A Clinical Decision Support System Based on Neural Symbol Fusion of Ancient Chinese Medical Texts

Abstract

This invention provides a clinical auxiliary decision-making system based on ancient Chinese medicine texts using neural symbol fusion. It belongs to the fields of natural language processing, knowledge graphs, artificial intelligence, and traditional Chinese medicine informatics. The system includes a multi-dimensional underlying knowledge base and a clinical semantic alignment module, a neural symbol dual-track reasoning architecture, and a neural symbol dual-track fusion and intelligent addition / subtraction module. It extracts knowledge from ancient texts, textbooks, and guidelines to construct a multi-source knowledge graph, converting medical record texts into standardized entity sequences. It employs a rule-based reasoning model and a heterogeneous graph neural network model in parallel to output reasoning results. Through heterogeneous spatial semantic alignment, it adaptively adjusts the dual-track fusion weights based on the case rule coverage and complexity. Through a three-level conflict safety detection, it extracts residual symptoms to complete intelligent drug addition / subtraction, outputting an interpretable auxiliary decision-making report with multi-source contribution attribution. This invention enables precise auxiliary decision-making in traditional Chinese medicine, considering both primary disease and primary prescription, as well as secondary diseases and secondary medications.

Description

Technical Field

[0001] This invention belongs to the fields of natural language processing, knowledge graph, artificial intelligence and traditional Chinese medicine informatics, and in particular refers to a clinical auxiliary decision-making system based on neural symbol fusion of ancient Chinese medicine books. Background Technology

[0002] Traditional Chinese medicine classics contain rich wisdom on syndrome differentiation and treatment; however, the clinical translation of knowledge from ancient Chinese medical texts and the construction of intelligent clinical decision support systems currently face three core challenges: First, the heterogeneity of underlying knowledge and the semantic gap between ancient and modern times lead to sparse feature representations at the underlying level of the knowledge base. There are significant temporal and individual differences in the expression methods of ancient Chinese medicine texts, modern textbooks, and real-world clinical medical records. Ancient texts are mostly unstructured natural language descriptions, and there are numerous synonyms and synonyms between highly condensed terminology and colloquial clinical symptom descriptions. Existing research often relies on simple string matching or shallow similarity calculations, which easily leads to semantic drift and matching failures when dealing with long-tail clinical terms. This severely affects the model's effective absorption and utilization of ancient knowledge, resulting in an inability of the underlying knowledge base to effectively support front-end clinical input.

[0003] Second, single data-driven models suffer from logical black boxes and gaps in dialectical logic, resulting in insufficient interpretability. Currently widely used deep learning or large language models are essentially end-to-end mappings based on statistical probability. This architecture reduces the complex process of dialectical treatment to a flattened multi-label classification task. While it can achieve a direct mapping from symptoms to prescriptions, it lacks explicit modeling of the TCM diagnostic logic of "cause, mechanism, symptom, and treatment." This not only easily creates illusions that contradict medical common sense but also severely disrupts the interconnected reasoning chain of TCM. This opaque reasoning process leads to a lack of interpretability in rigorous medical scenarios, making it difficult for clinicians to understand and trust the basis for their decisions, thus hindering the clinical application of the system.

[0004] Third, existing systems lack the ability to make dynamic inferences about complex chronic diseases and are insufficient in processing incomplete information. Real-world clinical reasoning is far from the static classification under ideal conditions. Chronic diseases such as coronary heart disease often have complex pathogenesis, a large number of symptom clusters, and present a complex state with both primary and secondary symptoms. Furthermore, clinical practice often encounters situations where information is incomplete or fragmented, such as during initial diagnosis, emergency care, or remote consultations. Existing research mostly models based on the assumption of complete information, using fixed feature inputs and a single output pattern, lacking the ability to analyze complex pathogenesis and recommend combined treatments. When faced with incomplete data and atypical cases, the accuracy and robustness are extremely poor, and it cannot simulate the flexible dynamic reasoning of human experts. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides a clinical auxiliary decision-making system based on neural symbol fusion of ancient Chinese medicine texts. The system constructs a reasoning architecture of "semantic alignment-logical serialization-decision parallelization". This architecture maximizes the search space of neural computation while retaining rule constraints, including: a multi-dimensional underlying knowledge base and clinical semantic alignment module, a neural symbol dual-track reasoning architecture, and a neural symbol dual-track fusion and intelligent addition and subtraction module. The multi-dimensional underlying knowledge base and clinical semantic alignment module are used to extract knowledge from ancient books, modern textbooks and guidelines, construct a multi-source knowledge graph, and transform natural language clinical medical record text into a standardized entity sequence. The neural symbolic dual-track reasoning architecture includes a rule-based reasoning model based on strong logical constraints that runs in parallel on both tracks and a heterogeneous graph neural network model based on Bert-Base-Chinese and HeteroGNN, which respectively outputs symbolic-side reasoning decision results and neural-side reasoning decision results. The neural symbol dual-track fusion and intelligent addition / subtraction module is used to first perform semantic alignment and shared perception between the heterogeneous spaces of the symbol side and the neural side, and then perform an asymmetric reasoning process. Under the premise of clear treatment constraints, dual-track parallel reasoning is performed, and the rule coverage and case complexity of the current input case are calculated through the metacognitive control hub. The fusion weight of the dual-track decision is adaptively adjusted. Finally, before the final decision is made, a three-level conflict safety detection is performed, and the uncovered residual symptoms are extracted to drive the intelligent prescription drug addition / subtraction calculation, so as to achieve precise treatment of "main disease main prescription, secondary disease secondary drug". Finally, a full-link interpretable auxiliary decision report including multi-source contribution attribution is generated.

[0006] The beneficial effects of the technical solution provided by this invention include at least the following: This invention breaks through the black-box limitations of traditional end-to-end models, deconstructing complex dialectical treatment thinking into computable feature perception, core reasoning, decision mapping, and logical quality control levels. Simultaneously, by combining pre-trained language models and heterogeneous graph neural networks, it exhibits good robustness when facing incomplete real-world data and atypical complex chronic diseases, thereby establishing a high-confidence, traceable auxiliary diagnosis and treatment mechanism. This provides clinicians with highly accurate and interpretable auxiliary decision-making solutions. Specifically: 1. This invention connects multiple sources of ancient books and modern medical knowledge through clinical semantic alignment, realizing the transformation of medical record natural language into standardized knowledge entities, and solving the problems of heterogeneous TCM knowledge and unstructured medical records that are difficult to calculate directly.

[0007] 2. The present invention employs a dual-track neural symbol reasoning method, which balances the strong logical constraints of rules with the ability of graph neural networks to mine complex cases, thereby expanding the model search space while preserving the rules of TCM diagnosis and treatment.

[0008] 3. This invention is based on metacognitive adaptive adjustment of dual-track decision weights to adapt to cases of varying complexity, and three-level conflict safety detection reduces the risk of diagnosis and treatment decisions.

[0009] 4. This invention enables intelligent addition and subtraction of prescriptions based on residual symptoms, implementing the main prescription for the primary disease and the additional medications for secondary diseases; it outputs an interpretable report with contribution attribution, and the entire reasoning process is traceable, improving the credibility and practicality of TCM AI diagnosis and treatment. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a block diagram of a clinical auxiliary decision-making system based on neural symbol fusion from ancient Chinese medical texts, provided in an embodiment of the present invention. Figure 2 This is an overall architecture diagram of the auxiliary decision-making model based on neural symbol dual-track fusion provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of dynamic policy routing (nine modes) based on the coverage-complexity matrix provided in the embodiments of the present invention. Detailed Implementation

[0012] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0013] The specific implementation of this invention mainly relies on a computer system to execute the above-mentioned technical solution. The following is an example of coronary heart disease.

[0014] like Figure 1 and 2 As shown, this embodiment of the invention provides a clinical auxiliary decision-making system based on neural symbol fusion of ancient Chinese medical texts. The system includes: The system constructs a "semantic alignment-logical serial-decision parallel" reasoning architecture, which maximizes the search space of neural computing while preserving rule constraints. This architecture includes: a multi-dimensional underlying knowledge base and clinical semantic alignment module, a neural symbol dual-track reasoning architecture, and a neural symbol dual-track fusion and intelligent addition and subtraction module. The multi-dimensional underlying knowledge base and clinical semantic alignment module are used to extract knowledge from ancient books, modern textbooks and guidelines, construct a multi-source knowledge graph, and transform natural language clinical medical record text into a standardized entity sequence. Optionally, in embodiments of the present invention, after performing complex negation logic parsing and semantic noise reduction, three-level cascade matching and conflict resolution on natural language clinical medical record text, it is transformed into a normalized entity sequence.

[0015] The neural symbolic dual-track reasoning architecture includes a rule-based reasoning model based on strong logical constraints that runs in parallel on both tracks and a heterogeneous graph neural network model based on Bert-Base-Chinese and HeteroGNN, which respectively outputs symbolic-side reasoning decision results and neural-side reasoning decision results. Optionally, the rule-based reasoning model is used to deconstruct complex dialectical logic into multi-level deductions. After inputting the clinically inputted TCM four diagnostic information or the standardized entity sequence into the rule-based reasoning model, the pathogenesis is deduced based on the four-level symptom triggering and weighted fusion mechanism. The three heterogeneous reasoning channels of direct symptom mapping, pathogenesis dialectical deduction and treatment logic reasoning are executed in parallel. The conflict detection of cold and heat, deficiency and excess attributes and the safety verification of compatibility and incompatibility are performed through the logic quality control layer, and the symbolic side decision result is output. Optionally, the heterogeneous graph neural network model is used to input clinical case text into the Bert-Base-Chinese pre-trained model for context reconstruction, extract deep semantic dynamic feature representations and use them as dynamic feature vectors of symptom nodes in the HeteroGNN heterogeneous graph neural network, and simultaneously adopt a static encoding strategy based on knowledge graph embedding for pathogenesis, treatment, and prescription nodes, and perform attention-based heterogeneous message passing in the HeteroGNN heterogeneous graph neural network to complete the full-chain feature aggregation including implicit logical constraints, realize high-dimensional feature representation of complex and atypical symptoms, and output neural side decision results.

[0016] The neural symbol dual-track fusion and intelligent addition / subtraction module is used to first perform semantic alignment and shared perception between the heterogeneous spaces of the symbol side and the neural side, and then perform an asymmetric reasoning process. Under the premise of clear treatment constraints, dual-track parallel reasoning is performed, and the rule coverage and case complexity of the current input case are calculated through the metacognitive control hub. The fusion weight of the dual-track decision is adaptively adjusted. Finally, before the final decision is made, a three-level conflict safety detection is performed, and the uncovered residual symptoms are extracted to drive the intelligent prescription drug addition / subtraction calculation, so as to achieve precise treatment of "main disease main prescription, secondary disease secondary drug". Finally, a full-link interpretable auxiliary decision report including multi-source contribution attribution is generated (the auxiliary decision report finally output by the embodiment of the present invention is only used as a reference for doctors' clinical diagnosis and is not the final treatment plan).

[0017] Optionally, the neural symbol dual-track fusion and intelligent addition / subtraction module first performs semantic alignment and shared perception between the heterogeneous spaces on the symbol side and the neural side to ensure that dual-track reasoning is based on the same clinical fact benchmark, specifically including: Construct a feature-sharing layer to ensure that dual-track reasoning is based on the same clinical fact benchmark (addressing the heterogeneity problem between the symbolic and neural input spaces): Let the clinical input cases be , including a set of symptoms For rule-based reasoning models, symbolic encoding is performed, and the input is mapped to a high-dimensional sparse Boolean vector. ,in The symptom set dimension is defined for the rule base. This vector, after being processed by pre-negation words and normalized by synonyms, serves as the direct input for rule reasoning. For neural models, embedded encoding is used, and the input is mapped to a dense real-valued vector. Using pre-trained symptom embedding matrices ,Will Each symptom in the map is mapped to The dimensional vectors are weighted and aggregated to capture the potential semantic relationships between symptoms; Both reasoning tracks use a unified, standardized symptom dictionary to avoid semantic gaps caused by inconsistent terminology.

[0018] Optionally, the neural symbol dual-track fusion and intelligent addition / subtraction module performs an asymmetric reasoning process, conducting dual-track parallel reasoning under the premise of clear governance constraints, specifically including: In Traditional Chinese Medicine (TCM) decision-making, treatment methods serve as the crucial link between pathogenesis and prescription. Treating method reasoning is treated as a preliminary sequential step, primarily relying on the logical judgments of rule-based models. This demonstrates the guiding role of TCM theoretical constraints on deep learning frameworks. Let the set of core solutions derived from the rule-based reasoning model be . Before entering the prescription recommendation stage, a treatment constraint filter is constructed. Its function is to force a reduction in weight or circuit breaking during the parallel phase if the prescription does not contain the currently derived core treatment, regardless of how confident the neural network model is in the effectiveness of a certain prescription. in For the full range of prescription space, Formula A collection of indications and therapeutic functions; Then, under the premise of clear legal constraints, conduct parallel reasoning along two tracks: Track A: Symbolic Logic Perform the preceding rule-based reasoning steps and output a set of candidate prescriptions based on rule matching. and its confidence level It has the advantage of high accuracy, but has a low recall rate in rare diseases or complex concurrent diseases; Track B: Neural Association Track Perform the forward propagation of the preceding heterogeneous graph neural network and output the probability distribution based on the fully connected layer. and its prediction probability Its characteristics include high recall, and the ability to discover non-linear implicit associations between symptoms and obscure remedies; Finally, the probability of heterogeneous confidence scores is calibrated and normalized: Will and Mapped to a unified space of evidence strength, so as to be used in subsequent stages for... and To achieve fusion, a temperature scaling function is employed to mitigate the overconfidence of the neural network in predicting probabilities far exceeding actual accuracy. A temperature parameter is introduced. : in The Logits output by the GNN Temperature hyperparameter (to be increased appropriately when a bias in the training data distribution is detected). , When the probability distribution of the output is smoother, it helps to improve the accuracy of confidence level calibration. The rule-based reasoning model outputs The values ​​are stepped, and a modified sigmoid function is used to smooth them out: in This is a hyperparameter used to adjust the steepness of the confidence curve; Finally, a unified evidence metric was implemented, and the dual-track outputs were uniformly represented as evidence pairs. : .

[0019] Optionally, the meta-cognition control hub in the neural symbol dual-track fusion and intelligent addition / subtraction module does not directly participate in diagnostic and treatment decisions. Instead, it is responsible for evaluating the interpretability of the input cases. By calculating indicators in two dimensions—rule coverage and case complexity—it perceives the relative position of the current case to the boundary of the rule base in real time and outputs a two-dimensional state vector. : in, Indicates rule coverage. Indicates the level of case complexity; This vector does not directly generate a prescription, but rather serves as a control signal input to the next step, dynamically determining the weights of the reasoning rules and the neural network in the final decision. .

[0020] For example, for In this state, the engine will classify it as low complexity and assign it a very high rule weight; while for When a certain state is reached, the engine will determine it to be of high complexity, thereby activating the dominant exploration function of the neural network.

[0021] Optionally, the metacognitive control hub calculates the rule coverage of the current input case, specifically including: Rule coverage This invention aims to measure the knowability of clinical characteristics of input cases in the rule base, and adopts a coverage algorithm based on weighted information gain (due to the inherent contradiction between the diversity of clinical descriptions and the finiteness of the rule base, simply calculating the symptom matching rate is difficult to reflect the true rule coverage, therefore, this embodiment of the invention designs a coverage algorithm based on weighted information gain). In traditional Chinese medicine diagnosis and treatment, the differential diagnostic value contained in the chief complaint is far greater than that of general concurrent symptoms. Therefore, the input symptom set... Perform hierarchical dynamic weight allocation: set up , Highlighting the contribution of core symptoms to coverage calculation; First, iterate through the input symptom set, checking whether each symptom can be mapped to the rule base, and define an indicator function. If symptoms Existing in the rule base Alternatively, it can be mapped through a thesaurus. ,otherwise ; Rule coverage Defined as the ratio of effective matching information to total input information: in As a key missing penalty item, if the unmatched symptom is a high-frequency core symptom, even if the overall matching rate is acceptable, the model will apply a non-linear penalty to forcibly reduce the coverage score, in order to warn of potential blind spots in key information. when This indicates that the case characteristics fall entirely within the knowledge scope of the rule-based reasoning model, and the reasoning engine enters rule-based reasoning; when This indicates that the cases contain a large number of rare symptoms or non-standard descriptions not included in the rule base, and the inference engine needs to rely on the generalization ability of the neural model.

[0022] Optionally, the metacognitive control hub calculates the case complexity of the currently input case, specifically including: Case complexity This is used to assess the complexity of a patient's condition. A case with few symptoms and a simple pathogenesis is relatively simple, while a case with complex symptoms and mixed cold and heat patterns indicates a higher level of complexity. A complexity scoring model is constructed based on three dimensions: information entropy, pathogenesis span, and logical conflict. The complexity score is... The calculation is as follows: (1) Feature space density score, used to measure the complexity of input information: Logarithmic smoothing was used in the calculation to prevent score overflow due to symptom accumulation. (2) Pathogenesis span score, used to measure the breadth of the disease location and nature involved. Based on the preliminary results of rule reasoning, the number of pathogenesis elements involved is counted. : For example, a case involving only "heart qi deficiency" has a relatively small scope, while a case involving the "heart, spleen, and kidney" organs and also having "phlegm, blood stasis, and fluid retention" as pathogenic factors, its It will increase significantly.

[0023] (3) Logical conflict factor score, which is the core of complexity assessment, is used to detect whether there are opposing features in the symptom set according to traditional Chinese medicine theory: (4) Complexity comprehensive score Final complexity score The weighted sum of the above items: (5) Set dual thresholds based on score distribution The cases were divided into three levels: Low complexity: The documents are clear and logically sound. Medium complexity: The symptoms are complex and slightly complicated. High complexity: It involves multiple organ dysfunctions or a mixture of cold and heat, and falls into the category of intractable diseases.

[0024] Optionally, the adaptive adjustment of the fusion weights of the dual-track decision-making in the neural symbol dual-track fusion and intelligent addition / subtraction module specifically includes: Construct a 3×3 dynamic fusion matrix (to achieve the optimal combination for neural symbolic dual-track inference), which is based on rule coverage. The horizontal axis represents the complexity of the cases. Using the vertical axis, the clinical diagnosis and treatment scenarios of traditional Chinese medicine are subdivided into nine modes, such as... Figure 3 As shown, it achieves a smooth transition from rigid reasoning to flexible reasoning, simulating the cognitive process of human experts flexibly adjusting the weight of experience, intuition, and logical analysis when facing cases of varying difficulty. According to the state vector Locate a specific unit in the matrix and activate the corresponding fusion strategy. Each strategy corresponds to a set of dynamic weight parameters. Used to adjust the results of rule-based reasoning With neural reasoning results Contribution to the final decision: Among them, the normalization constraint is satisfied. ; These nine patterns were clustered into three decision-making functional areas: (1) Rule-dominated areas with a focus on certainty Coverage area: The rule-dominated region is characterized by high rule coverage regardless of case complexity, representing textbook-style typical cases. In this case, the rule-based reasoning model can accurately capture the pathogenesis logic. The introduction of a neural model may introduce noise due to overgeneralization. This rule-dominated region includes the following three modes: Mode 1 (HL): High coverage + low complexity The input cases exhibit typical single-symptom presentations. The reasoning strategy is rule-exclusive, retaining only the neural model as background noise to prevent extremely rare logical loopholes in the rule base. Weights: ; Pattern 2 (HM): High Coverage + Medium Complexity The input case has a clear primary symptom and some secondary symptoms. The reasoning strategy is rule-based, and the rule-based reasoning model establishes the primary prescription. The neural model is only used to assist in verifying the rationality of medication for secondary symptoms. Weights: ; Mode 3 (HH): High coverage + high complexity The input cases are mostly multi-organ combined diseases, but the symptoms of each dimension are clearly defined in the rule base. The reasoning strategy is to prioritize rule verification. Due to the complexity of the diseases, the neural model may discover complex pathogenesis combinations that are not defined by the rules. Therefore, it is appropriate to enhance the rule base. Values ​​are introduced from a data-driven perspective, with weights: ; (2) Hybrid enhancement region with complementary advantages Coverage area: This region represents atypical or partially missing cases, with moderate rule coverage. The rules can only explain some clinical phenomena, and there are cognitive blind spots. The strategies tend to be synergistic, utilizing the generalization ability of neural models to fill the gaps in the rules. This mixed enhancement region contains the following three modes: Mode 4 (ML) / Mode 5 (MM): Medium coverage + medium to low complexity The input cases are characterized by non-standard symptom descriptions or clear chief complaints but vague supporting evidence. The reasoning strategy is weighted consensus, where the two tracks are in a relatively balanced state. The engine tends to select the intersection of the two or the union of high-confidence recommendations. If the two recommendations conflict, a subsequent arbitration mechanism is triggered. Weights: ; Mode 6 (MH): Medium Coverage + High Complexity The input cases present with mixed symptoms of chills and fever, and the symptom descriptions do not completely match the classic provisions. The reasoning strategy is a dynamic game, which is the most uncertain state. At this time, the reasoning result depends entirely on the subsequent competitive arbitration of the strength of evidence. Whoever has the stronger explanatory power wins. Weight: ; (3) Generalization-preferred neural complement areas Coverage area: The low rule coverage in this area represents complex or rare diseases encountered in clinical practice. The rule model fails due to knowledge boundary limitations, and the strategy tends to rely on the similarity retrieval ability of the neural model in the latent space, including the following three modes: Mode 7 (LL) / Mode 8 (LM): Low coverage + low to medium complexity The input cases are characterized by a large amount of modern medical descriptions, a lack of traditional Chinese medicine terminology, or a low correspondence with traditional Chinese medicine content. The reasoning strategy is neural-assisted. In this case, the rule model is downgraded to a safety filter, only responsible for eliminating options that obviously contradict common sense. The main recommendation power is given to the GNN, with weights as follows: ; Pattern 9 (LH): Low coverage + high complexity The input case is an extremely complex, unknown medical record where the rule base cannot match any valid path. The reasoning strategy is neural exploration. In this case, the GNN model generates exploratory solutions. To prevent hallucinations, solutions generated in this mode are marked with a "low confidence warning," prompting the user to manually review them. Weights: ; Finally, a sigmoid smoothing function is introduced to adaptively smooth the policy weights, thus avoiding the influence of... Tiny fluctuations cause the strategy to jump drastically between different modes: Let the target weight be The actual weight being executed at the current moment Defined as: This smoothing mechanism ensures that the inference engine maintains the continuity and stability of decision-making logic when transitioning from "typical proofs" to "difficult proofs," demonstrating the robustness of the architecture.

[0025] By constructing the above nine fusion models, a smooth transition from rigid reasoning to flexible reasoning was successfully achieved, simulating the cognitive process by which human experts flexibly adjust the weight of experience, intuition, and logical analysis when facing cases of varying difficulty.

[0026] During parallel reasoning on both sides, the symbolic track and the neural track often produce inconsistent Top-1 prescription recommendations due to the fundamental differences in their reasoning mechanisms. This conflict could stem from blind spots in the rule base or from statistical illusions in neural networks.

[0027] Optionally, the neural symbol dual-track fusion and intelligent addition / subtraction module performs a three-level conflict safety detection, specifically including: The three-tiered conflict arbitration protocol is designed to ensure the security and logical consistency of the final decision, with a strict priority order: Tier 1 arbitration: secure veto > Tier 2 arbitration: legal verification > Tier 3 arbitration: evidentiary debate; Among them, the primary arbitration, as a hard constraint of the engine, has a "veto power" based on security constraints. It is responsible for blocking all recommendations that violate medical common sense or medication contraindications (the primary principle of medical decision-making is "non-harm"). Regardless of how high the model's predicted probability is, the inference engine calls the "compatibility contraindication library" and "critical illness protocol module" built by the rule-based inference model to evaluate the prescriptions recommended by the neural pathway. Perform a full component scan and define a safety constraint function. : in The patient's safety sensitivity feature vector includes a set of clinical features that may trigger drug contraindications, such as pregnancy status, special physical conditions (e.g., bleeding tendency), allergy history, and special physiological stages. like The engine directly triggers the circuit breaker mechanism, at which point the neural weights are irrelevant. If you set a certain value, the recommendation scores of the neural network will all return to zero: At this point, the inference engine forcibly reverts to the rule recommendation result. And generate a "safety alert report"; Secondary arbitration: If the first-level security check is passed, the inference engine enters the second-level arbitration, which is a consistency check of treatment methods based on vector space. The recommended prescription must match the core treatment method (this stage mainly addresses the issue of mismatch between prescription and symptoms). First, a vectorized mapping of treatment methods is performed, which represents the core set of treatment methods determined by the rule-based reasoning model. and GNN recommended prescriptions Indications and uses Mapped to the same semantic vector space respectively: Then, the consistency coefficient is calculated, and the cosine similarity between the two is used as the consistency penalty coefficient. : in These are scaling and translation parameters used to... The cosine value is mapped to a reasonable penalty interval. ,like This indicates that although the prescriptions recommended by the GNN are statistically relevant, they deviate from the pathogenesis and treatment principles in terms of treatment direction. In this case, the inference engine applies a soft penalty to the neural scores. This mechanism ensures that even when data-driven approaches are introduced, the decision-making of the inference engine remains guided by the logic of dialectical consistency. Third-level arbitration: When candidate prescriptions are both safe and in line with the general direction of treatment, but the specific prescriptions still differ, the reasoning engine enters a three-level arbitration based on dynamic game theory of evidence strength—evidence competition: Based on the strategy weights determined above Calculate the final fusion score of the dual-track candidate prescriptions. : At this stage, the inference engine will include... and All candidate prescriptions were uniformly ranked; Introducing the Advantageous Range To avoid decision-making turbulence caused by slight differences in scores: When rule-based models have significant advantages: When, adopt ; When neural models have significant advantages: When, adopt (This usually occurs when rule coverage is low and neural confidence is extremely high). When the strength of evidence for both is roughly equal: At that time, the engine executes a conservative strategy, prioritizing the adoption of [certain methods / approaches]. And mark it as the baseline scheme, while Presented as an alternative, the user is advised that "either option is acceptable, option A is the classic option, while option B, suggested by data mining, may be more effective for concurrent symptoms X".

[0028] Through a three-level arbitration mechanism, the inference engine not only solves the problem of neural symbol conflict, but also realizes the decision-making wisdom of traditional Chinese medicine at the mathematical level: "treat the symptoms in acute cases and the root cause in chronic cases."

[0029] Traditional Chinese medicine clinical diagnosis and treatment is highly individualized. In order to achieve progress from general prescriptions to precise prescriptions and to solve the long-tailed distribution bias that may be generated by neural models, this invention constructs an intelligent addition and subtraction module based on pathogenesis elements and a clinically feasible replacement mechanism for obscure prescriptions.

[0030] Optionally, the neural symbol dual-track fusion and intelligent addition / subtraction module extracts uncovered residual symptoms to drive intelligent prescription drug addition / subtraction calculations, specifically including: (1) Intelligent addition and subtraction based on comorbidity-driven The principal side was determined through dual-track reasoning. Afterward, the engine does not directly output the result, but intelligently adds or subtracts from the prescription, identifying and addressing "residual symptoms" not covered by the core efficacy of the main prescription, including: (1.1) Extraction of residual feature set: First, calculate the input symptom set. Collection of symptoms treated by the main prescription The difference set is defined as the residual symptom set. : in, As symptom weights, To add or subtract trigger thresholds; This step ensures that medication adjustments are made only for concomitant symptoms of significant clinical importance, avoiding prescription clutter caused by irrelevant or trivial symptoms: (1.2) Pathogenesis-Medicinal Flavor Mapping Reasoning against For each element, the engine invokes the built-in "Microscopic Pathogenesis - Herbal Flavor" rule module. This rule module is built based on textbooks, guidelines, and the experience of renowned veteran TCM doctors, and includes rules such as... The triplet rule, the set of addition and subtraction suggestions The generation logic is as follows: If Furthermore, the main prescription lacks sedatives, and the rules for invoking [the medication / method] are not applicable. Based on this, suggestions for adding flavors are generated; this example belongs to the category of concurrent symptoms mapping. If the probability of "Qi deficiency" predicted by GNN is... When the main force's boost is insufficient, the engine triggers. The suggestion is that this case falls under the category of pathogenesis reinforcement. (1.3) Conflict elimination and final solution generation Generated addition and subtraction suggestions After two rounds of security checks, options that are incompatible with the main prescription's ingredients are eliminated, resulting in the final personalized decision support solution. Represented as: This mechanism enables precise treatment based on "primary prescription for primary disease and secondary medication for secondary disease," significantly improving the engine's clinical adaptability to complex concurrent syndromes.

[0031] Optionally, the neural symbol dual-track fusion and intelligent addition / subtraction module generates a full-link interpretable auxiliary decision-making report that includes multi-source contribution attribution, specifically including: This paper constructs an explanation generator independent of the inference module (to address the trust issues caused by the black-box nature of deep learning models in the medical field, thus endowing the inference engine with clinical interpretability). Based on the structured metadata retained during the inference process, it constructs end-to-end interpretable evidence by tracing back the complete decision-making path of "symptom-pathogenesis-treatment-prescription," generating end-to-end interpretable auxiliary decision-making reports. This not only achieves "knowing what" but also "knowing why," meeting the traceability requirements of medical scenarios, including: (1) Structured backtracking of heterogeneous reasoning paths During the dual-track parallel reasoning process, the engine maintains a dynamic evidence graph. Regardless of whether the final decision is driven by symbolic rules or a neural network, all activated intermediate nodes are recorded. A complete chain of evidence is defined as an ordered sequence of quadruplets. in A subset of symptoms that make a substantial contribution to reasoning; The core pathogenesis elements derived from this; The established treatment strategy; This is the final recommended prescription; After inference, the computation graph is traversed in reverse to extract the aforementioned key nodes, and nodes with weights below the threshold are filtered out. Noise nodes are eliminated to ensure the conciseness of the explanation. (2) Contribution-based multi-source attribution algorithm Design a multi-source contribution attribution algorithm to quantify the reasoning process. This algorithm integrates the matching weights of a rule engine and the attention weights of a graph neural network for a given pathogenesis. ,symptom Contribution The calculation is as follows: in For the rule hit indicator function; The attention coefficients from symptom nodes to pathogenesis nodes in HeteroGNN; The rule weights determined above ; Through this algorithm, the engine can clearly point out in the generated report: "The patient's chest tightness contribution of 0.42 and dark tongue contribution of 0.35 are the main basis for judging the syndrome of qi stagnation and blood stasis; (3) Transparent interpretation of conflict arbitration When a conflict arises in the dual-track reasoning, the reasoning behind the final output is explained to the user, and the explanation template is dynamically invoked based on the arbitration result of the three-level conflict arbitration mechanism. Safety veto scenario: "Although the neural model recommends prescription A with a confidence level of 0.88, the patient's bleeding tendency was detected, and the prescription contains leeches and other blood-breaking drugs, which triggered the first-level safety circuit breaker. Therefore, the rule-recommended prescription B was finally adopted." Evidence-based game scenario: "The rule recommends prescription A, the neurology recommends prescription B. After comprehensive evaluation, prescription B has a coverage of 85% for concurrent insomnia, which is significantly higher than prescription A's 40%, and it conforms to the core treatment principle. Therefore, prescription B is recommended first." This mechanism transforms the complex algorithmic game process into natural language logic that doctors can understand, enhancing the transparency of human-machine collaboration.

[0032] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A traditional Chinese medicine ancient book clinical auxiliary decision-making system based on neural symbol fusion, characterized in that, The system constructs a "semantic alignment-logical serial-decision parallel" reasoning architecture, which maximizes the search space of neural computing while preserving rule constraints. This architecture includes: a multi-dimensional underlying knowledge base and clinical semantic alignment module, a neural symbol dual-track reasoning architecture, and a neural symbol dual-track fusion and intelligent addition and subtraction module. The multi-dimensional underlying knowledge base and clinical semantic alignment module are used to extract knowledge from ancient books, modern textbooks and guidelines, construct a multi-source knowledge graph, and transform natural language clinical medical record text into a standardized entity sequence. The neural symbolic dual-track reasoning architecture includes a rule-based reasoning model based on strong logical constraints that runs in parallel on both tracks and a heterogeneous graph neural network model based on Bert-Base-Chinese and HeteroGNN, which respectively outputs symbolic-side reasoning decision results and neural-side reasoning decision results. The neural symbol dual-track fusion and intelligent addition / subtraction module is used to first perform semantic alignment and shared perception between the heterogeneous spaces of the symbol side and the neural side, and then perform an asymmetric reasoning process. Under the premise of clear treatment constraints, dual-track parallel reasoning is performed, and the rule coverage and case complexity of the current input case are calculated through the metacognitive control hub. The fusion weight of the dual-track decision is adaptively adjusted. Finally, before the final decision is made, a three-level conflict safety detection is performed, and the uncovered residual symptoms are extracted to drive the intelligent prescription drug addition / subtraction calculation, so as to achieve precise treatment of "main disease main prescription, secondary disease secondary drug". Finally, a full-link interpretable auxiliary decision report including multi-source contribution attribution is generated.

2. The system according to claim 1, characterized in that, The neural symbol dual-track fusion and intelligent addition / subtraction module first performs semantic alignment and shared perception between the heterogeneous spaces of the symbol side and the neural side to ensure that dual-track reasoning is based on the same clinical fact benchmark, specifically including: Construct a feature-sharing layer to ensure that dual-track reasoning is based on the same clinical fact benchmark: Let the clinical input cases be , including a set of symptoms For rule-based reasoning models, symbolic encoding is performed, and the input is mapped to a high-dimensional sparse Boolean vector. ,in The symptom set dimension is defined for the rule base. This vector, after being processed by pre-negation words and normalized by synonyms, serves as the direct input for rule reasoning. For neural models, embedded encoding is used, and the input is mapped to a dense real-valued vector. Using pre-trained symptom embedding matrices ,Will Each symptom in the map is mapped to The dimensional vectors are weighted and aggregated to capture the potential semantic relationships between symptoms; Both reasoning tracks use a unified, standardized symptom dictionary to avoid semantic gaps caused by inconsistent terminology.

3. The system according to claim 1, characterized in that, The neural symbol dual-track fusion and intelligent addition / subtraction module performs an asymmetric reasoning process. Under the premise of clear governance constraints, it conducts dual-track parallel reasoning, specifically including: In Traditional Chinese Medicine (TCM) decision-making, treatment methods serve as the crucial link between pathogenesis and prescription. Treating method reasoning is treated as a preliminary sequential step, primarily relying on the logical judgments of rule-based models. This demonstrates the guiding role of TCM theoretical constraints on deep learning frameworks. Let the set of core solutions derived from the rule-based reasoning model be . Before entering the prescription recommendation stage, a treatment constraint filter is constructed. Its function is to force a reduction in weight or circuit breaking during the parallel phase if the prescription does not contain the currently derived core treatment, regardless of how confident the neural network model is in the effectiveness of a certain prescription. in For the full range of prescription space, Formula A collection of indications and therapeutic functions; Then, under the premise of clear legal constraints, a dual-track parallel reasoning process is conducted: Track A: Symbolic Logic Perform the preceding rule-based reasoning steps and output a set of candidate prescriptions based on rule matching. and its confidence level It has the advantage of high accuracy, but has a low recall rate in rare diseases or complex concurrent diseases; Track B: Neural Association Track Perform the forward propagation of the preceding heterogeneous graph neural network and output the probability distribution based on the fully connected layer. and its prediction probability Its characteristics include high recall, and the ability to discover non-linear implicit associations between symptoms and obscure remedies; Finally, the probability of heterogeneous confidence scores is calibrated and normalized: Will and Mapped to a unified space of evidence strength, so as to be used in subsequent stages for... and To achieve fusion, a temperature scaling function is used to mitigate the overconfidence of the neural network's predicted probability, which is much higher than the actual accuracy. A temperature parameter is introduced. : in The Logits output by the GNN This refers to temperature hyperparameters. The rule-based reasoning model outputs The values ​​are stepped; a modified smoothing mapping function is used to smooth them out. in This is a hyperparameter used to adjust the steepness of the confidence curve; Finally, a unified evidence metric was implemented, and the dual-track outputs were uniformly represented as evidence pairs. : 。 4. The system according to claim 1, characterized in that, In the aforementioned neural symbol dual-track fusion and intelligent addition / subtraction module, the metacognitive control hub does not directly participate in diagnostic and treatment decisions. Instead, it is responsible for evaluating the interpretability of input cases. By calculating indicators in two dimensions—rule coverage and case complexity—it perceives the relative position of the current case to the rule base boundary in real time and outputs a two-dimensional state vector. : in, Indicates rule coverage. Indicates the level of case complexity; This vector does not directly generate a prescription, but rather serves as a control signal input to the next step, dynamically determining the weights of the reasoning rules and the neural network in the final decision. .

5. The system according to claim 4, characterized in that, The metacognitive control hub calculates the rule coverage of the current input case, specifically including: Rule coverage This study aims to measure the knowability of clinical characteristics of input cases in the rule base, employing a coverage algorithm based on weighted information gain. In traditional Chinese medicine diagnosis and treatment, the differential diagnostic value contained in the chief complaint is far greater than that of general concurrent symptoms. Therefore, the input symptom set... Perform hierarchical dynamic weight allocation: set up , Highlighting the contribution of core symptoms to coverage calculation; First, iterate through the input symptom set, checking whether each symptom can be mapped to the rule base, and define an indicator function. If symptoms Existing in the rule base Alternatively, it can be mapped through a thesaurus. ,otherwise ; Rule coverage Defined as the ratio of effective matching information to total input information: in As a key missing penalty item, if the unmatched symptom is a high-frequency core symptom, even if the overall matching rate is acceptable, the model will apply a non-linear penalty to forcibly reduce the coverage score, in order to warn of potential blind spots in key information. when This indicates that the case characteristics fall entirely within the knowledge scope of the rule-based reasoning model, and the reasoning engine enters rule-based reasoning; when This indicates that the cases contain a large number of rare symptoms or non-standard descriptions not included in the rule base, and the inference engine needs to rely on the generalization ability of the neural model.

6. The system according to claim 4, characterized in that, The metacognitive control hub calculates the case complexity of the currently input case, specifically including: Case complexity This is used to assess the complexity of a patient's condition. A case with few symptoms and a simple pathogenesis is relatively simple, while a case with complex symptoms and mixed cold and heat patterns indicates a higher level of complexity. A complexity scoring model is constructed based on three dimensions: information entropy, pathogenesis span, and logical conflict. The complexity score is... The calculation is as follows: (1) Feature space density score, used to measure the complexity of input information: Logarithmic smoothing was used in the calculation to prevent score overflow due to symptom accumulation. (2) Pathogenesis span score, used to measure the breadth of the disease location and nature involved. Based on the preliminary results of rule reasoning, the number of pathogenesis elements involved is counted. : (3) Logical conflict factor score, which is the core of complexity assessment, is used to detect whether there are opposing features in the symptom set according to traditional Chinese medicine theory: (4) Complexity comprehensive score Final complexity score The weighted sum of the above items: (5) Set dual thresholds based on score distribution The cases were divided into three levels: Low complexity: The documents are clear and logically sound. Medium complexity: The symptoms are complex and slightly complicated. High complexity: It involves multiple organ dysfunctions or a mixture of cold and heat, and falls into the category of intractable diseases.

7. The system according to claim 1, characterized in that, The adaptive adjustment of the fusion weights for dual-track decision-making in the neural symbol dual-track fusion and intelligent addition / subtraction module specifically includes: Construct a 3×3 dynamic fusion matrix, which has regular coverage. The horizontal axis represents the complexity of the cases. Using the vertical axis, the clinical diagnosis and treatment scenarios of traditional Chinese medicine are subdivided into nine modes, realizing a smooth transition from rigid reasoning to flexible reasoning, simulating the cognitive process of human experts flexibly adjusting the weight of experience, intuition and logical analysis when facing cases of different levels of difficulty. According to the state vector Locate a specific unit in the matrix and activate the corresponding fusion strategy. Each strategy corresponds to a set of dynamic weight parameters. Used to adjust the results of rule-based reasoning With neural reasoning results Contribution to the final decision: Among them, the normalization constraint is satisfied. ; These nine patterns were clustered into three decision-making functional areas: (1) Rule-dominated areas with a focus on certainty Coverage area: The rule-dominated region is characterized by high rule coverage regardless of case complexity, representing textbook-style typical cases. In this case, the rule-based reasoning model can accurately capture the pathogenesis logic. The introduction of a neural model may introduce noise due to overgeneralization. This rule-dominated region includes the following three modes: Mode 1: High coverage + low complexity The input cases exhibit typical single-symptom presentations. The reasoning strategy is rule-exclusive, retaining only the neural model as background noise to prevent extremely rare logical loopholes in the rule base. Weights: ; Pattern 2: High Coverage + Medium Complexity The input case has a clear primary symptom and some secondary symptoms. The reasoning strategy is rule-based, and the rule-based reasoning model establishes the primary prescription. The neural model is only used to assist in verifying the rationality of medication for secondary symptoms. Weights: ; Mode 3: High coverage + high complexity The input cases are mostly multi-organ combined diseases, but the symptoms of each dimension are clearly defined in the rule base. The reasoning strategy is to prioritize rule verification. Due to the complexity of the diseases, the neural model may discover complex pathogenesis combinations that are not defined by the rules. Therefore, it is appropriate to enhance the rule base. Values ​​are introduced from a data-driven perspective, with weights: ; (2) Hybrid enhancement region with complementary advantages Coverage area: This region represents atypical or partially missing cases, with moderate rule coverage. The rules can only explain some clinical phenomena, and there are cognitive blind spots. The strategies tend to be synergistic, utilizing the generalization ability of neural models to fill the gaps in the rules. This mixed enhancement region contains the following three modes: Mode 4 / Mode 5: Medium coverage + medium to low complexity The input cases are characterized by non-standard symptom descriptions or clear chief complaints but vague supporting evidence. The reasoning strategy is weighted consensus, where the two tracks are in a relatively balanced state. The engine tends to select the intersection of the two or the union of high-confidence recommendations. If the two recommendations conflict, a subsequent arbitration mechanism is triggered. Weights: ; Pattern 6: Medium Coverage + High Complexity The input cases present with mixed symptoms of chills and fever, and the symptom descriptions do not completely match the classic provisions. The reasoning strategy is a dynamic game, which is the most uncertain state. At this time, the reasoning result depends entirely on the subsequent competitive arbitration of the strength of evidence. Whoever has the stronger explanatory power wins. Weight: ; (3) Generalization-preferred neural complement areas Coverage area: The low rule coverage in this area represents complex or rare diseases encountered in clinical practice. The rule model fails due to knowledge boundary limitations, and the strategy tends to rely on the similarity retrieval ability of the neural model in the latent space, including the following three modes: Mode 7 / Mode 8: Low coverage + low to medium complexity The input cases are characterized by a large amount of modern medical descriptions, a lack of traditional Chinese medicine terminology, or a low correspondence with traditional Chinese medicine content. The reasoning strategy is neural-assisted. In this case, the rule model is downgraded to a safety filter, only responsible for eliminating options that obviously contradict common sense. The main recommendation power is given to the GNN, with weights as follows: ; Pattern 9: Low Coverage + High Complexity The input case is an extremely complex, unknown medical record where the rule base cannot match any valid path. The reasoning strategy is neural exploration. In this case, the GNN model generates exploratory solutions. To prevent illusions, solutions generated in this mode are marked with a "low confidence warning," prompting the user to manually review them. Weights: ; Finally, a sigmoid smoothing function is introduced to adaptively smooth the policy weights, thus avoiding the influence of... Tiny fluctuations cause the strategy to jump drastically between different modes: Let the target weight be The actual weight being executed at the current moment Defined as: This smoothing mechanism ensures that the inference engine maintains the continuity and stability of decision-making logic when transitioning from "typical proofs" to "difficult proofs," demonstrating the robustness of the architecture.

8. The system according to claim 1, characterized in that, The neural symbol dual-track fusion and intelligent addition / subtraction module performs a three-level conflict safety detection, specifically including: The three-tiered conflict arbitration protocol is designed to ensure the security and logical consistency of the final decision, with a strict priority order: Tier 1 arbitration: secure veto > Tier 2 arbitration: legal verification > Tier 3 arbitration: evidentiary debate; Among them, the first-level arbitration, as a hard constraint of the engine, has a "veto power" based on security constraints. It is responsible for blocking all recommendations that violate medical common sense or medication contraindications, regardless of how high the model's predicted probability is. The inference engine calls the "compatibility contraindication library" and "critical illness protocol module" built by the rule inference model to evaluate the prescriptions recommended by the neural pathway. Perform a full component scan and define a safety constraint function. : in The patient's safety sensitivity feature vector includes a set of clinical features that may trigger drug contraindications, such as pregnancy status, special physical condition, allergy history, and special physiological stage. like The engine directly triggers the circuit breaker mechanism, at which point the neural weights are irrelevant. If you set a certain value, the recommendation scores of the neural network will all return to zero: At this point, the inference engine forcibly reverts to the rule recommendation result. And generate a "safety alert report"; Secondary arbitration: If the first-level security check passes, the inference engine enters the second-level arbitration, which is a vector space-based consistency check of treatment methods. The recommended prescription must match the core treatment method. First, a vectorized mapping of treatment methods is performed, which represents the core set of treatment methods determined by the rule-based reasoning model. and GNN recommended prescriptions Indications and therapeutic effects Mapped to the same semantic vector space respectively: Then, the consistency coefficient is calculated, and the cosine similarity between the two is used as the consistency penalty coefficient. : in These are scaling and translation parameters used to... The cosine value is mapped to a reasonable penalty interval. ,like This indicates that although the prescriptions recommended by the GNN are statistically relevant, they deviate from the pathogenesis and treatment principles in terms of treatment direction. In this case, the inference engine applies a soft penalty to the neural scores. This mechanism ensures that even when data-driven approaches are introduced, the decision-making of the inference engine remains guided by the logic of dialectical consistency. Third-level arbitration: When candidate prescriptions are both safe and in line with the general direction of treatment, but the specific prescriptions still differ, the reasoning engine enters a three-level arbitration based on dynamic game theory of evidence strength—evidence competition: Based on the strategy weights determined above Calculate the final fusion score of the dual-track candidate prescriptions. : At this stage, the inference engine will include... and All candidate prescriptions were uniformly ranked; Introducing the Advantageous Range To avoid decision-making turbulence caused by slight differences in scores: When rule-based models have significant advantages: When, adopt ; When neural models have significant advantages: When, adopt ; When the strength of evidence for both is roughly equal: At that time, the engine executes a conservative strategy, prioritizing the adoption of [certain methods / approaches]. And mark it as the baseline scheme, while Presented as an alternative, the user is prompted that "both options are acceptable, with option A being the classic one, while data mining suggests that option B may be more effective for concurrent symptoms X".

9. The system according to claim 1, characterized in that, The neural symbol dual-track fusion and intelligent addition / subtraction module extracts uncovered residual symptoms to drive intelligent prescription drug addition / subtraction calculations, specifically including: (1) Intelligent addition and subtraction based on comorbidity-driven The principal side was determined through dual-track reasoning. Afterward, the engine does not directly output the result, but intelligently adds or subtracts from the prescription, identifying and addressing "residual symptoms" not covered by the core efficacy of the main prescription, including: (1.1) Extraction of residual feature set: First, calculate the input symptom set. Collection of symptoms treated by the main prescription The difference set is defined as the residual symptom set. : in, As symptom weights, To add or subtract trigger thresholds; This step ensures that medication adjustments are made only for concomitant symptoms of significant clinical importance, avoiding prescription clutter caused by irrelevant or trivial symptoms: (1.2) Pathogenesis-Medicinal Flavor Mapping Reasoning against For each element, the engine invokes the built-in "Microscopic Pathogenesis - Herbal Flavor" rule module. This rule module is built based on textbooks, guidelines, and the experience of renowned veteran TCM doctors, and includes rules such as... The triplet rule, the set of addition and subtraction suggestions The generation logic is as follows: If Furthermore, the main prescription lacks sedatives, and the rules for invoking [the medication / method] are not applicable. Based on this, suggestions for adding flavors are generated; this example belongs to the category of concurrent symptoms mapping. If the probability of "Qi deficiency" predicted by GNN is... When the main force's boost is insufficient, the engine triggers. The suggestion is that this case belongs to the category of pathogenesis reinforcement; (1.3) Conflict elimination and final solution generation Generated addition and subtraction suggestions After two rounds of security checks, options that are incompatible with the main prescription's ingredients are eliminated, resulting in the final personalized decision support solution. Represented as: This mechanism enables precise treatment based on "primary prescription for primary disease and secondary medication for secondary disease," significantly improving the engine's clinical adaptability to complex concurrent syndromes.

10. The system according to claim 1, characterized in that, The neural symbol dual-track fusion and intelligent addition / subtraction module generates a fully interpretable auxiliary decision-making report that includes multi-source contribution attribution, specifically including: An explanation generator, independent of the reasoning module, is constructed. Based on the structured metadata retained during the reasoning process, it builds end-to-end interpretable evidence by tracing back the complete decision-making path of "symptom-pathogenesis-treatment-prescription," generating an end-to-end interpretable auxiliary decision-making report. This not only achieves "knowing what" but also "knowing why," meeting the traceability requirements of medical scenarios, including: (1) Structured backtracking of heterogeneous reasoning paths During the dual-track parallel reasoning process, the engine maintains a dynamic evidence graph. Regardless of whether the final decision is dominated by symbolic rules or neural networks, all activated intermediate nodes are recorded. A complete chain of evidence is defined as an ordered sequence of quadruplets: in A subset of symptoms that make a substantial contribution to reasoning; The core pathogenesis elements derived from this; The established treatment strategy; This is the final recommended prescription; After inference, the computation graph is traversed in reverse to extract the aforementioned key nodes, and nodes with weights below the threshold are filtered out. Noise nodes are eliminated to ensure the conciseness of the explanation. (2) Contribution-based multi-source attribution algorithm Design a multi-source contribution attribution algorithm to quantify the reasoning process. This algorithm integrates the matching weights of a rule engine and the attention weights of a graph neural network for a given pathogenesis. ,symptom Contribution The calculation is as follows: in For the rule hit indicator function; The attention coefficients from symptom nodes to pathogenesis nodes in HeteroGNN; The rule weights determined above ; Through this algorithm, the engine can clearly point out in the generated report: "The patient's chest tightness contribution of 0.42 and dark tongue contribution of 0.35 are the main basis for judging it as Qi stagnation and blood stasis syndrome; (3) Transparent interpretation of conflict arbitration When a conflict arises in the dual-track reasoning, the reasoning behind the final output is explained to the user, and the explanation template is dynamically invoked based on the arbitration result of the three-level conflict arbitration mechanism. Safety veto scenario: "Although the neural model recommends prescription A with a confidence level of 0.88, the patient's bleeding tendency was detected, and the prescription contains blood-breaking drugs such as leeches, which triggered the first-level safety circuit breaker. Therefore, the rule-recommended prescription B was finally adopted." Evidence-based game scenario: "The rule recommends prescription A, the neurology recommends prescription B. After comprehensive evaluation, prescription B has a coverage of 85% for concurrent insomnia, which is significantly higher than prescription A's 40%, and it conforms to the core treatment principle. Therefore, prescription B is recommended first." This mechanism transforms the complex algorithmic game process into natural language logic that doctors can understand, enhancing the transparency of human-machine collaboration.

Images