A multi-disciplinary discussion automatic generation system based on evidence-based knowledge graph

Abstract

This invention relates to the field of medical information management technology and discloses an automatic generation system for multidisciplinary discussions based on evidence-based knowledge graphs. The system includes: a speech acquisition and speaker separation module, which acquires and separates speech signals in real time during multi-person discussions, outputting segmented speech data; a data fusion and correction module, which converts the segmented speech data into initial text via speech recognition, dynamically corrects the recognition results, and generates a patient treatment event sequence; a causal-enhanced entity recognition module, which receives the patient treatment event sequence and identifies a set of medical entities with timestamps and preset causal role attributes; an evidence-based verification relation extraction module, which queries relevant evidence for potential entity pairs in the medical entity set, determines or marks conflicts in the semantic relationships between entities, and outputs entity relation triples; and a treatment knowledge graph generation module, which integrates the medical entity set and entity relation triples to construct a patient treatment knowledge graph centered on the patient.

Description

Technical Field

[0001] This invention relates to the field of medical information management technology, and in particular to an automatic generation system for multidisciplinary discussions based on evidence-based knowledge graphs. Background Technology

[0002] In multidisciplinary team (MDT) consultations and other multi-person consultation scenarios, clinical discussions are primarily conducted in audio format. However, existing systems struggle to accurately convert mixed audio into structured medical records. On one hand, general speech recognition lacks effective processing of medical terminology, individual patient information, and speaker identity, leading to high transcription error rates and unclear information attribution. On the other hand, even when text is obtained, traditional medical information extraction methods still have significant limitations—entity recognition ignores the chronological order and causal logic of treatment, and relation extraction lacks evidence-based verification with authoritative guidelines, resulting in fragmented and unreliable knowledge graphs that cannot support traceable and compliant clinical decision-making. Chinese invention patent application CN121260341A discloses a preoperative discussion record generation system based on knowledge graphs and natural language processing. This system comprehensively integrates multimodal medical data and uses medical logic reasoning based on knowledge graphs to achieve intelligent, standardized, and personalized generation of preoperative discussion records. However, this technical solution does not provide in-depth explanations of medical entities, resulting in generated results that only summarize the meeting discussions, lacking interpretability.

[0003] Therefore, a multidisciplinary discussion automatic generation system based on evidence-based knowledge graphs is proposed to address the aforementioned problems. Summary of the Invention

[0004] The purpose of this invention is to solve the problems in the prior art.

[0005] The technical solution adopted by this invention to solve its technical problem is: to provide an automatic generation system for multidisciplinary discussions of evidence-based knowledge graphs, comprising:

[0006] The voice acquisition and speaker separation module is used to acquire voice signals in real time in multi-person discussion scenarios, and to distinguish the identities of different speakers and separate the voice streams based on voiceprint features, outputting track-by-track voice data with speaker identifiers.

[0007] The data fusion and correction module is used to convert the segmented speech data into initial text through speech recognition, and dynamically correct the medical terms, drug names and test indicators in the recognition results by combining the patient's clinical background information and medical professional terminology database, and generate a patient diagnosis and treatment event sequence that integrates structured and unstructured information.

[0008] The causal-enhanced entity recognition module receives patient diagnosis and treatment event sequences, decodes them using a neural network model with temporal causal constraints, and identifies a set of medical entities with timestamps and preset causal role attributes. The causal role attributes are used to distinguish the functional types of entities in the diagnosis and treatment events.

[0009] The evidence-based verification relation extraction module has a built-in authoritative medical knowledge base. It is used to query relevant evidence for potential entity pairs in the medical entity set, and judge or mark the semantic relationship between entities based on the authority score of the evidence, and output entity relationship triples with confidence labels. The entity relationship triples include entity pairs, semantic relationship types and confidence labels.

[0010] The diagnostic and treatment knowledge graph generation module is used to construct a dynamic patient diagnostic and treatment knowledge graph that contains causal and temporal information by integrating medical entity sets and entity relationship triples, with the patient at the center.

[0011] The structured diagnostic and treatment information output by the system complies with the current national standards for electronic medical records and medical documents, and is suitable for electronic medical record archiving, medical quality control, or MDT consultation record generation.

[0012] Preferably, the voice acquisition and speaker separation module adopts speaker log technology based on deep embedding clustering, and combines it with the voiceprint template or employee number information of the registered doctor to achieve speaker identity binding.

[0013] Preferably, the data fusion and correction module dynamically loads a patient-specific dictionary during the speech recognition process. The patient-specific dictionary consists of the diagnosis names, drug names, surgical names, and test items from the patient's historical medical records.

[0014] Preferably, in the causal-enhanced entity recognition module, the temporal neural network model that applies causal constraints is a Transformer model with a causal attention mask in the decoder part. The causal attention mask ensures that when decoding an entity at each time step, only the vector information of that time step and the previous time steps are considered.

[0015] Preferably, the preset causal role attributes include intervention type, reaction type and state type; wherein, intervention type corresponds to treatment measure type entity, reaction type corresponds to symptom, sign or test result type entity, and state type corresponds to diagnosis conclusion or disease stage type entity.

[0016] Preferably, in the evidence-based verification relationship extraction module, the authority score is calculated by weighting the level of the issuing institution of the evidence item, the time of publication, the strength of the evidence, and its matching degree with the current patient's clinical scenario.

[0017] Preferably, the evidence-based verification relation extraction module is further configured to: when the relation type inferred from patient data is inconsistent with the relation type recommended based on the highest authority score evidence, mark the output entity relation triple with a conflict flag and set its confidence level to low confidence.

[0018] Preferably, it also includes a key node analysis module, which is used to calculate the key score of each time point on the diagnosis and treatment time axis based on the change rate of text features of the medical entity set, the volatility of associated structured data indicators, and the change magnitude of treatment entities, and to identify the time points where the key score exceeds a preset threshold as key diagnosis and treatment turning points; the diagnosis and treatment knowledge graph generation module integrates the key diagnosis and treatment turning points as special nodes into the patient diagnosis and treatment knowledge graph.

[0019] Preferably, in the knowledge graph generated by the diagnostic and treatment knowledge graph generation module, the nodes contain medical entities and their causal role attributes, and the edges contain semantic relationship types and confidence level identifiers; the knowledge graph supports source tracing queries along the time dimension and causal relationship paths.

[0020] Preferably, it also includes a feedback learning module, which is used to collect correction information from users on entities, relationships or graphs automatically generated by the system, or to receive deviation signals between the actual outcome of patient diagnosis and treatment and the predicted outcome of the graph, and to use the case data corresponding to the correction information or deviation signals to perform incremental learning and parameter fine-tuning optimization on the models in the causal enhancement entity recognition module and / or the evidence-based verification relationship extraction module.

[0021] The present invention has the following beneficial effects:

[0022] (1) This invention is the first to realize end-to-end intelligent conversion from raw voice of multi-person consultation to structured and compliant medical records; through the voice acquisition and speaker separation module, different speakers are accurately distinguished, and the voice recognition results are dynamically corrected by combining the patient-specific dictionary and the national standard medical terminology database, which greatly improves the recognition accuracy of key information such as medical terms, drug names, and test indicators, and effectively solves the clinical pain points of traditional consultation recordings that are "unclear, incomplete, and difficult to archive"; on this basis, the system integrates the corrected text and structured time-series data, and introduces time-series causal constraints for medical entity recognition, so that the extracted entities not only have timestamps, but also label causal role attributes such as intervention type, reaction type, and state type, breaking through the fragmented limitations of traditional entity extraction; at the same time, through the built-in authoritative medical knowledge base, entity relationships are verified by evidence and scored with authority, and conflicting relationships are automatically labeled with low confidence, which significantly improves the professionalism and security of relationship extraction.

[0023] (2) This invention utilizes a causal-enhanced entity recognition module to completely eliminate the problem of the model 'peeking' at future diagnostic information when recognizing early events through a causal attention masking mechanism, effectively preventing the misjudgment of symptoms as diagnostic conclusions; at the same time, its output of temporal entities with causal role attributes provides directional constraints for subsequent relation extraction, significantly reducing the risk of AI illusion, and directly supporting the automatic generation of electronic medical records and medical quality control in accordance with national standards.

[0024] (3) This invention constructs a dynamic diagnosis and treatment knowledge graph centered on the patient and integrating causal logic and temporal evolution. It can automatically identify key diagnosis and treatment turning points, support source tracing queries by timeline or causal path, and generate consultation summaries that meet national requirements such as the "Electronic Medical Record Application Management Specifications" with one click. These summaries can be directly used for electronic medical record archiving and medical quality control, greatly improving the efficiency of MDT collaboration and the traceability of decisions. In addition, the system is equipped with a feedback learning mechanism, which can incrementally optimize the speech correction, entity recognition and relationship extraction models based on the doctor's corrected operation or the deviation of the actual diagnosis and treatment outcome, so as to achieve "more accurate the more it is used" clinical intelligent evolution and truly adapt to the complex and ever-changing real diagnosis and treatment environment.

[0025] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. Attached Figure Description

[0026] Figure 1 This is a diagram illustrating the overall functional module architecture of an embodiment of the present invention;

[0027] Figure 2 This is a flowchart illustrating the process of an embodiment of the present invention;

[0028] Figure 3 This is a schematic diagram of the causal attention masking mechanism in an embodiment of the present invention;

[0029] Figure 4 This is a schematic diagram of the visualization of the dynamic diagnostic and therapeutic knowledge graph in an embodiment of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0031] See Figure 1 The diagram shown is an overall functional module architecture diagram of an embodiment of the present invention, including:

[0032] The voice acquisition and speaker separation module is used to acquire voice signals in real time in multi-person discussion scenarios, and to distinguish the identities of different speakers and separate the voice streams based on voiceprint features, outputting track-by-track voice data with speaker identifiers.

[0033] The data fusion and correction module is used to convert the segmented speech data into initial text through speech recognition, and dynamically correct the medical terms, drug names and test indicators in the recognition results by combining the patient's clinical background information and medical professional terminology database, and generate a patient diagnosis and treatment event sequence that integrates structured and unstructured information.

[0034] The causal-enhanced entity recognition module receives patient diagnosis and treatment event sequences, decodes them using a neural network model with temporal causal constraints, and identifies a set of medical entities with timestamps and preset causal role attributes. The causal role attributes are used to distinguish the functional types of entities in the diagnosis and treatment events.

[0035] The evidence-based verification relation extraction module has a built-in authoritative medical knowledge base. It is used to query relevant evidence for potential entity pairs in the medical entity set, and judge or mark the semantic relationship between entities based on the authority score of the evidence, and output entity relationship triples with confidence labels. The entity relationship triples include entity pairs, semantic relationship types and confidence labels.

[0036] The diagnostic and treatment knowledge graph generation module is used to construct a dynamic patient diagnostic and treatment knowledge graph that contains causal and temporal information by integrating medical entity sets and entity relationship triples, with the patient at the center.

[0037] The key node analysis module is used to calculate the key score of each time point on the diagnosis and treatment time axis by weighting the change rate of the text features of the medical entity set, the volatility of the associated structured data indicators, and the change magnitude of the treatment entities, and to identify the time points where the key score exceeds a preset threshold as key diagnosis and treatment turning points; the diagnosis and treatment knowledge graph generation module integrates the key diagnosis and treatment turning points as special nodes into the patient diagnosis and treatment knowledge graph.

[0038] The feedback learning module is used to collect correction information from users regarding entities, relationships, or graphs automatically generated by the system, or to receive deviation signals between the actual outcomes of patient diagnosis and treatment and the predicted outcomes of the graphs. Using the case data corresponding to the correction information or deviation signals, the module performs incremental learning and parameter fine-tuning optimization on the models in the causal enhancement entity recognition module and / or the evidence-based verification relationship extraction module.

[0039] The system of this invention can be deployed on a local hospital server or a private cloud cluster, including a processor, memory, and network interface. Its core lies in deeply integrating "causal logic constraints" and "evidence-based medical verification" into the entire process from raw speech to structured knowledge. The system mainly consists of seven functional modules working collaboratively: speech acquisition and speaker separation, data fusion and correction, causal enhancement recognition, evidence-based verification extraction, graph generation, key node analysis, and feedback learning. (See [link to relevant documentation]). Figure 2 As shown, the specific implementation process is as follows:

[0040] First, the voice acquisition and speaker separation module serves as the system's perception entry point. In the MDT consultation room, it acquires multi-person discussion voice signals in real time through a microphone array. Based on deep embedding clustering speaker diarization technology, combined with the voiceprint templates or employee ID information of registered doctors, it binds the identities of different speakers and separates the voice streams, outputting track-by-track voice data with speaker identifiers.

[0041] Secondly, the data fusion and correction module converts the various audio streams into initial text using an Automatic Speech Recognition (ASR) engine. During this process, the system dynamically loads a patient-specific dictionary (containing diagnosis names, drug names, surgical names, and laboratory test items) built from the patient's historical medical records. It also combines standardized terminology databases such as the "Common Clinical Medical Terminology (2023 Edition)," "Surgical Operation Code National Clinical Version 3.0 (2022 Summary Edition)," and "National Heart Failure Guidelines (2023 Edition)" published by the National Health Commission and other relevant departments to perform context-aware dynamic correction on the medical terms, drug names, and laboratory test indicators in the recognition results, generating a highly accurate patient diagnosis and treatment event sequence.

[0042] Next, the causal-enhanced entity recognition module receives the sequence of diagnostic and treatment events, aligns it with structured time-series data (such as vital signs and test trends) synchronously acquired from systems like HIS and LIS, and fuses them to generate a uniformly encoded patient state vector sequence. The module incorporates a Transformer model with a causal attention mask (a neural network model imposing temporal causal constraints), allowing only current and historical information to be considered when decoding each time step, preventing "future information leakage." Simultaneously, it automatically labels identified medical entities with causal role attributes such as interventional (e.g., medication, surgery), reactive (e.g., symptoms, abnormal indicators), or state (e.g., diagnosis, disease stage). The causal-enhanced entity recognition module first processes the input text sequence... The initial contextual feature representation is extracted by feature encoders such as BERT or RoBERTa. ); then it enters the core causal attention mask Transformer model, the model uses the attention mechanism to construct a "dictionary of confusion factors" ( Addressing is performed in ")" and the calculation is adjusted using a backdoor. To cut off the confusing factors ( The spurious associations introduced by these features are used to extract "decongestion features". Next, the original contextual features are concatenated and fused with the aforementioned deobfuscated features through residual connections. This process preserves the original information and enhances feature representation; finally, these fused enhanced features are fed into a CRF sequence decoder for global decoding, ultimately outputting an accurate entity label sequence. The Transformer model of the causal attention mask first processes the observed historical and current time series data (…). Input to feature encoder ( In the process, high-dimensional representations are extracted and mapped to the latent variable space; subsequently, the model constructs complex dynamic dependencies within the latent space, dividing the features into a "historical state space" ( ) and "current state space" Finally, the latent features of the current state, updated through graph structure evolution, are fed into the feature decoder. The system performs feature mapping and reconstruction calculations, and finally outputs the prediction result for the data at the next time step. ).

[0043] Subsequently, the evidence-based verification relation extraction module performs relation inference on entity pairs and retrieves the built-in authoritative medical knowledge base (covering the latest clinical guidelines, pharmacopoeias, and high-quality evidence-based literature); based on the level of the issuing institution, the publication time, the strength of the evidence, and the patient scenario matching degree, it calculates the authority score; if the inferred relation conflicts with highly authoritative evidence (such as the inference that "vitamin C cures lung cancer"), it automatically marks the conflict and sets the confidence level to "low", outputting triples with confidence level.

[0044] In parallel, the key node analysis module integrates the rate of change of text features (such as the word density of "critical" and "deterioration"), the volatility of structured indicators (such as a sudden drop in blood pressure) and the magnitude of treatment changes (such as the activation of a ventilator), and calculates a weighted score for each time point. Those exceeding the threshold are marked as key diagnostic and treatment turning points.

[0045] Finally, the diagnosis and treatment knowledge graph generation module takes the patient as the center, integrates entities with causal roles, relationship triples with confidence, and key turning points to construct a dynamic diagnosis and treatment knowledge graph. This graph supports tracing queries by timeline or causal path, and can generate MDT consultation summaries that comply with the "Electronic Medical Record Application Management Specifications" with one click, which can be directly used for medical record archiving or quality control.

[0046] In addition, the feedback learning module silently collects doctors' manual corrections to the atlas (such as modifying entity categories and deleting incorrect relationships) and actual patient outcome data in the background; it uses incremental learning techniques such as LoRA to fine-tune the model, so that the system can complete parameter updates at night and become more accurate the next day, adapting to the personalized diagnosis and treatment habits of departments.

[0047] In the causal-enhanced entity recognition module, this invention constructs a strict causal attention mask mechanism to mathematically force the model to follow the arrow of time, thus solving the "causal inversion" problem that may occur in traditional bidirectional models. Specifically, let the length of the input patient state vector sequence be... The mask matrix generated internally by the model The definition is as follows:

[0048] ;

[0049] During the Transformer's self-attention calculation, this mask matrix is ​​directly added to the original attention logits, resulting in a corrected attention score. The calculation formula is:

[0050] ;

[0051] in, and These are the query matrix and the key matrix, respectively. This is the scaling factor.

[0052] Due to Softmax This formula mathematically ensures that in decoding the first... At any time step, the model is applicable to any The attention weight for future moments is always zero.

[0053] This mechanism completely eliminates the possibility of "future information leakage." For example, in the speech-to-text transcript of an MDT consultation, the system receives a sequence after terminology correction: '09:30 The patient complains of migratory right lower quadrant abdominal pain; 10:15 After physical examination and ultrasound, the diagnosis is acute appendicitis.' When analyzing the 'abdominal pain' at 09:30, the model is limited by causal masking and cannot access the 'diagnosis' information at 10:15. Therefore, it correctly labels it as a 'reactive' symptom rather than a 'state' diagnosis.

[0054] Furthermore, a specific embodiment of entity causal role labeling includes a system that, based on the aforementioned masking mechanism, pre-defines three causal role attributes: Action, Reaction, and State.

[0055] A specific example is as follows: During a cardiology consultation, the voice acquisition module recorded the doctor's statement: "Give me 300 mg of aspirin at 10 o'clock." The initial ASR output was "aspirin," but after the data fusion and correction module called the patient's exclusive dictionary (including the history of "aspirin" medication) and the national terminology database, it was automatically corrected to "aspirin." Subsequently, the causal enhancement entity recognition module recognized "aspirin" at t1 (10:00) and marked it as "intervention type"; at t2 (10:45), it recognized "burning sensation in the upper abdomen" (the original speech was transcribed as "burning sensation in the upper abdomen," and after contextual semantic correction) and marked it as "reactive type"... (10:00) The model reads "Aspirin...chewed"; protected by causal masking, no subsequent symptoms are observed at this time; the model identifies the entity "Aspirin" and labels it as [intervention type] based on the verb "give / chewed". (Timeframe) (10:45) The model reads "burning sensation in the upper abdomen"; at this point, the attention mechanism allows for review. The system identifies the entity "aspirin" at a specific time point; it also identifies the entity "upper abdominal burning sensation" and labels it as [reactive] based on the context of "complaint / accompaniment". The system automatically constructs a causal chain with time intervals: Aspirin [intervention] – (delayed by 45 min) –> Upper abdominal burning sensation [reactive]>. This process clarifies that the "drug" is the cause and the "symptom" is the effect, providing a structured data foundation for subsequent adverse drug reaction monitoring.

[0056] In this embodiment, the evidence verification module introduces a multi-dimensional authority scoring formula to quantify the credibility of evidence, and a conflict resolution mechanism for the retrieved evidence. Its rating The calculation is as follows:

[0057] ;

[0058] (By Institution): National Guidelines / Pharmacopoeia = 1.0, Academic Consensus = 0.8, General Literature = 0.5.

[0059] (Timeliness score): The past 2 years = 1.0, 2-5 years = 0.7, >5 years = 0.4.

[0060] (Intensity score): Meta-analysis / RCT=1.0, cohort study=0.8, case report=0.5.

[0061] (Match score): based on the semantic similarity (0~1) between patient characteristics (age, complications) and the applicable population of the evidence.

[0062] "It is worth noting that the original input for this relation inference may originate from a doctor's colloquial statement during a consultation (such as 'high doses of vitamin C are helpful for tumors'), which is then transcribed and standardized into the relation extraction module..."

[0063] Specifically, suppose the model incorrectly infers relationships from the context of the medical records. "<Vitamin C – cures –> advanced lung cancer>", initial confidence level 0.88.

[0064] The system retrieved the NCCN Clinical Practice Guidelines for Non-Small Cell Lung Cancer (Version 7.2025) and calculated the results. .

[0065] The guidelines explicitly state that "there is no evidence to support that vitamin C has a direct anti-tumor effect" (i.e., (Not related).

[0066] Processing logic: because and The system has identified this as a high-risk conflict.

[0067] Output: The confidence level of the triple is forcibly downgraded to "Low", connected by a red dotted line in the knowledge graph, and a warning label is displayed: "[Warning] This inference is contrary to the NCCN Guidelines (2025), please review"; This effectively intercepts AI illusions and ensures medical safety.

[0068] The input to the key node analysis module consists of two parts: one is the consultation text stream after voice correction, and the other is the synchronously accessed structured data from the monitor / EMR.

[0069] The multidimensional fusion algorithm for key diagnostic and treatment inflection points includes the following:

[0070] The key node analysis module dynamically calculates the key score by integrating the rate of change across three dimensions: text, data, and behavior. ).

[0071] The calculation process is as follows:

[0072] 1. Text dimension ( ): Calculate the cosine distance between the medical record text vectors at time t and time t−1, and count the density increment of high-risk words such as "critical", "deterioration", and "rescue".

[0073] 2. Data Dimensions ( : Extract key vital signs (heart rate, blood pressure, blood oxygen) and calculate the sum of the absolute values ​​of their first derivatives (i.e., the rate of change).

[0074] 3. Treatment Dimension ( ): Detect medical order change events, such as "upgrade nursing level", "activation of ventilator", "issuance of critical condition notice", etc., and assign corresponding weights; Comprehensive scoring formula:

[0075] ;

[0076] (where α, β, and γ are configurable weighting coefficients, all of which default to 0.33 and can be adjusted according to the characteristics of the department).

[0077] when When the time exceeds the preset threshold θ, the time point is marked as a key turning point in diagnosis and treatment. In the atlas visualization interface, these points will be presented as diamond-shaped highlighted nodes. Clicking on them will expand all the diagnosis and treatment details before and after that moment, helping doctors to locate the moment of sudden change in the condition in seconds in a large number of medical records.

[0078] Furthermore, the system possesses continuous evolution capabilities, and the workflow of its feedback learning module is as follows:

[0079] Data collection: The background silently records the doctor's correction operations on the graph (such as: manually correcting entity categories, deleting incorrect relation edges, confirming / rejecting conflict warnings) and the patient's final actual outcome data.

[0080] The correction operations include not only editing the graph, but also direct correction of the speech transcription results (such as clicking 'Aspirin' and selecting 'Aspirin'). This kind of feedback is used to optimize both the ASR language model and the NLP module.

[0081] Dataset construction: Transform manually corrected data into high-quality (Context, Correct_Label) training samples; in particular, construct "HardNegative" samples from cases where the system originally predicted incorrectly but were corrected by doctors, in order to strengthen the model's memory of error-prone points.

[0082] LoRA fine-tuning: Incremental learning is performed using Low-Rank Adaptation (LoRA) technology; the massive parameters of the pre-trained Transformer backbone network are frozen, and only trainable low-rank decomposition matrices (Rank=8 or 16) are injected into the Attention layer; this method has a very small number of parameters (less than 1% of the original model), and can be quickly trained on the hospital's local server during off-peak hours at night, without the need for expensive computing power clusters.

[0083] Hot update deployment: The trained LoRA weight package is dynamically loaded into the inference engine; the next day, when the system handles similar cases, it can absorb the "expert experience" from the previous day, significantly reducing the incidence of similar errors.

[0084] For example, if a doctor in a certain department frequently modifies the side effect label of a specific antibiotic, the system will learn the department's medication habits overnight, achieving personalized adaptation that "understands clinical practice better the more it is used".

[0085] Therefore, through the deep integration and refined implementation of the above seven modules, this invention not only solves the pain points of traditional medical NLP systems such as reversed causality, frequent hallucinations, and lack of evidence-based support, but also constructs a logically consistent, dynamically perceptive, and self-evolving intelligent ecosystem.

[0086] In real-world clinical settings, this system is able to:

[0087] Auxiliary diagnosis: Rapidly analyze complex medical histories through causal mapping to indicate potential adverse drug reactions or complication risks.

[0088] Quality control management: Automatically identify logical inconsistencies in medical record writing (such as the inversion of medication time and symptom onset time) to improve medical record quality.

[0089] Empowering research: Transforming unstructured medical records into high-quality causal structured data, providing a reliable data foundation for real-world research (RWS).

[0090] Teaching and training: By highlighting key turning points and conflict warnings, we help junior doctors quickly grasp the logic and standards of diagnosis and treatment.

[0091] In summary, the technical solution provided in this embodiment has extremely high practical value and promotion prospects, and can significantly improve the scientific nature and efficiency of multidisciplinary team (MDT) decision-making.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic generation system for multidisciplinary discussions based on evidence-based knowledge graphs, characterized in that, include: The voice acquisition and speaker separation module is used to acquire voice signals in real time in multi-person discussion scenarios, and to distinguish the identities of different speakers and separate the voice streams based on voiceprint features, outputting track-by-track voice data with speaker identifiers. The data fusion and correction module is used to convert the segmented speech data into initial text through speech recognition, and dynamically correct the medical terms, drug names and test indicators in the recognition results by combining the patient's clinical background information and medical professional terminology database, and generate a patient diagnosis and treatment event sequence that integrates structured and unstructured information. The causal-enhanced entity recognition module receives patient diagnosis and treatment event sequences, decodes them using a neural network model with temporal causal constraints, and identifies a set of medical entities with timestamps and preset causal role attributes. The causal role attributes are used to distinguish the functional types of entities in the diagnosis and treatment events. The evidence-based verification relation extraction module has a built-in authoritative medical knowledge base. It is used to query relevant evidence for potential entity pairs in the medical entity set, and judge or mark the semantic relationship between entities based on the authority score of the evidence, and output entity relationship triples with confidence labels. Entity relation triples include entity pairs, semantic relation type, and confidence level. The diagnostic and treatment knowledge graph generation module is used to construct a dynamic patient diagnostic and treatment knowledge graph that contains causal and temporal information by integrating medical entity sets and entity relationship triples, with the patient at the center. The structured diagnostic and treatment information output by the system complies with the current national standards for electronic medical records and medical documents, and is suitable for electronic medical record archiving, medical quality control, or MDT consultation record generation.

2. The automatic generation system for multidisciplinary discussions based on evidence-based knowledge graphs according to claim 1, characterized in that: The voice acquisition and speaker separation module uses speaker log technology based on deep embedding clustering, and combines it with the voiceprint template or employee ID information of registered doctors to achieve speaker identity binding.

3. The automatic generation system for multidisciplinary discussions based on evidence-based knowledge graphs according to claim 1, characterized in that: The data fusion and correction module dynamically loads a patient-specific dictionary during the speech recognition process. The patient-specific dictionary consists of the diagnosis name, drug name, surgery name, and test items from the patient's historical medical records.

4. The automatic generation system for multidisciplinary discussions based on evidence-based knowledge graphs according to claim 1, characterized in that: In the causal-enhanced entity recognition module, the neural network model that applies temporal causal constraints is a Transformer model with a causal attention mask in the decoder part. The causal attention mask ensures that when decoding an entity at each time step, only the vector information of that time step and the previous time steps is considered.

5. The automatic generation system for multidisciplinary discussions based on evidence-based knowledge graphs according to claim 4, characterized in that: The preset causal role attributes include intervention type, reaction type and state type; among them, intervention type corresponds to treatment measures, reaction type corresponds to symptoms, signs or test results, and state type corresponds to diagnosis conclusions or disease staging.

6. The automatic generation system for multidisciplinary discussions based on evidence-based knowledge graphs according to claim 1, characterized in that: In the evidence-based verification relationship extraction module, the authority score is calculated by weighting the level of the issuing institution of the evidence item, the time of publication, the strength of the evidence, and its matching degree with the current patient's clinical scenario.

7. The automatic generation system for multidisciplinary discussions based on evidence-based knowledge graphs according to claim 6, characterized in that: The evidence-based verification relation extraction module is further configured to: when the relation type inferred from patient data is inconsistent with the relation type recommended based on the highest authority score evidence, mark the output entity relation triple with a conflict flag and set its confidence level to low confidence.

8. The automatic generation system for multidisciplinary discussions based on evidence-based knowledge graphs according to claim 1, characterized in that: It also includes a key node analysis module, which is used to calculate the key scores of each time point on the diagnosis and treatment time axis based on the text feature change rate of the medical entity set, the volatility of the associated structured data indicators, and the change magnitude of the treatment entities, and to identify the time points where the key scores exceed a preset threshold as key diagnosis and treatment turning points; the diagnosis and treatment knowledge graph generation module integrates the key diagnosis and treatment turning points as special nodes into the patient diagnosis and treatment knowledge graph.

9. The automatic generation system for multidisciplinary discussions based on evidence-based knowledge graphs according to claim 1, characterized in that: In the knowledge graph generated by the diagnostic and treatment knowledge graph generation module, the nodes contain medical entities and their causal role attributes, and the edges contain semantic relationship types and confidence level identifiers; the knowledge graph supports source tracing queries along the time dimension and causal relationship paths.

10. The automatic generation system for multidisciplinary discussions based on evidence-based knowledge graphs according to claim 1, characterized in that: It also includes a feedback learning module, which is used to collect correction information from users on entities, relationships or graphs automatically generated by the system, or to receive deviation signals between the actual outcome of patient diagnosis and treatment and the outcome predicted by the graph. Using the case data corresponding to the correction information or deviation signals, incremental learning and parameter fine-tuning optimization are performed on the models in the causal enhancement entity recognition module and / or the evidence-based verification relationship extraction module.

Images