Treatment path planning method and system based on deep reinforcement learning

Abstract

The present application relates to the technical field of artificial intelligence, and more particularly to a treatment path planning method and system based on deep reinforcement learning, which converts patient multi-modal data into a structured medical concept graph by constructing a double-layer representation space, generates and clips a candidate treatment path set using a differentiable logical reasoning engine, and finally constructs a neural-symbolic attribution graph to trace the decision basis, generate a hierarchical interpretable report and optimize the treatment path scheme. The present application realizes the fusion of data-driven and knowledge reasoning, and improves the rationality of path planning and the transparency of the decision-making process.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, and in particular to a treatment path planning method and system based on deep reinforcement learning. Background Technology

[0002] In the field of critical care medicine, treatment pathway planning is a crucial step in influencing patient prognosis. Traditional treatment pathway planning relies primarily on clinical guidelines and expert experience. Clinical guidelines provide standardized treatment procedures, while expert experience is used to handle complex or special cases not covered by the guidelines. This model has formed a relatively stable decision-making framework through long-term medical practice.

[0003] In existing technologies, a common approach is automated decision support based on predefined rule systems. These systems encode clinical guidelines and expert knowledge into a series of "if-then" logical rules. After receiving the patient's clinical data, the system performs forward chaining reasoning through a rule engine, matching rules that meet the conditions to output suggested treatment steps or pathways. This approach achieves standardization and automation of the process to a certain extent, reducing human error.

[0004] Another increasingly popular approach is purely data-driven predictive models, such as deep learning-based sequence models. These models learn temporal relationships and patterns between clinical events by analyzing large amounts of historical electronic medical record data. Given a patient's current condition data, the model can predict its future evolution or recommend the next intervention. This approach can uncover potential complex relationships from massive amounts of data, aiding clinical decision-making.

[0005] However, the aforementioned conventional approaches have significant limitations. While systems based on predefined rules possess clear logic and interpretability, they suffer from a severe lack of flexibility. Medical knowledge updates rapidly, and clinical scenarios are complex and ever-changing; rigid rule systems struggle to cover all marginal cases and emerging treatment options. When encountering situations not explicitly defined in the rule base, the system cannot provide effective suggestions and may even offer incorrect guidance. The maintenance and expansion of rules heavily rely on frequent collaboration between knowledge engineers and domain experts, which is costly and time-consuming. Summary of the Invention

[0006] This invention provides a treatment path planning method and system based on deep reinforcement learning, which can solve the problems in the prior art.

[0007] A first aspect of this invention provides a treatment pathway planning method based on deep reinforcement learning, comprising:

[0008] The system acquires patients' multimodal clinical data streams and dynamic medical resource status, and constructs a two-layer representation space. The neural coding layer converts the multimodal clinical data streams into continuous vector representations, and the symbolic abstraction layer decodes the continuous vector representations into structured medical concept graphs.

[0009] Based on a differentiable logic reasoning engine, the structured medical concept graph is subjected to forward deduction and backward causation to generate a multi-branch treatment path hypothesis space. The differentiable logic reasoning engine transforms the logical operations of symbolic rules into gradient-propagable tensor operations to establish a path-constraint mapping relationship. Each path in the multi-branch treatment path hypothesis space is transformed into a formal representation of a temporal constraint satisfaction problem. The feasible region is then pruned using a constraint propagation algorithm to obtain a set of candidate paths.

[0010] A neural symbol attribution map is constructed. By tracing the generation process of each path in the candidate path set in reverse, the mapping relationship between the activated feature subspace in the neural network and the rule subset triggered in symbolic reasoning is extracted. Based on the mapping relationship, the dual-source contribution decomposition of each decision node is calculated. A hierarchical interpretable report is generated based on the dual-source contribution decomposition, and an optimized treatment path scheme is output.

[0011] A two-layer representation space is constructed, wherein the neural coding layer converts the multimodal clinical data stream into continuous vector representations, and the symbolic abstraction layer decodes the continuous vector representations into structured medical concept graphs, including:

[0012] In the neural coding layer, multi-scale feature extraction is performed on the multimodal clinical data stream through hierarchical semantic embedding to generate a multi-level vector representation containing local detail features and global context features. Based on the variational inference framework, the multi-level vector representation is compressed into a continuous vector representation that follows a preset prior distribution.

[0013] In the symbolic abstraction layer, a continuous-discrete transformation mechanism is established. The dimensional intervals in the continuous vector representation are mapped to a set of candidate medical concept nodes through a probabilistic decoder. The candidate medical concept node set is then subjected to topological inference based on a graph structure generation model. The graph structure generation model identifies causal relationships and temporal dependencies between medical concept nodes by analyzing the covariance structure and gradient flow of the continuous vector representation in the latent space. Based on this, a structured medical concept graph containing node semantic attributes and edge relationship types is constructed.

[0014] In the symbolic abstraction layer, a continuous-discrete transformation mechanism is established. A probabilistic decoder maps the dimensional intervals in the continuous vector representation to a set of candidate medical concept nodes. Based on a graph structure generation model, topological inference is performed on the set of candidate medical concept nodes, including:

[0015] When establishing the continuous-discrete transformation mechanism, the semantic segmentation hyperplane is identified in the latent space of the continuous vector representation by the variational boundary localization method. The continuous vector representation is divided into multiple dimensional intervals according to the semantic segmentation hyperplane, where each dimensional interval corresponds to a concept category in a predefined medical ontology. The probability decoder is used to perform Bayesian decoding on each dimensional interval to generate a set of candidate medical concept nodes containing concept identifiers and posterior probabilities.

[0016] The candidate medical concept node set is modeled based on a graph structure generation model. Each node in the candidate medical concept node set is mapped to a relation embedding space through latent relation embedding. The affinity matrix between node pairs is calculated in the relation embedding space, and edge generation decisions are made based on the numerical distribution in the affinity matrix. At the same time, the edge generation decisions are weighted with confidence based on the posterior probability of the nodes in the candidate medical concept node set, forming a topology structure with probability labeling.

[0017] The topology is optimized by iteratively updating node representations and edge weights in the topology using a message passing algorithm, so that the topology converges to a stable state that satisfies the medical ontology constraints, thus completing the topology inference.

[0018] Based on a differentiable logic reasoning engine, the structured medical concept graph is subjected to forward deduction and backward causation to generate a multi-branch treatment path hypothesis space. The differentiable logic reasoning engine establishes path-constraint mapping relationships by transforming the logical operations of symbolic rules into gradient-propagable tensor operations.

[0019] During forward inference, the concept nodes and edge relationships in the structured medical concept graph are encoded as tensor representations, and the predefined medical diagnosis and treatment symbol rules are transformed into tensor operation operators. The tensor operation operators are applied to the tensor representations through a differentiable logic reasoning engine to perform logical conjunction, disjunction, and implication differentiable computations. Multi-step reasoning propagation is carried out from the initial symptom node along the directed edges to generate a set of forward inference paths.

[0020] When performing reverse causation, starting from the target treatment node in the structured medical concept graph, the differentiable logic reasoning engine performs reverse tensor operations along the reverse direction of the edge to identify the combination of precursor concept nodes that can lead to the target treatment node. For each combination of precursor concept nodes, the confidence score of satisfying medical causal constraints is calculated, and the combination of precursor concept nodes with confidence scores exceeding a preset confidence threshold is selected as the set of reverse causation paths.

[0021] The forward inference path set and the reverse causal path set are bidirectionally matched and fused. By calculating the semantic consistency and logical coherence between the paths, a path-constraint mapping relationship is constructed.

[0022] By applying the tensor operators to the tensor representation through a differentiable logic reasoning engine, performing logical conjunction, disjunction, and implication differentiable computations, and conducting multi-step reasoning propagation along directed edges from the initial symptom node, a set of forward inference paths is generated, including:

[0023] When performing differentiable computation, the differentiable logic reasoning engine transforms logical conjunction, disjunction, and implication operations into their corresponding tensor operation forms, and constructs reasoning state tensors based on the tensor representations of nodes in the structured medical concept graph.

[0024] During multi-step inference propagation, the inference state tensor is initialized from the initial symptom node. In each inference step, a downstream node to be activated is selected according to the directed edge topology of the structured medical concept graph. Tensor operation operators are applied to the downstream node. The updated activation value of the downstream node is calculated by combining the activation strength of its upstream node in the inference state tensor and the weight of the directed edge. The updated activation value is then written into the inference state tensor.

[0025] After completing a preset number of inference steps, the node sequence with activation intensity exceeding the activation threshold is extracted from the inference state tensor. The complete path trajectory of inference propagation is reconstructed based on the directed edge connection relationship between nodes and the historical record of the inference state tensor in each inference step, generating a forward inference path set.

[0026] Each path in the multi-branch treatment path hypothesis space is transformed into a formal representation of a temporal constraint satisfaction problem, and the feasible region is pruned using a constraint propagation algorithm to obtain a candidate path set including:

[0027] During the formal representation transformation, the sequence of medical intervention nodes contained in each path in the multi-branch treatment path hypothesis space is extracted, each node in the medical intervention node sequence is mapped to a time-series constraint variable, and a constraint domain is constructed for each time-series constraint variable. At the same time, a constraint relationship network is established based on the causal relationship between nodes.

[0028] When performing feasible domain pruning, the constraint propagation algorithm performs bidirectional propagation in the constraint relationship network. When the constraint domain of a certain time-series constraint variable shrinks, the corresponding shrinkage effect is propagated to the successor time-series constraint variable along the constraint relationship network to update the lower bound of the constraint domain. At the same time, the corresponding shrinkage effect is propagated back to the predecessor time-series constraint variable to update the upper bound of the constraint domain, forming a cascaded shrinkage process of the constraint domain.

[0029] During the cascading contraction process, the constraint domain status of all time-series constraint variables is monitored. When it is detected that the constraint domain of a time-series constraint variable in a certain path has shrunk to an empty set, the corresponding path is marked as an infeasible path and eliminated. The paths with non-empty constraint domains are retained to form a candidate path set.

[0030] Constructing a neural symbolic attribution map involves tracing the generation process of each path in the candidate path set backwards, and extracting the mapping relationship between the activated feature subspace in the neural network and the subset of rules triggered in symbolic reasoning.

[0031] During the reverse tracing, each path in the candidate path set is traversed in reverse along its generation link. In the neural coding layer, the set of neurons that contribute to the path is located through gradient backpropagation. The activated feature subspace is delineated according to the coordinate distribution of the set of neurons in the latent space. In the symbolic abstraction layer, the sequence of logical rules called during the reasoning process of the path is traced back. The core rules are selected to form a rule subset based on the frequency of rule calls and the reasoning depth.

[0032] When extracting mapping relationships, a cross-layer association matrix is ​​established, using the feature dimensions in the feature subspace as row indices and the logical rules in the rule subset as column indices. By calculating the degree of matching between the activation mode of the feature dimensions and the preconditions of the logical rules, association strength values ​​are assigned to the matrix elements.

[0033] Based on the cross-layer association matrix, a neural symbol attribution graph is constructed, and the feature dimension and logical rules are represented as two types of nodes respectively. Directed edges are established between corresponding nodes according to the association strength value to complete the extraction of mapping relationships.

[0034] A second aspect of this invention provides a treatment pathway planning system based on deep reinforcement learning, comprising:

[0035] The representation construction unit is used to acquire the patient's multimodal clinical data stream and dynamic medical resource status, and construct a two-layer representation space, wherein the neural coding layer converts the multimodal clinical data stream into continuous vector representations, and the symbolic abstraction layer decodes the continuous vector representations into a structured medical concept graph.

[0036] The path deduction unit is used to perform forward deduction and backward causation on the structured medical concept map based on a differentiable logic reasoning engine to generate a multi-branch treatment path hypothesis space. The differentiable logic reasoning engine transforms the logical operations of symbolic rules into gradient-propagable tensor operations, establishes a path-constraint mapping relationship, converts each path in the multi-branch treatment path hypothesis space into a formal representation of a temporal constraint satisfaction problem, and performs feasible domain pruning through a constraint propagation algorithm to obtain a candidate path set.

[0037] The attribution graph unit is used to construct a neural symbol attribution graph. By tracing back the generation process of each path in the candidate path set, it extracts the mapping relationship between the activated feature subspace in the neural network and the rule subset triggered in symbolic reasoning. Based on the mapping relationship, it calculates the dual-source contribution decomposition of each decision node, generates a hierarchical interpretable report based on the dual-source contribution decomposition, and outputs an optimized treatment path scheme.

[0038] A third aspect of the present invention provides an electronic device, comprising:

[0039] processor;

[0040] Memory used to store processor-executable instructions;

[0041] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0042] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0043] The two-layer representation space constructed by this method can effectively integrate multimodal clinical data streams and dynamic medical resource information. The neural coding layer transforms temporal, high-dimensional multimodal data into low-dimensional continuous vectors, preserving the complex temporal relationships and latent patterns in the data. The symbolic abstraction layer further decodes these vectors into structured medical concept graphs, realizing a symbolic mapping from data representation to medical knowledge. This two-layer structure utilizes the representational capabilities of deep learning while introducing an interpretable framework of prior medical knowledge.

[0044] By using a differentiable logic reasoning engine to deduce medical concept graphs, multi-branch treatment path hypotheses that conform to medical logic can be generated. This engine transforms the logical operations of symbolic rules into differentiable tensor operations, enabling the symbolic reasoning process to be co-optimized with neural networks. The generated paths are formalized as temporal constraint satisfaction problems, and feasible region pruning is performed using a constraint propagation algorithm. This process significantly reduces the search space, efficiently eliminating a large number of invalid paths that violate medical resource constraints or clinical guidelines, thereby quickly focusing on a set of feasible candidate paths.

[0045] The construction of a neural symbolic attribution map enables dual-source tracing and contribution decomposition of the decision-making process. By establishing a mapping relationship between neural network activation patterns and symbolic reasoning rule triggers, it quantifies the relative contributions of data-driven features and knowledge-driven rules at each decision node. The resulting hierarchical interpretable report clearly demonstrates the complete decision-making chain from raw data to the final path. This provides clinicians with intuitive decision-making support, enhances their trust in AI-recommended solutions, and helps identify potential data biases or rule conflicts. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the treatment pathway planning method based on deep reinforcement learning, as described in an embodiment of the present invention.

[0047] Figure 2 This is a flowchart illustrating the extraction of candidate path mapping relationships based on neural symbol attribution maps in an embodiment of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0050] Figure 1 This is a flowchart illustrating the treatment pathway planning method based on deep reinforcement learning, as described in an embodiment of the present invention. Figure 1 As shown, the treatment pathway planning method based on deep reinforcement learning includes:

[0051] The system acquires patients' multimodal clinical data streams and dynamic medical resource status, and constructs a two-layer representation space. The neural coding layer converts the multimodal clinical data streams into continuous vector representations, and the symbolic abstraction layer decodes the continuous vector representations into structured medical concept graphs.

[0052] Based on a differentiable logic reasoning engine, the structured medical concept graph is subjected to forward deduction and backward causation to generate a multi-branch treatment path hypothesis space. The differentiable logic reasoning engine transforms the logical operations of symbolic rules into gradient-propagable tensor operations to establish a path-constraint mapping relationship. Each path in the multi-branch treatment path hypothesis space is transformed into a formal representation of a temporal constraint satisfaction problem. The feasible region is then pruned using a constraint propagation algorithm to obtain a set of candidate paths.

[0053] A neural symbol attribution map is constructed. By tracing the generation process of each path in the candidate path set in reverse, the mapping relationship between the activated feature subspace in the neural network and the rule subset triggered in symbolic reasoning is extracted. Based on the mapping relationship, the dual-source contribution decomposition of each decision node is calculated. A hierarchical interpretable report is generated based on the dual-source contribution decomposition, and an optimized treatment path scheme is output.

[0054] Upon receiving information about critically injured traffic accident victims, the multi-source data acquisition module simultaneously acquires data from portable monitoring equipment at the emergency scene, vehicle-mounted imaging equipment, and the emergency personnel's voice input system. The monitoring equipment collects heart rate, blood pressure, blood oxygen saturation, and respiratory rate once per second. Data fields are encoded using the IEEE 11073 standard, with heart rate ranging from 30 to 250 beats per minute, blood pressure ranging from 50 to 250 mmHg systolic and 30 to 150 mmHg diastolic, and blood oxygen saturation ranging from 70% to 100%. The vehicle-mounted CT equipment acquires images of the head and chest / abdomen at a resolution of 512 x 512 pixels, with slice thickness of 5 mm, and encapsulated in DICOM format. The voice input is converted to text by a pre-trained speech recognition module in the medical field, achieving an accuracy rate of over 95%.

[0055] Data standardization maps data from different sources to a unified medical terminology coding system. Raw vital signs values ​​are standardized by subtracting the population mean and then dividing by the standard deviation. The mean heart rate is set at 75 beats per minute with a standard deviation of 15, and the mean systolic blood pressure is set at 120 mmHg with a standard deviation of 20 mmHg. Pixel values ​​in image data are normalized to the 0-1 range by dividing by 255. Text descriptions extract symptoms, signs, and injury sites using a medical entity recognition module. Entity recognition employs a sequence labeling model based on conditional random fields, achieving an accuracy of at least 90%. The extracted medical entities are mapped to the ICD-10 coding system and the SNOMED-CT ontology, establishing standardized medical concept identifiers.

[0056] The neural coding layer extracts features from the standardized multimodal data. Vital sign time series are processed through three parallel one-dimensional convolutional branches: the short-term branch has a window length of 600 sampling points corresponding to 10 minutes, the medium-term branch has a window of 3600 sampling points corresponding to 1 hour, and the long-term branch has a window of 21600 sampling points corresponding to 6 hours. Each branch contains three convolutional layers, with the number of convolutional kernels increasing from 32 to 64 and then to 128. The kernel size is fixed at 3, the stride is 2, and zero padding is used to maintain time alignment. The output feature maps of the three branches are concatenated along the channel dimension and then subjected to global average pooling to obtain a 256-dimensional temporal feature vector. Image data is input into a 50-layer three-dimensional residual network, where multi-scale features are extracted at layers 10, 20, 35, and 50. The feature maps of each layer are uniformly sized to 8x8x8 using adaptive average pooling and then flattened into a 512-dimensional image feature vector after concatenation along the channel dimension. The text data is encoded using a 768-dimensional pre-trained language model in the medical field. The maximum length of the input sequence is 512 words, and the pooling output of the last layer is taken as the text feature vector.

[0057] A cross-attention mechanism integrates multimodal features, using temporal feature vectors as queries and concatenating image and text feature vectors as key-value pairs. The dot product of the query and key vectors, divided by the square root of the dimension (16), is then normalized using softmax to obtain attention weights. These weights are then weighted and summed with the value vectors to output the context vector. The temporal encoding vector converts the number of minutes between the data acquisition time and the admission time into a combination of sine and cosine functions with a period parameter of 10000. This encoding vector is added to the feature vectors to incorporate temporal information. The fused 2048-dimensional feature vector is input to a variational encoder, which contains two parallel fully connected networks predicting the mean and log-variance vectors, respectively, both with an output dimension of 128. Sampling is achieved through a reparameterization technique: first, a noise vector is sampled from a standard normal distribution, multiplied by the square root of the variance, and then added to the mean to obtain the latent vector. The loss function includes a reconstruction loss and a KL divergence term, with a KL divergence weight coefficient of 0.001 to avoid posterior collapse.

[0058] The symbolic abstraction layer transforms a 128-dimensional continuous latent vector into discrete medical concepts. The latent vectors are divided into four intervals based on their dimensions: the first 32 dimensions correspond to symptom categories, the next 32 to sign categories, dimensions 65-96 to diagnostic indicator categories, and the last 32 to diagnostic categories. Bayesian decoding is performed on each interval to calculate the posterior probability of each candidate medical concept. The prior probability is derived from a historical case database, calculated by dividing the frequency of a concept's occurrence in the database by the total number of cases. A Gaussian kernel with a bandwidth parameter of 0.1 is used to calculate the similarity between the latent vector activation value and the concept center value. The posterior probability is proportional to the product of the prior probability and the likelihood function, and is normalized for all candidate concepts. Concept nodes with a posterior probability greater than 0.4 are retained; each node contains a concept identifier, a posterior probability value, and a source dimension index.

[0059] The graph structure generation module maps candidate concept nodes to a relation embedding space. A three-layer graph convolutional network processes node features: the first layer has a hidden dimension of 256, the second 192, and the third 128. The activation function is ReLU. Affinity is calculated by dividing the dot product of the embedding vectors of any two nodes by the product of their vector norms, with values ​​ranging from -1 to +1. Edges with an affinity greater than 0.5 are strong, while those between 0.2 and 0.5 are weak. Edge types are determined based on the combination of concept categories of nodes in the medical ontology: edges from symptom to diagnosis are labeled as indicative relationships, and edges from diagnosis to treatment are labeled as indication relationships. Edge confidence is calculated as the geometric mean of the posterior probabilities of the two endpoints multiplied by the absolute value of the affinity; edges with confidence below 0.3 are filtered out.

[0060] The message-passing algorithm iteratively optimizes the graph structure, with an iteration limit of 10. In each iteration, the node representation is updated by aggregating the representations of neighboring nodes. The aggregation uses a weighted summation, with the weights being the confidence scores of the connecting edges. The edge weights are adjusted based on the cosine similarity of the representations of the two endpoints; the weights increase as the similarity increases and decrease as the similarity decreases, with the decrease rate decreasing by 10% each iteration. Medical ontology constraints are used as regularization terms; edges that violate the constraints are forcibly deleted, such as symptom nodes not being directly connected to treatment nodes. The termination condition is when the maximum norm of the change in node representations is less than 0.01 for two consecutive iterations or when the iteration limit is reached. The converged graph is the structured medical concept graph.

[0061] The differentiable logic reasoning engine performs forward inference and backward causation on the concept graph. Nodes are encoded as 512-dimensional tensors, with the first 8 dimensions being one-hot encodings of node types and the 9th dimension representing the posterior probability. Edge relations are encoded as 512x512 relation tensors, learned through a multilayer perceptron. The rule base predefines 500 clinical pathway rules, each including preconditions, inference conclusions, and confidence levels. Logical conjunction is transformed into element-wise minimization of the tensor, with a smoothing parameter of 10 to ensure differentiability. Logical disjunction is transformed into element-wise maximization. Logical implication is transformed into multiplication of the relation tensor and node tensor matrix, activated by a sigmoid function.

[0062] Forward inference begins with the current set of symptom nodes, assuming the observation of fever, headache, and neck stiffness, with posterior probabilities of 0.85, 0.72, and 0.68, respectively. The matching rule "fever, headache, and neck stiffness imply meningitis" is retrieved with a confidence level of 0.9. A smoothed minimum operation is performed on the activation vectors of the three symptom nodes to obtain the conjunctive result. This result is then right-multiplied by the causal tensor and subjected to a sigmoid function to obtain the activation vector for the meningitis diagnosis node with a confidence level of 0.61. The antibiotic treatment node is derived from the meningitis node with a confidence level of 0.55. The inference depth is capped at 8 steps. After completion, node sequences with activation strengths exceeding 0.4 are extracted, and the path is reconstructed by backtracking the state tensor history, generating a set of forward inference paths.

[0063] Reverse causation proceeds backward from the target treatment node, with each dimension of the target activation vector for the antibiotic treatment node being 1. The expected activation distribution of the preceding diagnostic nodes is calculated using the relation tensor transpose. The reverse operation employs pseudo-inverse matrix multiplication, with the pseudo-inverse calculated through singular value decomposition, retaining components with singular values ​​greater than 0.01. A confidence score is calculated for each combination of preceding diagnostic nodes, comprising three components: statistical confidence, guideline support, and patient fit, with weighting coefficients of 0.4, 0.3, and 0.3, respectively. Combinations with confidence scores exceeding 0.5 are retained, forming the reverse causation path set.

[0064] Bidirectional path matching uses a bidirectional long short-term memory network to encode paths with a hidden dimension of 256. The cosine similarity between the forward and backward path vectors is used as the semantic consistency score. Path pairs with a similarity greater than 0.7 are considered semantically consistent and are merged into candidate paths.

[0065] The temporal constraint satisfaction problem transformation maps the medical intervention nodes of the candidate paths to temporal variables. The antibiotic administration node variable v1 has a constraint domain of 0 to 120 minutes after admission; the blood culture sampling node variable v2 has a constraint domain of 0 to 30 minutes; the fluid resuscitation node variable v3 has a constraint domain of 0 to 60 minutes; and the hemodynamic monitoring node variable v4 has a constraint domain of v3 plus 30 minutes. Causal constraints are established between nodes, requiring that blood culture must occur before antibiotic administration, with v2 corresponding to v1 being less than v1.

[0066] Constraint propagation is bidirectional. If the upper bound of v2 shrinks from 30 to 20, since v2 is less than v1, the upper bound of v1 also shrinks to 20. The shrinkage of the upper bound of v1 affects subsequent variables, and cascading propagation occurs along the constraint network. The constraint domain status of all variables is monitored. If the lower bound of v2 is delayed to 25 due to resource conflicts, the constraint domain becomes 25 to 20. The lower bound exceeding the upper bound results in an empty constraint domain. An empty set is detected, marking the path as infeasible and removing it from the candidate path set. Paths with non-empty constraint domains are retained.

[0067] Real-time resource status is obtained from the hospital information system and the regional emergency command platform, updated every minute. This includes operating room occupancy status, number of intensive care unit beds, blood product inventory, specialist on-call status, and medical equipment availability indicators. When an operating room is occupied, v1 delays occupancy by 45 minutes; if the safe window period is exceeded, the path is marked as infeasible and the reason for removal is recorded.

[0068] The neural symbol attribution graph locates contributing neurons through gradient backpropagation, using the activation vector at the path endpoint as the gradient source to calculate the gradient relative to the activation values ​​of each neuron in the neural coding layer. Neurons with an absolute gradient value exceeding 0.05 are marked as contributing neurons, and the convex hull of the coordinates is calculated to determine the activation feature subspace. The sequence of logical rules invoked during the back-reasoning process is traced back, and an importance score is calculated by combining the call frequency weight (0.6) and the inference depth weight (0.4). A threshold of 2.0 is used to filter core rules. A cross-layer association matrix is ​​established, with the number of rows representing the feature subspace dimension and the number of columns representing the core rules. Matrix elements represent the matching degree between the activation interval of the feature dimension and the preconditions of the rule. An association strength threshold of 0.4 is applied; directed edges are established from feature dimension nodes to logical rule nodes at locations exceeding the threshold. The attribution graph is stored using a directed bipartite graph. Node attributes include type, identifier, and data source, while edge attributes include weight and timestamp, supporting transparent interpretation of the decision.

[0069] In one alternative implementation, a two-layer representation space is constructed, wherein a neural coding layer converts the multimodal clinical data stream into continuous vector representations, and a symbolic abstraction layer decodes the continuous vector representations into structured medical concept graphs, including:

[0070] In the neural coding layer, multi-scale feature extraction is performed on the multimodal clinical data stream through hierarchical semantic embedding to generate a multi-level vector representation containing local detail features and global context features. Based on the variational inference framework, the multi-level vector representation is compressed into a continuous vector representation that follows a preset prior distribution.

[0071] In the symbolic abstraction layer, a continuous-discrete transformation mechanism is established. The dimensional intervals in the continuous vector representation are mapped to a set of candidate medical concept nodes through a probabilistic decoder. The candidate medical concept node set is then subjected to topological inference based on a graph structure generation model. The graph structure generation model identifies causal relationships and temporal dependencies between medical concept nodes by analyzing the covariance structure and gradient flow of the continuous vector representation in the latent space. Based on this, a structured medical concept graph containing node semantic attributes and edge relationship types is constructed.

[0072] In practical applications, patients' multimodal clinical data streams typically include various heterogeneous data types such as vital sign time series, medical imaging data, electronic medical record text, and laboratory test indicators. To transform these heterogeneous data into a unified vector representation space, a hierarchical semantic embedding mechanism is employed in the neural coding layer for multi-scale feature extraction. Specifically, for vital sign time series data, a multi-resolution temporal convolutional network is used to extract dynamic features under different time windows, for example, setting window lengths of 10 minutes, 1 hour, and 6 hours to capture short-term fluctuations, medium-term trends, and long-term evolution features, respectively. For medical imaging data, a pre-trained 3D convolutional neural network is used to extract spatial features, sampling at multiple feature layers. Shallow features retain local details such as texture and edges, while deep features encode global semantic information such as lesion morphology and anatomical structure. For electronic medical record text, a domain-adapted pre-trained medical language model is used for encoding, converting clinical descriptions into context vectors while retaining the location indices of key medical entities such as disease diagnosis, symptom descriptions, and treatment plans. For laboratory testing indicators, after standardizing each numerical indicator, they are mapped to the embedding space through a fully connected network, and the correlation structure information between indicators is introduced.

[0073] After obtaining preliminary vector representations of data from different modalities, these heterogeneous features are integrated through a multimodal attention fusion mechanism. The calculation of attention weights considers not only the importance of the feature values ​​themselves but also introduces a time alignment module to address the timestamp mismatch problem caused by inconsistent acquisition frequencies of different modalities. For example, imaging data is acquired every 12 hours, while vital sign data is recorded every minute; time interpolation and dynamic weighting are needed to ensure that the fused representation accurately reflects the patient's overall clinical status at a specific moment. The multi-level vector representation generated after fusion contains multiple abstraction levels from the original signal level to a high-level semantic level. The vector dimension of each level is typically set between 256 and 1024, with the specific value adjusted according to data complexity and computational resources.

[0074] To improve the generalization ability of vector representations and mitigate the curse of dimensionality, a variational inference framework is used to compress multi-level vector representations into continuous vector representations that follow a predefined prior distribution. During variational inference, it is assumed that the latent vectors follow a multivariate Gaussian distribution, and a neural network parameterized encoder simultaneously predicts the mean vector and covariance matrix of this distribution. To ensure numerical stability, the log-variance is predicted instead of the variance itself in actual computation, avoiding numerical underflow caused by negative or excessively small values ​​in the covariance matrix. During training, latent vectors are sampled from this distribution using reparameterization techniques, making the sampling process differentiable and allowing gradient backpropagation to update the encoder parameters. The objective function consists of two parts: a reconstruction loss and a regularization term. The reconstruction loss measures the fidelity of decoding the latent vectors back to the original data space, while the regularization term uses KL divergence to constrain the difference between the latent distribution and the standard normal distribution, preventing excessive dispersion of the representation space. The dimensionality of the compressed continuous vector representation is typically set between 64 and 128, achieving significant dimensionality reduction while preserving information.

[0075] The core task of the symbolic abstraction layer is to establish a mapping relationship between the continuous vector space and the discrete medical concept space. Each dimension interval in the continuous vector representation corresponds to a specific clinical semantic. These continuous values ​​are converted into discrete medical concept nodes by a probabilistic decoder. The decoder employs a strategy combining gating mechanisms and threshold judgments. When the value of a certain dimension exceeds a preset threshold, the corresponding candidate medical concept node is activated. For example, if the values ​​of the 15th to 20th dimensions of the potential vector all exceed 0.7, the concept node "acute myocardial infarction" is activated; if the values ​​of the 30th to 35th dimensions are between 0.4 and 0.6, the node "moderate risk of heart failure" is activated. The threshold settings are based on statistical analysis of large-scale clinical data, and concept nodes of different disease domains and different severities have differentiated activation thresholds.

[0076] After generating the set of candidate medical concept nodes, it is necessary to further infer the topological relationships between these nodes and construct a structured medical concept graph. The graph structure generation model identifies statistical dependencies between different dimensions by analyzing the covariance structure of continuous vector representations in the latent space. The off-diagonal elements in the covariance matrix reflect the linear correlation between different feature dimensions; high correlation indicates that the corresponding medical concepts have a causal relationship or a common pathological basis. For example, if there is a strong positive correlation between the vector dimension representing "coronary artery stenosis" and the vector dimension representing "ST-segment elevation," then a directed edge from "coronary artery stenosis" to "ST-segment elevation" is established in the concept graph, with the edge type labeled "causing."

[0077] Beyond covariance structure analysis, gradient flow tracking provides a dynamic perspective for identifying causal relationships. By calculating the partial derivatives of the loss function with respect to each dimension of the latent vector, the contribution gradient of each dimension to the final prediction result is obtained. The gradient direction indicates the direction of the influence on the prediction when that dimension increases or decreases, while the gradient magnitude reflects the strength of the influence. If the gradient of the "blood pressure" dimension and the gradient of the "cerebral hemorrhage risk" dimension have the same direction and similar magnitude, it indicates that they have a synergistic effect in the current patient state, and an edge of the "synergistic influence" type should be established in the concept graph. Gradient flow analysis can also identify temporal dependencies. By comparing gradient changes at different time steps, it can determine whether the emergence of certain medical concepts depends on changes in the state of previous concepts.

[0078] In constructing a concept graph, the semantic attributes of nodes include not only the concept name but also meta-information such as confidence score, timestamp, and data source. The confidence score is calculated based on the stability of the vector dimension values ​​activating the node and the amount of supporting evidence; if data from multiple modalities point to the same concept, the node has a higher confidence score. The timestamp records the moment the concept was first identified, used for subsequent temporal reasoning. The data source indicates which type of clinical data the concept is primarily based on; for example, the concept of "pulmonary infection" mainly comes from imaging data and inflammatory markers, while the concept of "disorder of consciousness" mainly comes from neurological scoring data.

[0079] The types of edge relationships are finely categorized, including but not limited to "causal relationship," "temporal dependency," "mutual exclusion relationship," "synergistic effect," and "subordinate relationship." A causal relationship edge connects a cause concept and a result concept; for example, "bacterial infection" leads to "elevated white blood cell count." A temporal dependency edge indicates that one concept must occur after another; for example, "cardiac arrest" must precede "cardiopulmonary resuscitation." A mutual exclusion relationship edge connects two concepts that cannot be true simultaneously; for example, "cardiogenic shock" and "anaphylactic shock" are usually mutually exclusive at the same time. A synergistic effect edge connects concepts that have a cumulative effect on treatment effects; for example, "anticoagulation therapy" and "thrombolytic therapy" are used synergistically under specific conditions. A subordinate relationship edge represents a hierarchical structure of concepts; for example, "acute myocardial infarction" is subordinate to "coronary heart disease."

[0080] To ensure that the generated structured medical concept graph conforms to medical knowledge standards, domain knowledge constraints are introduced during the graph structure generation process. A medical ontology knowledge base is pre-built, containing standardized disease classification systems, symptom-disease association rules, drug-disease indications, and other knowledge graphs. When generating candidate edges, the knowledge base is queried to verify whether the edge violates known medical common sense; if it does, the edge is filtered or its confidence is reduced. Simultaneously, typical case patterns in the knowledge base are used to guide the generation of the concept graph; when a concept combination that highly matches a typical pattern is identified, a graph structure conforming to that pattern is generated first.

[0081] The resulting structured medical concept graph is stored in a graph database format, with nodes and edges represented as JSON objects, supporting efficient graph query and traversal operations. This concept graph not only provides structured input for subsequent path planning but also serves as a visualization interface for clinical decision support, helping doctors understand the evolution of a patient's condition and key pathogenic factors.

[0082] In one optional implementation, a continuous-discrete transformation mechanism is established in the symbolic abstraction layer. A probabilistic decoder maps the dimensional intervals in the continuous vector representation to a set of candidate medical concept nodes. Topological inference of the candidate medical concept node set is then performed based on a graph structure generation model, including:

[0083] When establishing the continuous-discrete transformation mechanism, the semantic segmentation hyperplane is identified in the latent space of the continuous vector representation by the variational boundary localization method. The continuous vector representation is divided into multiple dimensional intervals according to the semantic segmentation hyperplane, where each dimensional interval corresponds to a concept category in a predefined medical ontology. The probability decoder is used to perform Bayesian decoding on each dimensional interval to generate a set of candidate medical concept nodes containing concept identifiers and posterior probabilities.

[0084] The candidate medical concept node set is modeled based on a graph structure generation model. Each node in the candidate medical concept node set is mapped to a relation embedding space through latent relation embedding. The affinity matrix between node pairs is calculated in the relation embedding space, and edge generation decisions are made based on the numerical distribution in the affinity matrix. At the same time, the edge generation decisions are weighted with confidence based on the posterior probability of the nodes in the candidate medical concept node set, forming a topology structure with probability labeling.

[0085] The topology is optimized by iteratively updating node representations and edge weights in the topology using a message passing algorithm, so that the topology converges to a stable state that satisfies the medical ontology constraints, thus completing the topology inference.

[0086] In the process of converting continuous vector representations to discrete medical concepts, it is necessary to address the modal differences between the high-dimensional continuous representations output by the neural coding layer and the structured concepts required by the symbolic abstraction layer. Assume the continuous vector representation output by the neural coding layer is... ,in The vector dimension is typically between 256 and 1024 in medical applications. This vector is distributed on a continuous manifold in the latent space, while the concept categories defined in the medical concept ontology are discrete sets; therefore, a reliable continuous-to-discrete transformation mechanism is needed.

[0087] Variational boundary localization methods search for hyperplanes in the latent space that optimally segment different semantic regions. Specifically, for vectors in the latent space... First, its local density gradient in each dimension is calculated to identify boundary regions where significant density changes occur. These boundary regions correspond to the natural boundaries of different medical concept categories in the latent space. Assuming that in the... Identify in each dimension boundary points These boundary points divide this dimension into The potential space is divided into multiple hyperrectangular regions through multi-dimensional interval combinations. Each hyperrectangular region is associated with a specific concept category in a predefined medical ontology.

[0088] In practice, we assume that the medical ontology includes basic concept categories such as "symptoms," "signs," "laboratory indicators," and "diagnosis," with each category containing several specific concept nodes. For example, the "symptoms" category might include nodes like "fever," "cough," and "chest pain." The first 64 dimensions of the continuous vector representation correspond to the "symptoms" category, the next 64 dimensions correspond to the "signs" category, and so on. When the partitioned dimensional intervals are obtained, the probabilistic decoder performs Bayesian decoding on each interval.

[0089] The probabilistic decoder employs a variational inference framework for cases falling into the first... Vector components of each dimension interval Through prior distribution and likelihood function Calculate the posterior probability distribution ,in This represents the candidate concept node corresponding to the interval. The posterior probability is calculated by combining the relative position of the vector component within the interval with the frequency of the concept node in the training data. For each dimensional interval, the decoder generates a list of candidate concept nodes, with each node in the list accompanied by its posterior probability value. For example, for the dimensional interval corresponding to the "symptom" category, a set of candidate nodes is generated. , where the numerical value represents the posterior probability.

[0090] After summarizing the decoding results across all dimensional intervals, a candidate node set containing multiple medical concept nodes and their probability annotations is obtained. This set serves as the foundational input for subsequent topology inference. To construct the relationship structure between these discrete nodes, a graph structure generation model is introduced for topology inference. The core task of the graph structure generation model is to determine whether there should be edge connections between candidate nodes, as well as the type and strength of the edges.

[0091] Latent relation embedding techniques map each candidate concept node to an embedding space specifically designed for relation modeling. Assume the candidate node set contains... There are nodes, denoted as . Each node Through relational embedding functions Mapped to relation embedding vector ,in Typically, 128 or 256 are chosen. The construction of the relation embedding space considers the semantic relation types defined in the medical ontology, such as "causal relationship," "concurrency relationship," and "temporal relationship." The embedding function is learned through pre-training, ensuring that pairs of concept nodes with specific relationships in the ontology have high similarity in the embedding space.

[0092] In the relation embedding space, compute any two nodes and The affinity between them is used to construct an affinity matrix. Matrix elements Reflecting nodes and The tendency to establish edge connections between nodes is calculated by comprehensively considering the inner product of relation embedding vectors, the co-occurrence statistics of concepts defined in the medical ontology, and the contextual information of the current clinical scenario. The affinity matrix exhibits different numerical distribution patterns, with high numerical regions corresponding to strongly associated node pairs and low numerical regions corresponding to weakly associated or unassociated node pairs.

[0093] Edge generation decisions are based on the numerical distribution of the affinity matrix, for each element in the matrix. Set adaptive threshold This threshold combines the statistical properties of the global affinity distribution with the node and Each of their respective local neighborhood structures. When At the node and Edges are established between the nodes. The type of edge is determined by analyzing the combination of concept categories of the two nodes in the medical ontology. For example, the "symptom-diagnosis" category combination usually corresponds to "support relationship" or "indication relationship".

[0094] In the edge generation process, the posterior probability of candidate nodes provides important confidence information for edges that have already been generated. Its confidence level is determined by combining nodes. and The posterior probability is calculated. Specifically, edge confidence can be expressed using a geometric mean or a weighted combination, ensuring that an edge receives a high confidence only when both endpoint nodes have high posterior probabilities. This confidence weighting mechanism prevents topological uncertainty introduced by low-confidence nodes. After edge generation and confidence weighting, a topological structure with probability labels is formed, containing both node probability labels and edge confidence labels.

[0095] The initially generated topology exhibits inconsistencies with medical ontology constraints, such as medically impossible causal loops or missing necessary associations defined in the ontology. The structure optimization phase iteratively adjusts the topology using a message-passing algorithm to gradually satisfy the medical ontology constraints. The message-passing algorithm treats the topology as a graph neural network, where each node maintains a representation vector and each edge maintains a weight value. During iteration, nodes receive messages from neighboring nodes via edges, updating their own representations; simultaneously, they adjust edge weights based on changes in node representations.

[0096] The specific process of message passing is as follows: in the first... In the next iteration, the node The representation By aggregating its set of neighboring nodes The representation is updated. The aggregation function considers a weighted combination of edge weights and neighbor node representations, while introducing constraint rules defined in the medical ontology as regularization terms. The edge weight is updated based on the consistency of the representations of the two nodes. If the distance between the representations of the two nodes in the semantic space increases, it indicates that the edge should not exist, and its weight is reduced accordingly; otherwise, the weight is increased.

[0097] The iterative update process continues until the topology converges to a stable state. Stability criteria include node representation changes below a preset threshold, edge weight distribution stabilizing, and the structure satisfying the hard and soft constraints defined in the medical ontology. Hard constraints include prohibited concept combinations and relation types, while soft constraints include recommended concept association patterns and typical topological features. The converged topology becomes the structured medical concept graph used for inference. This graph accurately reflects the medical semantics of the patient's current clinical state while maintaining consistency with the predefined medical ontology, providing a structured input representation for the subsequent differentiable logic reasoning engine. The entire transformation and inference process achieves a reliable mapping from continuous neural encoding representations to discrete symbolic abstraction concept graphs, establishing a bridge between neural and symbolic representation modalities.

[0098] In one optional implementation, the structured medical concept graph is forward-engineered and backward-engineered based on a differentiable logic reasoning engine to generate a multi-branch treatment path hypothesis space. The differentiable logic reasoning engine establishes path-constraint mapping relationships by transforming the logical operations of symbolic rules into gradient-propagable tensor operations.

[0099] During forward inference, the concept nodes and edge relationships in the structured medical concept graph are encoded as tensor representations, and the predefined medical diagnosis and treatment symbol rules are transformed into tensor operation operators. The tensor operation operators are applied to the tensor representations through a differentiable logic reasoning engine to perform logical conjunction, disjunction, and implication differentiable computations. Multi-step reasoning propagation is carried out from the initial symptom node along the directed edges to generate a set of forward inference paths.

[0100] When performing reverse causation, starting from the target treatment node in the structured medical concept graph, the differentiable logic reasoning engine performs reverse tensor operations along the reverse direction of the edge to identify the combination of precursor concept nodes that can lead to the target treatment node. For each combination of precursor concept nodes, the confidence score of satisfying medical causal constraints is calculated, and the combination of precursor concept nodes with confidence scores exceeding a preset confidence threshold is selected as the set of reverse causation paths.

[0101] The forward inference path set and the reverse causal path set are bidirectionally matched and fused. By calculating the semantic consistency and logical coherence between the paths, a path-constraint mapping relationship is constructed.

[0102] When processing structured medical concept graphs using a differentiable logic reasoning engine, a mapping mechanism from symbolic logic to tensor operations needs to be established. Specifically, each concept node in the structured medical concept graph is encoded as a fixed-dimensional tensor representation, such as a 512-dimensional real-number vector. Disease nodes, symptom nodes, examination item nodes, and treatment measure nodes are mapped to different vector subspaces to ensure that different types of nodes are distinguishable in the semantic space. The edge relationships between concept nodes also need to be converted into tensor form; for example, the causal relationship "symptom A leads to diagnosis B" can be represented as a relation tensor. The mapping from symptom vectors to diagnostic vectors is achieved through matrix multiplication.

[0103] After tensor encoding, medical diagnostic symbolic rules need to be transformed into differentiable tensor operators. Traditional logical conjunction in symbolic reasoning is represented as "If symptom A and symptom B occur simultaneously, then the diagnosis is disease C." In a differentiable framework, this is converted into element-wise multiplication or minimization operations of tensors. Specifically, assuming the activation vector of symptom A is... The activation vector for symptom B is The differentiable form of the conjunction operation can be expressed as: Alternatively, the product norm form in fuzzy logic can be used. ,in This represents element-wise multiplication. The logical disjunction operation corresponds to "the occurrence of symptom A or symptom B triggers diagnosis C," and is implemented in tensor form as follows: Or, in the form of probability summation The logical implication operation "If symptom A then diagnosis B" is implemented in a differentiable framework through matrix multiplication of relation tensors, i.e. ,in For activation function, This is the bias vector.

[0104] The forward inference process begins with the patient's current set of symptom nodes. Assuming the patient is admitted with fever, cough, and chest pain, the initial activation vectors for these symptom nodes are denoted as follows: , and The differentiable logic reasoning engine first retrieves diagnostic rules related to these symptoms, such as the tensor operation corresponding to the rule "fever and cough imply respiratory infection". The operation corresponding to the rule "chest pain and fever indicate a cardiovascular event" is: Through multi-step reasoning propagation, the path moves from the symptom node through intermediate diagnostic nodes to the final treatment node. Each step of the reasoning involves the propagation and updating of tensor activation values. In the... In step-by-step reasoning, nodes activation value Based on all its predecessor nodes The activation values ​​and corresponding relation tensors are updated, and the update formula can be expressed as follows: This iterative propagation process continues until the preset maximum inference depth is reached or the node activation value converges, and all the activation paths generated in the end constitute the forward inference path set.

[0105] Reverse abduction proceeds by reasoning backward from the target treatment node. Assuming the clinical goal is to determine the "antibiotic treatment plan" node, it's necessary to trace back to identify which antecedent causes and symptom combinations support this treatment decision. Within the tensor operations framework, reverse abduction is achieved through the transpose of relation tensors. If relation tensors are used in forward reasoning... From node Derivation to nodes Then, when tracing back to the cause, the transposed tensor is used. From node Reverse Node Derivation The state. Specifically, let the target activation vector of the treatment node be... The expected activation distribution of its predecessor diagnostic node is calculated through inverse tensor operations. Furthermore, tracing back from the diagnostic node to the symptom node yields a combination of symptoms that can explain the target treatment.

[0106] When identifying precursor concept node combinations, it is necessary to consider the degree to which medical causal constraints are met. The medical field contains numerous validated causal relationships, such as "bacterial infection leads to elevated white blood cell count" and "lung inflammation causes changes in chest imaging." These causal constraints serve as validation mechanisms in the reverse causation process, calculating a confidence score for each precursor concept node combination. The calculation of the confidence score comprehensively considers multiple factors, including the statistical correlation between the symptom combination and the diagnosis, the strength of recommendations in clinical guidelines, and the degree of matching with the current patient characteristics. In practice, for precursor node combinations… Its confidence score This can be represented as a weighted sum, which is a linear combination of three components: statistical confidence, guideline support, and patient fit. Statistical confidence is calculated from the conditional probability in the historical case database, guideline support is extracted from the recommendation level in the standardized clinical pathway knowledge base, and patient fit is obtained by comparing the cosine similarity between the current patient feature vector and the typical patient profile of this combination.

[0107] When selecting precursor concept node combinations with confidence scores exceeding a preset threshold, the confidence threshold setting needs to balance sensitivity and specificity. In emergency scenarios, to avoid missing critical causes, the confidence threshold can be set to a lower value, such as 0.3, allowing more causal pathways to be retained. In scenarios such as elective surgery planning, the confidence threshold can be increased to above 0.7 to ensure the reliability of the causal results. The set of precursor concept node combinations obtained after threshold screening constitutes the set of reverse causal pathways. Each path in this set starts from a symptom node, passes through a diagnosis node, and ultimately points to the target treatment node.

[0108] The bidirectional matching and fusion stage requires integrating the forward inference path set and the reverse abduction path set. Forward inference provides a causal reasoning chain from symptoms to treatment, while reverse abduction provides the necessary preconditions for treatment decisions. The intersection of the two represents a logically consistent and causally complete treatment path. When calculating the semantic consistency between paths, a similarity measure is performed using the vector representation of the path node sequence. Specifically, the forward paths... and reverse path Each path vector is encoded separately. Path vectors can be generated by average pooling of all node vectors in the path or by using a sequence encoder. Semantic consistency scores are obtained by calculating the cosine similarity between two path vectors. Path pairs with high similarity scores indicate that they describe similar diagnostic and treatment logic in the semantic space.

[0109] The calculation of logical coherence focuses on the integrity of the causal chain between nodes within the path. For a matching forward-backward path pair, it is necessary to verify whether the transitions between adjacent nodes in the path conform to medical logic rules.

[0110] In one optional implementation, the tensor operation operators are applied to the tensor representation using a differentiable logic reasoning engine to perform logical conjunction, disjunction, and implied differentiable computations. Multi-step reasoning propagation is performed from the initial symptom node along directed edges to generate a set of forward inference paths, including:

[0111] When performing differentiable computation, the differentiable logic reasoning engine transforms logical conjunction, disjunction, and implication operations into their corresponding tensor operation forms, and constructs reasoning state tensors based on the tensor representations of nodes in the structured medical concept graph.

[0112] During multi-step inference propagation, the inference state tensor is initialized from the initial symptom node. In each inference step, a downstream node to be activated is selected according to the directed edge topology of the structured medical concept graph. Tensor operation operators are applied to the downstream node. The updated activation value of the downstream node is calculated by combining the activation strength of its upstream node in the inference state tensor and the weight of the directed edge. The updated activation value is then written into the inference state tensor.

[0113] After completing a preset number of inference steps, the node sequence with activation intensity exceeding the activation threshold is extracted from the inference state tensor. The complete path trajectory of inference propagation is reconstructed based on the directed edge connection relationship between nodes and the historical record of the inference state tensor in each inference step, generating a forward inference path set.

[0114] In the reasoning process of a differentiable logic reasoning engine, the first step is to establish an isomorphic mapping relationship between symbolic logic and tensor operations. This involves transforming logical conjunction operations into element-wise minimization operations of tensors, i.e., for two nodes... and tensor representation and The tensor representation of its conjunction result is as follows The minimum value operation is performed independently in each dimension. Logical disjunction is transformed into element-wise maximum value operation of the tensor, i.e. For logical implication operations Using the implication definition of Łukasiewicz in fuzzy logic, it is transformed into ,in This represents a tensor where all elements are 1. These tensor operations are all highly differentiable and support backpropagation of gradients. In practical applications, to enhance numerical stability, smooth approximation functions are often used to replace discrete minimum and maximum operations; for example, the minimum operation is replaced with... The smoothing parameter The value is typically between 5 and 20. The larger the value of this parameter, the more accurate the approximation, but the numerical stability of the gradient will decrease.

[0115] The construction of the inference state tensor is based on the set of all nodes in a structured medical concept graph, assuming that the concept graph contains... There are n nodes, and the tensor representation of each node has a dimension of n. Then the shape of the inference state tensor is During initialization, the inference state tensor assigns zero vectors to all node positions except the initial symptom node, while the initial symptom node's position is filled with its tensor representation output from the neural coding layer. In multi-step inference propagation, the inference state tensor dynamically records the activation state of each node in the current inference step; its value reflects both the confidence level of node activation and the semantic information of the corresponding medical concept. To support path reconstruction, a three-dimensional history tensor needs to be maintained, with the shape... ,in This indicates the preset number of inference steps. This tensor stores a complete copy of the current inference state tensor at the end of each inference step.

[0116] The execution of multi-step inference propagation adopts an iterative update mechanism, in the first step... In each reasoning step, the first step is to determine the set of downstream nodes to be activated based on the adjacency matrix of the structured medical concept graph. Specifically, this involves traversing the tensor of the current reasoning state where the activation strength is greater than the propagation threshold. For all nodes in the concept graph, the threshold is typically set between 0.3 and 0.5. For each such node, all its successor nodes directly connected by directed edges in the concept graph are found, and these successor nodes are aggregated to form the activation set. For each downstream node in the activation set... This requires aggregating the contributions of all its upstream nodes. Let node... The set of upstream nodes is For upstream nodes Extracting connecting edges from the concept graph weight This weight reflects the medical concept To medical concept The causal strength or diagnostic correlation is determined by weights, which are typically extracted from medical knowledge bases or derived from clinical data statistics, ranging from 0 to 1. (Calculation node) When updating the activation value, first check each upstream node. Represent its current activation in the reasoning state tensor With edge weight Multiply to obtain a weighted representation Then, based on the logical type of the connected edges, the corresponding tensor operators are applied for aggregation.

[0117] When node When multiple upstream nodes have a disjunctive relationship, aggregation is performed using the maximum value operation, i.e. This applies to scenarios where a diagnosis can be derived from any of multiple symptoms. When multiple upstream nodes are in a conjunction relationship, the minimum value operation is used, i.e. This corresponds to medical inferences that require multiple conditions to be met simultaneously. In actual concept graphs, the logical relationships between nodes are usually labeled as edge attributes during the graph construction phase, and the inference engine selects the corresponding aggregation method based on these attributes. After aggregation, the resulting... With nodes In the activation representation of the previous step To achieve fusion, a weighted summation method is used. , where the fusion coefficient Typically set to 0.6 to 0.8, this design ensures that new information is absorbed during inference while retaining historical activation states, avoiding drastic oscillations in activation signals. Updated The state is written back to the corresponding position in the inference state tensor, completing the update of this node in the current step.

[0118] Preset number of reasoning steps The determination of the solution requires comprehensive consideration of the concept graph's depth and computational resource limitations. In common disease diagnosis and treatment scenarios, the reasoning chain from symptoms to diagnosis to treatment plan typically consists of 3 to 6 steps. The number of inference steps is typically set between 5 and 10. Too few inference steps result in insufficient reasoning, failing to cover the deeper diagnostic logic, while too many steps introduce computational redundancy and cause activation signals to propagate repeatedly in the graph, creating circular dependencies. After completing all inference steps, activation strength exceeding the activation threshold is extracted from the final inference state tensor. The threshold for the nodes is typically set between 0.4 and 0.6, with the specific value adjustable according to the sensitivity requirements of different clinical scenarios. The extracted nodes are arranged in descending order of activation intensity, forming an activated node sequence.

[0119] Path trajectory reconstruction is based on the historical record of the inference state tensor, for each node in the sequence of activated nodes. Tracing back to the reasoning steps that were first activated in historical records. That is, satisfying and The minimum step index, where This represents the L2 norm of a vector. By comparing the order of the first activation times between nodes and combining this with the topological structure of the directed edges in the concept graph, the inference dependencies between nodes can be identified. Specifically, if a node... In the steps First activation, node In the steps First activation, and Meanwhile, the concept graph contains directed edges. Then the node is considered Activation is performed by the node This is derived from the following logic: starting from the initial symptom node, a complete reasoning path is constructed along directed edges and the order of activation times. Since the concept graph has a branching structure, a single starting node can derive multiple parallel paths. Therefore, the reconstruction process employs a depth-first search strategy, traversing all path combinations.

[0120] In one optional implementation, each path in the multi-branch treatment path hypothesis space is transformed into a formal representation of a temporal constraint satisfaction problem, and the feasible region is pruned using a constraint propagation algorithm to obtain a candidate path set including:

[0121] During the formal representation transformation, the sequence of medical intervention nodes contained in each path in the multi-branch treatment path hypothesis space is extracted, each node in the medical intervention node sequence is mapped to a time-series constraint variable, and a constraint domain is constructed for each time-series constraint variable. At the same time, a constraint relationship network is established based on the causal relationship between nodes.

[0122] When performing feasible domain pruning, the constraint propagation algorithm performs bidirectional propagation in the constraint relationship network. When the constraint domain of a certain time-series constraint variable shrinks, the corresponding shrinkage effect is propagated to the successor time-series constraint variable along the constraint relationship network to update the lower bound of the constraint domain. At the same time, the corresponding shrinkage effect is propagated back to the predecessor time-series constraint variable to update the upper bound of the constraint domain, forming a cascaded shrinkage process of the constraint domain.

[0123] During the cascading contraction process, the constraint domain status of all time-series constraint variables is monitored. When it is detected that the constraint domain of a time-series constraint variable in a certain path has shrunk to an empty set, the corresponding path is marked as an infeasible path and eliminated. The paths with non-empty constraint domains are retained to form a candidate path set.

[0124] like Figure 2 As shown, the method includes:

[0125] Before transforming the hypothesis space of multi-branch treatment pathways into a formal representation, it is necessary to understand the structural characteristics of the pathway hypothesis space. The hypothesis space of multi-branch treatment pathways is essentially a set of multiple parallel paths, each representing a treatment plan. These paths consist of a series of medical intervention nodes, such as medication nodes, examination nodes, surgical nodes, and monitoring nodes. Each node not only contains intervention type information but also is associated with constraints such as time windows, resource requirements, and preconditions.

[0126] The formal representation transformation process begins with path extraction, which involves extracting any path from the hypothesis space of multi-branch treatment pathways. The path contains a sequence of medical intervention nodes arranged in chronological order. The extraction process needs to retain complete attribute information of the nodes, including node type, execution time, duration, and resource consumption. For example, a sepsis treatment pathway may include nodes for antibiotic administration, fluid resuscitation, hemodynamic monitoring, and organ function assessment.

[0127] When mapping medical intervention nodes to time-series constraint variables, it is necessary to do so for each node. Construct the corresponding time-series constraint variables The core attribute of a time-series constraint variable is its value range, that is, the time interval within which the node can be scheduled for execution. (This refers to a time-series constraint variable.) Constructed constraint domain The initial state is usually set to a relatively wide time interval. ,in This indicates the earliest time that the node can be executed. This indicates the latest possible execution time. The boundaries of the constraint domain are influenced by various factors, including the time window requirements of the patient's condition, the availability of medical resources, and the completion time of preliminary treatments.

[0128] Establishing a constraint network requires identifying various types of constraints between nodes. Sequential constraints require certain nodes to be executed before or after other nodes; for example, blood culture sampling must be completed before antibiotic administration. Time interval constraints stipulate that a specific time distance must be maintained between two nodes; for example, the dosing interval for certain drugs must be greater than [a certain value]. Hours. Resource mutual exclusion constraints indicate that certain nodes cannot execute simultaneously because they are competing for the same medical resource; for example, only one patient can be served in the same operating room at a time. These constraints are represented in the network as directed edges between nodes, with the constraint type and constraint parameters labeled on the edges.

[0129] The constraint propagation algorithm's execution mechanism is based on the principle of local consistency of constraints. When a certain temporal constraint variable... Constraint domain When contraction occurs, this change affects neighboring variables that have constraints on it. The process of propagating the contraction effect to subsequent variables can be understood through a specific scenario: suppose a node... and nodes There are temporal constraints between them, requiring... Must At least an interval after completion Only then can execution begin. If The constraint domain is from Shrink to ,So The earliest completion time from Postponed to ,lead to The earliest start time must also be postponed accordingly. Therefore The lower bound of the constraint domain needs to be updated to This enables backpropagation of the lower bound of the constraint domain.

[0130] The backpropagation process deals with the reverse constraints of the predecessor node. Continuing the example above, if... The shrinking of the upper bound of the constraint domain means It must be completed earlier, which in turn requires It must also be completed earlier to meet the time interval requirements. Specifically, if The constraint domain is from Shrink to ,but The latest start time from early , and by reverse deduction The latest completion time cannot exceed ,therefore The upper bound of the constraint domain needs to be updated to This enables backpropagation of the upper bound of the constraint domain.

[0131] Cascaded contraction exemplifies the global effect of constraint propagation. Contraction of the constraint domain of one variable triggers updates to the constraint domains of adjacent variables, which in turn influence their neighbors, creating a domino-like chain reaction. This process propagates bidirectionally along the edges of the constraint network until the constraint domains of all variables reach a stable state, meaning no further contraction occurs. The depth and breadth of cascaded contraction depend on the topology of the constraint network and the tightness of the constraints.

[0132] When monitoring the state of the constraint domain, a real-time detection mechanism needs to be established for the path. Each time-series constraint variable in In each iteration step of constraint propagation, its constraint domain is checked. The validity of the constraint domain. The criterion for shrinking the constraint domain to an empty set is that the lower bound of the constraint domain exceeds the upper bound, i.e. This situation indicates that the node cannot find a feasible execution time and path while satisfying all constraints. There are inherent unsatisfiable constraints. Empty set detection not only targets individual variables but also checks the compatibility between pairs of variables. For example, if two nodes must satisfy a sequential relationship, and the earliest completion time of the preceding node is later than the latest start time of the following node, it also constitutes an infeasibility condition.

[0133] The mechanism for marking and removing infeasible paths needs to accurately record the reasons for failure. When a path is deemed infeasible, the system records the key constraint conflicts that led to the failure, such as which nodes' constraint domains first shrank to an empty set and which constraint relationships became contradictory. This information is valuable for guiding subsequent optimization of path generation strategies. The elimination operation removes the path from the multi-branch treatment path hypothesis space, preventing it from participating in subsequent evaluation and selection processes.

[0134] The candidate path set is constructed according to the non-empty constraint domain criterion. After global processing by the constraint propagation algorithm, paths whose constraint domains for all temporal constraint variables remain non-empty are retained. This candidate path set possesses an important theoretical property: each path in the set has at least one scheduling scheme that allows all medical intervention nodes to be executed while satisfying the constraints. The degree of constraint domain contraction reflects the path's flexibility; a narrower constraint domain indicates fewer alternative scheduling schemes and a weaker robustness to temporal perturbations.

[0135] In practical applications, the combined application of formal representation transformation and feasible region pruning can significantly reduce the number of paths that need to be evaluated. For example, when dealing with the treatment pathways for patients with acute myocardial infarction, the initial hypothesis space contains dozens of paths, involving various combinations of thrombolytic therapy, interventional surgery, and drug therapy. By transforming these paths into a temporal constraint satisfaction problem and introducing constraints such as catheterization lab availability time, drug contraindication time windows, and patient arrival time, the constraint propagation algorithm can automatically identify those paths with fundamental time conflicts, such as operations that need to be completed before the patient arrives or schemes that exceed the longest waiting time in the catheterization lab. These infeasible paths are eliminated in advance, allowing the subsequent deep reinforcement learning module to optimize and select from a set of high-quality candidate paths, greatly improving computational efficiency and decision quality.

[0136] The iteration termination conditions of the constraint propagation algorithm include two cases: first, the constraint domains of all variables no longer change, reaching a stable state; second, an empty constraint domain is detected in a path, which is marked as infeasible. The algorithm uses a queue mechanism to manage the variables to be processed. Initially, all variables are added to the queue. Each time, a variable is taken from the queue for constraint propagation. If propagation causes a change in the constraint domain of adjacent variables, these adjacent variables are added back to the queue. This process continues until the queue is empty, indicating that the entire constraint network has reached a locally consistent state.

[0137] In one optional implementation, a neural symbol attribution map is constructed. By tracing back the generation process of each path in the candidate path set, the mapping relationship between the activated feature subspace in the neural network and the rule subset triggered in symbolic reasoning is extracted, including:

[0138] During the reverse tracing, each path in the candidate path set is traversed in reverse along its generation link. In the neural coding layer, the set of neurons that contribute to the path is located through gradient backpropagation. The activated feature subspace is delineated according to the coordinate distribution of the set of neurons in the latent space. In the symbolic abstraction layer, the sequence of logical rules called during the reasoning process of the path is traced back. The core rules are selected to form a rule subset based on the frequency of rule calls and the reasoning depth.

[0139] When extracting mapping relationships, a cross-layer association matrix is ​​established, using the feature dimensions in the feature subspace as row indices and the logical rules in the rule subset as column indices. By calculating the degree of matching between the activation mode of the feature dimensions and the preconditions of the logical rules, association strength values ​​are assigned to the matrix elements.

[0140] Based on the cross-layer association matrix, a neural symbol attribution graph is constructed, and the feature dimension and logical rules are represented as two types of nodes respectively. Directed edges are established between corresponding nodes according to the association strength value to complete the extraction of mapping relationships.

[0141] After obtaining the candidate path set, it is necessary to delve into the collaborative mechanism between the neural network and the symbolic reasoning system during path generation. For each specific path in the candidate path set, a reverse tracing mechanism is initiated. This mechanism starts from the endpoint decision node of the path and traverses it backward along its generation link. During the traversal, each decision node stores pointers to its predecessor nodes and intermediate state information when generating the decision. Through this information, the complete reasoning process can be reconstructed.

[0142] In the analysis phase of the neural coding layer, gradient backpropagation is used to locate the set of neurons that substantially contribute to the current path. Specifically, the decision output at the path's endpoint is used as the gradient source, and the gradient value of this output relative to the parameters of each layer in the network is calculated. Neurons with larger absolute gradient values ​​are considered to have a significant impact on the generation of the path. To avoid interference from gradient noise, a gradient threshold is set, and only neurons with absolute gradient values ​​exceeding this threshold are retained. These selected neurons have specific coordinate positions in the multidimensional latent space. By collecting these coordinate positions, a region can be delineated in the latent space; this region is the activated feature subspace. The boundary of the feature subspace is determined by calculating the convex hull of the neuron coordinates or by using density clustering methods, ensuring that the region covers all key neurons while excluding irrelevant regions.

[0143] During the backtracking process at the symbolic abstraction layer, the focus is on tracing the sequence of logical rules invoked during the reasoning process of that path. Each logical rule leaves a record in the inference log when it is invoked, including the rule's identifier, trigger time, matching status of preconditions, and inference conclusion. By analyzing these log records, the complete sequence of rule invocations can be reconstructed. For the same rule, if it is invoked multiple times during inference, its invocation frequency is accumulated. Inference depth reflects the hierarchical position of a rule in the inference chain; rules located at key inference nodes typically have higher importance. Considering both invocation frequency and inference depth, an importance score is calculated for each rule. Rules with high invocation frequency and high inference depth are assigned higher scores. A score threshold is set, and rules with scores exceeding this threshold are selected; these rules constitute a subset of rules. The size of this subset is usually much smaller than the complete rule base, but it contains the core rules that play a decisive role in the generation of the current path.

[0144] After completing the independent analysis of the neural coding layer and the symbolic abstraction layer, it is necessary to establish the correlation between the two layers. A cross-layer correlation matrix is ​​constructed as a bridge connecting the two layers. The number of rows in this matrix equals the number of feature dimensions in the feature subspace, and the number of columns equals the number of logical rules in the rule subset. Each element of the matrix represents the correlation strength between a specific feature dimension and a specific logical rule. To calculate the correlation strength, it is first necessary to analyze the activation patterns of the feature dimensions. The activation pattern of a feature dimension refers to the distribution characteristics of its activation values ​​when processing different input samples. This can be characterized by statistically analyzing the activation values ​​of the dimension on multiple samples and extracting its mean, variance, and activation value range. The preconditions of logical rules are usually expressed as constraints on clinical indicators, such as a biochemical indicator needing to be within a specific numerical range, or multiple vital signs needing to simultaneously meet specific conditions.

[0145] When matching activation patterns of feature dimensions with the preconditions of logical rules, a similarity-based method is used. For numerical conditions, the overlap between the activation value range of the feature dimension and the numerical range required by the rule preconditions is calculated. The overlap can be quantified by calculating the ratio of the intersection length to the union length of the two intervals. For categorical conditions, it is checked whether the activation patterns of the feature dimension can distinguish the different categories required by the rule preconditions, and the matching degree is evaluated by calculating the ratio of inter-class separation to intra-class tightness. For composite conditions, i.e., where the rule preconditions contain multiple sub-conditions connected by logical operators, the matching degree of each sub-condition is calculated separately, and then they are combined according to the type of logical operator. A high overall matching degree is given only when the matching degree of all sub-conditions is high, while for an OR operation, a high matching degree of any one sub-condition is sufficient.

[0146] Through the above calculations, each element of the cross-layer association matrix is ​​assigned an association strength value. This value is typically normalized to the range of zero to one; the closer the value is to one, the stronger the association between the corresponding feature dimension and the logical rule. After filling the matrix, rich cross-layer association information can be read from it. A row of the matrix represents the association between a specific feature dimension and all rules. By analyzing the numerical distribution of that row, the set of rules most relevant to that feature dimension can be identified. A column of the matrix represents the association between a specific logical rule and all feature dimensions. By analyzing this column, the neural feature basis supporting the rule can be discovered.

[0147] A neural symbolic attribution graph is constructed based on a cross-layer association matrix. This graph adopts a bipartite graph structure, containing two types of nodes. The first type of nodes represents feature dimensions in the feature subspace, and the second type of nodes represents logical rules in the rule subset. The number of feature dimension nodes is equal to the number of rows in the matrix, and the number of logical rule nodes is equal to the number of columns in the matrix. Directed edges are established between the two types of nodes to represent their associations. A directed edge from a feature dimension node to a logical rule node indicates that the feature dimension provides neural-level support for the logical rule.

[0148] The rules for establishing directed edges are based on the association strength values ​​in the cross-layer association matrix. A threshold for association strength is set, and a directed edge is established between the corresponding feature dimension node and logical rule node only when the association strength value of a matrix element exceeds this threshold. Setting the threshold requires a trade-off between the sparsity and information integrity of the graph. A threshold that is too high will lead to an overly sparse graph that loses important associations, while a threshold that is too low will introduce a large number of weak associations, increasing the difficulty of analysis. The weight of the directed edge is directly set to the association strength value. The thickness or color depth of the edge can be used to visualize the magnitude of the weight; edges with larger weights are more prominent in the graph.

[0149] In neural symbolic attribution maps, additional attribute information can be added to nodes. Feature dimension nodes can be labeled with the original clinical data type corresponding to that dimension, such as whether the dimension primarily encodes biochemical indicators or imaging features. Logical rule nodes can be labeled with the type of rule, such as diagnostic rules, treatment selection rules, or risk assessment rules. This attribute information helps to understand the structure of the attribution map at a higher level. By analyzing the topological characteristics of the map, key feature dimension nodes and logical rule nodes can be identified. Feature dimension nodes with high out-degree indicate that the feature dimension is associated with multiple logical rules and has a wide influence in the reasoning process. Logical rule nodes with high in-degree indicate that the rule is supported by multiple feature dimensions, and its triggering has a stronger neural evidence basis.

[0150] The construction of the neural symbolic attribution map completes the extraction of mapping relationships, establishing a traceable correspondence between the abstract high-dimensional features in the neural network and the explicit logical rules in symbolic reasoning. This mapping relationship provides the necessary basic data structure for subsequent dual-source contribution decomposition and also provides an intuitive visualization tool for understanding and verifying the generation mechanism of the treatment path.

[0151] A second aspect of this invention provides a treatment pathway planning system based on deep reinforcement learning, comprising:

[0152] The representation construction unit is used to acquire the patient's multimodal clinical data stream and dynamic medical resource status, and construct a two-layer representation space, wherein the neural coding layer converts the multimodal clinical data stream into continuous vector representations, and the symbolic abstraction layer decodes the continuous vector representations into a structured medical concept graph.

[0153] The path deduction unit is used to perform forward deduction and backward causation on the structured medical concept map based on a differentiable logic reasoning engine to generate a multi-branch treatment path hypothesis space. The differentiable logic reasoning engine transforms the logical operations of symbolic rules into gradient-propagable tensor operations, establishes a path-constraint mapping relationship, converts each path in the multi-branch treatment path hypothesis space into a formal representation of a temporal constraint satisfaction problem, and performs feasible domain pruning through a constraint propagation algorithm to obtain a candidate path set.

[0154] The attribution graph unit is used to construct a neural symbol attribution graph. By tracing back the generation process of each path in the candidate path set, it extracts the mapping relationship between the activated feature subspace in the neural network and the rule subset triggered in symbolic reasoning. Based on the mapping relationship, it calculates the dual-source contribution decomposition of each decision node, generates a hierarchical interpretable report based on the dual-source contribution decomposition, and outputs an optimized treatment path scheme.

[0155] A third aspect of the present invention provides an electronic device, comprising:

[0156] processor;

[0157] Memory used to store processor-executable instructions;

[0158] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0159] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0160] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0161] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A treatment pathway planning method based on deep reinforcement learning, characterized in that, include: The system acquires patients' multimodal clinical data streams and dynamic medical resource status, and constructs a two-layer representation space. The neural coding layer converts the multimodal clinical data streams into continuous vector representations, and the symbolic abstraction layer decodes the continuous vector representations into structured medical concept graphs. Based on a differentiable logic reasoning engine, the structured medical concept graph is subjected to forward deduction and backward causation to generate a multi-branch treatment path hypothesis space. The differentiable logic reasoning engine transforms the logical operations of symbolic rules into gradient-propagable tensor operations to establish a path-constraint mapping relationship. Each path in the multi-branch treatment path hypothesis space is transformed into a formal representation of a temporal constraint satisfaction problem. The feasible region is then pruned using a constraint propagation algorithm to obtain a set of candidate paths. A neural symbol attribution map is constructed. By tracing the generation process of each path in the candidate path set in reverse, the mapping relationship between the activated feature subspace in the neural network and the rule subset triggered in symbolic reasoning is extracted. Based on the mapping relationship, the dual-source contribution decomposition of each decision node is calculated. A hierarchical interpretable report is generated based on the dual-source contribution decomposition, and an optimized treatment path scheme is output.

2. The method according to claim 1, characterized in that, A two-layer representation space is constructed, wherein the neural coding layer converts the multimodal clinical data stream into continuous vector representations, and the symbolic abstraction layer decodes the continuous vector representations into structured medical concept graphs, including: In the neural coding layer, multi-scale feature extraction is performed on the multimodal clinical data stream through hierarchical semantic embedding to generate a multi-level vector representation containing local detail features and global context features. Based on the variational inference framework, the multi-level vector representation is compressed into a continuous vector representation that follows a preset prior distribution. In the symbolic abstraction layer, a continuous-discrete transformation mechanism is established. The dimensional intervals in the continuous vector representation are mapped to a set of candidate medical concept nodes through a probabilistic decoder. The candidate medical concept node set is then subjected to topological inference based on a graph structure generation model. The graph structure generation model identifies causal relationships and temporal dependencies between medical concept nodes by analyzing the covariance structure and gradient flow of the continuous vector representation in the latent space. Based on this, a structured medical concept graph containing node semantic attributes and edge relationship types is constructed.

3. The method according to claim 2, characterized in that, In the symbolic abstraction layer, a continuous-discrete transformation mechanism is established. A probabilistic decoder maps the dimensional intervals in the continuous vector representation to a set of candidate medical concept nodes. Based on a graph structure generation model, topological inference is performed on the set of candidate medical concept nodes, including: When establishing the continuous-discrete transformation mechanism, the semantic segmentation hyperplane is identified in the latent space of the continuous vector representation by the variational boundary localization method. The continuous vector representation is divided into multiple dimensional intervals according to the semantic segmentation hyperplane, where each dimensional interval corresponds to a concept category in a predefined medical ontology. The probability decoder is used to perform Bayesian decoding on each dimensional interval to generate a set of candidate medical concept nodes containing concept identifiers and posterior probabilities. The candidate medical concept node set is modeled based on a graph structure generation model. Each node in the candidate medical concept node set is mapped to a relation embedding space through latent relation embedding. The affinity matrix between node pairs is calculated in the relation embedding space, and edge generation decisions are made based on the numerical distribution in the affinity matrix. At the same time, the edge generation decisions are weighted with confidence based on the posterior probability of the nodes in the candidate medical concept node set, forming a topology structure with probability labeling. The topology is optimized by iteratively updating node representations and edge weights in the topology using a message passing algorithm, so that the topology converges to a stable state that satisfies the medical ontology constraints, thus completing the topology inference.

4. The method according to claim 1, characterized in that, Based on a differentiable logic reasoning engine, the structured medical concept graph is subjected to forward deduction and backward causation to generate a multi-branch treatment path hypothesis space. The differentiable logic reasoning engine establishes path-constraint mapping relationships by transforming the logical operations of symbolic rules into gradient-propagable tensor operations. During forward inference, the concept nodes and edge relationships in the structured medical concept graph are encoded as tensor representations, and the predefined medical diagnosis and treatment symbol rules are transformed into tensor operation operators. The tensor operation operators are applied to the tensor representations through a differentiable logic reasoning engine to perform logical conjunction, disjunction, and implication differentiable computations. Multi-step reasoning propagation is carried out from the initial symptom node along the directed edges to generate a set of forward inference paths. When performing reverse causation, starting from the target treatment node in the structured medical concept graph, the differentiable logic reasoning engine performs reverse tensor operations along the reverse direction of the edge to identify the combination of precursor concept nodes that can lead to the target treatment node. For each combination of precursor concept nodes, the confidence score of satisfying medical causal constraints is calculated, and the combination of precursor concept nodes with confidence scores exceeding a preset confidence threshold is selected as the set of reverse causation paths. The forward inference path set and the reverse causal path set are bidirectionally matched and fused. By calculating the semantic consistency and logical coherence between the paths, a path-constraint mapping relationship is constructed.

5. The method according to claim 4, characterized in that, By applying the tensor operators to the tensor representation through a differentiable logic reasoning engine, performing logical conjunction, disjunction, and implication differentiable computations, and conducting multi-step reasoning propagation along directed edges from the initial symptom node, a set of forward inference paths is generated, including: When performing differentiable computation, the differentiable logic reasoning engine transforms logical conjunction, disjunction, and implication operations into their corresponding tensor operation forms, and constructs reasoning state tensors based on the tensor representations of nodes in the structured medical concept graph. During multi-step inference propagation, the inference state tensor is initialized from the initial symptom node. In each inference step, a downstream node to be activated is selected according to the directed edge topology of the structured medical concept graph. Tensor operation operators are applied to the downstream node. The updated activation value of the downstream node is calculated by combining the activation strength of its upstream node in the inference state tensor and the weight of the directed edge. The updated activation value is then written into the inference state tensor. After completing a preset number of inference steps, the node sequence with activation intensity exceeding the activation threshold is extracted from the inference state tensor. The complete path trajectory of inference propagation is reconstructed based on the directed edge connection relationship between nodes and the historical record of the inference state tensor in each inference step, generating a forward inference path set.

6. The method according to claim 1, characterized in that, Each path in the multi-branch treatment path hypothesis space is transformed into a formal representation of a temporal constraint satisfaction problem, and the feasible region is pruned using a constraint propagation algorithm to obtain a candidate path set including: During the formal representation transformation, the sequence of medical intervention nodes contained in each path in the multi-branch treatment path hypothesis space is extracted, each node in the medical intervention node sequence is mapped to a time-series constraint variable, and a constraint domain is constructed for each time-series constraint variable. At the same time, a constraint relationship network is established based on the causal relationship between nodes. When performing feasible domain pruning, the constraint propagation algorithm performs bidirectional propagation in the constraint relationship network. When the constraint domain of a certain time-series constraint variable shrinks, the corresponding shrinkage effect is propagated to the successor time-series constraint variable along the constraint relationship network to update the lower bound of the constraint domain. At the same time, the corresponding shrinkage effect is propagated back to the predecessor time-series constraint variable to update the upper bound of the constraint domain, forming a cascaded shrinkage process of the constraint domain. During the cascading contraction process, the constraint domain status of all time-series constraint variables is monitored. When it is detected that the constraint domain of a time-series constraint variable in a certain path has shrunk to an empty set, the corresponding path is marked as an infeasible path and eliminated. The paths with non-empty constraint domains are retained to form a candidate path set.

7. The method according to claim 1, characterized in that, Constructing a neural symbolic attribution map involves tracing the generation process of each path in the candidate path set backwards, and extracting the mapping relationship between the activated feature subspace in the neural network and the subset of rules triggered in symbolic reasoning. During the reverse tracing, each path in the candidate path set is traversed in reverse along its generation link. In the neural coding layer, the set of neurons that contribute to the path is located through gradient backpropagation. The activated feature subspace is delineated according to the coordinate distribution of the set of neurons in the latent space. In the symbolic abstraction layer, the sequence of logical rules called during the reasoning process of the path is traced back. The core rules are selected to form a rule subset based on the frequency of rule calls and the reasoning depth. When extracting mapping relationships, a cross-layer association matrix is ​​established, using the feature dimensions in the feature subspace as row indices and the logical rules in the rule subset as column indices. By calculating the degree of matching between the activation mode of the feature dimensions and the preconditions of the logical rules, association strength values ​​are assigned to the matrix elements. Based on the cross-layer association matrix, a neural symbol attribution graph is constructed, and the feature dimension and logical rules are represented as two types of nodes respectively. Directed edges are established between corresponding nodes according to the association strength value to complete the extraction of mapping relationships.

8. A treatment pathway planning system based on deep reinforcement learning, used to implement the method as described in any one of claims 1-7, characterized in that, include: The representation construction unit is used to acquire the patient's multimodal clinical data stream and dynamic medical resource status, and construct a two-layer representation space, wherein the neural coding layer converts the multimodal clinical data stream into continuous vector representations, and the symbolic abstraction layer decodes the continuous vector representations into a structured medical concept graph. The path deduction unit is used to perform forward deduction and backward causation on the structured medical concept map based on a differentiable logic reasoning engine to generate a multi-branch treatment path hypothesis space. The differentiable logic reasoning engine transforms the logical operations of symbolic rules into gradient-propagable tensor operations, establishes a path-constraint mapping relationship, converts each path in the multi-branch treatment path hypothesis space into a formal representation of a temporal constraint satisfaction problem, and performs feasible domain pruning through a constraint propagation algorithm to obtain a candidate path set. The attribution graph unit is used to construct a neural symbol attribution graph. By tracing back the generation process of each path in the candidate path set, it extracts the mapping relationship between the activated feature subspace in the neural network and the rule subset triggered in symbolic reasoning. Based on the mapping relationship, it calculates the dual-source contribution decomposition of each decision node, generates a hierarchical interpretable report based on the dual-source contribution decomposition, and outputs an optimized treatment path scheme.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images