An ontology and knowledge graph-based pulmonary embolism AI-assisted diagnosis intelligent agent construction method and system

Abstract

This invention discloses a method and system for constructing an AI-assisted diagnostic agent for pulmonary embolism based on ontology and knowledge graph, belonging to the fields of medical information systems and medical artificial intelligence technology. By integrating a pulmonary embolism-specific medical ontology, a clinically aligned knowledge graph, and thought chain reasoning, this invention leverages the hard constraints of ontology knowledge to achieve interpretability analysis of the entire process of AI-assisted diagnosis for pulmonary embolism, precise suppression of AI illusions specific to pulmonary embolism, end-to-end quality control, and closed-loop optimization. It is particularly suitable for rapid diagnosis of acute and critical pulmonary embolism, AI-assisted identification of pulmonary artery filling defects in CTPA images, thrombosis severity stratification, and emergency treatment decision tracing and regulatory compliance quality control scenarios. It can be adapted to the implementation, regulatory verification, and clinical optimization of various pulmonary embolism AI imaging diagnostic and emergency auxiliary diagnosis and treatment systems, specifically addressing core illusion problems in pulmonary embolism AI diagnosis such as false positive embolus identification and disordered risk stratification.

Description

Technical Field

[0001] This invention relates to the fields of medical information systems and medical artificial intelligence, specifically to a method and system for constructing an AI-assisted diagnostic agent for pulmonary embolism based on ontology and knowledge graph. Background Technology

[0002] Pulmonary embolism is a common acute and critical illness in clinical practice, characterized by rapid onset, high rates of missed or misdiagnosis, and high mortality. AI-assisted diagnostic technology has been widely applied to CTPA (Computed Tomography Pulmonary Angiography) image recognition, lesion localization, and risk stratification for pulmonary embolism. Relying on deep learning models, it can quickly identify the core diagnostic sign of pulmonary artery filling defects, significantly improving the efficiency of emergency diagnosis. However, existing AI-assisted diagnostic systems for pulmonary embolism have significant shortcomings, especially regarding the unique clinical and imaging characteristics of pulmonary embolism. The hallucination problem is prominent and difficult to manage, failing to meet the requirements of emergency clinical practice and regulatory compliance. Firstly, most models are purely data-driven black-box architectures, only outputting embolus location, risk level, and probability values. They can only provide superficial visualization through attention heatmaps and cannot generate a complete reasoning chain that aligns with the specific diagnostic logic of "imaging signs - clinical indicators - risk stratification - treatment plan" for pulmonary embolism. The decision-making basis is untraceable, failing to meet the mandatory requirements of the National Medical Products Administration and other institutions regarding the interpretability of AI medical devices, resulting in low acceptance among emergency clinicians. Secondly, the specific hallucination problem in pulmonary embolism CTPA images is frequent. The models are prone to generating false-positive embolism diagnoses and incorrect risk assessments in pulmonary embolism-specific imaging scenarios such as vascular flow artifacts, pulmonary artery anatomical variations, non-embolic filling defects, and interference from small emboli below the segmental level, which contradict clinical logic. First, the system suffers from several flaws. First, it exhibits poor risk stratification and misjudgments of high / low risk levels, lacking a specific mechanism for identifying and managing hallucinations in pulmonary embolism scenarios. This can easily mislead critical treatment decisions such as emergency thrombolysis and interventional procedures. Second, quality control is fragmented, focusing only on image quality and model accuracy, failing to cover the entire process from data input and model inference to decision output and clinical adaptability. This makes it unable to identify model drift, annotation bias, conflicts between decisions and pulmonary embolism treatment guidelines, and AI hallucinations in real time, and lacks a closed-loop feedback correction mechanism. Third, there is a severe disconnect between medical knowledge and the algorithm model. It fails to deeply integrate pulmonary embolism-specific pulmonary artery anatomy, emergency treatment guidelines, embolus identification rules, and expert experience. Relying solely on data fitting can easily lead to hallucination errors that contradict the logic of pulmonary embolism treatment. Finally, the knowledge utilization is limited, employing only medical ontology or knowledge graphs without incorporating thought chains to achieve step-by-step logical derivation that aligns with the pulmonary embolism treatment process, resulting in insufficient knowledge adaptability and inference accuracy.

[0003] In summary, existing AI-assisted diagnostic technologies for pulmonary embolism cannot address the unique imaging and clinical characteristics of pulmonary embolism, and therefore cannot achieve clinically interpretable reasoning, precise suppression of pulmonary embolism-specific AI hallucinations, closed-loop quality control throughout the entire process, or full traceability of diagnostic and treatment logic. There are significant technological gaps in the vertical AI diagnostic optimization scheme for acute and critical pulmonary embolism. Therefore, a complete intelligent agent system is urgently needed to solve the above pain points. Summary of the Invention

[0004] To overcome the above-mentioned technical problems, the present invention provides a method and system for constructing an AI-assisted diagnostic agent for pulmonary embolism based on ontology and knowledge graph.

[0005] To achieve the above technical solution, the present invention includes two aspects. The first aspect provides a method for constructing an AI-assisted diagnostic agent for pulmonary embolism based on ontology and knowledge graph, as detailed below: S1. Collect multimodal data for pulmonary embolism diagnosis from open-source databases and perform preprocessing operations; S2. Construct a medical ontology specifically for pulmonary embolism based on the preprocessed data; S3. Construct a pulmonary embolism knowledge graph based on a dedicated medical ontology framework for pulmonary embolism; S4. Based on the pulmonary embolism-specific medical ontology, construct an ontology-based algorithm for identifying the core signs of pulmonary artery filling defects, in order to identify whether pulmonary artery filling defects in pulmonary embolism CTPA images are true emboli and suppress the generation of false positive illusions. S5. Based on the signs and symptoms in S4, a stratified assessment algorithm for the severity of pulmonary embolism is constructed to stratify the severity of pulmonary embolism. S6, a pulmonary embolism-specific medical ontology based on S2, a pulmonary embolism knowledge graph based on S3, and algorithms based on S4 and S5 to generate a complete thought chain, and to strengthen the suppression of pulmonary embolism hallucinations through gradient penalty. S7. Based on S1~S6, conduct full-process quality control and AI hallucination control scoring, and use the scoring results to provide AI hallucination warnings and interceptions. S8. Closed-loop feedback correction based on AI illusion warning and interception results; The construction of the intelligent agent is completed by executing S1~S8.

[0006] Specifically, the pulmonary embolism-specific medical ontology is defined as follows: In the formula, A collection of medical concepts specifically for pulmonary embolism; A is the set of logical relationships between concepts; A is the set of concept attributes; T is the data type constraint. A specific logical axiom for the emergency diagnosis and treatment of pulmonary embolism.

[0007] Specifically, the pulmonary embolism knowledge graph is defined as follows: In the formula, For the set of entities related to pulmonary embolism; Relationships between entities; For entity attribute triples; Score the credibility of knowledge sources.

[0008] Specifically, the expression for the ontology-based algorithm for recognizing core signs of pulmonary artery filling defects is as follows: In the formula, The score indicates the true nature of the filling defect. If the score is higher than the threshold, it is judged as a true embolic sign of pulmonary embolism. If the score is lower than the threshold, it is judged as an artifact or normal vascular variation, directly intercepting false positive illusions. Indicates the morphological characteristics of filling defects; Indicates the density characteristics of filling defects; Indicates anatomical location matching features; and Represents the dynamic weights of each feature term; Indicates the artifact penalty coefficient; Confidence level for image artifacts.

[0009] Specifically, in S5, the expression for the stratified assessment algorithm for pulmonary embolism severity is as follows: In the formula, This indicates the comprehensive score for the severity of pulmonary embolism; Indicates the image feature score; Indicates clinical signs and symptoms score; Indicates the score of the test index; These represent the dynamic weights of each scoring item; The criteria for stratifying the severity of pulmonary embolism are as follows: In the formula, These represent the thresholds for the first, second, and third strata, respectively.

[0010] Specifically, in S6, the formula for the thought chain reasoning process is as follows: In the formula, This is a single-step reasoning node; This is the final diagnosis plus risk stratification conclusion. This serves as the basis for corresponding guideline clauses or case studies within the knowledge graph; To verify the results of single-step hallucination, each step of reasoning must be matched with the axioms of the pulmonary embolism-specific medical ontology and the facts of the pulmonary embolism knowledge graph, so as to eliminate hallucinations induced by unfounded reasoning. The expression for strengthening the inhibitory effect of pulmonary embolism hallucinations through gradient penalty is as follows: In the formula, Indicates the dynamic weight of a single-step inference node; Indicates the credibility of single-step reasoning; Indicates a single-step hallucination risk score; This represents the penalty coefficient for hallucination.

[0011] Specifically, in S7, the expression for the AI ​​illusion control score is as follows: Satisfy dynamic weight normalization constraints: In the formula, This indicates the quality control score of the input data; This represents the quality control score of the model inference; This indicates the quality control score of the decision-making results; AI Illusion Special Control Score; These represent the dynamic adaptive weights of each item; The rules for AI-based hallucination warning and interception are as follows: In the formula, The quality control threshold for pulmonary embolism diagnosis; The threshold for the risk of pulmonary embolism-specific hallucinations; when Immediately trigger hallucination warnings, intercept erroneous diagnostic results, and simultaneously initiate a closed-loop correction process to completely prevent hallucination diagnoses from entering clinical settings, ensuring the safety of emergency treatment. Results that pass quality control are then entered into the results output stage.

[0012] Specifically, in S8, the closed-loop feedback correction iteration process is as follows: In the formula, For the corrected model / knowledge; To revise the previous model / knowledge; The learning rate; The quality control scoring gradient for pulmonary embolism; Risk gradient for pulmonary embolism hallucinations.

[0013] On the other hand, the present invention provides a system for constructing an AI-assisted diagnostic intelligent agent for pulmonary embolism based on ontology and knowledge graph. The system is used to execute the method as described in claims 1 to 8. The system is characterized by comprising five core modules: a knowledge base construction module, a thought chain interpretable reasoning module, a full-process quality control and hallucination control module, a closed-loop feedback correction module, and a visual interaction module.

[0014] The knowledge foundation construction module is used to perform the following steps: S1. Multi-source data acquisition and preprocessing; S2. Construct a dedicated medical ontology for pulmonary embolism; S3. Construct a multi-source fusion pulmonary embolism knowledge graph; The thought chain interpretable reasoning module is used to perform the following steps: S4. Construct an algorithm for recognizing core signs of pulmonary artery filling defects based on ontology knowledge; S5. Construct a stratified evaluation algorithm for the severity of pulmonary embolism; S6, Mind Chain Reasoning and Credibility Scoring; The full-process quality control and hallucination control module is used to perform the following steps: S7, Full-process quality control and AI illusion control; The closed-loop feedback correction module is used to perform the following steps: S8, closed-loop feedback correction.

[0015] The beneficial effects of this invention are: This invention develops two core ontological constraints: pulmonary artery filling defect sign recognition and pulmonary embolism severity stratification. These algorithms fully align with the specific characteristics of pulmonary embolism, abandoning the purely data-driven approach. They achieve accurate identification of true emboli and compliant output of four-level risk stratification. Relying on hard ontological constraints, they suppress false positives and false negatives from the source, generating a traceable chain of evidence that fits the logic of emergency diagnosis and treatment, fully complying with regulatory compliance requirements.

[0016] This invention optimizes the credibility of the thought chain and the full-process quality control scoring mechanism, abandoning the conventional manual fixed weighting mode. It innovatively adopts a dynamic adaptive weight driven by the confidence of pulmonary embolism ontology knowledge plus nonlinear hallucination penalty. The weight is dynamically adjusted in real time according to the compliance of pulmonary embolism diagnosis and treatment and the risk of hallucination, rather than a simple static numerical superposition. This breaks through the innovation limitations of existing technologies. At the same time, through four means of ontology axiom verification, knowledge graph fact matching, step-by-step risk scoring, and closed-loop correction, it accurately identifies and intercepts pulmonary embolism-specific hallucinations, eliminating misdiagnosis from the source.

[0017] This invention establishes a closed-loop quality control and feedback correction system covering four core aspects: data, model, results, and hallucinations. It dynamically adjusts the weighting of quality control and specifically addresses industry pain points such as frequent hallucinations in pulmonary embolism AI, fragmented quality control, and lack of feedback correction. It is fully adapted to the stringent requirements of emergency diagnosis and treatment of pulmonary embolism.

[0018] This invention is the first to construct a dual-layer knowledge base for pulmonary embolism, consisting of a customized medical ontology and a guideline-aligned knowledge graph. The ontology bears the hard constraints of the diagnosis and treatment logic of pulmonary embolism, while the knowledge graph provides factual support. It blocks specific AI illusions such as CTPA artifacts, vascular variations, and disordered risk stratification in pulmonary embolism from both logical and factual perspectives, and completely solves the pain point of the disconnect between medical knowledge and algorithm models. Attached Figure Description

[0019] Figure 1 This is a flowchart of the process of the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to specific embodiments.

[0021] This invention aims to provide an AI-assisted diagnostic agent specifically designed for the clinical diagnosis and treatment of acute and critical pulmonary embolism, with ontology knowledge constraints as the core hallucination suppression mechanism. The entire process is customized to address the unique characteristics of pulmonary embolism, including its rapid onset, high risk of missed or misdiagnosed cases, susceptibility to artifacts in CTPA images, the critical need for timely emergency treatment, and stringent requirements for diagnostic accuracy and traceability. It constructs a dual-layer knowledge base: a medical ontology specific to the pulmonary embolism vertical domain and a knowledge graph strictly aligned with clinical guidelines. Using ontology knowledge as the core hard constraint, it incorporates two key algorithms: pulmonary artery filling defect recognition and embolism severity stratification. Combined with a step-by-step reasoning technology tailored to the logic of emergency pulmonary embolism diagnosis, it completely breaks through the existing black-box dilemma of AI diagnosis for pulmonary embolism. It specifically addresses false-positive emboli caused by CTPA image artifacts, misjudgment of vascular variations, and risk classification in pulmonary embolism. Addressing the pain points of specific AI hallucinations such as layer confusion and unfounded treatment recommendations, this approach leverages ontological knowledge constraints to block the source of hallucinations, verify the process, and correct them in a closed loop. This enables end-to-end interpretability and traceability of pulmonary embolism (PE) AI-assisted diagnosis, precise control of PE-specific hallucinations, full-link quality control in emergency care, and closed-loop correction and optimization for compliance with treatment guidelines. It fundamentally eliminates hallucinatory maldiagnoses that violate the clinical logic of PE, fully adapting to the stringent requirements of emergency thrombolysis and interventional treatment for PE. It meets the mandatory requirements of medical AI regulatory agencies such as the National Medical Products Administration for interpretability and quality control compliance, significantly improving the clinical credibility, emergency practicality, and regulatory compliance of PE AI-assisted diagnosis. This effectively reduces the rates of missed, misdiagnosed, and mistreated PE in acute and critical cases, providing reliable, verifiable, and traceable AI-assisted support for rapid and accurate emergency diagnosis.

[0022] See Figure 1 A method for constructing an AI-assisted diagnostic agent for pulmonary embolism based on ontology and knowledge graph includes the following steps: S1. Multi-source data acquisition and preprocessing; Data was collected from multiple data sources, including hospital PACS imaging systems, EMR electronic medical record systems, LIS laboratory systems, pulmonary embolism-specific clinical guideline databases, and emergency department expert experience databases. The core focus was on the essential needs of pulmonary embolism diagnosis, including lung CTPA imaging data, patient D-dimer specific indicators, emergency vital signs such as heart rate and blood pressure, anatomical structures of pulmonary artery branches at all levels, embolus location / morphology / range of involvement, pulmonary embolism risk stratification criteria, emergency quality control rules, and hallucination identification rules. The data cleaning process specifically removes duplicate, erroneous, and inconsistent data items. It also specifically labels and filters interfering data in pulmonary embolism CTPA images that are prone to causing specific AI hallucinations, such as vascular flow artifacts, anatomical variations, and non-embolic filling defects. Missing values ​​are filled using emergency medicine statistical methods, and CTPA images are standardized and preprocessed to reduce the probability of pulmonary embolism-specific AI hallucinations from the data source, thus ensuring the data quality for subsequent knowledge construction and model inference.

[0023] S2. Construct a dedicated medical ontology for pulmonary embolism; Using the S1-cleaned data as a hard constraint on knowledge, a customized medical ontology for pulmonary embolism is constructed. This ontology defines standardized medical concepts, hierarchical relationships, attributes, and emergency treatment logic axioms related to pulmonary embolism, achieving semantic and logical hard constraints on knowledge. This blocks the pulmonary embolism-specific AI illusion from the root of reasoning logic. The ontology is formally defined as follows: In the formula, This is a collection of medical concepts specifically for pulmonary embolism, precisely covering core anatomical and pathological concepts such as the main / lobe / segment / subsegmental branches of the pulmonary artery, pulmonary emboli, core signs of CTPA filling defects, D-dimer specific indicators, four-level risk stratification, and emergency thrombolysis / interventional treatment protocols. A is a set of logical relationships between concepts, including strong constraints such as inclusion, adjacency, cause, diagnostic basis, association, and exclusion, which strictly conform to the logic of pulmonary embolism diagnosis and treatment; A is a set of conceptual attributes, involving key attributes such as embolus location, size, shape, degree of filling defect, D-dimer threshold, high-risk threshold values ​​for heart rate and blood pressure, and core indicators of risk stratification; T is a data type constraint. The system employs proprietary logical axioms for the emergency diagnosis and treatment of pulmonary embolism. For example, a normal D-dimer level can rule out low-risk pulmonary embolism, emboli below the segmental level without high-risk indicators are classified as low-risk, CTPA imaging artifacts are not considered true emboli, and high-risk pulmonary embolism requires accompanying signs of shock / hypotension. Through these hard constraints of ontological axioms, the system prevents AI from violating the clinical logic of pulmonary embolism and generating illusions.

[0024] S3. Construct a multi-source fusion pulmonary embolism knowledge graph; Based on the S2 ontology framework, a knowledge graph is constructed by integrating pulmonary embolism diagnosis and treatment guidelines, high-quality annotated emergency cases, core medical literature, AI-specific hallucination judgment rules for pulmonary embolism, and expert quality control experience. This enables structured knowledge storage and associative querying, providing factual support for interpretable reasoning while simultaneously verifying the authenticity of model reasoning, thus doubly controlling the risk of hallucinations. The knowledge graph is formally defined as follows: In the formula, This is a collection of entities related to pulmonary embolism, including emboli, CTPA filling defect signs, laboratory indicators, emergency protocols, guideline clauses, hallucination criteria, etc. The relationships between entities correspond one-to-one with the relationships on the medical ontology, ensuring the consistency of knowledge logic and avoiding hallucinations induced by knowledge conflicts. For entity attribute triples; To assess the credibility of knowledge sources, priority is given to clinical guidelines and authoritative emergency case data to ensure the authority of the explanations, while filtering out low-credibility knowledge to avoid inducing AI illusions.

[0025] S4. Construct an algorithm for recognizing core signs of pulmonary artery filling defects based on ontology knowledge; For the core diagnostic sign of pulmonary embolism CTPA imaging, which is also a high-risk area for AI false positives, a pulmonary embolism-specific sign recognition algorithm is designed based on the hard constraints of anatomical and pathological logic of the pulmonary embolism-specific medical ontology. This algorithm abandons the simple data-driven recognition mode and combines ontology semantic verification to accurately distinguish between true emboli and interference items, thereby suppressing false positive embolism hallucinations caused by CTPA artifacts and vascular variations from the source.

[0026] The core logic of the algorithm is as follows: Using a medical ontology specific to S2 pulmonary embolism as a strong constraint, it first performs anatomical localization of pulmonary artery branches at all levels on CTPA images. It then accurately matches the hierarchical concepts and anatomical location attributes of the pulmonary artery trunk, lobar arteries, segmental arteries, and subsegmental arteries within the ontology. Next, it extracts the morphology, density, edge, and location-related features of filling defects. Finally, it uses ontology axioms to determine the authenticity of defects. The core recognition formula is: In the formula, The score indicates the true nature of the filling defect. If the score is higher than the threshold, it is judged as a true embolic sign of pulmonary embolism. If the score is lower than the threshold, it is judged as an artifact or normal vascular variation, directly intercepting false positive illusions. It represents the morphological characteristics of filling defects, including typical morphological scores of pulmonary emboli such as eccentricity, centrality, and occlusion, and matches the morphological axioms of the emboli in the body; It represents the density characteristics of filling defects, compares them with the blood density of normal pulmonary arteries, quantifies the low density score of emboli, and distinguishes between pulmonary embolism emboli and non-embolic filling defects; It indicates the anatomical location matching features, matching the location of the pulmonary artery branches in the body. The more accurate the positioning, the higher the score, eliminating the illusion of misjudgment across anatomical locations; and The dynamic weights of each item are determined by the confidence level of the match between the signs and the pulmonary embolism ontology. Adaptive calculation, weight allocation is based on the anatomical logic of pulmonary artery in pulmonary embolism, and non-fixed manual assignment avoids misjudgment illusion caused by weights without basis; Indicates the artifact penalty coefficient; The confidence score for image artifacts is calculated using the ontology artifact judgment rule. The higher the score, the greater the penalty. This thoroughly distinguishes between true emboli and pulmonary embolism. CTPA-specific interference items suppress false positive illusions from the source. The algorithm output is synchronously linked to the corresponding signs in the knowledge graph as a guideline, serving as the core basis for subsequent thought chain reasoning. This enables sign recognition to be interpretable and traceable, with no unfounded reasoning throughout the process, further reducing the risk of hallucinations.

[0027] S5. Construct a stratified evaluation algorithm for the severity of pulmonary embolism; Based on pulmonary embolism diagnosis and treatment guidelines and the pulmonary embolism medical ontology-specific axioms of risk stratification, combined with S4 sign recognition results, patient clinical laboratory indicators, and vital sign data, a pulmonary embolism-specific severity assessment algorithm was designed. This algorithm achieves accurate four-level stratification: high-risk, intermediate-high-risk, intermediate-low-risk, and low-risk. It abandons the traditional simple weighted scoring model and adopts ontology-driven dynamic weighting and logical judgment to specifically suppress core pulmonary embolism illusions such as risk stratification confusion and misjudgment of high / low risk. The core formula of the algorithm is: In the formula, The comprehensive score indicates the severity of pulmonary embolism. To ensure the final risk stratification results are fully aligned with clinical guidelines for pulmonary embolism and to avoid unfounded stratification illusions; The imaging feature score is based on the pulmonary artery filling defect identification results. The more critical the extent and location of the embolus, the higher the score, which matches the axiom of the embolic risk level. It represents the clinical signs score, covering emergency vital signs such as heart rate, blood pressure, and respiratory rate, which is consistent with the axiom of high-risk judgment and avoids the illusion of high-risk misjudgment without the support of signs. The scoring of the test indicators includes core pulmonary embolism-specific indicators such as D-dimer, troponin, and BNP, and conforms to the interpretation rules of the test indicators. Each represents a dynamic weight, which is adaptively generated by the matching confidence of imaging signs, clinical signs, laboratory indicators and ontology guidelines. The weight of high-risk indicators increases dynamically with the degree of risk, thus preventing weight allocation that violates the logic of pulmonary embolism diagnosis and treatment. The stratification criteria (strictly adhering to the axioms of pulmonary embolism diagnosis and treatment, eliminating logical fallacies) are as follows: In the formula, These represent the first, second, and third stratification thresholds, respectively. They are set based on guidelines and ontology axioms and can be fine-tuned according to clinical scenarios. The thresholds are fixed and based on evidence, eliminating the illusion of random stratification. This algorithm achieves complete alignment between risk stratification and clinical guidelines and ontology axioms for pulmonary embolism. The stratification results are directly linked to the reasoning basis of the thought chain. At the same time, it incorporates a hallucination verification mechanism to intercept erroneous stratifications that violate clinical logic in real time, thereby achieving precise control of hallucinations in the stratification process.

[0028] S6, Mind Chain Reasoning and Credibility Scoring; Based on the pulmonary embolism-specific medical ontology of S2 and the knowledge graph of S3, combined with the two core algorithms of S4 and S5, a step-by-step thought chain reasoning technology that fits the emergency diagnosis and treatment process of pulmonary embolism is used to generate a complete thought chain. This ensures that the reasoning process is traceable, verifiable, and auditable throughout. At the same time, through single-step hallucination verification, pulmonary embolism-specific hallucinations are controlled throughout the process. The thought chain reasoning process formula is as follows: In the formula, For single-step reasoning nodes, We strictly follow the emergency diagnosis and treatment process for pulmonary embolism: "data input → sign recognition / initial assessment of severity → ontology logic verification → knowledge graph matching → logical judgment". This is the final diagnosis plus risk stratification conclusion. For the corresponding guidelines and case studies in the knowledge graph; To verify the results of a single-step illusion, each step of reasoning must match the ontology axioms with the facts of the graph, thus preventing illusions induced by unfounded reasoning. To address the limitations of the original simple weighted summation method and enhance the suppression of hallucinations in pulmonary embolism, this invention optimizes the formula for the credibility of the thought chain and the hallucination risk score. It introduces a dynamically adaptive weight based on the confidence level of pulmonary embolism ontology knowledge, and adjusts the weight allocation in real time based on ontology matching degree and compliance of pulmonary embolism diagnosis and treatment logic. Simultaneously, a non-linear hallucination penalty mechanism is superimposed, using gradient penalties to strengthen the suppression of high-risk hallucinations in pulmonary embolism. The optimized formula is as follows: In the formula, This represents the dynamic weight of a single-step inference node, not a manually preset fixed value, and is determined by the matching confidence level between that step and the pulmonary embolism ontology and knowledge graph. The adaptive calculation yields values ​​ranging from [0,1]. Higher compliance results in greater weight. The calculation formula is as follows: Where k is the weight adjustment coefficient and b is the confidence threshold for emergency diagnosis and treatment of pulmonary embolism, both of which are determined through clinical sample training and optimization. The weights are dynamically allocated in accordance with the logic of pulmonary embolism diagnosis and treatment to avoid the illusion caused by weight imbalance. It represents the credibility of single-step reasoning, and characterizes the rationality and accuracy of core steps in pulmonary embolism such as sign recognition and severity stratification; This represents a single-step hallucination risk score, which quantifies pulmonary embolism-specific hallucination problems such as false positive emboli, erroneous risk stratification, and logical conflicts. The value range is [0,1]. The hallucination penalty coefficient is represented by a non-linear squared penalty term, which is used in conjunction with the stringency setting of emergency treatment for pulmonary embolism. The system employs a gradient-enhanced penalty for high-risk hallucination reasoning. Once a reasoning step violates the ontological axioms of pulmonary embolism or lacks factual basis in the diagram, the penalty intensity is significantly increased, thereby completely suppressing the core hallucination problem of pulmonary embolism from a mechanistic perspective.

[0029] S7, Full-process quality control and AI illusion control; The entire process of output is subject to quality control, covering four core aspects: data input, model inference, decision output, and clinical adaptability. For the entire diagnosis and treatment process of pulmonary embolism, AI-powered real-time quality control, quantitative scoring, and accurate identification and interception of hallucinations are implemented, managing hallucination risks across the entire inference chain. Addressing the lack of innovation in the original fixed-weight quality control scoring, the entire process quality control scoring formula is optimized. Dynamic weights driven by pulmonary embolism ontology knowledge replace fixed manual weights, increasing the dynamic proportion of hallucination control dimensions. This aligns with the clinical quality control priorities for pulmonary embolism, focusing on strengthening the control of specific hallucinations such as false-positive emboli and disordered risk stratification. The optimized entire process quality control and hallucination control scoring formulas are as follows: Satisfy dynamic weight normalization constraints: In the formula, This represents the input data quality control score, covering core indicators such as pulmonary embolism CTPA artifacts, annotation errors, interference data, and data integrity, thus controlling the causes of hallucinations at their source. It represents the quality control score of model inference, monitors problems such as model drift, overfitting, inference anomalies, and feature extraction bias, and avoids pulmonary embolism hallucinations caused by model anomalies; It indicates the quality control score of the decision-making results, verifies whether the diagnostic conclusions and risk stratification are consistent with the guidelines for the diagnosis and treatment of pulmonary embolism and clinical logic, and intercepts terminal hallucination conclusions; AI-based hallucination management score, which specifically scores hallucination problems specific to pulmonary embolism, such as false positive emboli, false negative missed diagnosis, unfounded reasoning, and disordered risk stratification; Dynamic adaptive weights are calculated in real-time based on the knowledge confidence levels of the corresponding links and the pulmonary embolism ontology, rather than being manually fixed. Addressing the high risk of hallucinations in pulmonary embolism emergency scenarios, when high-risk hallucinations such as false positive emboli or stratification errors are detected, [the system will respond accordingly]. Automatically increases and strengthens the weighting of hallucination control, enabling the quality control focus to be dynamically adjusted according to the risk of pulmonary embolism hallucinations, which is completely different from conventional fixed weighting. The AI-based hallucination warning and interception judgment formula is adapted to the pulmonary embolism quality control logic, and the judgment rules are as follows: In the formula, The quality control threshold for pulmonary embolism diagnosis; The threshold for the risk of pulmonary embolism-specific hallucinations; when Immediately trigger hallucination warnings, intercept erroneous diagnostic results, and simultaneously initiate a closed-loop correction process to completely prevent hallucination diagnoses from entering clinical settings, ensuring the safety of emergency treatment. Results that pass quality control are then entered into the results output stage.

[0030] S8, closed-loop feedback correction; Based on the output of S9, when the quality control module determines that the quality control is unqualified or detects pulmonary embolism-specific AI hallucinations, the closed-loop feedback module automatically links the thought chain reasoning and the two-layer knowledge module to accurately locate abnormal reasoning nodes and hallucination triggers, distinguish different hallucination types such as false positive emboli and disordered risk stratification, and complete model parameter fine-tuning or knowledge base updates to achieve targeted correction and eliminate similar hallucination risks from the root. The iterative formula is: In the formula, For the corrected model / knowledge; To revise the previous model / knowledge; The learning rate; The quality control scoring gradient for pulmonary embolism; To establish a risk gradient for pulmonary embolism hallucinations, ensuring accurate and efficient correction, and continuously reducing the incidence of pulmonary embolism-specific AI hallucinations.

[0031] A system for constructing an AI-assisted diagnostic agent for pulmonary embolism based on ontology and knowledge graphs comprises five core modules: a knowledge base construction module, a thought chain interpretable reasoning module, a full-process quality control and hallucination control module, a closed-loop feedback correction module, and a visualization interaction module. The modules are interconnected and logically linked, and can seamlessly connect with various existing AI diagnostic models for pulmonary embolism without reconstructing the underlying algorithm, enabling rapid interpretable upgrades and precise control of pulmonary embolism-specific hallucinations.

[0032] Furthermore, the knowledge foundation building module is used to perform the following steps: S1. Multi-source data acquisition and preprocessing; S2. Construct a dedicated medical ontology for pulmonary embolism; S3. Construct a multi-source fusion pulmonary embolism knowledge graph; Furthermore, the thought chain interpretable reasoning module is used to perform the following steps: S4. Construct an algorithm for recognizing core signs of pulmonary artery filling defects based on ontology knowledge; S5. Construct a stratified evaluation algorithm for the severity of pulmonary embolism; S6, Mind Chain Reasoning and Credibility Scoring; Furthermore, the end-to-end quality control and hallucination control module is used to perform the following steps: S7, Full-process quality control and AI illusion control; Furthermore, the closed-loop feedback correction module is used to perform the following steps: S8, closed-loop feedback correction; The visualization interaction module is used for visual human-computer interaction.

[0033] Furthermore, the process implemented in the system includes: Data input: Input multimodal data of pulmonary embolism emergency, such as patient lung CTPA images, D-dimer, heart rate and blood pressure, and complete preprocessing and artifact annotation; Knowledge matching: It invokes a two-layer knowledge base of pulmonary embolism medical ontology and knowledge graph to complete a strong logical matching between data and authoritative medical knowledge; AI diagnostic integration: Invoke the AI ​​diagnostic model to obtain preliminary results of embolus identification and hazard stratification; Explainable reasoning based on the chain of thought: It conducts step-by-step reasoning and explanation based on two layers of knowledge, relies on two core algorithms to complete sign recognition and criticality stratification, and simultaneously completes single-step pulmonary embolism hallucination verification to generate a complete chain of diagnostic and treatment evidence and credibility score. Full-process quality control and hallucination management: Calculate the total quality control score and pulmonary embolism hallucination risk value based on the optimized full-process quality control and hallucination management scoring formula, and determine whether it is qualified and whether there is specific hallucination according to the AI ​​hallucination warning and interception judgment formula; Closed-loop correction: For unqualified results or results that present pulmonary embolism hallucinations, an early warning is triggered and the model / knowledge is corrected, and the reasoning process is re-executed; Output results: Generate an interpretable pulmonary embolism diagnosis process, hallucination verification results, and full-process quality control analysis, and push them to the emergency medical staff terminal; Knowledge Updates: Accumulate data on emergency treatment and hallucination correction, regularly update the knowledge base, and continuously optimize the ability to manage hallucinations in pulmonary embolism.

[0034] This invention focuses on the critical clinical needs of pulmonary embolism, leveraging its core advantages of interpretable diagnosis throughout the entire process, precise control of pulmonary embolism-specific hallucinations, and dynamic quality control optimization. It has broad application prospects in two core scenarios: clinical emergency care and medical AI regulatory compliance. Clinically, it can be directly adapted to emergency medicine departments, radiology departments, and chest pain centers at all levels of hospitals. It can assist primary hospitals in overcoming diagnostic experience gaps, accurately identifying true emboli, and intercepting false positive / false negative hallucinations, thus reducing the rate of missed or misdiagnosed pulmonary embolism. It can also help tertiary hospitals and chest pain centers achieve rapid stratified diagnosis in emergency care, shorten treatment time, and gain valuable time for thrombolysis and interventional therapy. The interpretable report also facilitates clinical traceability and doctor-patient communication. From a regulatory perspective, this invention can serve as a core supporting technology for applying for Class III medical device registration certificates from the National Medical Products Administration for pulmonary embolism AI diagnostic products. It fills the industry's gaps in interpretability and pulmonary embolism hallucination control technology, while breaking through the limitations of conventional simple weighted average technology, meeting the innovation requirements of invention patents, and helping products to pass review and enter the market quickly. At the same time, it can be connected to hospital AI quality control platforms to realize routine monitoring and anomaly tracing of pulmonary embolism AI models, meeting the hard requirements of medical quality and safety and medical insurance compliance, and becoming a key support for the compliant implementation of pulmonary embolism AI medical products.

[0035] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for constructing an AI-assisted diagnostic agent for pulmonary embolism based on ontology and knowledge graph, characterized in that, Includes the following steps: S1. Collect multimodal data for pulmonary embolism diagnosis from open-source databases and perform preprocessing operations; S2. Construct a medical ontology specifically for pulmonary embolism based on the preprocessed data; S3. Construct a pulmonary embolism knowledge graph based on a dedicated medical ontology framework for pulmonary embolism; S4. Based on the pulmonary embolism-specific medical ontology, construct an ontology-based algorithm for identifying the core signs of pulmonary artery filling defects, in order to identify whether pulmonary artery filling defects in pulmonary embolism CTPA images are true emboli and suppress the generation of false positive illusions. S5. Based on the signs and symptoms in S4, a stratified assessment algorithm for the severity of pulmonary embolism is constructed to stratify the severity of pulmonary embolism. S6, a pulmonary embolism-specific medical ontology based on S2, a pulmonary embolism knowledge graph based on S3, and algorithms based on S4 and S5 to generate a complete thought chain, and to strengthen the suppression of pulmonary embolism hallucinations through gradient penalty. S7. Based on S1~S6, conduct full-process quality control and AI hallucination control scoring, and use the scoring results to provide AI hallucination warnings and interceptions. S8. Closed-loop feedback correction based on AI illusion warning and interception results; The construction of the intelligent agent is completed by executing S1~S8.

2. The method for constructing an AI-assisted diagnostic agent for pulmonary embolism based on ontology and knowledge graph as described in claim 1, characterized in that, The specific medical ontology for pulmonary embolism is defined as follows: In the formula, A collection of medical concepts specifically for pulmonary embolism; A is the set of logical relationships between concepts; A is the set of concept attributes; T is the data type constraint. A specific logical axiom for the emergency diagnosis and treatment of pulmonary embolism.

3. The method for constructing an AI-assisted diagnostic agent for pulmonary embolism based on ontology and knowledge graph as described in claim 1, characterized in that, The pulmonary embolism knowledge graph is defined as follows: In the formula, For the set of entities related to pulmonary embolism; Relationships between entities; For entity attribute triples; Score the credibility of knowledge sources.

4. The method for constructing an AI-assisted diagnostic agent for pulmonary embolism based on ontology and knowledge graph as described in claim 1, characterized in that, The expression for the ontology-based algorithm for recognizing core signs of pulmonary artery filling defects is as follows: In the formula, The score indicates the true nature of the filling defect. If the score is higher than the threshold, it is judged as a true embolic sign of pulmonary embolism. If the score is lower than the threshold, it is judged as an artifact or normal vascular variation, directly intercepting false positive illusions. Indicates the morphological characteristics of filling defects; Indicates the density characteristics of filling defects; Indicates anatomical location matching features; and Represents the dynamic weights of each feature term; Indicates the artifact penalty coefficient; Confidence level for image artifacts.

5. The method for constructing an AI-assisted diagnostic agent for pulmonary embolism based on ontology and knowledge graph as described in claim 1, characterized in that, In S5, the algorithm expression for the stratified assessment of pulmonary embolism severity is as follows: In the formula, This indicates the comprehensive score for the severity of pulmonary embolism; Indicates the image feature score; Indicates clinical signs and symptoms score; Indicates the score of the test index; These represent the dynamic weights of each scoring item; The criteria for stratifying the severity of pulmonary embolism are as follows: In the formula, These represent the thresholds for the first, second, and third strata, respectively.

6. The method for constructing an AI-assisted diagnostic agent for pulmonary embolism based on ontology and knowledge graph as described in claim 1, characterized in that, In S6, the formula for the thought chain reasoning process is as follows: In the formula, This is a single-step reasoning node; This is the final diagnosis plus risk stratification conclusion. This serves as the basis for corresponding guideline clauses or case studies within the knowledge graph; To verify the results of single-step hallucination, each step of reasoning must be matched with the axioms of the pulmonary embolism-specific medical ontology and the facts of the pulmonary embolism knowledge graph, so as to eliminate hallucinations induced by unfounded reasoning. The expression for strengthening the inhibitory effect of pulmonary embolism hallucinations through gradient penalty is as follows: In the formula, Indicates the dynamic weight of a single-step inference node; Indicates the credibility of single-step reasoning; Indicates a single-step hallucination risk score; This represents the penalty coefficient for hallucination.

7. The method for constructing an AI-assisted diagnostic agent for pulmonary embolism based on ontology and knowledge graph as described in claim 1, characterized in that, In S7, the expression for the AI ​​hallucination control score is as follows: Satisfy dynamic weight normalization constraints: In the formula, This indicates the quality control score of the input data; This represents the quality control score of the model inference; This indicates the quality control score of the decision-making results; AI Illusion Special Control Score; These represent the dynamic adaptive weights of each item; The rules for AI-based hallucination warning and interception are as follows: In the formula, The quality control threshold for pulmonary embolism diagnosis; The threshold for the risk of pulmonary embolism-specific hallucinations; when Immediately trigger hallucination warnings, intercept erroneous diagnostic results, and simultaneously initiate a closed-loop correction process to completely prevent hallucination diagnoses from entering clinical settings, ensuring the safety of emergency treatment. Results that pass quality control are then entered into the results output stage.

8. The method for constructing an AI-assisted diagnostic agent for pulmonary embolism based on ontology and knowledge graph as described in claim 1, characterized in that, In S8, the closed-loop feedback correction iteration process is as follows: In the formula, For the corrected model / knowledge; To revise the previous model / knowledge; The learning rate; The quality control scoring gradient for pulmonary embolism; Risk gradient for pulmonary embolism hallucinations.

9. A system for constructing an AI-assisted diagnostic agent for pulmonary embolism based on ontology and knowledge graph, the system being used to execute the method as described in claims 1 to 8, characterized in that, The system comprises five core modules: a knowledge foundation construction module, a thought chain interpretable reasoning module, a full-process quality control and illusion control module, a closed-loop feedback correction module, and a visual interaction module.

10. A system for constructing an AI-assisted diagnostic agent for pulmonary embolism based on ontology and knowledge graph as described in claim 9, characterized in that: The knowledge foundation construction module is used to perform the following steps: S1. Multi-source data acquisition and preprocessing; S2. Construct a dedicated medical ontology for pulmonary embolism; S3. Construct a multi-source fusion pulmonary embolism knowledge graph; The thought chain interpretable reasoning module is used to perform the following steps: S4. Construct an algorithm for recognizing core signs of pulmonary artery filling defects based on ontology knowledge; S5. Construct a stratified evaluation algorithm for the severity of pulmonary embolism; S6, Mind Chain Reasoning and Credibility Scoring; The full-process quality control and hallucination control module is used to perform the following steps: S7, Full-process quality control and AI illusion control; The closed-loop feedback correction module is used to perform the following steps: S8, closed-loop feedback correction.

Images