Medical decision-making methods, devices, and storage media based on neural symbolic mixture models

Abstract

This application relates to a medical decision-making method, device, and storage medium based on a neural symbolic fusion model, belonging to the field of medical information processing technology. First, this application fuses multimodal clinical data of the target object through an attention weighting mechanism to obtain a multimodal fusion representation. Then, based on a high-risk indicator medical mask enhancement mechanism, the high-risk indicators of the structure-aware cross-modal attention network constructed based on the multimodal fusion representation are associated with medical indicator masks to obtain a joint feature vector of the target object. The joint feature vector is input into a neural network decision layer to obtain a neural network recommendation vector, and then input into a medical knowledge graph to obtain a symbolic rule recommendation vector. Finally, the neural network recommendation vector, the symbolic rule recommendation vector, and the joint feature vector are input into a three-layer neural symbolic fusion network to obtain a target decision suggestion. This improves the accuracy, interpretability, and clinical adaptability of intelligent medical decision-making.

Description

Technical Field

[0001] This application relates to the field of medical information processing technology, specifically to a medical decision-making method, device, and storage medium based on a neural symbolic hybrid model. Background Technology

[0002] With the continuous increase in the number of patients in intensive care and emergency departments, clinical medical staff are facing increasing pressure in diagnosis and intervention when dealing with high concurrency, complex conditions, and multi-source data. Traditional artificial intelligence (AI) model-assisted decision-making systems are mostly based on data-driven deep learning models. The "black box" structure of deep learning models makes it difficult to provide traceable recommendations to clinicians, resulting in low trust and adoption rates among medical staff. At the same time, patients in acute and critical conditions often experience a surge of data from multiple modalities (structured, text, images, etc.). Traditional AI models have bottlenecks in data fusion and semantic consistency modeling, affecting the overall recommendation effectiveness. Summary of the Invention

[0003] The purpose of this application is to provide a medical decision-making method, device, and storage medium based on a neural symbolic hybrid model, in order to solve the problem of low accuracy in traditional AI model-assisted medical decision-making.

[0004] To achieve the above objectives, the first aspect of this application provides a medical decision-making method based on a neural symbolic mixture model, comprising:

[0005] The multimodal clinical data of the target object is acquired, and the multimodal clinical data is fused through an attention weighting mechanism to obtain a multimodal fusion representation. The multimodal clinical data includes structured time-series data, free text data, and medical image data.

[0006] Based on the multimodal fusion representation, a structure-aware cross-modal attention network is constructed. Based on the medical mask enhancement mechanism of high-risk indicators, the high-risk indicators of the structure-aware cross-modal attention network are associated with medical indicator masks to obtain the joint feature vector of the target object.

[0007] The joint feature vector is input into the decision layer of the neural network to obtain the neural network recommendation vector, and the joint feature vector is input into the medical knowledge graph to obtain the symbol rule recommendation vector;

[0008] The neural network recommendation vector, the symbol rule recommendation vector, and the joint feature vector are input into a three-layer neural symbol fusion network to obtain target decision suggestions.

[0009] A second aspect of this application provides a medical decision-making device based on a neural symbolic hybrid model, comprising:

[0010] The fusion module is used to acquire multimodal clinical data of the target object and fuse the multimodal clinical data through an attention weighting mechanism to obtain a multimodal fusion representation. The multimodal clinical data includes structured time-series data, free text data, and medical image data.

[0011] The association module is used to construct a structure-aware cross-modal attention network based on the multimodal fusion representation, and based on the medical mask enhancement mechanism of high-risk indicators, associate the high-risk indicators of the structure-aware cross-modal attention network with medical indicator masks to obtain the joint feature vector of the target object.

[0012] The recommendation module is used to input the joint feature vector into the neural network decision layer to obtain the neural network recommendation vector, and to input the joint feature vector into the medical knowledge graph to obtain the symbol rule recommendation vector;

[0013] The generation module is used to input the neural network recommendation vector, the symbol rule recommendation vector, and the joint feature vector into a three-layer neural symbol fusion network to obtain target decision suggestions.

[0014] A third aspect of this application provides a computer-readable storage medium storing a program that can be loaded by a processor and executed as described above regarding the medical decision-making method based on a neural symbolic hybrid model.

[0015] The beneficial effects of this application are:

[0016] This application fuses multimodal clinical data of the target object through an attention-weighted mechanism to obtain a multimodal fusion representation. By dynamically allocating weights to structured time-series data, free text data, and medical image data, it can adaptively capture the semantic contributions of different modalities in different scenarios. Then, based on a high-risk indicator-based medical mask enhancement mechanism, the high-risk indicators are associated with medical indicator masks to obtain a joint feature vector of the target object, which can capture key information and improve the ability to identify high-risk indicators.

[0017] Next, a dual-path recommendation model, employing both a neural network decision layer and a medical knowledge graph, was used to obtain neural network recommendation vectors and symbolic rule recommendation vectors. The neural network recommendation vectors utilize data patterns to generate recommendations tailored to the individual differences of the target audience, improving the accuracy of decision-making recommendations. The symbolic rule recommendation vectors output recommendations consistent with medical knowledge, providing interpretability for the decision-making recommendations.

[0018] Finally, a three-layer neural symbol fusion network is used to verify and integrate the dual-path recommendation results from multiple dimensions. This not only amplifies the complementary advantages of the neural network recommendation vector and the symbol rule recommendation vector, but also identifies the recommendation conflicts between the two, making the output of the final target decision suggestion more reliable.

[0019] In summary, this application improves the accuracy, interpretability, and clinical adaptability of intelligent medical decision-making through dynamic fusion of multimodal clinical data, feature enhancement guided by medical priors, dual-pathway recommendation (neural and symbolic), and three-layer neural-symbolic fusion. It can also be applied to individualized intervention recommendations in complex critical care scenarios.

[0020] Other features and advantages of this application will be described in detail in the following detailed description section. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating an application scenario of a medical decision-making method based on a neural symbolic hybrid model provided in this application embodiment;

[0022] Figure 2 This is a flowchart illustrating a medical decision-making method based on a neural symbolic hybrid model provided in an embodiment of this application.

[0023] Figure 3 This is a schematic diagram of the structure of a medical decision-making device based on a neural symbolic hybrid model provided in an embodiment of this application. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified. In this application, the term "exemplary" is used to mean "used as an example, illustration, or description." Any embodiment described as "exemplary" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to implement and use this application. In the following description, details are set forth for illustrative purposes. It should be understood that those skilled in the art will recognize that this application can be implemented without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid unnecessary detail that would obscure the description of this application. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0026] like Figure 1 As shown, Figure 1 This is a schematic diagram illustrating an application scenario of a medical decision-making method based on a neural symbolic mixture model provided in this application embodiment. The application scenario includes an electronic device 110 for the medical decision-making method based on a neural symbolic mixture model. The electronic device 110 integrates a computer-readable storage medium capable of running the medical decision-making method based on the neural symbolic mixture model, in order to execute the steps of the medical decision-making method based on the neural symbolic mixture model.

[0027] Understandable Figure 1 The electronic devices in the application scenarios of the medical decision-making method based on the neural symbolic hybrid model, or the devices contained in the electronic devices, do not constitute a limitation on the embodiments of this application. That is, the number or type of devices in the application scenarios of the medical decision-making method based on the neural symbolic hybrid model, or the number or type of devices contained in each device, do not affect the overall implementation of the technical solution in the embodiments of this application, and can all be considered as equivalent substitutions or derivatives of the technical solutions claimed in the embodiments of this application.

[0028] In this application embodiment, the electronic device 110 can be an independent device, or a device network or device cluster composed of devices. For example, the electronic device 110 described in this application embodiment includes, but is not limited to, a computer, a network host, a single network device, a set of multiple network devices, or a cloud device composed of multiple devices. Among them, the cloud device is composed of a large number of computers or network devices based on cloud computing.

[0029] Those skilled in the art will understand that Figure 1 The application scenarios shown are merely one application scenario corresponding to the technical solution of this application, and do not constitute a limitation on the application scenarios of the technical solution of this application. Other application scenarios may include more than one application scenario. Figure 1 The number of more or fewer electronic devices shown, or the network connections of electronic devices, for example Figure 1 Only one electronic device is shown in the diagram. It is understood that the scenario of the medical decision-making method based on the neural symbolic mixture model may also include one or more other electronic devices, which are not specifically limited here. The electronic device 110 may also include a memory and a processor. The memory is used to store information related to the medical decision-making method based on the neural symbolic mixture model.

[0030] Furthermore, in the application scenario of the medical decision-making method based on the neural symbolic mixture model in this application embodiment, the electronic device 110 can be equipped with a display device, or the electronic device 110 can be without a display device but connected to an external display device 120. The display device 120 is used to output the results of the medical decision-making method based on the neural symbolic mixture model executed by the electronic device. The electronic device 110 can access the background database 130. The background database 130 can be the local storage of the electronic device 110 or a cloud database located in the cloud. The background database 130 stores information related to the medical decision-making method based on the neural symbolic mixture model.

[0031] It should be noted that, Figure 1 The application scenario of the medical decision-making method based on the neural symbolic mixture model shown is merely an example. The application scenario of the medical decision-making method based on the neural symbolic mixture model described in the embodiments of this application is to more clearly illustrate the technical solution of the embodiments of this application, and does not constitute a limitation on the technical solution provided in the embodiments of this application.

[0032] Based on the application scenarios of the aforementioned medical decision-making method based on neural symbolic mixture models, an embodiment of the medical decision-making method based on neural symbolic mixture models is proposed. A detailed description is provided below with reference to the accompanying drawings.

[0033] Figure 2 This application provides a medical decision-making method based on a neural symbolic mixture model. For example... Figure 2 As shown, this medical decision-making method may include steps 201-204, which will be described in detail below.

[0034] Step 201: Obtain multimodal clinical data of the target object, and fuse the multimodal clinical data through an attention weighting mechanism to obtain a multimodal fusion representation.

[0035] In this embodiment, the target object refers to a patient requiring assisted treatment decision-making, such as a critically ill patient. Multimodal clinical data refers to various types of clinical data of the target object, which may include structured time-series data, free text data, and medical image data. Structured time-series data refers to quantitative data that changes over time, including vital signs (such as blood pressure, heart rate, blood oxygen, etc.) and laboratory indicators (such as lactate, white blood cell count, etc.). Free text data refers to unstructured text descriptions, including admission medical records, progress notes, and consultation opinions. Medical image data refers to image data within the target object's medical data, such as chest X-rays, computed tomography (CT) scans, and ultrasound imaging data. Structured time-series data can quantify the dynamic indicators of the target object. Free text data can be used to characterize the semantic description of the target object. Medical image data can express the visual characteristics of the target object. By acquiring a collection of health information of the target object from different sources and in different formats, a comprehensive reflection of the target object's medical data can be achieved.

[0036] Attention-weighted fusion is a dynamic weighting method that assigns higher weights to modalities that are more relevant to the target patient's medical data by calculating the similarity between features of each modality. For example, the "lactate level" modality in patients with septic shock may have a higher weight than the "body temperature" modality. The unified feature vector obtained through attention-weighting can integrate key information from different modalities, eliminating the differences in expression between heterogeneous data.

[0037] As an example, feature extraction can be performed separately for the three modalities of data. For instance, structured time-series data can have dynamic trend features extracted using Temporal Convolutional Networks (TCNs), Long Short-Term Memory (LSTMs), or Gated Recurrent Units (GRUs). Free text data can have semantic features extracted using a pre-trained medical BERT (Bidirectional Encoder Representations from Transformers) model. Medical image data can have visual features extracted using Convolutional Neural Networks (CNNs). Then, a cross-modal attention mechanism is constructed, calculating the weight coefficients of different modalities based on the semantic association strength between modalities. A weighted summation yields a multimodal fusion representation. This multimodal fusion representation is a unified feature vector of the target object obtained through attention weighting. This approach addresses the problem of fixed-weight fusion failing to adapt to dynamic changes in patient conditions, improves the specificity of feature representation, and reduces the feature bias caused by single-modal information.

[0038] Step 202: Construct a structure-aware cross-modal attention network based on multimodal fusion representation, and based on the medical mask enhancement mechanism of high-risk indicators, associate the high-risk indicators of the structure-aware cross-modal attention network with medical indicator masks to obtain the joint feature vector of the target object.

[0039] The Structural Cross-Modality Attention Network (S-CMAN) is a network that simultaneously captures pathological associations both intramodal (i.e., between features of the same type of data) and intermodal (i.e., between features of different types of data), enabling it to understand the intrinsic logic of indicator changes, symptom descriptions, and imaging manifestations. By constructing a two-layer attention mechanism based on multimodal fusion representation, it can capture both intramodal feature associations and intermodal cross-associations.

[0040] Medical masking enhancement mechanisms based on high-risk indicators utilize prior knowledge derived from clinical consensus. They force the model to focus on key indicators strongly correlated with the patient's condition through a mask matrix. Associating high-risk indicators with medical indicator masks, for example, by weighting the feature locations associated with high-risk indicators, can reduce decision bias caused by data noise or secondary features.

[0041] The joint feature vector is the final feature vector that integrates multimodal structural relationships with enhanced weights for high-risk indicators. It can include both data-driven association patterns and focus on key medical priors. This provides more accurate feature data for subsequent recommendation decisions.

[0042] Step 203: Input the joint feature vector into the neural network decision layer to obtain the neural network recommendation vector, and input the joint feature vector into the medical knowledge graph to obtain the symbol rule recommendation vector.

[0043] The neural network decision layer is a recommendation model based on deep neural networks. By inputting a joint feature vector into the neural network decision layer and learning patterns from historical medical data, it can output intervention recommendation probabilities tailored to the individual differences of the target population, thus obtaining the neural network recommendation vector. The neural network recommendation vector is the output of the neural network decision layer, reflecting the correlation between features in the data and interventions. For example, it can determine the proportion of effective interventions using vasopressors in a certain combination of lactic acid and blood pressure.

[0044] A medical knowledge graph is a structured knowledge base that stores medical entities (such as diseases, symptoms, and interventions) and relationships (such as causality, constraints, and guidelines). Inputting a joint feature vector into the medical knowledge graph allows for the output of rule recommendation vectors based on the association between indicator thresholds and intervention suggestions. Symbolic rule recommendation vectors, on the other hand, are outputs based on the medical knowledge graph, reflecting the degree to which medical rules recommend interventions. For example, they can recommend drug priorities based on medical knowledge.

[0045] The dual-path recommendation strategy overcomes the limitations of a single recommendation model. The neural network recommendation path corresponding to the neural network decision layer utilizes data patterns for personalized recommendations. The symbolic rule recommendation path based on the medical knowledge graph determines the medical compliance of the recommendation decision based on medical knowledge. This solves the black-box problem of uninterpretable recommendations solely through neural networks, as the symbolic rule recommendation vectors trace the specific recommendation path, providing an interpretable foundation for subsequent integration. Furthermore, it addresses the rigidity and lack of adaptability inherent in recommendations solely based on knowledge graphs, as the neural network recommendation path can adapt to new scenarios not covered by existing rules, improving the flexibility of the recommendation process.

[0046] Step 204: Input the neural network recommendation vector, symbol rule recommendation vector, and joint feature vector into a three-layer neural symbol fusion network to obtain target decision suggestions.

[0047] A three-layer neural symbolic fusion network refers to a network structure that integrates dual-path recommendations in stages, including confidence alignment, logical association, and consistency verification. For example, the first layer of the neural symbolic fusion network can be a confidence alignment layer, which can scale the neural network recommendation vector and the symbolic rule recommendation vector, unifying their confidence distribution to eliminate scale differences between the outputs of different paths. The second layer can be a logical association layer, dynamically adjusting the weights by calculating the association strength between the neural network recommendation vector and the symbolic rule recommendation vector. The third layer can be a consistency verification layer, outputting the final target decision suggestion by calculating the consistency between the neural network recommendation vector and the symbolic rule recommendation vector. This solves the problem that traditional single-weight fusion cannot handle the differences between dual paths, allowing the fused result to retain the advantages of both paths. The output includes decision suggestions with reasoning basis, which, while maintaining accuracy, also has clinical interpretability, enhancing doctors' trust in intelligent recommendations.

[0048] The embodiments of this application improve the accuracy, interpretability, and clinical adaptability of intelligent medical decision-making through dynamic fusion of multimodal clinical data, feature enhancement guided by medical priors, dual-path recommendation of neural and symbolic approaches, and three-layer neural-symbolic fusion. It can also be applied to individualized intervention recommendations in complex critical care scenarios.

[0049] In step 201, the multimodal clinical data can first be encoded into multimodal feature vectors. Specifically, feature extraction can be performed on the three types of multimodal clinical data to obtain single-modal feature vectors. For example, a bidirectional GRU can be used to extract single-modal feature vectors from structured time-series data; a medical-specific vector model (such as Med-BERT or Clinical-Longformer) can be used to extract single-modal feature vectors from free text data; and a transfer learning model (such as the EfficientNet-B3+SE module) can be used to extract single-modal feature vectors from medical image data. The sampling period for structured time-series data is... , forming a matrix A bidirectional GRU-based sequence encoder is used to extract the state representation: Free text data is embedded into context using a medical-specific word vector model (such as Med-BERT or Clinical-Longformer) to extract the [CLS] bit vector as a global representation. Medical image data is input into a transfer learning model (such as the EfficientNet-B3+SE module) to extract image feature vectors with consistent dimensions. .

[0050] Then, attention weights are calculated on the multimodal feature vectors using a learnable gating function. Based on the multimodal feature vectors and attention weights, a multimodal fusion representation is then calculated. The learnable gating function is a dynamically generated weight function with learnable parameters, which adaptively adjusts the weights of each modality according to the input multimodal feature vectors. For example, the attention weights for the three modalities can be defined as follows: Satisfying the constraints: The weight vector is generated by the following learnable gating function: The final multimodal fusion is represented as: Compared to traditional splicing or averaging, this weighted fusion structure has stronger modal adaptability and individualized expression capabilities, providing more refined representation vectors for subsequent model processing.

[0051] The embodiments of this application can construct a multidimensional vector representation of the state of the target object by integrating multi-source heterogeneous multimodal clinical data, and achieve unified embedding.

[0052] In step 202, the multimodal fusion representation can first be projected into multiple subspace representations using a learnable parameter matrix. Each subspace representation projects the multimodal fusion representation onto the query vector, key vector, and value vector of a low-dimensional subcontrol. The subspace representation can include query vectors, key vectors, and value vectors, used to calculate the correlation strength between features. Next, an initial attention matrix is ​​constructed based on the query vectors and key vectors. This initial attention matrix includes multiple feature channels or indicator dimensions. The initial attention matrix reflects the original correlations between feature channels or indicator dimensions, capturing the implicit correlation patterns in the data.

[0053] For example, suppose the input is a fused representation vector: ,in, For the feature dimensions after multimodal fusion, assume =128. Multi-head attention projection includes... These are mapped to query (Q), key (K), and value (V) vectors, respectively: ,in, These are learnable parameters. If... =128, then can be set Among them This subspace dimension is usually a hyperparameter set manually. These vectors will enter the attention calculation module, which determines the degree of weighted aggregation between features of different modalities.

[0054] Next, high-risk indicators are masked and enhanced across multiple feature channels or indicator dimensions to obtain the target attention matrix. Some medical indicators (such as lactate, SBP, and SpO2) have significant weight in critical illness assessment and need to be prioritized by the attention mechanism. This mechanism reinforces this preference by introducing a structural mask matrix. To improve the model's ability to perceive key medical indicators, a structural mask matrix can be introduced. If dimension If the corresponding indicator is a high-risk item (such as lactic acid, SBP<90, low PaO2 / FiO2 ratio), then Otherwise, it is 0, where This is a time step index, indicating the monitoring time.

[0055] Introduce a masking enhancement term when calculating attention weights:

[0056] ;

[0057] in, To control the coefficient of mask influence, Ensure that normalization pays attention to weights. , where is the attention matrix.

[0058] For example, a certain multimodal fusion representation The 12th dimension represents lactate level (5.6 mmol / L, elevated), and the 35th dimension represents SpO2 (86%, decreased). When constructing the mask matrix: During attention calculation, the interaction weights of these two metrics will be amplified, thereby significantly improving the model's focus on "lactic acid + hypoxia".

[0059] Finally, the value vector is weighted by the target attention matrix, then residual-connected and layer-normalized with the multimodal fusion representation to obtain the joint feature vector. After obtaining the attention-weighted representation, it is input into the feedforward neural network and integrated into the final representation through residual and normalization processes.

[0060] ;

[0061] in, It is a two-layer feedforward network. To avoid gradient explosion or vanishing, and improve stability, It is the final patient representation vector.

[0062] Table 1 shows an example of a fusion feature.

[0063] Table 1

[0064] ,

[0065] For example, suppose a patient has the fusion features shown in Table 1. The S-CMAN model notices abnormalities in blood pressure and lactate, but also notices body temperature. The MIMA mask tells the S-CMAN model, "Don't focus too much on body temperature, but pay close attention to blood pressure and lactate," thus forcibly biasing the attention weights towards the key indicators. Finally, the S-CMAN model integrates the original fusion features with the "focused features" to output a comprehensive representation. This is used for subsequent graph reasoning or recommendation.

[0066] The purpose of this application embodiment is to construct a cross-modal attention model with structural cognitive ability, which is used to further extract key deep semantics from the fused patient features, and to introduce an importance constraint mechanism for medical features to enhance the model's focusing ability and interpretability.

[0067] In step 203, the joint feature vector is input into the decision layer of a neural network composed of a Multilayer Perceptron (MLP) to generate the probability distribution of the intervention plan, thus obtaining the neural network recommendation vector. An MLP is a neural network composed of fully connected layers that uses a non-linear activation function to capture the complex correlation between features and interventions, making it suitable for mining interaction patterns of multi-dimensional features. The MLP network structure is as follows: the hidden layers gradually increase the dimension from... Mapped to ,in The number of intervention options; the activation function for each layer uses To prevent gradient vanishing; the output layer is connected This forms a recommendation probability distribution.

[0068] Therefore, the neural network recommendation vector satisfies the following formula: ;in, This is the weight matrix; This represents the LeakyReLU activation function; h is the hidden layer dimension (e.g., h=64); k is the recommendation label dimension (e.g., 10 intervention methods). This is the bias term. A Dropout regularization layer is introduced into the traditional MLP structure to prevent overfitting, and Top-K priority learning is achieved through clinical intervention using a sparsity regulator.

[0069] In step 203, the joint feature vector is input into the medical knowledge graph, and the cosine similarity between the joint feature vector and the nodes of the medical knowledge graph is calculated to filter entity nodes that are related to the target object. The medical knowledge graph is a weighted multi-type heterogeneous graph. ,in, It is a set of entity nodes, including patient status (such as low perfusion, high lactate), intervention procedures (such as vasopressor, mechanical ventilation), and test indicators. The set of edges represents relationships such as causality, taboo, promotion, and substitution. It is a set of relation types, such as "Cause", "Trigger", "Contraindicate", etc. Encode a static semantic embedding vector (annotated by a knowledge base or clinical experts) for each entity node. The clinical weights of the edges are represented by rule confidence / guideline recommendation level. This graph can be constructed offline from guidelines, consensus literature, and real prescription data, and is continuously updatable. For example, calculating the joint feature vector. With each entity node in the graph Cosine similarity: Retain items with similarity exceeding a threshold. Nodes (such as) =0.65), which is considered a medical entity associated with the state of the target object.

[0070] Then, based on the selected entity nodes, a reasoning subgraph is dynamically constructed, including relevant paths to these entity nodes. These relevant paths can include rule paths. Entity nodes are the basic units in a knowledge graph, representing clinical concepts, and each node is quantized using embedding vectors. The reasoning subgraph is a sub-network dynamically trimmed from the knowledge graph that is related to the state of the target object. It can contain core entities, rule paths, and intervention nodes, and is used to focus on key reasoning paths. For example, starting with a selected entity node, the reasoning subgraph is traced back to its upstream and downstream rule nodes, attribute nodes, and intervention nodes. A reasoning subgraph containing only relevant paths is constructed. To avoid redundant traversal of the entire graph, symbolic path constraints are supported (e.g., pathology = severe + indicators = lactate↑ → recommended to go to ICU).

[0071] Next, the rule paths in the inference subgraph are traversed, the matching degree between each rule path and the state of the target object is calculated, and the symbolic rule recommendation vector is obtained through normalization. For example, each rule path in the inference subgraph is traversed. , and the current state Matching. Calculate the match confidence score. For example, based on the range of indicator landing points or the distance of state ambiguity. Normalize all recommendations to form a rule vector output: By employing a dynamic graph pruning mechanism, the complexity of symbolic reasoning is significantly reduced while maintaining rule completeness. Furthermore, by introducing a state-driven fuzzy rule matching and scoring mechanism (SoftRuleMatching), fuzzy reasoning that deviates from the rules is supported.

[0072] Finally, step 203 outputs two recommendation results in parallel: neural network recommendation vector: Symbolic rule recommendation vector: Both will serve as inputs to the three-layer fusion mechanism in step four, further modeling complementary relationships and collaboratively enhancing recommendation accuracy.

[0073] In step 204, after obtaining the neural network recommendation vector... Recommendation vectors based on symbolic rules Building upon this foundation, a three-layer fusion mechanism (e.g., including surface, middle, and deep layers) is introduced to weight the two recommendation vectors, perform graph embedding expansion, and conduct consistency checks to ultimately generate the optimal intervention recommendation, i.e., the target decision recommendation. It also provides conflict analysis capabilities for human-computer interaction review.

[0074] Specifically, the first fusion vector can be obtained by integrating the neural network recommendation vector and the symbolic rule recommendation vector through a dynamically weighted confidence network. The dynamically weighted confidence network is a fusion network that dynamically adjusts the weights based on data features, and the weights adaptively change with the reliability of the two recommendation vectors.

[0075] As an example, parameters can be controlled through policies. By integrating data-driven and knowledge-driven recommendations, the first fusion vector is formed: , The initial value can be set to 0.5 and dynamically adjusted during subsequent reinforcement learning. If an intervention has a high confidence level (e.g., ≥0.3) in both channels, it will be retained first. Weak signals (confidence level <0.1) are suppressed during fusion.

[0076] The joint feature vector is then injected into the nodes of the medical knowledge graph to construct a neural symbolic graph neural network, and a second fusion vector is obtained through graph attention propagation. The neural symbolic graph neural network is a network that integrates graph neural networks and the symbolic logic of knowledge graphs. By injecting the joint feature vector into the paper graph nodes and learning the structured relationships between data features, symbolic entities, and intervention nodes through attention propagation, deep integration of data and knowledge can be achieved. Through the construction of the neural symbolic graph neural network, consistency verification is performed on the graph attention propagation between the target object state and the recommendation node.

[0077] For example, joint feature vectors Inject the knowledge graph node space as input to the graph neural network. Perform a graph attention mechanism (such as GAT) on the node state propagation: .in, Represents a node The neighbor set. After obtaining the graph attention vector of each intervention node, the recommendation score is recalculated to obtain the second fusion vector: Graph structure propagation can identify potential collaborative interventions (such as the co-occurrence pattern of "expansion + boost"), preserve structural causal path information, and enhance interpretability.

[0078] Finally, the first and second fused vectors are input into the consistency discriminator, which outputs the probability that the first and second fused vectors conflict. This is achieved through comparison. and Whether there are significant differences and conflicting recommendations between the two results is screened using a consistency discriminant, and explanations for conflicts are provided. For example, the KL (Kullback-Leibler) divergence or cosine distance between the two fusion results can be calculated. .

[0079] Specifically, when the probability is less than or equal to a set conflict threshold, a target decision suggestion is generated based on the first fusion vector and the second fusion vector. When the probability is greater than the conflict threshold, a conflict triage mechanism is triggered, generating a confirmation prompt. The conflict threshold is a critical value used to define whether the difference in fusion vectors requires manual intervention, and is usually set based on the clinical risk level. For example, it can be set to 0.2 for critical care scenarios and 0.4 for ordinary scenarios. The smaller the conflict threshold, the lower the tolerance for conflict, and the more likely it is to trigger manual review. The confirmation prompt is structured information that requires manual review, and can include the reason for the conflict between the first and second fusion vectors. Therefore, the confirmation prompt can be sent to the display device in a structure of conflict point, reason, and suggestion to reduce the decision-making burden on doctors. In response to receiving a confirmation instruction for the confirmation prompt, a target decision suggestion is generated based on the first and second fusion vectors.

[0080] For example, if Triggering conflict routing mechanism (e.g.) =0.2), trigger the interpretation module to mark "conflicting intervention items" and prompt doctors to pay attention. If there is no conflict or the conflict item is identified, the final recommendation vector is output: Conflicting items can be prompted for secondary confirmation from doctors through the human-computer interaction interface. The system can record the selection of conflicting items, using them as samples for subsequent model calibration and writing them into the strategy update module.

[0081] This application's embodiments achieve a balance between recommendation accuracy and robustness through a three-layer fusion mechanism (probabilistic layer, graph structure layer, and consistency layer). By employing a conflict triage and interpretation mechanism, high-discrepancy items are marked and diagnostic suggestions are output, making it suitable for high-risk clinical scenarios. Furthermore, the graph structure reprojection embedding method allows for enhancement of recommendations at the structural semantic level.

[0082] In this embodiment, a modal-level causal explanation mechanism can be introduced to trace the causal contribution path between each recommendation result and multimodal data, enabling doctors to intuitively understand the causes of the recommendations and improving clinical adoption and system credibility. Therefore, this embodiment also includes a structured recommendation output and an interpretable engine generation mechanism.

[0083] Specifically, the target recommendation item is first determined from the candidate intervention items of the target decision suggestion based on the recommendation probability. Then, based on the final recommendation vector... (in The total number of interventions is used to generate a structured recommendation list, sorted by confidence level. The list may include: intervention name, recommendation confidence level, source of recommendation (e.g., neural pathway / symbolic pathway / fusion), supporting medical indicators (e.g., elevated lactate, decreased SBP, etc.), and recommendation level (high / medium / low, based on grading rules).

[0084] Then, explanatory information corresponding to the target recommendation item is constructed. This explanatory information may include neural feature contribution weights, symbolic path tracing explanations, and modality source tracing information.

[0085] The contribution weights of neural features are used to identify the influence of input features on the recommendation results. Gradient weighting, SHAP, or CAM-like mechanisms can be employed for each dimension of the input. Calculate its influence score And map it back to the corresponding medical variables. ;in This is an index of currently recommended interventions.

[0086] Symbolic path tracing interpretation constructs a visual, traceable, and clinically semantically consistent symbolic logical reasoning chain to explain "where the intervention recommended by the symbolic path comes from" and "why it is recommended," enhancing doctors' trust and understanding of AI decision-making. This interpretation can rely on a medical knowledge graph, proposing a subgraph related to the current state based on the semantic similarity between the joint feature vector and nodes in the medical knowledge graph, and performing layer-by-layer logical reasoning. For each activated recommendation path, a structured explanatory text (Rule-to-Text) is generated: Format: [Rule R number]: If condition X (entity + relation chain) is met, then intervention Y is recommended. Matching score: S_j; Key entities: {e_1,e_2,...}. Taking the recommended intervention as the use of vasopressors as an example, the reasoning path is R45: If SBP < 90 mmHg and LAC > 4.0 mmol / L, then "rapid fluid resuscitation + vasopressors" is recommended. Currently, SBP = 87, and LAC = 4.1 indicates a successful match. Matching score: 0.78. Inference path entity: {SBP, lactate, hypoperfusion state}. Output rationale: Inferring that the patient has a "hypoperfusion state," and the symbolic rules support rapid fluid resuscitation + vasopressor treatment.

[0087] Modal source tracing information is used to explain the independent causal contribution of each modal input (such as tests, physiology, medical history, imaging, etc.) to the intervention recommendations in the final recommendation results, and supports tracing back to the original data source, enhancing the responsibility apportionment and accountability interpretability of the model's decision-making. Let the original multimodal inputs be: : Test index vector; : Vital signs vector; : Medical history coding vector; Image representation vector. The corresponding encoder is... Output modal representation: The final fusion is represented as: .

[0088] Then, Shapley weights or gradient attribution are introduced to estimate modal contributions. For example, Shapley value estimation or the Integrated Gradients method can be used to estimate the final output. Back-estimation of the marginal contribution value of each mode: .in, Representing modes For the causal contribution value of the current decision, a threshold can be set to filter the dominant modality. For the dominant dimension within each modality (such as lactate, CRP, etc. in the test modality), the contribution of specific variables can be further deduced using attention scores or feature contribution values. .

[0089] As shown in Table 2, Table 2 is an example of a result generation explanation structure.

[0090] Table 2

[0091] ,

[0092] The system supports structured outputs such as: Decision Interpretation Summary, which can also be linked to the electronic medical record system, automatically linking to original record segments. For example, the recommendation of the "vasopressor + fluid resuscitation" regimen is mainly based on the following key features:

[0093] Laboratory indicators (contribution weight: 41%): elevated lactate (4.2 mmol / L), elevated CRP (96 mg / L); vital signs (contribution weight: 27%): low systolic blood pressure (SBP=87), rapid heart rate (HR=123); past medical history (contribution weight: 18%): advanced age + diabetes, indicating a high-risk host.

[0094] The neural feature contribution weight outputs a structured recommendation vector, i.e., a clinically interpretable recommendation result, such as "Recommended medication regimen A (probability 0.78) + treatment B (probability 0.65)". Symbolic path tracing explains "why this recommendation was given", i.e., identifying the influence of key feature factors on the recommendation. Modality source tracing traces which modalities (data sources) dominated the decision and the key features within each modality. For example, recommendation result: vasopressors (0.78), fluid resuscitation (0.65); recommendation explanation: because of high lactate (4.2 mmol / L) and low SBP (87 mmHg); modality source tracing: data from: laboratory indicators (contribution 41%, dominant items: lactate, CRP), vital signs (contribution 27%, dominant items: SBP, HR).

[0095] Finally, the target recommendations and explanatory information are encapsulated into a data package according to the set structure to generate a modal source tracing report. This modal source tracing report is a structured report that can be read and displayed in the clinical system by physicians.

[0096] Traditional decision-making systems often struggle to continuously optimize by combining physician feedback with the actual outcomes of target subjects, and are difficult to adapt to the practice preferences of different hospitals or individual physicians. Therefore, this application also designs a continuously learning and optimizing model architecture that can perform reinforcement training and strategy updates based on physician feedback and the clinical outcomes of target subjects, achieving system adaptation and dynamic evolution. This optimization model integrates human-computer interaction and reinforcement learning feedback mechanisms into a closed-loop self-optimization mechanism, enabling iterative performance improvement of the decision-making model, continuous enhancement of recommendation rationality, and dynamic adjustment of recommendation criteria.

[0097] First, a physician-participatory feedback pathway is constructed, enabling the model to perceive physicians' adoption of recommendations and accordingly perform "effectiveness reinforcement" and "error correction." Interaction methods can include: recommendation acceptance (marked as a positive sample); recommendation rejection (marked as a negative sample); and physician modification suggestions (marked as partially effective samples and interpretability backtracking). Based on the feedback information, collaborative optimization of reinforcement learning and contrastive learning can be performed.

[0098] First, feedback information regarding the target decision-making suggestions and the clinical outcomes of the target subjects are obtained to acquire interaction trajectory data. Then, the network parameters of the neural network decision layer are optimized based on the interaction trajectory data until the neural network decision layer converges, yielding the reinforcement learning result.

[0099] Specifically, scenario modeling is performed using Markov decision processes. (State) For the joint feature vector Z, the action For the recommended intervention combination (such as vasopressors + antibiotics), a reward will be given. Based on the definition of doctor adoption behavior, a recommendation accepted is assigned +1, a recommendation modified is assigned 0, and a recommendation rejected is assigned -1. The policy update method uses either the REINFORCE algorithm or the Proximal Policy Optimization (PPO) method to update the policy network parameters. . For the policy network (i.e., the decision layer of the neural network), the reward signal can be designed as a combination: By introducing a partial adoption reward discount mechanism and adding a causal origination stability regularization term, we can prevent the generalization of recommendation paths with high rewards but no interpretability.

[0100] Then, interventions with feedback information indicating adoption and effective clinical outcomes are used as positive sample features, while interventions with feedback information indicating rejection and / or ineffectiveness are used as negative sample features. A contrastive loss function is constructed based on the positive and negative sample features, and this function is backpropagated to the multimodal feature encoder corresponding to the multimodal clinical data. The weight parameters of the multimodal feature encoder are then updated using gradient descent to obtain the contrastive learning results.

[0101] Specifically, the encoder (feature extraction layer) is optimized to narrow the distance between "adopted + valid" recommendation vectors and widen the distance between "rejected / erroneous" samples. Positive sample pairs: similar status of the same patient → doctor adoption → positive recommendation; negative sample pairs: different patients → recommendation rejection → heterogeneous features. Contrast loss (such as InfoNCE or TripletLoss) is used. in, Encode the current sample. As a positive sample, Here, is the temperature coefficient (hyperparameter), and is the cosine similarity function. The update involves the encoder (feature extraction part) weights.

[0102] Next, error correction and credibility calibration are performed based on the reinforcement learning and contrastive learning results. The reinforcement learning and contrastive learning results are then fed back to the recommendation credibility calibration module. For items with large fluctuations in adoption rate, the Softmax confidence level is adjusted; for controversial items, they are marked as "low-confidence recommendations" and added to the doctor's priority review list.

[0103] Finally, the first loss of the reinforcement learning result is calculated, and compared with the second loss of the learning result and the consistency regularization term of the neural network recommendation vector and the symbol rule recommendation vector to construct a joint loss function. The joint loss function is then backpropagated to the multimodal feature encoder, the neural network decision layer, and the medical knowledge graph to update the parameters of these components. The final joint optimization objective is: . To adjust the hyperparameters for different subtask weights, This is a regularization term representing the consistency difference between symbolic recommendation paths and neural recommendations.

[0104] For example, suppose the target decision recommendation is as follows: Recommendation: "Vapor resuscitation + vasopressor"; Doctor's behavior: Only adopt "vasopressor", refuse fluid resuscitation; Doctor's reason: "No volume responsiveness (no fluid response)"; Feedback process: Reinforcement learning: The model is partially adopted, rewarded with 0.5, and the policy layer is updated; Contrastive learning: The current vector is compared with the history of "rejected fluid resuscitation", and the encoder parameters are updated. Calibration module: The confidence of "fluid resuscitation" decreases in this state.

[0105] The medical decision-making method based on a neural symbolic hybrid model provided in this application integrates multimodal state modeling of the target object, structured medical knowledge reasoning, and deep neural network expression capabilities, and has the following beneficial technical effects.

[0106] First, improve the quality of multimodal data fusion. By constructing a structure-aware modal attention fusion mechanism, the semantic contribution weights of structured, text, and image inputs can be dynamically allocated, significantly enhancing the comprehensive recognition ability of complex pathological states, especially suitable for early identification of critical illnesses and analysis of atypical clinical manifestations.

[0107] Secondly, it achieves collaborative reasoning between clinical medical rules and neural networks. By proposing a three-layer neural-symbolic fusion strategy, including dynamic weighting of confidence, interactive fusion of mid-level graph neural networks, and consistency conflict discrimination mechanism, it overcomes the problem of lack of interpretability in traditional black-box neural networks, ensuring that the recommended content has medical rationality and behavioral constraints.

[0108] Furthermore, it supports personalized, structured, and interpretable recommendation outputs. By introducing structured logical path backtracking, SHAP feature weight display, and modality source tracing mechanisms, the system can output transparent, accountable, and traceable intervention suggestions, helping clinicians understand the basis for recommendations and improving the acceptance and implementation efficiency of suggestions.

[0109] Furthermore, it possesses closed-loop learning and continuous optimization capabilities. By establishing reinforcement learning and comparative learning mechanisms based on physician operational feedback and patient outcome annotations, the model strategy can be continuously optimized according to the practical habits of different hospitals, departments, or physicians, thereby improving the system's generalization ability and supporting local personalized deployment.

[0110] In addition, it significantly improves the efficiency of critical care identification and intervention. In typical application scenarios (such as early identification of ARDS and recommendations for vasopressor strategies in septic shock), it can significantly shorten the clinical intervention decision-making time, improve the level of medical quality control, and reduce the risk of adverse outcomes due to delays or misjudgments.

[0111] In summary, the embodiments of this application not only resolve the structural contradictions of existing AI-assisted decision-making methods, namely "strong data-driven but lacking constraints" and "rule systems that are interpretable but lack learning capabilities," but also provide a deployable, iterative, and traceable intelligent recommendation system solution, which has broad clinical application value and good prospects for industrial transformation.

[0112] Figure 3 This is a schematic diagram of the structure of a medical decision-making device based on a neural symbolic hybrid model provided in an embodiment of this application. Figure 3 As shown, the medical decision-making device 300 may include a fusion module 301, an association module 302, a recommendation module 303, and a generation module 304.

[0113] The fusion module 301 is used to acquire multimodal clinical data of the target object and fuse the multimodal clinical data through an attention weighting mechanism to obtain a multimodal fusion representation. The multimodal clinical data includes structured time-series data, free text data and medical image data.

[0114] The association module 302 is used to construct a structure-aware cross-modal attention network based on multimodal fusion representation, and based on the medical mask enhancement mechanism of high-risk indicators, associate the high-risk indicators of the structure-aware cross-modal attention network with medical indicator masks to obtain the joint feature vector of the target object.

[0115] The recommendation module 303 is used to input the joint feature vector into the neural network decision layer to obtain the neural network recommendation vector, and input the joint feature vector into the medical knowledge graph to obtain the symbol rule recommendation vector.

[0116] The generation module 304 is used to input the neural network recommendation vector, the symbol rule recommendation vector, and the joint feature vector into the three-layer neural symbol fusion network to obtain the target decision suggestion.

[0117] The generation module 304 may include a first integration unit, a second integration unit, and an output unit. The first integration unit integrates the neural network recommendation vector and the symbolic rule recommendation vector through a dynamic confidence weighted network to obtain a first fused vector. The second integration unit injects the joint feature vector into the nodes of the medical knowledge graph, constructs a neural symbolic graph neural network, and obtains a second fused vector through graph attention propagation. The output unit inputs the first fused vector and the second fused vector into a consistency discriminator and outputs the probability that the first fused vector and the second fused vector conflict.

[0118] The generation module 304 may further include a suggestion unit, a prompting unit, and a response unit. The suggestion unit is used to generate a target decision suggestion based on a first fusion vector and a second fusion vector when the probability is less than or equal to a set conflict threshold. The prompting unit is used to trigger a conflict splitting mechanism and generate a confirmation prompt message when the probability is greater than the conflict threshold. The confirmation prompt message includes the reason for the conflict between the first fusion vector and the second fusion vector. The response unit is used to generate a target decision suggestion based on the first fusion vector and the second fusion vector in response to receiving a confirmation instruction for the confirmation prompt message.

[0119] The fusion module 301 may include an encoding unit, a gating unit, and a computation unit. The encoding unit encodes multimodal clinical data into multimodal feature vectors. The gating unit calculates attention weights on the multimodal feature vectors using a learnable gating function. The computation unit calculates the multimodal fused representation based on the multimodal feature vectors and the attention weights.

[0120] The association module 302 may include a projection unit, a first construction unit, an enhancement unit, and a connection unit. The projection unit projects the multimodal fusion representation into multiple subspace representations using a learnable parameter matrix. These subspace representations include a query vector, a key vector, and a value vector. The first construction unit constructs an initial attention matrix based on the query vector and the key vector. This initial attention matrix includes multiple feature channels or indicator dimensions. The enhancement unit performs mask enhancement on high-risk indicators across the multiple feature channels or indicator dimensions to obtain a target attention matrix. The connection unit weights the value vector using the target attention matrix and then performs residual connections and layer normalization with the multimodal fusion representation to obtain a joint feature vector.

[0121] The recommendation module 303 may include a first recommendation unit, a filtering unit, a second construction unit, and a second recommendation unit. The first recommendation unit inputs the joint feature vector into the decision layer of a neural network composed of a multilayer perceptron to generate a probability distribution of intervention plans, thereby obtaining a neural network recommendation vector. The filtering unit inputs the joint feature vector into a medical knowledge graph, calculates the cosine similarity between the joint feature vector and the nodes of the medical knowledge graph, and filters entity nodes that are associated with the target object. The second construction unit dynamically constructs a reasoning subgraph including relevant paths of the entity nodes based on the filtered entity nodes; these relevant paths include rule paths. The second recommendation unit traverses the rule paths of the reasoning subgraph, calculates the matching degree between each rule path and the state of the target object, and obtains a symbolic rule recommendation vector through normalization.

[0122] In this embodiment, the medical decision-making device 300 may further include a determination module, an interpretation module, and an encapsulation module. The determination module determines a target recommendation item from among the candidate intervention items for the target decision suggestion based on the recommendation probability. The interpretation module constructs interpretation information corresponding to the target recommendation item, including neural feature contribution weights, symbolic path tracing interpretations, and modality source tracing information. The encapsulation module encapsulates the target recommendation item and interpretation information into a data packet according to a predefined structure to generate a modality source tracing report. This modality source tracing report is a structured report that can be read by doctors and displayed in the clinical system.

[0123] In this embodiment, the medical decision-making device 300 may further include an acquisition module, a reinforcement learning module, a feature determination module, a contrastive learning module, a calibration module, a joint module, and an update module. The acquisition module acquires feedback information on the target decision recommendation and the clinical outcome of the target object to obtain interaction trajectory data. The reinforcement learning module optimizes the network parameters of the neural network decision layer based on the interaction trajectory data until the neural network decision layer converges, obtaining the reinforcement learning result. The feature determination module uses interventions with adopted feedback information and effective clinical outcomes as positive sample features, and interventions with rejected and / or invalid feedback information as negative sample features. The contrastive learning module constructs a contrastive loss function based on the positive and negative sample features, backpropagates the contrastive loss function to the multimodal feature encoder corresponding to the multimodal clinical data, and updates the weight parameters of the multimodal feature encoder using gradient descent to obtain the contrastive learning result. The calibration module performs error correction and credibility calibration operations based on the reinforcement learning result and the contrastive learning result. The joint module calculates the first loss of the reinforcement learning result, the second loss of the contrastive learning result, and the consistency regularization term of the neural network recommendation vector and the symbol rule recommendation vector to construct a joint loss function. The update module is used to backpropagate the joint loss function to the multimodal feature encoder, the neural network decision layer, and the medical knowledge graph to update the parameters of the multimodal feature encoder, the neural network decision layer, and the medical knowledge graph.

[0124] This application also provides a computer-readable storage medium storing a program that can be loaded by a processor and executed by any of the medical decision-making methods based on a neural symbolic hybrid model in this application.

[0125] Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the program can also be stored in a server, another computer, disk, optical disk, flash drive, or external hard drive, etc., and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved.

[0126] The above examples illustrate this application only to aid understanding and are not intended to limit its scope. Those skilled in the art to which this application pertains can make various simple deductions, modifications, or substitutions based on the ideas presented.

Claims

1. A medical decision-making method based on a neural symbolic mixture model, characterized in that, include: The multimodal clinical data of the target object is acquired, and the multimodal clinical data is fused through an attention weighting mechanism to obtain a multimodal fusion representation. The multimodal clinical data includes structured time-series data, free text data, and medical image data. The multimodal fusion representation is projected into multiple subspace representations using a learnable parameter matrix, wherein the subspace representations include query vectors, key vectors, and value vectors. An initial attention matrix is ​​constructed based on the query vector and the key vector, and the initial attention matrix includes multiple feature channels or indicator dimensions; In multiple feature channels or indicator dimensions, high-risk indicators are masked to enhance their appearance, thus obtaining a target attention matrix. After weighting the value vector using the target attention matrix, it is then subjected to residual connection and layer normalization with the multimodal fusion representation to obtain a joint feature vector; The joint feature vector is input into the decision layer of a neural network composed of a multilayer perceptron to generate the probability distribution of the intervention scheme, so as to obtain the neural network recommendation vector. The joint feature vector is input into the medical knowledge graph, the cosine similarity between the joint feature vector and the nodes of the medical knowledge graph is calculated, and entity nodes that are associated with the target object are filtered out. Based on the selected entity nodes, a reasoning subgraph including the relevant paths of the entity nodes is dynamically constructed, and the relevant paths include rule paths; Traverse the rule paths of the reasoning subgraph, calculate the matching degree between each rule path and the state of the target object, and obtain the symbol rule recommendation vector through normalization; The first fusion vector is obtained by integrating the neural network recommendation vector and the symbol rule recommendation vector through a dynamic confidence weighted network; The joint feature vector is injected into the nodes of the medical knowledge graph to construct a neural symbolic graph neural network, and a second fusion vector is obtained through graph attention propagation; The first fusion vector and the second fusion vector are input into the consistency discriminator, which outputs the probability that the first fusion vector and the second fusion vector conflict. If the probability is less than or equal to a set conflict threshold, a target decision suggestion is generated based on the first fusion vector and the second fusion vector. If the probability is greater than the conflict threshold, a conflict splitting mechanism is triggered to generate a confirmation prompt message, which includes the reason for the conflict between the first fusion vector and the second fusion vector. In response to receiving a confirmation instruction for the confirmation prompt information, the target decision suggestion is generated based on the first fusion vector and the second fusion vector.

2. The medical decision-making method according to claim 1, characterized in that, The process of fusing the multimodal clinical data through an attention-weighted mechanism to obtain a multimodal fusion representation includes: The multimodal clinical data is encoded into multimodal feature vectors; The attention weights of the multimodal feature vectors are calculated using a learnable gating function; The multimodal fusion representation is calculated based on the multimodal feature vector and the attention weight.

3. The medical decision-making method according to claim 1, characterized in that, Also includes: The target recommendation is determined from the candidate interventions of the target decision suggestion based on the recommendation probability; Construct explanatory information corresponding to the target recommendation item, the explanatory information including neural feature contribution weights, symbolic path tracing explanations, and modality source tracing information; The target recommendations and the explanatory information are encapsulated into a data packet according to the set structure to generate a modal source tracing report, which is a structured report that can be read by doctors and displayed in the clinical system.

4. The medical decision-making method according to claim 1, characterized in that, Also includes: Obtain feedback information on the target decision-making suggestions and the clinical outcomes of the target subjects to obtain interaction trajectory data; The network parameters of the neural network decision layer are optimized based on the interaction trajectory data until the neural network decision layer converges, thus obtaining the reinforcement learning result. Interventions whose feedback information is adopted and whose clinical outcome is effective are considered positive sample features, while interventions whose feedback information is rejected and / or ineffective are considered negative sample features. A contrastive loss function is constructed based on the positive and negative sample features, and the contrastive loss function is backpropagated to the multimodal feature encoder corresponding to the multimodal clinical data. The weight parameters of the multimodal feature encoder are then updated using the gradient descent method to obtain the contrastive learning result. Error correction and credibility calibration operations are performed based on the reinforcement learning results and the contrastive learning results; Calculate the first loss of the reinforcement learning result, the second loss of the contrastive learning result, and the consistency regularization term of the neural network recommendation vector and the symbol rule recommendation vector to construct a joint loss function; The joint loss function is backpropagated to the multimodal feature encoder, the neural network decision layer, and the medical knowledge graph to update the parameters of the multimodal feature encoder, the neural network decision layer, and the medical knowledge graph.

5. A medical decision-making device based on a neural symbolic hybrid model, characterized in that, include: The fusion module is used to acquire multimodal clinical data of the target object and fuse the multimodal clinical data through an attention weighting mechanism to obtain a multimodal fusion representation. The multimodal clinical data includes structured time-series data, free text data, and medical image data. The association module includes a projection unit, a first construction unit, an enhancement unit, and a connection unit; the projection unit is used to project the multimodal fusion representation into multiple subspace representations through a learnable parameter matrix, the subspace representations including query vectors, key vectors, and value vectors; The first construction unit is used to construct an initial attention matrix based on the query vector and the key vector, the initial attention matrix including multiple feature channels or indicator dimensions; the enhancement unit is used to perform mask enhancement on high-risk indicators in the multiple feature channels or indicator dimensions to obtain a target attention matrix; the connection unit is used to weight the value vector through the target attention matrix, and then perform residual connection and layer normalization with the multimodal fusion representation to obtain a joint feature vector; The recommendation module includes a first recommendation unit, a filtering unit, a second construction unit, and a second recommendation unit. The first recommendation unit is used to input the joint feature vector into the decision layer of a neural network composed of a multilayer perceptron to generate a probability distribution of intervention schemes, thereby obtaining a neural network recommendation vector. The filtering unit is used to input the joint feature vector into a medical knowledge graph, calculate the cosine similarity between the joint feature vector and the nodes of the medical knowledge graph, and filter entity nodes that are associated with the target object. The second construction unit is used to dynamically construct a reasoning subgraph including the relevant paths of the entity nodes based on the filtered entity nodes, wherein the relevant paths include rule paths. The second recommendation unit is used to traverse the rule paths of the reasoning subgraph, calculate the matching degree between each rule path and the state of the target object, and obtain the symbol rule recommendation vector through normalization; The generation module includes a first integration unit, a second integration unit, an output unit, a suggestion unit, a prompting unit, and a response unit. The first integration unit integrates the neural network recommendation vector and the symbolic rule recommendation vector through a dynamic confidence weighted network to obtain a first fusion vector. The second integration unit injects the joint feature vector into the nodes of the medical knowledge graph to construct a neural symbolic graph neural network and obtains a second fusion vector through graph attention propagation. The output unit inputs the first fusion vector and the second fusion vector into a consistency discriminator and outputs the probability that the first fusion vector and the second fusion vector conflict. The suggestion unit generates a target decision suggestion based on the first fusion vector and the second fusion vector when the probability is less than or equal to a set conflict threshold. The prompting unit is used to trigger a conflict splitting mechanism and generate a confirmation prompt message when the probability is greater than the conflict threshold. The confirmation prompt message includes the conflict reason between the first fusion vector and the second fusion vector. The response unit is used to generate the target decision suggestion based on the first fusion vector and the second fusion vector in response to receiving a confirmation instruction for the confirmation prompt message.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that can be loaded by a processor and executed as described in any one of claims 1 to 4.

Images