatment plan and the patient's individual characteristics is quantified by combining grey relational analysis. Grey relational analysis measures the matching degree between the two by calculating the grey relational degree between the characteristic vector of the treatment plan and the individual characteristic vector of the patient. Through grey relational analysis, the treatment plan that best matches the patient's individual characteristics can be screened out from the Pareto optimal solution set.

[0050] In step 6, the local model training of federated learning adopts the adaptive gradient clipping strategy. The heterogeneous distributed model parameters are aggregated through the elastic averaging algorithm. The elastic averaging algorithm introduces a dynamic weight allocation mechanism, and calculates the aggregation weight according to the data distribution difference and contribution degree of the local models. In this way, the local models with small data distribution differences and large contribution degrees have greater weights in parameter aggregation. At the same time, the network load is reduced by sparsifying communication, only part of the important model parameters are transmitted, the data transmission volume is reduced, and the efficiency of federated learning is improved.

[0051] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "including", "comprising" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0052] Although embodiments of the present invention have been shown and described, those of ordinary skill in the art will appreciate that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A cardiovascular disease medical decision support method based on a large model, characterized in that: The following steps are involved: Step 1: Obtain multimodal cardiovascular disease medical data, including electronic health records, medical imaging data, traditional Chinese medicine diagnosis and treatment texts, biomarker test results, and real-time monitoring data from wearable devices, and construct a multi-source heterogeneous dataset; Step 2: Preprocess the multi-source heterogeneous data, use the adaptive noise robust alignment algorithm to fill the missing values, use the deep adversarial generative network to generate synthetic data to balance the category distribution, and realize the feature alignment and fusion between different data modalities through the cross-modal attention mechanism; Step 3: Based on the pre-trained large language model, the knowledge of TCM syndrome diagnosis and the knowledge graph of Western medicine pathology are integrated to build a joint embedding representation layer of TCM and Western medicine. Dynamic knowledge distillation technology is used to inject domain expert experience into model parameters to form a multimodal joint representation model of cardiovascular disease. Step 4: Design a hierarchical spatiotemporal graph convolutional network to extract local features of temporal physiological signals and generate a dynamic physiological state code for the patient by combining global spatiotemporal dependency modeling; Step 5: Construct a multi-objective optimization decision tree, integrate reinforcement learning strategy and fuzzy logic reasoning, and generate treatment recommendations based on the individual characteristics of the patient, including Western medicine recommendations, Chinese medicine prescription compatibility, and non-drug intervention strategies; Step 6: Use the federated learning framework to collaboratively train local models of decentralized medical institutions, protect patient data security through differential privacy, and use the model parameter aggregation algorithm to achieve continuous optimization and iteration of the global model.

2. The cardiovascular disease medical decision support method based on a large model according to claim 1, characterized in that: The analysis of the TCM diagnosis and treatment text described in step 1 adopts a BERT-based TCM terminology entity recognition model, combined with a bidirectional long short-term memory network and conditional random fields to extract syndrome elements, and performs semantic enhancement through a TCM classic prescription knowledge base.

3. The cardiovascular disease medical decision support method based on a large model according to claim 1, characterized in that: The cross-modal attention mechanism described in step 2 adopts a multi-head self-attention structure to assign weights to feature vectors of data of different modalities, screens features with high cross-modal correlation through a gating mechanism, and uses a contrast loss function to optimize inter-modal consistency.

4. The cardiovascular disease medical decision support method based on a large model according to claim 1, characterized in that: The dynamic knowledge distillation technology described in step 3 includes a teacher-student architecture. The teacher model is a symbolic reasoning system that integrates the rules of Chinese and Western medicine experts. The student model is a large pre-trained model based on Transformer. The soft label distribution is adjusted through an adaptive temperature coefficient to achieve knowledge transfer and model lightweighting.

5. The cardiovascular disease medical decision support method based on a large model according to claim 1, characterized in that: The hierarchical spatiotemporal graph convolutional network described in step 4 includes a local spatiotemporal module and a global relationship module. The local module uses a dilated convolution to capture high-frequency physiological signal features, and the global module models the interaction relationship between organ systems through a graph attention network.

6. The cardiovascular disease medical decision support method based on a large model according to claim 1, characterized in that: The multi-objective optimization decision tree described in step 5 takes patient survival rate, minimization of treatment side effects and control of medical costs as optimization objectives, uses a non-dominated sorting genetic algorithm to generate a Pareto optimal solution set, and selects the final solution through fuzzy comprehensive evaluation.

7. The cardiovascular disease medical decision support method based on a large model according to claim 6, characterized in that: The membership function of the fuzzy comprehensive evaluation is dynamically adjusted according to clinical guidelines, and the matching degree between the treatment plan and the individual characteristics of the patient is quantified by combining grey correlation analysis.

8. The cardiovascular disease medical decision support method based on a large model according to claim 1, characterized in that: The local model training of federated learning described in step 6 adopts an adaptive gradient clipping strategy, constrains the parameter update amplitude to prevent privacy leakage, and aggregates heterogeneous distributed model parameters through an elastic averaging algorithm.

9. The cardiovascular disease medical decision support method based on a large model according to claim 8, characterized in that: The elastic averaging algorithm introduces a dynamic weight allocation mechanism, calculates the aggregation weight according to the data distribution difference and contribution of the local model, and reduces the network load by sparse communication.

10. The cardiovascular disease medical decision support method based on a large model according to claim 1, characterized in that: The method further includes: using a counterfactual neural network to simulate potential outcomes under different treatment options and correcting decision biases through causal effect estimation.

Images