Myocardial infarction risk assessment method, system and equipment based on multi-modal interpretability

Abstract

The invention discloses a myocardial infarction risk assessment method, system and equipment based on multi-modal interpretability. The method comprises the following steps: constructing a dynamic weighted fuzzy cognitive map model for myocardial infarction risk assessment; inputting the aggregated feature vector into a dynamic weighted fuzzy cognitive map model, updating the activation value of each concept node by adopting the finely-adjusted weight and the activation function until a preset condition is met to obtain a final activation value of each concept node, and taking the final activation value of the myocardial infarction risk concept node as a myocardial infarction risk assessment result; calculating the contribution degree of each physiological index in the aggregated feature vector according to a myocardial infarction risk assessment result; searching a maximum influence propagation path from each source node to the target node to obtain at least one maximum influence propagation path; and constructing an interpretable myocardial infarction risk assessment report based on the contribution degree of each physiological index and the at least one maximum influence propagation path. According to the invention, interpretable and high-precision model evaluation is realized.

Description

Technical Field

[0001] The present application relates to the field of medical health technology, and in particular to a method, system and device for assessing the risk of myocardial infarction based on multimodal interpretability. Background Art

[0002] Early and accurate diagnosis of cardiovascular diseases, such as myocardial infarction, is crucial for improving patient survival. With the advancement of information technology, intelligent risk assessment using multimodal data such as electrocardiograms (ECGs), blood markers, and medical imaging has become a research hotspot. Currently, intelligent diagnostic methods based on machine learning, particularly deep learning, are able to automatically learn complex features from massive amounts of data, achieving significant progress in recognition accuracy. However, in the high-risk field of healthcare, diagnostic decisions require not only accurate results but also a transparent, logically clear process that can be understood and trusted by clinicians. The "black box" nature of existing data-driven models, i.e., the opacity of their decision-making logic, constitutes a major obstacle to their widespread application in clinical practice.

[0003] While existing technologies for myocardial infarction identification have achieved some progress in diagnostic accuracy, they still face widespread challenges, including high model complexity, high computational overhead, and an inherent conflict between model accuracy and interpretability. In particular, existing interpretability methods, while supplementing "black box" models, do not fundamentally address the opaque nature of the decision-making process. This makes it difficult for models to achieve the reliability and credibility of their results when processing real-world medical data with noise or missing information, meeting the stringent clinical requirements for transparency and traceability. Summary of the Invention

[0004] This application aims to propose a myocardial infarction risk assessment method, system and equipment based on multimodal interpretability, which can achieve interpretable and high-precision model evaluation, thereby improving the reliability and credibility of myocardial infarction risk assessment results.

[0005] In a first aspect, an embodiment of the present application provides a method for assessing myocardial infarction risk based on multimodal interpretability, the method comprising: Extract multimodal physiological index features including electrocardiogram clinical features, blood marker features, and imaging radiomics features; Splicing each feature in the multimodal physiological indicator feature to obtain a spliced ​​feature vector; Calculating the attention score of each feature in the multimodal physiological indicator feature according to the spliced ​​feature vector, and performing weighted summation on each feature in the multimodal physiological indicator feature based on the attention score to obtain an aggregated feature vector; A dynamic weighted fuzzy cognitive graph model for myocardial infarction risk assessment was constructed. The dynamic weighted fuzzy cognitive graph model included a weight adjustment mechanism, multiple concept nodes, weighted edges, and an activation function. The weight adjustment mechanism was used to fine-tune the basic weights based on the context vector to obtain fine-tuned weights. The weighted edge was an edge connecting two concept nodes and was used to represent the influence of one concept node on another concept node. Inputting the aggregated feature vector into the dynamic weighted fuzzy cognitive graph model, updating the activation value of each of the concept nodes using the fine-tuned weights and activation function until a final activation value of each of the concept nodes is obtained after a preset condition is met, and using the final activation value of the myocardial infarction risk concept node as the myocardial infarction risk assessment result; Calculating the contribution of each physiological indicator in the aggregated feature vector according to the myocardial infarction risk assessment result; Finding the maximum impact propagation path between each source node and the target node based on the preset concept node sensitivity coefficient, the path context attenuation factor, the multiple concept nodes, and the weighted edges, and obtaining at least one maximum impact propagation path; Based on the contribution of each physiological indicator and the at least one maximum impact transmission path, an interpretable myocardial infarction risk assessment report is constructed.

[0006] Compared with the prior art, the first aspect of the present application has the following beneficial effects: This method extracts multimodal physiological indicator features including electrocardiogram clinical features, blood marker features and imaging radiomics features; splices each feature in the multimodal physiological indicator features to obtain a spliced ​​feature vector; calculates the attention score of each feature in the multimodal physiological indicator features based on the spliced ​​feature vector, and based on the attention score, performs weighted summation on each feature in the multimodal physiological indicator features to obtain an aggregated feature vector; constructs a dynamic weighted fuzzy cognitive graph model for myocardial infarction risk assessment, the dynamic weighted fuzzy cognitive graph model includes a weight adjustment mechanism, multiple concept nodes, weighted edges and activation functions, the weight adjustment mechanism is used to fine-tune the basic weight based on the context vector to obtain the fine-tuned weight, the weighted edge is an edge connecting two concept nodes, and is used to represent a concept node. The influence strength of a concept node on another concept node is calculated; the aggregated feature vector is input into the dynamic weighted fuzzy cognitive graph model, and the fine-tuned weight and activation function are used to update the activation value of each concept node until the preset conditions are met to obtain the final activation value of each concept node, and the final activation value of the myocardial infarction risk concept node is used as the myocardial infarction risk assessment result; according to the myocardial infarction risk assessment result, the contribution of each physiological indicator in the aggregated feature vector is calculated; according to the preset concept node sensitivity coefficient, path context attenuation factor, multiple concept nodes and weighted edges, the maximum influence propagation path between each source node and the target node is found, and at least one maximum influence propagation path is obtained; based on the contribution of each physiological indicator and at least one maximum influence propagation path, an interpretable myocardial infarction risk assessment report is constructed. In this way, comprehensive consideration of the characteristics of multimodal physiological indicators can improve the accuracy of myocardial infarction risk assessment results. By inputting the aggregated feature vector into the dynamic weighted fuzzy cognitive graph model and using the fine-tuned weights and activation functions to update the activation value of each concept node, the myocardial infarction risk assessment results can be obtained. The dynamic weighted fuzzy cognitive graph model directly constructs model parameters through expert knowledge rather than data-driven training, which effectively reduces the computational overhead of the model. Finally, by calculating the contribution of each physiological indicator and the maximum impact propagation path, an interpretable myocardial infarction risk assessment report is constructed, which can make the model interpretable, thereby improving the reliability and credibility of the myocardial infarction risk assessment results.

[0007] In some embodiments, the extracting of multimodal physiological indicator features comprising electrocardiogram clinical features, blood marker features, and imaging radiomics features comprises: Performing baseline correction on an electrocardiogram signal in the acquired electrocardiogram to obtain a corrected electrocardiogram signal; performing harmonic wavelet packet decomposition on the corrected ECG signal to obtain a wavelet packet decomposition tree; Calculating the time-frequency entropy of each node in the wavelet packet decomposition tree according to the node coefficients in the wavelet packet decomposition tree, and obtaining the optimal wavelet basis by minimizing the time-frequency entropy; extracting electrocardiogram clinical features including ST segment deviation and T wave symmetry from the optimal wavelet basis; The extracted blood markers are standardized to obtain blood marker characteristics; Extract radiomic features including texture features from medical images to obtain image radiomic features; The electrocardiogram clinical features, the blood marker features and the imaging radiomics features are constructed into multimodal physiological indicator features.

[0008] In some embodiments, updating the activation value of each of the concept nodes using the fine-tuned weights and activation functions includes: ; in, Represents a concept node In the The new activation value after iterations, Indicates in In the iteration, the concept node All other concept nodes that have an impact The activation value of Represents a concept node Pointing to concept node After fine-tuning the weights, Represents a concept node The initial activation value of represents the activation function, Indicates the total number of concept nodes.

[0009] In some embodiments, calculating the contribution of each physiological indicator in the aggregated feature vector based on the myocardial infarction risk assessment result includes: Replacing each physiological indicator in the aggregated feature vector with a clinical baseline value to construct a counterfactual input vector corresponding to each physiological indicator; Inputting the counterfactual input vector into the dynamic weighted fuzzy cognitive graph model to obtain a counterfactual risk value corresponding to each physiological indicator; The counterfactual risk value is differentially calculated from the myocardial infarction risk assessment result to obtain the contribution of each physiological indicator in the aggregated feature vector.

[0010] In some embodiments, the step of finding the maximum impact propagation path between each source node and the target node based on a preset concept node sensitivity coefficient, a path context attenuation factor, the multiple concept nodes, and the weighted edges to obtain at least one maximum impact propagation path includes: Build at least one path from each source node to the target node; If each of the paths includes multiple concept nodes and weighted edges, the influence flow between two adjacent concept nodes is calculated based on the preset concept node sensitivity coefficient, the final activation values ​​corresponding to the sub-source nodes of the two adjacent concept nodes, and the fine-tuned weights of the weighted edges between the two adjacent concept nodes; performing weighted summation of the influence flows between all two adjacent concept nodes in each path according to the path context attenuation factor to obtain the total influence of each path; According to the total influence, a maximum influence propagation path between each source node and the target node is found to obtain at least one maximum influence propagation path.

[0011] In some embodiments, constructing an interpretable myocardial infarction risk assessment report based on the contribution of each physiological indicator and the at least one maximum impact propagation path includes: By sorting the contribution of each physiological indicator and screening out multiple key physiological indicators, a quantitative attribution summary including the indicator name corresponding to each key physiological indicator, the specific value derived from the aggregated feature vector, and the contribution is obtained; Matching the maximum impact propagation path corresponding to each of the key physiological indicators, and obtaining a core logic path including the name and final activation value of each concept node in the maximum impact propagation path; Visually encode the final activation value of each concept node and the maximum impact propagation path corresponding to each key physiological indicator to obtain a visual decision map; The quantitative attribution summary, the core logic path, and the visual decision map are integrated to construct an interpretable myocardial infarction risk assessment report.

[0012] In some embodiments, after constructing an interpretable myocardial infarction risk assessment report, the method further comprises: constructing the base weights as a prior probability distribution of weights; According to the prior probability distribution of the weight, the posterior probability distribution of the weight is obtained using Bayes' theorem; The weight distribution mean obtained from the posterior probability distribution of the weight is used as a new weight in the dynamic weighted fuzzy cognitive map model to update the dynamic weighted fuzzy cognitive map model.

[0013] In a second aspect, an embodiment of the present application further provides a myocardial infarction risk assessment system based on multimodal interpretability, the system comprising: A data extraction unit is used to extract multimodal physiological index features including electrocardiogram clinical features, blood marker features, and imaging radiomics features; A feature splicing unit, configured to splice each feature in the multimodal physiological indicator feature to obtain a spliced ​​feature vector; a feature aggregation unit, configured to calculate an attention score of each feature in the multimodal physiological indicator feature according to the spliced ​​feature vector, and perform a weighted summation of each feature in the multimodal physiological indicator feature based on the attention score to obtain an aggregated feature vector; A model construction unit, configured to construct a dynamic weighted fuzzy cognitive graph model for myocardial infarction risk assessment, wherein the dynamic weighted fuzzy cognitive graph model comprises a weight adjustment mechanism, a plurality of concept nodes, weighted edges, and an activation function. The weight adjustment mechanism is configured to fine-tune a basic weight based on a context vector to obtain a fine-tuned weight. The weighted edge is an edge connecting two concept nodes and is configured to characterize the strength of influence of one concept node on another concept node. a risk assessment unit, configured to input the aggregated feature vector into the dynamic weighted fuzzy cognitive graph model, update the activation value of each of the concept nodes using the fine-tuned weights and activation function, obtain a final activation value of each of the concept nodes after a preset condition is met, and use the final activation value of the myocardial infarction risk concept node as a myocardial infarction risk assessment result; a contribution calculation unit, configured to calculate the contribution of each physiological indicator in the aggregated feature vector according to the myocardial infarction risk assessment result; A path finding unit, configured to find a maximum-influence propagation path between each source node and a target node based on a preset concept node sensitivity coefficient, a path context attenuation factor, the plurality of concept nodes, and the weighted edges, and obtain at least one maximum-influence propagation path; An assessment report construction unit is used to construct an interpretable myocardial infarction risk assessment report based on the contribution of each physiological indicator and the at least one maximum impact propagation path.

[0014] In a third aspect, an embodiment of the present application also provides an electronic device comprising at least one control processor and a memory for communicating with the at least one control processor; the memory stores instructions that can be executed by the at least one control processor, and the instructions are executed by the at least one control processor so that the at least one control processor can execute a myocardial infarction risk assessment method based on multimodal interpretability as described above.

[0015] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute a myocardial infarction risk assessment method based on multimodal interpretability as described above.

[0016] It can be understood that the beneficial effects of the above-mentioned second to fourth aspects compared with the relevant technologies are the same as the beneficial effects of the above-mentioned first aspect compared with the relevant technologies. Please refer to the relevant description in the above-mentioned first aspect and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which: Figure 1 1 is a flowchart of an embodiment of a method for assessing myocardial infarction risk based on multimodal interpretability provided by the present application; Figure 2 1 is a schematic structural diagram of an embodiment of a myocardial infarction risk assessment system based on multimodal interpretability provided by the present application; Figure 3 It is a structural diagram of an embodiment of the electronic device provided by this application. DETAILED DESCRIPTION

[0018] The following describes in detail embodiments of the present application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application and are not to be construed as limiting the present application.

[0019] In the description of this application, if there is a description of first, second, etc., it is only for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0020] In the description of this application, it should be understood that descriptions involving orientation, such as the orientation or positional relationship indicated by up, down, etc., are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on this application.

[0021] In the description of this application, it should be noted that, unless otherwise clearly defined, terms such as setting, installing, and connecting should be understood in a broad sense, and technical personnel in the relevant technical field can reasonably determine the specific meaning of the above terms in this application based on the specific content of the technical solution.

[0022] While existing technologies for myocardial infarction identification have achieved some progress in diagnostic accuracy, they still face widespread challenges, including high model complexity, high computational overhead, and an inherent conflict between model accuracy and interpretability. In particular, existing interpretability methods, while supplementing "black box" models, do not fundamentally address the opaque nature of the decision-making process. This makes it difficult for models to achieve the reliability and credibility of their results when processing real-world medical data with noise or missing information, meeting the stringent clinical requirements for transparency and traceability.

[0023] In order to solve the problem of low reliability and credibility of the results of the above-mentioned existing technologies, the present application proposes a method, system and device for myocardial infarction risk assessment based on multimodal interpretability.

[0024] Reference Figure 1 , a flow chart of a method for assessing risk of myocardial infarction based on multimodal interpretability provided in an embodiment of the present application. The method for assessing risk of myocardial infarction based on multimodal interpretability is applied to an electronic device, which may be a server or a mobile terminal. Figure 1 As shown, the myocardial infarction risk assessment method based on multimodal interpretability may include the following steps: Step S101: extracting multimodal physiological index features including electrocardiogram clinical features, blood marker features, and imaging radiomics features; Step S102: splicing each feature in the multimodal physiological indicator feature to obtain a spliced ​​feature vector; Step S103: Calculate the attention score of each feature in the multimodal physiological indicator feature according to the spliced ​​feature vector, and perform weighted summation of each feature in the multimodal physiological indicator feature based on the attention score to obtain an aggregated feature vector; Step S104: Constructing a dynamic weighted fuzzy cognitive graph model for myocardial infarction risk assessment. The dynamic weighted fuzzy cognitive graph model includes a weight adjustment mechanism, multiple concept nodes, weighted edges, and an activation function. The weight adjustment mechanism is used to fine-tune the basic weights based on the context vector to obtain fine-tuned weights. The weighted edge is an edge connecting two concept nodes and is used to represent the influence strength of one concept node on another concept node. Step S105: Input the aggregated feature vector into the dynamic weighted fuzzy cognitive graph model, and use the fine-tuned weights and activation function to update the activation value of each concept node until the preset conditions are met to obtain the final activation value of each concept node. The final activation value of the myocardial infarction risk concept node is used as the myocardial infarction risk assessment result. Step S106: Calculate the contribution of each physiological indicator in the aggregated feature vector based on the myocardial infarction risk assessment result; Step S107: Find the maximum influence propagation path between each source node and the target node based on the preset concept node sensitivity coefficient, the path context attenuation factor, the multiple concept nodes, and the weighted edges, and obtain at least one maximum influence propagation path; Step S108: Based on the contribution of each physiological indicator and at least one maximum impact propagation path, an interpretable myocardial infarction risk assessment report is constructed.

[0025] In this embodiment, multimodal physiological indicator features including electrocardiogram clinical features, blood marker features and imaging radiomics features are extracted; each feature in the multimodal physiological indicator features is spliced ​​to obtain a spliced ​​feature vector; based on the spliced ​​feature vector, the attention score of each feature in the multimodal physiological indicator features is calculated, and based on the attention score, each feature in the multimodal physiological indicator features is weighted and summed to obtain an aggregated feature vector; a dynamic weighted fuzzy cognitive graph model for myocardial infarction risk assessment is constructed, and the dynamic weighted fuzzy cognitive graph model includes a weight adjustment mechanism, multiple concept nodes, weighted edges and activation functions, and the weight adjustment mechanism is used to fine-tune the basic weight based on the context vector to obtain the fine-tuned weight, and the weighted edge is an edge connecting two concept nodes, which is used to represent a concept The influence strength of a concept node on another concept node; the aggregated feature vector is input into the dynamic weighted fuzzy cognitive graph model, and the fine-tuned weight and activation function are used to update the activation value of each concept node until the preset conditions are met to obtain the final activation value of each concept node, and the final activation value of the myocardial infarction risk concept node is used as the myocardial infarction risk assessment result; according to the myocardial infarction risk assessment result, the contribution of each physiological indicator in the aggregated feature vector is calculated; according to the preset concept node sensitivity coefficient, path context attenuation factor, multiple concept nodes and weighted edges, the maximum influence propagation path between each source node and the target node is found, and at least one maximum influence propagation path is obtained; based on the contribution of each physiological indicator and at least one maximum influence propagation path, an interpretable myocardial infarction risk assessment report is constructed. In this way, comprehensive consideration of the characteristics of multimodal physiological indicators can improve the accuracy of myocardial infarction risk assessment results. By inputting the aggregated feature vector into the dynamic weighted fuzzy cognitive graph model and using the fine-tuned weights and activation functions to update the activation value of each concept node, the myocardial infarction risk assessment results can be obtained. The dynamic weighted fuzzy cognitive graph model directly constructs model parameters through expert knowledge rather than data-driven training, which effectively reduces the computational overhead of the model. Finally, by calculating the contribution of each physiological indicator and the maximum impact propagation path, an interpretable myocardial infarction risk assessment report is constructed, which can make the model interpretable, thereby improving the reliability and credibility of the myocardial infarction risk assessment results.

[0026] The above-mentioned dynamic weighted fuzzy cognitive graph model for myocardial infarction risk assessment can be constructed by taking each component in the aggregated feature vector as a concept node for constructing a dynamic weighted fuzzy cognitive graph. The dynamic weighted fuzzy cognitive graph also includes concept nodes such as "ST segment abnormality", "myocardial enzyme level", "coronary artery stenosis degree" and finally "myocardial infarction risk". The edge connecting two concept nodes is called a weighted edge. The activation value of each concept node is updated through the activation function. A weight adjustment mechanism is introduced in the model reasoning process to fine-tune the basic weight of the dynamic weighted fuzzy cognitive graph itself based on the context vector. The final activation value of the concept node "myocardial infarction risk" is used as the myocardial infarction risk assessment result. Therefore, the dynamic weighted fuzzy cognitive graph constitutes the dynamic weighted fuzzy cognitive graph model of this embodiment.

[0027] In some embodiments, extracting multimodal physiological indicator features comprising electrocardiogram clinical features, blood marker features, and imaging radiomics features comprises: Performing baseline correction on an electrocardiogram signal in the acquired electrocardiogram to obtain a corrected electrocardiogram signal; Perform harmonic wavelet packet decomposition on the corrected ECG signal to obtain a wavelet packet decomposition tree; According to the node coefficients in the wavelet packet decomposition tree, the time-frequency entropy of each node in the wavelet packet decomposition tree is calculated, and the optimal wavelet basis is obtained by minimizing the time-frequency entropy. Extracting electrocardiographic clinical features including ST segment deviation and T wave symmetry from the optimal wavelet basis; The extracted blood markers are standardized to obtain blood marker characteristics; Extract radiomic features including texture features from medical images to obtain image radiomic features; The electrocardiogram clinical characteristics, blood marker characteristics and imaging radiomics characteristics are constructed into multimodal physiological indicator characteristics.

[0028] In this embodiment, baseline correction can eliminate the baseline drift and low-frequency noise of the ECG signal, ensuring that subsequent features reflect the true physiological state. Through harmonic wavelet packet decomposition and time-frequency entropy guidance, the optimal decomposition basis that best characterizes the signal structure can be adaptively selected, thereby more finely capturing transient change information related to myocardial infarction, and extracting accurate ECG clinical characteristics, blood marker characteristics, and imaging radiomics characteristics, laying a good data foundation for the subsequent dynamic weighted fuzzy cognitive map model to perform myocardial infarction risk assessment, which can improve the accuracy of myocardial infarction risk assessment results.

[0029] In some embodiments, updating the activation value of each concept node using the fine-tuned weights and activation functions includes: ; in, Represents a concept node In the The new activation value after iterations, Indicates in In the iteration, the concept node All other concept nodes that have an impact The activation value of Represents a concept node Pointing to concept node After fine-tuning the weights, Represents a concept node The initial activation value of represents the activation function, Indicates the total number of concept nodes.

[0030] In this embodiment, by considering the influence of all other concept nodes on the current concept node, the dynamic weighted fuzzy cognitive graph model can obtain a more accurate myocardial infarction risk assessment result.

[0031] In some embodiments, calculating the contribution of each physiological indicator in the aggregated feature vector based on the myocardial infarction risk assessment result includes: Each physiological indicator in the aggregated feature vector is replaced with a clinical baseline value to construct a counterfactual input vector corresponding to each physiological indicator; The counterfactual input vector is input into the dynamic weighted fuzzy cognitive graph model to obtain the counterfactual risk value corresponding to each physiological indicator; The difference between the counterfactual risk value and the myocardial infarction risk assessment result is calculated to obtain the contribution of each physiological indicator in the aggregated feature vector.

[0032] In this embodiment, by introducing counterfactual thinking, the marginal effect of each physiological indicator is systematically evaluated, so that the contribution of each physiological indicator to the risk outcome can be quantitatively evaluated and the key risk drivers can be identified.

[0033] In some embodiments, based on a preset concept node sensitivity coefficient, a path context attenuation factor, multiple concept nodes, and weighted edges, finding a maximum impact propagation path between each source node and a target node to obtain at least one maximum impact propagation path includes: Build at least one path from each source node to the target node; If each path contains multiple concept nodes and weighted edges, the influence flow between two adjacent concept nodes is calculated based on the preset concept node sensitivity coefficient, the final activation value corresponding to the sub-source node of the two adjacent concept nodes, and the fine-tuned weight of the weighted edge between the two adjacent concept nodes; According to the path context attenuation factor, the influence flows between all two adjacent concept nodes in each path are weighted and summed to obtain the total influence of each path; According to the total influence, the maximum influence propagation path between each source node and the target node is found to obtain at least one maximum influence propagation path.

[0034] In this embodiment, by deeply analyzing the internal reasoning process of the dynamic weighted fuzzy cognitive graph model, the medical logic chain of important causes is clearly revealed, that is, why the indicators are important is revealed; by introducing the concept node sensitivity coefficient, the complex and nonlinear causal effects in medicine can be simulated, and the contribution of the activation value to the influence flow can be adjusted; by introducing the path context attenuation factor, the contribution of the single-step influence between two adjacent concept nodes to the total influence can be correspondingly attenuated, which is more in line with the loss characteristics of information transmission in medical logic.

[0035] In some embodiments, based on the contribution of each physiological indicator and at least one maximum impact propagation path, an interpretable myocardial infarction risk assessment report is constructed, including: By ranking the contribution of each physiological indicator and screening multiple key physiological indicators, a quantitative attribution summary containing the indicator name corresponding to each key physiological indicator, the specific value derived from the aggregated feature vector, and the contribution is obtained; Match the maximum impact propagation path corresponding to each key physiological indicator, and obtain the core logic path containing the name and final activation value of each concept node in the maximum impact propagation path; Visually encode the final activation value of each concept node and the maximum impact propagation path corresponding to each key physiological indicator to obtain a visual decision map; Integrate the quantitative attribution summary, core logic path and visual decision map to construct an interpretable myocardial infarction risk assessment report.

[0036] In this embodiment, contribution analysis is integrated with path activation information, and key indicators are weighted and ranked based on medical semantics to generate a comprehensive explainable report that is well-organized and in line with clinical logic, thereby enhancing the system's decision-making support capabilities for medical personnel.

[0037] In some embodiments, after constructing an interpretable myocardial infarction risk assessment report, the method further comprises: Construct the basis weights as a prior probability distribution of weights; According to the prior probability distribution of weights, the posterior probability distribution of weights is obtained using Bayes' theorem; The weight distribution mean obtained from the posterior probability distribution of the weight is used as a new weight in the dynamic weighted fuzzy cognitive map model to update the dynamic weighted fuzzy cognitive map model.

[0038] In this embodiment, the weight distribution mean obtained from the posterior probability distribution of the weight is used as a new weight in the dynamic weighted fuzzy cognitive map model to update the dynamic weighted fuzzy cognitive map model. This enables the dynamic weighted fuzzy cognitive map model to have an uncertainty measurement of its own knowledge. One-time feedback from an expert will not subversively change the dynamic weighted fuzzy cognitive map model, but will be smoothly integrated with the original knowledge. In addition, each update has a clear record of the Bayesian reasoning process, which has data traceability.

[0039] To facilitate understanding by those skilled in the art, a set of best embodiments is provided below: Existing technologies include data-driven black-box evaluation models and post-process interpretability analysis methods for black-box models. Specifically, data-driven black-box evaluation models are represented by deep learning algorithms such as deep neural networks (DNNs) and convolutional neural networks (CNNs). Their core approach is to learn the nonlinear mapping relationship between raw input data (such as ECG waveforms) and risk conclusions through training on large-scale annotated datasets. While such models can achieve high predictive accuracy, their decision-making process is hidden within millions of model parameters, creating a "black box." Doctors cannot understand the specific clinical indicators used by the model or the medical logic chain used to reach its final conclusion, making it difficult to fully trust the model's output in critical clinical decisions.

[0040] Post-process interpretability analysis methods for black-box models, such as gradient attribution analysis and latent semantic analysis, attempt to explain the model's behavior by calculating the contribution or sensitivity of input features to the output after the model makes a prediction. Their limitation lies in the fact that these explanations are essentially external observations and approximate inferences of a known, opaque model, rather than direct representations of the model's internal decision-making logic. They can reveal "what the model pays attention to," but cannot clearly answer "why the model makes such a decision." Furthermore, the computational costs of these explanation methods often incur additional system overhead, increasing the overall complexity of the model.

[0041] The above prior art has the following disadvantages: 1. The decision-making process lacks inherent explainability and medical logic.

[0042] The core flaw of existing technologies lies in their data-driven nature, which dictates that their decision-making logic is statistical rather than grounded in medical causal knowledge. The model itself is a complex mathematical function, lacking explicit medical concepts and inference rules for review. This fundamentally deviates from the clinician's thinking paradigm, making it difficult to establish trust. Post-processing interpretation techniques can only provide limited clues and fail to fundamentally make the model transparent.

[0043] 2. It is highly dependent on large-scale data training and has high computational costs.

[0044] Building deep learning models requires significant computing resources and time to train on large-scale, high-quality, annotated medical data. Once the model is trained, its knowledge is embedded in the network weights. When medical knowledge is updated or needs to be fine-tuned for specific situations, costly data collection and model retraining are often necessary.

[0045] 3. Knowledge updating and calibration are difficult.

[0046] In data-driven models, medical knowledge is implicitly encoded and cannot be modified directly and precisely. If a model's logic is found to be inconsistent with the latest clinical guidelines, researchers cannot directly modify the model as they would by modifying rules. Consequently, there is a lack of a flexible and efficient, expert-led mechanism for iterating and maintaining model knowledge.

[0047] In order to address the shortcomings of the above-mentioned existing technologies, this embodiment aims to provide a myocardial infarction risk assessment method based on multimodal interpretability. By proposing a set of innovative, expert knowledge-driven instead of data-driven complete technical solutions, it fundamentally overcomes the inherent defects of existing technologies in interpretability, computational efficiency, and knowledge maintenance. This embodiment aims to provide an intelligent risk assessment tool that is both highly accurate and fully transparent, ensuring that doctors can understand and trust every decision-making basis of the model, thereby significantly improving the clinical application value and reliability of the intelligent auxiliary diagnosis system for myocardial infarction. The main purpose of the method of this embodiment is to: 1. Provide a risk assessment framework that is inherently transparent and driven by medical knowledge.

[0048] This implementation aims to build an assessment engine centered around an expert-defined Dynamic Weighted Fuzzy Cognitive Map (DW-FCM). The weights of the engine's concept nodes and weighted edges directly correspond to explicit medical concepts and causal relationships, making the entire risk reasoning process completely transparent and aligning its logic with clinical thinking, fundamentally addressing the "black box" problem.

[0049] 2. Achieved risk assessment with zero training and low computational overhead.

[0050] This embodiment aims to completely eliminate the reliance on large-scale data and computationally intensive training processes by directly transforming medical expert knowledge into model structures. This not only significantly reduces the cost of building and deploying system models, but also enables real-time evaluation in resource-constrained environments.

[0051] 3. Establish a dynamic knowledge calibration mechanism led by experts.

[0052] This embodiment aims to introduce an expert feedback-driven parameter calibration loop (EFPCL). This mechanism allows medical experts to directly modify and improve the concepts, causal relationships, and their strengths in the evaluation model based on the latest clinical evidence or practical experience, ensuring that the model can keep pace with the times and achieve continuous and traceable evolution of knowledge.

[0053] 4. Achieved multi-level and in-depth interpretation of risk assessment results.

[0054] This embodiment not only makes the model itself transparent but also aims to provide a systematic explanatory mechanism. It quantifies the importance of each input indicator through structural perturbation counterfactual contribution analysis (SP-CCA), and clearly displays the complete logical chain of conclusions through decision path tracing and concept activation visualization (DPT-CAV), providing comprehensive, in-depth, and reliable auxiliary information for clinical decision-making.

[0055] This implementation addresses the needs of MI risk assessment and interpretability by constructing a two-stage framework based entirely on expert knowledge, directly constructing model parameters through expert knowledge rather than data-driven training. The first stage (MPIR-KRA) utilizes multimodal physiological indicator refinement (MPIR) techniques, including adaptive baseline correction, multiscale wavelet transform (ABC-MWT), and radiomics feature extraction (PRFE), to transform raw electrocardiogram (ECG), blood markers, and imaging data into structured clinical indicators. These are then fed into a dynamic weighted fuzzy cognitive map (DW-FCM) model for risk inference. The second phase (MIMA-KC) focuses on the output results of the DW-FCM model. It adopts two complementary modules, structural perturbation counterfactual contribution analysis (SP-CCA) and decision path traceability concept activation visualization (DPT-CAV), to quantify "which indicator is important" and "why it is so important", and integrates the two through semantic guided contribution focus (SGCF) to generate a comprehensive explanation report; finally, it introduces an expert feedback driven parameter calibration loop (EFPCL) to continuously optimize the DW-FCM structure and weights without data retraining, to ensure the accuracy and credibility of the system model in the evolution of medical knowledge.

[0056] The method of this embodiment specifically includes the following contents: Phase 1: Multimodal physiological indicator refinement and knowledge-driven risk assessment (MPIR-KRA).

[0057] This phase constructed an end-to-end knowledge flow from raw signals to risk assessment. First, the Multimodal Physiological Index Extraction and Quantification (MPIR) module generated a unified indicator vector through steps such as adaptive baseline correction, multiscale wavelet decomposition, clinical feature extraction, blood marker standardization, and radiomics extraction. Subsequently, the risk assessment engine, based on the Dynamic Weighted Fuzzy Cognitive Map (DW-FCM), leveraged expert-defined medical concepts and causal weights to iteratively activate and propagate between concept nodes, outputting a myocardial infarction risk value. This established the model framework and input / output anchors for subsequent interpretable analysis.

[0058] Step 1: Adaptive Harmonic Wavelet Packet Transform (AHWPT).

[0059] (1) Purpose: Traditional multi-scale wavelet decomposition uses a fixed wavelet basis, which is difficult to optimally adapt to the complex and non-stationary pathological features in electrocardiogram (ECG) signals. This step uses an adaptive harmonic wavelet packet transform (AHWPT) to adaptively select the optimal decomposition basis (i.e., the optimal wavelet basis) that best represents the signal structure through a mechanism guided by time-frequency entropy, thereby more precisely capturing transient changes related to myocardial infarction.

[0060] (2) Formula and description: Baseline correction: ,in, Represents the adaptive baseline correction operator.

[0061] Adaptive harmonic wavelet packet transform: First, the baseline-corrected ECG signal Perform harmonic wavelet packet decomposition to obtain a complete wavelet packet decomposition tree. For any node in the decomposition tree ( is the number of decomposition layers, is the frequency band index of the layer), its node coefficient is .

[0062] Then, calculate the time-frequency entropy of each node ,Time-frequency entropy is used to measure the uncertainty or complexity of a signal in the corresponding frequency band.

[0063] ; in, , is the normalized energy probability distribution.

[0064] Finally, according to the principle of minimizing time-frequency entropy, pruning is performed on the wavelet packet decomposition tree to find a path that minimizes the sum of the time-frequency entropy of all terminal nodes. This path constitutes the optimal wavelet basis.

[0065] .

[0066] in, Represents a complete basis in the wavelet packet decomposition tree (i.e., a set of terminal nodes covering all signal frequency bands). The optimization process is to find the optimal solution among all possible basis Find the basis that minimizes the sum of the time-frequency entropy of the terminal nodes in the set.

[0067] (3) Input / output and variables: enter: .

[0068] Output: and optimal wavelet basis .

[0069] symbol: is the original ECG vector, representing the original ECG signal vector collected at time t; is the corrected vector.

[0070] Step 2: Electrocardiogram (ECG) clinical feature extraction.

[0071] (1) Purpose: From the optimal wavelet basis Clinically relevant physiological indicators (including ST segment deviation and T wave symmetry) are extracted from the data, and these indicators are quantified to provide input for subsequent knowledge reasoning.

[0072] (2) Formula and description: ST segment deviation: ,in, Fixed time interval after the J-level point The potential at It represents the time point of ST segment deviation measurement starting from the J point according to the preset clinical criteria. is the baseline potential.

[0073] T wave symmetry: , in, 、 and They represent the time coordinates of the starting point, peak point, and ending point of the T wave respectively.

[0074] (3) Input / output and variables: Input: The optimal wavelet basis generated in step 1 .

[0075] Output: Clinical feature vector , which is a vector consisting of one or more ECG clinical features ( and ) splicing composition.

[0076] symbol: Indicates ST segment deviation; Indicates T wave symmetry; Indicates J-point time: the time coordinate of the turning point where the QRS complex ends and the ST segment begins in the electrocardiogram; represents the baseline potential, Represents the number of sampling points of the original ECG signal vector.

[0077] Step 3: Blood marker standardization.

[0078] (1) Purpose: To eliminate dimensions and remove anomalies to improve the comparability of various markers.

[0079] (2) Formula and description: ; in, is the raw data vector collected at the current moment, which contains one or more blood markers related to myocardial infarction assessment. In order to eliminate the influence of different marker dimensions and perform normalization, this embodiment adopts the Z-score normalization method. The statistical parameters required by this method (mean and standard deviation ), which is obtained by statistical calculation of the same blood marker data set collected from a large number of historical cases. The function limits the normalized result to the interval [-3, 3] to handle potential extreme outliers.

[0080] (3) Input / output and variables: enter: , Indicates the number of blood markers collected, Represents the original blood marker data vector collected at time t.

[0081] Output: Represents the original blood marker data vector collected at time t.

[0082] symbol: It represents the statistical parameters of blood markers calculated from a large amount of historical case data and is used for Z-score normalization.

[0083] Step 4: Image radiomics feature extraction.

[0084] (1) Purpose: This step aims to extract quantitative texture features from medical images. This is achieved by calculating the Gray-Level Co-occurrence Matrix (GLCM) and deriving multiple features based on this matrix. The elements in the GLCM are Indicates that the gray value in the image is The pixel and gray value are The probability that pixels of θ and θ co-occur in a specific spatial relationship.

[0085] (2) Formula and description: A series of radiomic features can be extracted from GLCM, such as: Contrast: Measures the intensity of local changes in an image, reflecting the clarity of the texture and the depth of the grooves.

[0086] ; Energy (Angular Second Moment, ASM): Measures the uniformity of image texture. A larger energy value indicates a more regular and uniform texture distribution.

[0087] ; (3) Input / output and variables: Input: Original medical images , Represents the dimensions of the original medical image.

[0088] Output: Radiomics feature vector , which is a concatenation of one or more extracted texture features, e.g. .

[0089] symbol: Represents the gray-level co-occurrence matrix composed of gray-level pairs Defined element values ​​(probability); Represents a set (vector) containing all extracted image texture features.

[0090] Step 5: Adaptive feature fusion based on attention mechanism.

[0091] (1) Purpose: The role of different modal physiological indicators in indicative of myocardial infarction risk is not static; their importance may vary depending on individual differences. Simple feature concatenation cannot reflect this dynamic nature. This step introduces an attention mechanism module to automatically learn the weights of each modal feature and achieve adaptive weighted fusion of information.

[0092] (2) Formula and description: The obtained electrocardiogram clinical characteristics, blood marker characteristics and imaging characteristics Concatenate into a feature vector .

[0093] Then, a small neural network (such as a fully connected layer) is used to calculate the attention score for each feature. : ; in, and represents the learnable parameters, The function ensures that the sum of all scores is 1.

[0094] The final aggregated feature vector is the weighted sum of the original eigenvectors:

[0095] in, 、 and Represent the attention scores of ECG clinical characteristics, blood marker characteristics, and imaging characteristics, respectively.

[0096] The parameters of the attention module can be obtained by pre-training on relevant datasets, or optimized in an expert feedback loop.

[0097] (3) Input / output and variables: enter: .

[0098] Output: Aggregated feature vector : A comprehensive feature vector formed by weighted summation of ECG clinical features, blood marker features, and imaging features.

[0099] Step 6: DW-FCM risk reasoning.

[0100] (1) Purpose: This step is the core of risk assessment and aims to use a dynamic weighted fuzzy cognitive graph model (DW-FCM model) constructed from medical expert knowledge to perform comprehensive reasoning on the multimodal feature vectors aggregated in step 5. The DW-FCM model simulates the thinking process of clinical experts when diagnosing myocardial infarction. It converts the input physiological indicators into the final quantitative risk value through the causal relationship and dynamic weights between concept nodes. Its purpose is to achieve a transparent, explainable and data-free risk assessment process.

[0101] (2) Formula and description: DW-FCM is a directed graph consisting of a set of concept nodes and weighted edges connecting these nodes constitute.

[0102] Concepts ): represents key concepts in the medical field, such as "ST segment abnormality", "myocardial enzyme level", "coronary artery stenosis degree" and the final "myocardial infarction risk" node. Each concept node Each has an activation value , indicating that at time The activity or quantification degree of the concept node is usually in the range of .

[0103] Weighted edges ): Basic weight Represents a concept node Concept Node The strength and nature of the effect (positive promotion, negative inhibition, or no effect).

[0104] Context-adaptive weight modulation mechanism: Unlike traditional static FCM, the weights in this embodiment can be dynamically fine-tuned based on specific input indicators (such as contextual information such as patient age, gender, and medical history), making it more personalized. However, its core reasoning process is still based on the following iterative calculation. Specifically, To achieve personalized evaluation, this embodiment introduces a weight modulation mechanism into the standard DW-FCM reasoning. The fine-tuned weight matrix Including fine-tuned weights, the system receives a context vector containing patient background information (such as age, gender, and medical history) . Basic weight According to the vector Perform fine-tuning to obtain the fine-tuned weights: ; in, Represents a concept node Pointing to concept node The basic weight, which is pre-defined by experts, is the weight of the dynamic weighted fuzzy cognitive map model itself. is a modulation function (e.g., a small neural network) that is tuned according to the context vector Output a specific connection In this way, for patients of different ages or genders, the causal reasoning strength within the model will change dynamically, making it more consistent with clinical reality. Subsequent iterative calculations will use the modulated weights .

[0105] (3) Risk reasoning process: Risk reasoning is an iterative calculation process that simulates the propagation of the causal chain by continuously updating the activation value of each concept node until the system model reaches a stable state. Specifically: 1) Initialization: The aggregated feature vector generated in step 5 Each component of the aggregated feature vector is mapped to the corresponding input layer concept node in DW-FCM, that is, each component in the aggregated feature vector is used as a node in the DW-FCM graph, and each component in the aggregated feature vector is used as the initial activation value .

[0106] 2) Iterative calculation: At each iteration step , each concept node Activation value Update according to the following formula: ; in, Represents a concept node In the The new activation value after iterations, represents the total number of concept nodes, Indicates in In the iteration, the concept node All other concept nodes that have an impact The activation value of Represents a concept node Pointing to concept node The fine-tuned weights represent right The strength of the causal influence, Represents a concept node The initial activation value of (if it is an input node), which ensures that the external input information continues to play a role in each iteration. Represents a nonlinear conversion function (or activation function) used to compress the calculation results into a specific range (such as ), common choices are Sigmoid function or tanh function, and other activation functions known to those skilled in the art can also be used, which are not specifically limited in this embodiment. For example, using Sigmoid function: ; in, is a parameter that controls the slope of the function and adjusts its sensitivity.

[0107] 3) Termination condition: The iteration process continues until the activation value change of all nodes is less than a preset minimum threshold , or the maximum number of iterations is reached , at this point the system is considered to have reached a stable state. The final "heart attack risk" node The activation value of is the final risk result of this assessment (i.e. myocardial infarction risk assessment result). .

[0108] (4) Input / output and variables: enter: The aggregated feature vector .

[0109] Predefined DW-FCM model, i.e. model , including: concept node set The basic weight matrix of expert knowledge . Output: Final myocardial infarction risk assessment .

[0110] The activation vectors of all concept nodes when reaching a stable state , which will serve as the input for the second stage of interpretability analysis. symbol: Concept Node In the The activation value of the iteration; Represents a concept node To concept node The basic weight of . Represents an activation function (such as Sigmoid). Represents the final output of myocardial infarction risk assessment value.

[0111] Summary: At this stage, this embodiment implements the refinement and risk assessment process of multimodal data (such as electrocardiogram signals, blood markers, and imaging data).

[0112] Phase 2: Construction of multi-level explainability mechanisms and knowledge calibration.

[0113] Building on the DW-FCM risk assessment engine constructed in Phase 1, this phase focuses on providing comprehensive, actionable, and interpretable risk assessment reports and establishing a non-data-driven system maintenance mechanism. The core of this approach is an interpretability mechanism that integrates feature contribution and decision path analysis. This mechanism is implemented through two complementary analytical techniques: First, Structural Perturbation Counterfactual Contribution Analysis (SP-CCA) quantifies the impact of each extracted input physiological indicator (i.e., the various features extracted in Phase 1), precisely identifying which indicator is more important; second, Decision Path Tracing and Concept Activation Visualization (DPT-CAV) deeply analyzes the internal reasoning process of the constructed DW-FCM model, clearly revealing the medical logic chain of key causes. The results of these two dimensions are intelligently integrated through the innovative Semantic Guided Contribution Focusing (SGCF) mechanism to generate a comprehensive interpretive report (i.e., the MI risk assessment report) that connects quantitative attribution with symbolic logic. Finally, to ensure the long-term effectiveness and accuracy of the DW-FCM model, this embodiment introduces an expert feedback-based parameter calibration loop (EFPCL), which allows medical experts to make direct and informed manual adjustments to the weights of weighted edges and concept nodes in the DW-FCM model based on clinical feedback, thereby achieving adaptive optimization of the system during the evolution of medical knowledge. This process is completely independent of data retraining. Specifically: 1. An interpretability mechanism that integrates feature contribution and decision path analysis.

[0114] 1.1. Structural perturbation-based counterfactual contribution analysis (SP-CCA).

[0115] (1) Purpose: To obtain the final risk assessment value in the first stage (MPIR-KRA) Finally, the primary explanatory requirement is to quantify the contribution of each raw input physiological indicator to the outcome. This step introduces counterfactual thinking to systematically evaluate the marginal effect of each indicator, thereby identifying the risk drivers of key indicators.

[0116] (2) Formula and description: The core of this embodiment is to compare the difference between the risk assessment results under the original input and the risk assessment results after removing or neutralizing the influence of a specific indicator. The final risk value is ,in, Represents the aggregated feature vector Include To calculate the physiological index Physiological indicators Contribution , construct a counterfactual input vector , among which Each indicator is replaced by its clinical baseline value (or neutral value, such as the median of the normal range) Replace, while other indicators remain unchanged, the specific calculation formula is: ; This counterfactual is then fed into the vector Re-input into the DW-FCM risk assessment engine, perform a complete iterative calculation until it stabilizes, and obtain a new counterfactual risk value . No. Contribution of physiological indicators Defined as the difference between the original risk and the counterfactual risk: ; The output value of this formula Intuitively quantify physiological indicators The net impact of the presence of the physiological indicator on the final risk value. A positive value indicates that the indicator increases the risk, a negative value indicates that it reduces the risk, and the absolute value represents the intensity of the impact of the physiological indicator.

[0117] 3. Input / Output and Variables: enter: The original feature vector after aggregation .

[0118] The predefined DW-FCM model includes a concept node set C and a basic weight matrix W.

[0119] Final risk assessment value and the final activation value of the concept node in the stable state .

[0120] Output: An M-dimensional contribution vector , where each element corresponds to the contribution of an input physiological indicator.

[0121] 1.2. Decision Path Tracing and Concept Activation Visualization (DPT-CAV).

[0122] (1) Purpose: SP-CCA reveals “which indicator is important” but fails to clarify the internal logical pathway of its impact on risk assessment. This step aims to Starting from this, through a novel nonlinear impact propagation model, we trace back and identify the key causal chain from high-contribution input nodes to the final risk output nodes, thereby answering the deep mechanism of "why it is important".

[0123] (2) Formula and description: This example searches for the maximum influence propagation path on the DW-FCM graph structure. Instead of using simple linear multiplication to calculate the influence flow, a nonlinear influence propagation function is defined. This function can better simulate complex, nonlinear causal effects in medicine, such as threshold effects and saturation effects. (which can be regarded as the source node) is passed to the concept node The influence flow of a node (which can be regarded as a target node) can be quantified as: ; in, Is the source node The final activation value of It is a concept node Concept Node The impact strength of and The connection weights between (i.e., the weights after fine-tuning), Represents the sensitivity coefficient of the concept node, which is preset by experts based on the physiological characteristics of the node (for example, a small change in some myocardial enzyme markers may cause a violent reaction, and its The higher the value, the higher the contribution of the activation value to the influence flow). The introduction of the (hyperbolic tangent function) makes the influence flow show nonlinear growth and tends to saturation when the activation value is very high, which is consistent with the physiological limit.

[0124] A complete decision path Is a concept node sequence ,in Usually it is an input concept node with high contribution (i.e. it can be regarded as a sub-source node). is the final risk concept node (which can be regarded as a sub-goal node), and the decision path Include concept nodes, Indicates the concept nodes , Indicates the concept nodes The total impact of the path It is defined as the weighted accumulation of the impact flow of each segment on the path, for example, the sum: ; in, express Path context decay factor of concept nodes This factor makes the causal chain The farther the distance between concept nodes is, the less contribution their single-step influence will make to the total influence, which is more in line with the loss characteristics of information transmission in medical logic.

[0125] This step adopts a method from the risk output node The greedy best-first search algorithm starts by tracing back. At each step, the algorithm selects the flow with the greatest influence Precursor concept node As the previous concept node of the path, until it is traced back to an input concept node. Through this method, one or more decision paths that contribute most to the final risk activation value can be identified. If we only need to find a path from a source node to a target node, there will be a decision path that contributes the most to the final risk activation value. If we need to find paths from multiple source nodes to target nodes, there will be multiple decision paths that contribute the most to the final risk activation value.

[0126] (3) Input / output and variables: enter: The weight matrix after fine-tuning and conceptual node sensitivity coefficients .

[0127] The final activation value of the concept node in the stable state .

[0128] High-contribution input indicators identified by SP-CCA and their corresponding concept nodes.

[0129] Output: One or more most influential propagation paths Each path consists of an ordered list of medical concept nodes and their corresponding influence flow values.

[0130] 1.3. Semantic-guided Contribution Focusing (SGCF) mechanism.

[0131] (1) Motivation: The separate contribution values ​​(from SP-CCA) and logical pathways (from DPT-CAV) are still fragmented information for clinicians. This step aims to deeply integrate the quantitative attribution results of SP-CCA with the qualitative logical pathways of DPT-CAV and transform the analysis results into a structured, semantically coherent, and clinically readable comprehensive interpretation report through a standardized report generation process.

[0132] (2) Formula and description: The report generation process of this mechanism begins with obtaining the contribution vector Processing is done by taking its absolute value Sort in descending order and based on the preset threshold Screening out a set of key physiological indicators The process then builds the three core modules of the report: First, the "Quantitative Attribution Summary" module is generated, which targets For each key physiological indicator in the input vector, clearly list its indicator name, The specific values ​​in and the contribution score calculated by SP-CCA ; Next, to provide deep logic analysis, the system creates a "core logic path" module, which Each key physiological indicator in the decision path set Search and match the total influence The largest path (i.e., one key physiological indicator corresponds to a maximum impact propagation path), and is presented in the form of a text causal chain, with the names of the intermediate concept nodes on the path and their final activation values ​​clearly marked. To further provide a global perspective, the system continues to generate a "visual decision map" module, which renders the complete DW-FCM graph and uses visual encoding (such as mapping activation values ​​by concept node size or color depth) to visualize the decision map. Finally, the three modules of quantitative attribution summary, core logic path and visual decision map are integrated into a report with the final risk assessment value. Output a comprehensive explanatory report with titles and unified format.

[0133] (3) Input / output and variables: enter: Contribution vector .

[0134] Maximum impact propagation path set .

[0135] Original input vector , the basic weight matrix of DW-FCM and the final activation value of the concept node in the stable state .

[0136] Output: A standardized, comprehensive explainability report with quantitative attribution, textualized logic paths, and visual maps.

[0137] 2. Expert knowledge calibration loop based on Bayesian fusion.

[0138] (1) Motivation: Any model based on prior knowledge must be able to evolve with the development of medical knowledge. To make the knowledge updating process more rigorous and mathematically sound, this embodiment proposes an expert knowledge calibration mechanism based on Bayesian fusion.

[0139] (2) Formula and description: In this framework, each basic weight in the dynamic weighted fuzzy cognitive graph model It is no longer a fixed value, but is regarded as a random variable that obeys a specific prior probability distribution (such as Gaussian distribution): ,in, represents the mean of the current weight prior distribution, Represents the variance of the current weight prior distribution.

[0140] When an expert provides feedback (e.g., “I think the effect of indicator A on indicator B should be stronger”), the feedback is formalized as a likelihood function ,in, Representative expert feedback evidence.

[0141] According to Bayes' theorem, we can integrate prior knowledge and new evidence to obtain the posterior probability distribution of weights: ; in, represents the posterior probability distribution of the weights, Indicates a direct proportion, represents the prior probability distribution of weights.

[0142] Updated weight distribution mean It will be used as the new weight value in the dynamic weighted fuzzy cognitive map model. The advantages of this method are: 1) Quantified uncertainty: The dynamic weighted fuzzy cognitive graph model has uncertainty measurement for its own knowledge.

[0143] 2) Incremental Updates: Through Bayesian reasoning, expert feedback is treated as new evidence and combined with the model's existing prior knowledge, allowing model weights to be adjusted incrementally and quantifiably in the form of a posteriori probabilities rather than abruptly.

[0144] 3) Traceability: Each update has a clear record of the Bayesian reasoning process.

[0145] The updating process of the dynamic weighted fuzzy cognitive graph model can be expressed as: ; in, is the current dynamic weighted fuzzy cognitive graph model (defined by the weight distribution), It's an expert in Evidence provided at all times.

[0146] Compared with the existing technology, the technical solution of this embodiment has the following advantages: 1. Utilize adaptive baseline correction (ABC), adaptive harmonic wavelet packet transform, clinical feature extraction, blood marker standardization, and radiomics feature extraction to achieve unified quantitative representation of raw ECG signals, blood markers, and medical images, build a comprehensive and structured clinical indicator input, and provide a reliable foundation for subsequent risk assessment.

[0147] 2. An innovative fuzzy cognitive map with a dynamic weight adjustment mechanism is designed. Through concept nodes and causal weights predefined by medical experts, it enables iterative reasoning of complex nonlinear causal relationships between multimodal indicators, outputs accurate and interpretable myocardial infarction risk assessment results, and has knowledge-driven characteristics that are independent of training data.

[0148] 3. A multi-level interpretability mechanism combining Structural Perturbation Counterfactual Contribution Analysis (SP-CCA) and Decision Path Tracing Activation Visualization (DPT-CAV). This mechanism applies counterfactual perturbations to the input features of the risk assessment model, quantifies the marginal contribution of each indicator, and combines graph-based optimal path reverse search with node activation visualization to reveal "which indicator is important" and "why it is important," enabling in-depth and transparent explanations of risk assessment results.

[0149] 4. Integrate contribution analysis with path activation information, perform weighted ranking of key indicators based on medical semantics, generate a comprehensive explanatory report that is clear and clinically logical, and enhance the system model's ability to support medical staff in auxiliary decision-making.

[0150] 5. Innovatively introduce an expert knowledge calibration loop mechanism based on Bayesian fusion, eliminating the need to retrain the model. This ensures that the system model can continuously optimize risk assessment performance and interpretation accuracy as medical knowledge is updated, thereby enhancing the long-term adaptability and credibility of the system model.

[0151] Reference Figure 2 The embodiment of the present application further provides a myocardial infarction risk assessment system based on multimodal interpretability, the system comprising a data extraction unit 201, a feature splicing unit 202, a feature aggregation unit 203, a model construction unit 204, a risk assessment unit 205, a contribution calculation unit 206, a path finding unit 207, and an assessment report construction unit 208, wherein: The data extraction unit 201 is used to extract multimodal physiological index features including electrocardiogram clinical features, blood marker features and imaging radiomics features; A feature splicing unit 202 is used to splice each feature in the multimodal physiological indicator feature to obtain a spliced ​​feature vector; A feature aggregation unit 203 is configured to calculate an attention score of each feature in the multimodal physiological indicator feature based on the spliced ​​feature vector, and perform a weighted summation of each feature in the multimodal physiological indicator feature based on the attention score to obtain an aggregated feature vector; A model construction unit 204 is configured to construct a dynamic weighted fuzzy cognitive graph model for myocardial infarction risk assessment. The dynamic weighted fuzzy cognitive graph model includes a weight adjustment mechanism, multiple concept nodes, weighted edges, and an activation function. The weight adjustment mechanism is configured to fine-tune a basic weight based on a context vector to obtain a fine-tuned weight. A weighted edge is an edge connecting two concept nodes and is configured to represent the strength of influence of one concept node on another concept node. The risk assessment unit 205 is configured to input the aggregated feature vector into the dynamic weighted fuzzy cognitive graph model, update the activation value of each concept node using the fine-tuned weight and activation function, and obtain the final activation value of each concept node after a preset condition is met. The final activation value of the myocardial infarction risk concept node is used as the myocardial infarction risk assessment result. A contribution calculation unit 206 is configured to calculate the contribution of each physiological indicator in the aggregated feature vector according to the myocardial infarction risk assessment result; A path finding unit 207 is configured to find a maximum-influence propagation path between each source node and a target node based on a preset concept node sensitivity coefficient, a path context attenuation factor, multiple concept nodes, and weighted edges, and obtain at least one maximum-influence propagation path; The assessment report construction unit 208 is used to construct an interpretable myocardial infarction risk assessment report based on the contribution of each physiological indicator and at least one maximum impact propagation path.

[0152] It should be noted that since the myocardial infarction risk assessment system based on multimodal interpretability in this embodiment and the above-mentioned myocardial infarction risk assessment method based on multimodal interpretability are based on the same inventive concept, the corresponding contents in the method embodiment are also applicable to the system embodiment and will not be described in detail here.

[0153] Reference Figure 3 , an embodiment of the present application further provides an electronic device, the electronic device comprising: at least one memory; at least one processor; at least one program; The program is stored in the memory, and the processor executes at least one program to implement the above-mentioned myocardial infarction risk assessment method based on multimodal interpretability implemented in the present disclosure.

[0154] The electronic device may be any intelligent terminal including a mobile phone, a tablet computer, a personal digital assistant (PDA), a car computer, etc.

[0155] The electronic device according to the embodiment of the present application is described in detail below.

[0156] The processor 1600 may be implemented as a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is configured to execute relevant programs to implement the technical solutions provided by the embodiments of the present disclosure. Memory 1700 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). Memory 1700 can store an operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in memory 1700 and is called by processor 1600 to execute the myocardial infarction risk assessment method based on multimodal interpretability in the embodiments of this disclosure.

[0157] Input / output interface 1800, used for information input and output; Communication interface 1900, used to implement communication interaction between this device and other devices, which can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WiFi, Bluetooth, etc.); Bus 2000 , which transmits information between various components of the device (e.g., processor 1600 , memory 1700 , input / output interface 1800 , and communication interface 1900 ); The processor 1600 , the memory 1700 , the input / output interface 1800 , and the communication interface 1900 are connected to each other in communication within the device via the bus 2000 .

[0158] An embodiment of the present disclosure also provides a storage medium, which is a computer-readable storage medium. The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the above-mentioned myocardial infarction risk assessment method based on multimodal interpretability.

[0159] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0160] The embodiments described in the embodiments of the present disclosure are intended to more clearly illustrate the technical solutions of the embodiments of the present disclosure and do not constitute a limitation on the technical solutions provided by the embodiments of the present disclosure. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of the present disclosure are also applicable to similar technical problems.

[0161] Those skilled in the art will understand that the technical solutions shown in the drawings do not constitute a limitation on the embodiments of the present disclosure, and may include more or fewer steps than shown in the drawings, or a combination of certain steps, or different steps.

[0162] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, i.e., they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

[0163] Those skilled in the art will appreciate that all or some of the steps in the methods, systems, and functional modules / units in the devices disclosed above may be implemented as software, firmware, hardware, or appropriate combinations thereof.

[0164] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0165] It should be understood that in this application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following items" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0166] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0167] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0168] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0169] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including multiple instructions for enabling an electronic device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the various embodiments of the present application. The aforementioned storage medium includes various media that can store programs, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk. The embodiments of the present application are described in detail above in conjunction with the accompanying drawings, but the present application is not limited to the above embodiments. Various changes can be made within the scope of knowledge possessed by ordinary technicians in the relevant technical field without departing from the purpose of the present application.

[0170] The embodiments of the present application are described in detail above in conjunction with the accompanying drawings, but the present application is not limited to the above embodiments. Various changes can be made within the scope of knowledge possessed by ordinary technicians in the relevant technical field without departing from the purpose of the present application.

Claims

1. A method for assessing myocardial infarction risk based on multimodal interpretability, characterized in that: The method comprises: Extract multimodal physiological index features including electrocardiogram clinical features, blood marker features, and imaging radiomics features; Splicing each feature in the multimodal physiological indicator feature to obtain a spliced ​​feature vector; Calculating the attention score of each feature in the multimodal physiological indicator feature according to the spliced ​​feature vector, and performing weighted summation on each feature in the multimodal physiological indicator feature based on the attention score to obtain an aggregated feature vector; A dynamic weighted fuzzy cognitive graph model for myocardial infarction risk assessment was constructed. The dynamic weighted fuzzy cognitive graph model included a weight adjustment mechanism, multiple concept nodes, weighted edges, and an activation function. The weight adjustment mechanism was used to fine-tune the basic weights based on the context vector to obtain fine-tuned weights. The weighted edge was an edge connecting two concept nodes and was used to represent the influence of one concept node on another concept node. Inputting the aggregated feature vector into the dynamic weighted fuzzy cognitive graph model, updating the activation value of each of the concept nodes using the fine-tuned weights and activation function until a final activation value of each of the concept nodes is obtained after a preset condition is met, and using the final activation value of the myocardial infarction risk concept node as the myocardial infarction risk assessment result; Calculating the contribution of each physiological indicator in the aggregated feature vector according to the myocardial infarction risk assessment result; Finding the maximum impact propagation path between each source node and the target node based on the preset concept node sensitivity coefficient, the path context attenuation factor, the multiple concept nodes, and the weighted edges, and obtaining at least one maximum impact propagation path; Based on the contribution of each physiological indicator and the at least one maximum impact transmission path, an interpretable myocardial infarction risk assessment report is constructed.

2. The myocardial infarction risk assessment method based on multimodal interpretability according to claim 1, characterized in that: The extraction of multimodal physiological indicator features including electrocardiogram clinical features, blood marker features and imaging radiomics features includes: Performing baseline correction on an electrocardiogram signal in the acquired electrocardiogram to obtain a corrected electrocardiogram signal; performing harmonic wavelet packet decomposition on the corrected ECG signal to obtain a wavelet packet decomposition tree; Calculating the time-frequency entropy of each node in the wavelet packet decomposition tree according to the node coefficients in the wavelet packet decomposition tree, and obtaining the optimal wavelet basis by minimizing the time-frequency entropy; extracting electrocardiogram clinical features including ST segment deviation and T wave symmetry from the optimal wavelet basis; The extracted blood markers are standardized to obtain blood marker characteristics; Extract radiomic features including texture features from medical images to obtain image radiomic features; The electrocardiogram clinical features, the blood marker features and the imaging radiomics features are constructed into multimodal physiological indicator features.

3. The myocardial infarction risk assessment method based on multimodal interpretability according to claim 1, characterized in that: The updating of the activation value of each of the concept nodes using the fine-tuned weights and activation functions includes: ； in, Represents a concept node In the The new activation value after iterations, Indicates in In the iteration, the concept node All other concept nodes that have an impact The activation value of Represents a concept node Pointing to concept node After fine-tuning the weights, Represents a concept node The initial activation value of represents the activation function, Indicates the total number of concept nodes.

4. The myocardial infarction risk assessment method based on multimodal interpretability according to claim 1, characterized in that: Calculating the contribution of each physiological indicator in the aggregated feature vector according to the myocardial infarction risk assessment result includes: Replacing each physiological indicator in the aggregated feature vector with a clinical baseline value to construct a counterfactual input vector corresponding to each physiological indicator; Inputting the counterfactual input vector into the dynamic weighted fuzzy cognitive graph model to obtain a counterfactual risk value corresponding to each physiological indicator; The counterfactual risk value is differentially calculated from the myocardial infarction risk assessment result to obtain the contribution of each physiological indicator in the aggregated feature vector.

5. The myocardial infarction risk assessment method based on multimodal interpretability according to claim 1, characterized in that: The step of finding the maximum impact propagation path between each source node and the target node based on the preset concept node sensitivity coefficient, the path context attenuation factor, the multiple concept nodes, and the weighted edges to obtain at least one maximum impact propagation path includes: Build at least one path from each source node to the target node; If each of the paths includes multiple concept nodes and weighted edges, the influence flow between two adjacent concept nodes is calculated based on the preset concept node sensitivity coefficient, the final activation values ​​corresponding to the sub-source nodes of the two adjacent concept nodes, and the fine-tuned weights of the weighted edges between the two adjacent concept nodes; performing weighted summation of the influence flows between all two adjacent concept nodes in each path according to the path context attenuation factor to obtain the total influence of each path; According to the total influence, a maximum influence propagation path between each source node and the target node is found to obtain at least one maximum influence propagation path.

6. The myocardial infarction risk assessment method based on multimodal interpretability according to claim 1, characterized in that: The step of constructing an interpretable myocardial infarction risk assessment report based on the contribution of each physiological indicator and the at least one maximum impact transmission path includes: By sorting the contribution of each physiological indicator and screening out multiple key physiological indicators, a quantitative attribution summary including the indicator name corresponding to each key physiological indicator, the specific value derived from the aggregated feature vector, and the contribution is obtained; Matching the maximum impact propagation path corresponding to each of the key physiological indicators, and obtaining a core logic path including the name and final activation value of each concept node in the maximum impact propagation path; Visually encode the final activation value of each concept node and the maximum impact propagation path corresponding to each key physiological indicator to obtain a visual decision map; The quantitative attribution summary, the core logic path, and the visual decision map are integrated to construct an interpretable myocardial infarction risk assessment report.

7. The myocardial infarction risk assessment method based on multimodal interpretability according to claim 1, characterized in that: After constructing the interpretable myocardial infarction risk assessment report, the method further includes: constructing the base weights as a prior probability distribution of weights; According to the prior probability distribution of the weight, the posterior probability distribution of the weight is obtained using Bayes' theorem; The weight distribution mean obtained from the posterior probability distribution of the weight is used as a new weight in the dynamic weighted fuzzy cognitive map model to update the dynamic weighted fuzzy cognitive map model.

8. A myocardial infarction risk assessment system based on multimodal interpretability, characterized in that: The system comprises: A data extraction unit is used to extract multimodal physiological index features including electrocardiogram clinical features, blood marker features, and imaging radiomics features; A feature splicing unit, configured to splice each feature in the multimodal physiological indicator feature to obtain a spliced ​​feature vector; a feature aggregation unit, configured to calculate an attention score of each feature in the multimodal physiological indicator feature according to the spliced ​​feature vector, and perform a weighted summation of each feature in the multimodal physiological indicator feature based on the attention score to obtain an aggregated feature vector; A model construction unit, configured to construct a dynamic weighted fuzzy cognitive graph model for myocardial infarction risk assessment, wherein the dynamic weighted fuzzy cognitive graph model comprises a weight adjustment mechanism, a plurality of concept nodes, weighted edges, and an activation function. The weight adjustment mechanism is configured to fine-tune a basic weight based on a context vector to obtain a fine-tuned weight. The weighted edge is an edge connecting two concept nodes and is configured to characterize the strength of influence of one concept node on another concept node. a risk assessment unit, configured to input the aggregated feature vector into the dynamic weighted fuzzy cognitive graph model, update the activation value of each of the concept nodes using the fine-tuned weights and activation function, obtain a final activation value of each of the concept nodes after a preset condition is met, and use the final activation value of the myocardial infarction risk concept node as a myocardial infarction risk assessment result; a contribution calculation unit, configured to calculate the contribution of each physiological indicator in the aggregated feature vector according to the myocardial infarction risk assessment result; A path finding unit, configured to find a maximum-influence propagation path between each source node and a target node based on a preset concept node sensitivity coefficient, a path context attenuation factor, the plurality of concept nodes, and the weighted edges, and obtain at least one maximum-influence propagation path; An assessment report construction unit is used to construct an interpretable myocardial infarction risk assessment report based on the contribution of each physiological indicator and the at least one maximum impact propagation path.

9. An electronic device, characterized in that: It includes at least one control processor and a memory for communicating with the at least one control processor; the memory stores instructions that can be executed by the at least one control processor, and the instructions are executed by the at least one control processor to enable the at least one control processor to execute the myocardial infarction risk assessment method based on multimodal interpretability as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the myocardial infarction risk assessment method based on multimodal interpretability as described in any one of claims 1 to 7.

Images