An agent-based diagnosis report content review system and method
By employing an agent-based diagnostic report content review method, which utilizes multi-round confidence games and agent clusters for three-dimensional review of diagnostic reports, the problem of low efficiency and easy error occurrence in the quality control of diagnostic reports in existing technologies is solved, thus achieving efficient and accurate report review.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SUZHOU YUNTU HEALTH TECH CO LTD
- Filing Date
- 2026-05-07
- Publication Date
- 2026-07-31
AI Technical Summary
In existing technologies, the quality control of electronic diagnostic reports handwritten or entered by doctors relies on individual carefulness or subsequent spot checks by senior physicians. This is inefficient, prone to errors, and fails to cover complex clinical semantic logic and contextual relationships. It also lacks flexibility and intelligence.
A diagnostic report content review method based on intelligent agents is adopted. Composite feature vectors are extracted through a pre-trained medical big language word segmentation model. A review cluster of intelligent agents with language standardization, logical consistency and arbitration is constructed. Multi-round confidence game is carried out. Error correction and logical consistency verification are carried out through three channels of phonetics, prosody and semantics. Combined with medical knowledge graph, a three-dimensional review is carried out.
It enables efficient identification of anomalies in diagnostic reports, reduces false alarms and missed alarms, improves computational efficiency by more than 30%, ensures the accuracy and standardization of reports, and enables rapid review of simple cases and thorough review of complex cases through dynamic resource allocation.
Smart Images
Figure CN122494104A_ABST
Abstract
Description
Technical Field
[0001] This invention proposes a method and system for reviewing the content of diagnostic reports based on intelligent agents, which relates to the field of data processing technology. Background Technology
[0002] In the medical setting, handwritten or entered electronic diagnostic reports by doctors are a core part of patient medical records, and their accuracy and standardization directly affect subsequent treatment and legal liability. Currently, report quality control typically relies on the individual doctor's diligence or subsequent spot checks by senior physicians. This approach is inefficient, incomplete, and prone to errors (such as typos, logical inconsistencies, omissions of key information, and inaccurate descriptions) due to human fatigue and negligence. While some rule-based document verification tools exist, they struggle to cover complex clinical semantic logic and contextual relationships, lacking flexibility and intelligence. Summary of the Invention
[0003] This invention provides a method and system for reviewing the content of diagnostic reports based on intelligent agents, in order to solve the problems mentioned above: This invention proposes a method for reviewing the content of a diagnostic report based on an intelligent agent, the method comprising: The diagnostic report content to be reviewed is obtained, and the diagnostic report content is input into a pre-trained medical big language segmentation model to extract composite feature vectors. The composite feature vectors include: patient basic information embedding vector, chief complaint text embedding vector, auxiliary examination result embedding vector, preliminary diagnosis embedding vector, and medical order text embedding vector. Construct a review cluster consisting of multiple specialized intelligent agents, the review cluster including: a language specification intelligent agent, a logical consistency intelligent agent, and an arbitration intelligent agent; The composite feature vector set is scanned by the language specification agent and the logical consistency agent in the verification cluster to detect anomalies; For each candidate revision, each agent engages in multiple rounds of confidence games under the coordination of the arbitration agent; in each round, each agent publishes a confidence score for the revision and adjusts its own score after mutual review and argumentation. The arbitration agent calculates the divergence of the confidence distribution in adjacent rounds and predicts the marginal payoff for continuing the game based on the historical divergence sequence. When it is determined that the confidence distribution has converged and there is no substantial room for improvement in continuing the game, the game is terminated, and the final corrected confidence and corresponding correction suggestions are output.
[0004] Furthermore, the pre-trained medical large-scale language segmentation model includes: Obtain a training set, which includes multiple training data sets, each of which includes a diagnostic report and labeled data of the composite feature vector corresponding to the diagnostic report; For each training data point in the training set, the following processing is performed: The diagnostic reports from the training data are input into a preset deep learning model to obtain the predicted data of the composite feature vector corresponding to the diagnostic reports; Based on the predicted data and labeled data of the composite feature vector corresponding to the diagnostic report, the model parameters of the deep learning model are updated. The system checks whether the preset training termination condition is met; if so, the trained deep learning model is used as the medical big data word segmentation model; if not, the deep learning model is trained again using the next set of training data.
[0005] Furthermore, the composite feature vector set is scanned by agents in the verification cluster, including: The probability of internal logical conflicts in the diagnostic report is calculated by the logical consistency agent in the review cluster. in, This represents the probability of internal logical conflicts in the diagnostic report, where k represents the total number of symptom entries extracted from the chief complaint text embedding vector. This represents the i-th symptom embedding vector generated from the chief complaint text embedding vector. This represents the initial diagnostic embedding vector. This represents the clinical association weight coefficient between the i-th symptom embedding vector and the preliminary diagnosis embedding vector.
[0006] Furthermore, the composite feature vector set is scanned by agents in the verification cluster, including: The embedded vector of the auxiliary inspection results contains structured coded information including the name of the inspection item, the observed value, the unit, and the reference range; The physiological consistency deviation index of the auxiliary inspection result embedding vector is calculated by the logical consistency agent in the review cluster. in, This indicates the deviation index from physiological consistency. Represents the observed values. Indicates the standard reference value. The semantic embedding vector representing the numerical severity is generated by the logical consistency agent, which maps the observed numerical value to a preset clinical severity level semantic space based on the degree of deviation between the observed numerical value and the reference numerical value. This represents the initial diagnostic embedding vector. and These represent the specific values of the two high-dimensional vectors in the j-th dimension.
[0007] Furthermore, the composite feature vector set is scanned by agents in the verification cluster, including: The language standardization agent calculates the probability of correcting suspected typos based on the error correction confidence algorithm. ,in, Indicates the probability of correcting a suspected typo. This indicates the error correction confidence level of the candidate correction term.
[0008] Furthermore, the error correction confidence is calculated in the following manner: in, Indicates candidate correction words Error correction confidence, Indicates the original observation word, Indicates candidate correction words, Indicating the similarity between sounds and rhymes, Indicates the similarity of character shapes. Represents the contextual semantic probability of a large language model. This represents the first weighting coefficient. This represents the second weighting coefficient. This represents the third weighting coefficient.
[0009] Furthermore, the phonetic similarity is calculated in the following manner: in, This represents the pinyin encoding sequence of candidate correction words. The pinyin encoding sequence representing the original observed word. This represents the edit distance of the Pinyin encoding sequence. This indicates the length of the pinyin encoding sequence of the candidate correction word. Indicates the length of the pinyin encoding sequence of the original observed word; in, Indicates the number of characters in the candidate correction words. This represents the a-th Chinese character in the candidate correction word. This represents the a-th Chinese character in the original observation word. The glyph embedding vector of the a-th Chinese character in the candidate correction word is generated by a pre-trained convolutional neural network. This shows the glyph embedding vector of the a-th Chinese character in the original observed word.
[0010] Furthermore, the language standardization agent obtains the contextual semantic probability of candidate correction words through the following steps: The suspected typos in the text to be verified The original word element is replaced with a pre-defined special mask word element, forming a masked text sequence; The masked text sequence is input into the pre-trained medical large language model, and forward propagation calculation is performed; wherein, the medical large language model is a bidirectional encoder model based on a transformer architecture, with multiple transformer coding layers; Extract the final output vector sequence of the coding layer of the multilayer transformer, and obtain the specific dimension dense vector corresponding to the position of the special mask word from the sequence as the context semantic feature representation vector. ; Obtain the candidate correction words Word embedding vectors in the word embedding matrix of the medical big language model ; Based on the contextual semantic feature representation vector and the word embedding vector, the contextual semantic probability of the candidate corrected word at the current position is calculated using the Softmax function.
[0011] Furthermore, after the t-th round of the game, the arbitrator agent collects the confidence score vectors of each agent for the same revised opinion. ; The confidence score vector is normalized into a probability distribution. ; Calculate the confidence distribution for round t-1. Compared with the confidence distribution of round t The Jansen-Shannon divergence between : in, , Indicates the Kulbeck-Leibler divergence; Arbitration agents based on historical divergence sequences Construct a linear prediction model to predict the expected divergence reduction if the next round of the game is played. : in, Indicates the time decay factor; The arbitral agent decides whether to terminate the game based on the following two conditions: when and When the arbitration agent determines that the confidence distribution has reached a stable state and there is no substantial room for improvement by continuing the game, it terminates the game process in advance and outputs the weighted median of the current confidence score vector as the final corrected confidence.
[0012] This invention proposes a diagnostic report content review system based on intelligent agents, the system comprising: The vector extraction module is used to obtain the content of the diagnostic report to be reviewed, input the content of the diagnostic report into the pre-trained medical big language word segmentation model, and extract composite feature vectors. The composite feature vectors include: patient basic information embedding vector, chief complaint text embedding vector, auxiliary examination result embedding vector, preliminary diagnosis embedding vector, and medical order text embedding vector. A smart agent module is used to construct a review cluster consisting of multiple dedicated smart agents, the review cluster including: a language specification smart agent, a logical consistency smart agent, and an arbitration smart agent; The anomaly detection module is used to scan the composite feature vector set by the language specification agent and the logical consistency agent in the verification cluster to detect anomalies. The dialogue game module is used for each candidate revision suggestion. Under the coordination of the arbitration agent, each agent conducts multiple rounds of confidence game. In each round of game, each agent publishes a confidence score for the revision suggestion and reviews and discusses each other's opinions before adjusting their own scores. The convergence module is used to calculate the divergence of the confidence distribution in adjacent rounds through the arbitral agent, and predict the marginal payoff of continuing the game based on the historical divergence sequence. When it is determined that the confidence distribution has converged and there is no substantial room for improvement in continuing the game, the game is terminated, and the final corrected confidence and the corresponding correction suggestions are output.
[0013] The beneficial effects of this invention are as follows: Traditional review tools typically rely on single-dimensional rule matching or statistical models, making it difficult to detect deep-seated errors across dimensions. This solution employs a language standardization agent to jointly correct errors from three channels: phonetics, character form, and semantics; a logical consistency agent to verify from two dimensions: symptom-diagnosis semantic matching and numerical-description severity matching; and a medical knowledge graph agent to ensure compliance with clinical guidelines, forming a comprehensive review network for diagnostic reports. The detection results of each agent mutually corroborate and complement each other's blind spots in multiple rounds of game theory, effectively reducing false positives and false negatives caused by a single algorithm. Unlike fixed-round iterations or simple weighted fusion of multiple agents, this scheme introduces an adaptive convergence control mechanism based on the evolution of confidence distribution divergence. For common outliers where all agents have highly consistent opinions and confidence converges rapidly, the system can terminate early with only 2 to 3 rounds of play, saving computational resources. For high-risk outliers with significant disputes and divergent opinions, the system automatically extends the number of rounds or introduces external knowledge to corroborate the findings, ensuring that the problem is fully demonstrated. Experimental results show that approximately 70% of common outliers meet the convergence condition before reaching the preset maximum number of rounds, improving overall computational efficiency by more than 30% and achieving dynamic resource allocation for both quick review of simple cases and thorough review of complex cases. Attached Figure Description
[0014] Figure 1 This is a schematic diagram of a diagnostic report content verification method based on intelligent agents according to the present invention. Detailed Implementation
[0015] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.
[0016] Numerous specific details are set forth in the following description to provide a thorough understanding of the invention. The described embodiments are only a part of the embodiments of the invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.
[0018] One embodiment of the present invention provides a method for reviewing the content of a diagnostic report based on an intelligent agent, the method comprising: The diagnostic report content to be reviewed is obtained, and the diagnostic report content is input into a pre-trained medical big language segmentation model to extract composite feature vectors. The composite feature vectors include: patient basic information embedding vector, chief complaint text embedding vector, auxiliary examination result embedding vector, preliminary diagnosis embedding vector, and medical order text embedding vector. Construct a review cluster consisting of multiple specialized intelligent agents, the review cluster including: a language specification intelligent agent, a logical consistency intelligent agent, and an arbitration intelligent agent; The composite feature vector set is scanned by the language specification agent and the logical consistency agent in the verification cluster to detect anomalies; For each candidate revision, each agent engages in multiple rounds of confidence games under the coordination of the arbitration agent; in each round, each agent publishes a confidence score for the revision and adjusts its own score after mutual review and argumentation. The arbitration agent calculates the divergence of the confidence distribution in adjacent rounds and predicts the marginal payoff for continuing the game based on the historical divergence sequence. When it is determined that the confidence distribution has converged and there is no substantial room for improvement in continuing the game, the game is terminated, and the final corrected confidence and corresponding correction suggestions are output.
[0019] The system first inputs the diagnostic report to be reviewed into a pre-trained medical text structured semantic encoding model. This model, based on a transformer architecture, undergoes further pre-training and supervised fine-tuning using domain-specific corpora such as medical textbooks, clinical guidelines, and desensitized electronic medical records, enabling it to achieve deep semantic understanding of medical text. After performing word segmentation, word embedding, positional encoding, and multi-layer self-attention encoding on the report text, the model outputs five types of structured embedding vectors: patient basic information embedding vector, chief complaint text embedding vector, auxiliary examination results embedding vector, preliminary diagnosis embedding vector, and medical order text embedding vector. These vectors transform discrete textual information into a dense semantic representation that can be computed by the machine, providing a unified data foundation for subsequent intelligent agents.
[0020] The system constructs a review cluster consisting of a language standardization agent, a logical consistency agent, and an arbitration agent. Each agent executes specialized detection tasks in parallel based on a composite feature vector set. The language standardization intelligent agent integrates a three-channel error correction algorithm encompassing phonology, glyphics, and semantics. For suspected misspellings in the main complaint text, it calculates the phonological similarity (based on the edit distance of the pinyin encoding sequence) and glyphic similarity (based on the cosine distance of the glyphic embedding vectors extracted by a convolutional neural network) between the candidate correction word and the original observed word, as well as the semantic probability of the large language model in the masked context. The three are then weighted and fused to obtain the error correction confidence score, identifying and correcting common errors in medical texts such as homophones and similar-looking characters.
[0021] The logical consistency agent is responsible for two types of logical verification. First, it calculates the probability of logical conflicts within the report: it calculates the cosine similarity between the semantic embedding vectors of each symptom extracted from the chief complaint and the initial diagnosis embedding vector, multiplies it by the clinical association weight coefficient, and then multiplies them together to quantify the semantic matching degree between the symptom combination and the diagnostic conclusion. Second, it calculates the physiological consistency deviation index: it extracts the observed values and reference values from the embedding vectors of auxiliary examination results. On the one hand, it captures the cross-order-of-magnitude deviation caused by unit errors through logarithmic operations; on the other hand, it maps the degree of numerical deviation and the severity of the symptom description in the chief complaint to the same clinical severity level semantic space, and detects semantic contradictions such as "slightly abnormal numerical value but critical description" or "significantly abnormal numerical value but mild description" through cosine similarity.
[0022] One embodiment of the present invention includes a pre-trained medical large-scale language segmentation model, comprising: Obtain a training set, which includes multiple training data sets, each of which includes a diagnostic report and labeled data of the composite feature vector corresponding to the diagnostic report; For each training data point in the training set, the following processing is performed: The diagnostic reports from the training data are input into a preset deep learning model to obtain the predicted data of the composite feature vector corresponding to the diagnostic reports; Based on the predicted data and labeled data of the composite feature vector corresponding to the diagnostic report, the model parameters of the deep learning model are updated. The system checks whether the preset training termination condition is met; if so, the trained deep learning model is used as the medical big data word segmentation model; if not, the deep learning model is trained again using the next set of training data.
[0023] One embodiment of the present invention involves scanning the composite feature vector set by an agent in a verification cluster, including: The probability of internal logical conflicts in the diagnostic report is calculated by the logical consistency agent in the review cluster. in, This represents the probability of internal logical conflicts in the diagnostic report, where k represents the total number of symptom entries extracted from the chief complaint text embedding vector. This represents the i-th symptom embedding vector generated from the chief complaint text embedding vector. This represents the initial diagnostic embedding vector. This represents the clinical association weight coefficient between the i-th symptom embedding vector and the preliminary diagnosis embedding vector.
[0024] One embodiment of the present invention involves scanning the composite feature vector set by an agent in a verification cluster, including: The embedded vector of the auxiliary inspection results contains structured coded information including the name of the inspection item, the observed value, the unit, and the reference range; The physiological consistency deviation index of the auxiliary inspection result embedding vector is calculated by the logical consistency agent in the review cluster. in, This indicates the deviation index from physiological consistency. Represents the observed values. Indicates the standard reference value. The semantic embedding vector representing the numerical severity is generated by the logical consistency agent, which maps the observed numerical value to a preset clinical severity level semantic space based on the degree of deviation between the observed numerical value and the reference numerical value. This represents the initial diagnostic embedding vector. and These represent the specific values of the two high-dimensional vectors in the j-th dimension.
[0025] One embodiment of the present invention involves scanning the composite feature vector set by an agent in a verification cluster, including: The language standardization agent calculates the probability of correcting suspected typos based on the error correction confidence algorithm. ,in, Indicates the probability of correcting a suspected typo. This indicates the error correction confidence level of the candidate correction term.
[0026] In one embodiment of the present invention, the error correction confidence is calculated in the following manner: in, Indicates candidate correction words Error correction confidence, Indicates the original observation word, Indicates candidate correction words, Indicating the similarity between sounds and rhymes, Indicates the similarity of character shapes. Represents the contextual semantic probability of a large language model. This represents the first weighting coefficient. This represents the second weighting coefficient. This represents the third weighting coefficient.
[0027] In one embodiment of the present invention, the phonetic similarity is calculated in the following manner: in, This represents the pinyin encoding sequence of candidate correction words. The pinyin encoding sequence representing the original observed word. This represents the edit distance of the Pinyin encoding sequence. This indicates the length of the pinyin encoding sequence of the candidate correction word. Indicates the length of the pinyin encoding sequence of the original observed word; in, Indicates the number of characters in the candidate correction words. This represents the a-th Chinese character in the candidate correction word. This represents the a-th Chinese character in the original observation word. The glyph embedding vector of the a-th Chinese character in the candidate correction word is generated by a pre-trained convolutional neural network. This shows the glyph embedding vector of the a-th Chinese character in the original observed word.
[0028] In one embodiment of the present invention, the language standardization agent obtains the contextual semantic probability of candidate correction words through the following steps: The suspected typos in the text to be verified The original word element is replaced with a pre-defined special mask word element, forming a masked text sequence; The masked text sequence is input into the pre-trained medical large language model, and forward propagation calculation is performed; wherein, the medical large language model is a bidirectional encoder model based on a transformer architecture, with multiple transformer coding layers; Extract the final output vector sequence of the coding layer of the multilayer transformer, and obtain the specific dimension dense vector corresponding to the position of the special mask word from the sequence as the context semantic feature representation vector. ; Obtain the candidate correction words Word embedding vectors in the word embedding matrix of the medical big language model ; Based on the contextual semantic feature representation vector and the word embedding vector, the contextual semantic probability of the candidate corrected word at the current position is calculated using the Softmax function.
[0029] In one embodiment of the present invention, after the t-th round of the game, the arbitrator agent collects the confidence score vectors of each agent for the same revised opinion. ; The confidence score vector is normalized into a probability distribution. ; Calculate the confidence distribution for round t-1. Compared with the confidence distribution of round t The Jansen-Shannon divergence between : in, , Indicates the Kulbeck-Leibler divergence; Arbitration agents based on historical divergence sequences Construct a linear prediction model to predict the expected divergence reduction if the next round of the game is played. : in, Indicates the time decay factor; The arbitral agent decides whether to terminate the game based on the following two conditions: when and When the arbitration agent determines that the confidence distribution has reached a stable state and there is no substantial room for improvement by continuing the game, it terminates the game process in advance and outputs the weighted median of the current confidence score vector as the final corrected confidence.
[0030] One embodiment of the present invention provides a diagnostic report content review system based on intelligent agents, the system comprising: The vector extraction module is used to obtain the content of the diagnostic report to be reviewed, input the content of the diagnostic report into the pre-trained medical big language word segmentation model, and extract composite feature vectors. The composite feature vectors include: patient basic information embedding vector, chief complaint text embedding vector, auxiliary examination result embedding vector, preliminary diagnosis embedding vector, and medical order text embedding vector. A smart agent module is used to construct a review cluster consisting of multiple dedicated smart agents, the review cluster including: a language specification smart agent, a logical consistency smart agent, and an arbitration smart agent; The anomaly detection module is used to scan the composite feature vector set by the language specification agent and the logical consistency agent in the verification cluster to detect anomalies. The dialogue game module is used for each candidate revision suggestion. Under the coordination of the arbitration agent, each agent conducts multiple rounds of confidence game. In each round of game, each agent publishes a confidence score for the revision suggestion and reviews and discusses each other's opinions before adjusting their own scores. The convergence module is used to calculate the divergence of the confidence distribution in adjacent rounds through the arbitral agent, and predict the marginal payoff of continuing the game based on the historical divergence sequence. When it is determined that the confidence distribution has converged and there is no substantial room for improvement in continuing the game, the game is terminated, and the final corrected confidence and the corresponding correction suggestions are output.
[0031] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A method for reviewing the content of a diagnostic report based on an intelligent agent, characterized in that, The method includes: The diagnostic report content to be reviewed is obtained, and the diagnostic report content is input into a pre-trained medical big language segmentation model to extract composite feature vectors. The composite feature vectors include: patient basic information embedding vector, chief complaint text embedding vector, auxiliary examination result embedding vector, preliminary diagnosis embedding vector, and medical order text embedding vector. Construct a review cluster consisting of multiple specialized intelligent agents, the review cluster including: a language specification intelligent agent, a logical consistency intelligent agent, and an arbitration intelligent agent; The composite feature vector set is scanned by the language specification agent and the logical consistency agent in the verification cluster to detect anomalies; For each candidate revision, each agent engages in multiple rounds of confidence games under the coordination of the arbitration agent; in each round, each agent publishes a confidence score for the revision and adjusts its own score after mutual review and argumentation. The arbitration agent calculates the divergence of the confidence distribution in adjacent rounds and predicts the marginal payoff for continuing the game based on the historical divergence sequence. When it is determined that the confidence distribution has converged and there is no substantial room for improvement in continuing the game, the game is terminated, and the final corrected confidence and corresponding correction suggestions are output.
2. The method for reviewing the content of a diagnostic report based on an intelligent agent according to claim 1, characterized in that, Pre-trained medical large-scale language segmentation model, including: Obtain a training set, which includes multiple training data sets, each of which includes a diagnostic report and labeled data of the composite feature vector corresponding to the diagnostic report; For each training data point in the training set, the following processing is performed: The diagnostic reports from the training data are input into a preset deep learning model to obtain the predicted data of the composite feature vector corresponding to the diagnostic reports; Based on the predicted data and labeled data of the composite feature vector corresponding to the diagnostic report, the model parameters of the deep learning model are updated. The system checks whether the preset training termination condition is met; if so, the trained deep learning model is used as the medical big data word segmentation model; if not, the deep learning model is trained again using the next set of training data.
3. The method for reviewing the content of a diagnostic report based on an intelligent agent according to claim 1, characterized in that, The composite feature vector set is scanned by agents in the verification cluster, including: The probability of internal logical conflicts in the diagnostic report is calculated by the logical consistency agent in the review cluster. in, This represents the probability of internal logical conflicts in the diagnostic report, where k represents the total number of symptom entries extracted from the chief complaint text embedding vector. This represents the i-th symptom embedding vector generated from the chief complaint text embedding vector. This represents the initial diagnostic embedding vector. This represents the clinical association weight coefficient between the i-th symptom embedding vector and the preliminary diagnosis embedding vector.
4. The method for reviewing the content of a diagnostic report based on an intelligent agent according to claim 1, characterized in that, The composite feature vector set is scanned by agents in the verification cluster, including: The embedded vector of the auxiliary inspection results contains structured coded information including the name of the inspection item, the observed value, the unit, and the reference range; The physiological consistency deviation index of the auxiliary inspection result embedding vector is calculated by the logical consistency agent in the review cluster. in, This indicates the deviation index from physiological consistency. Represents the observed values. Indicates the standard reference value. The semantic embedding vector representing the numerical severity is generated by the logical consistency agent, which maps the observed numerical value to a preset clinical severity level semantic space based on the degree of deviation between the observed numerical value and the reference numerical value. This represents the initial diagnostic embedding vector. and These represent the specific values of the two high-dimensional vectors in the j-th dimension.
5. The method for reviewing the content of a diagnostic report based on an intelligent agent according to claim 1, characterized in that, The composite feature vector set is scanned by agents in the verification cluster, including: The language standardization agent calculates the probability of correcting suspected typos based on the error correction confidence algorithm. ,in, Indicates the probability of correcting a suspected typo. This indicates the error correction confidence level of the candidate correction term.
6. The method for reviewing the content of a diagnostic report based on an intelligent agent according to claim 5, characterized in that, The error correction confidence level is calculated in the following way: in, Indicates candidate correction words Error correction confidence, Indicates the original observation word, Indicates candidate correction words, Indicating the similarity between sounds and rhymes, Indicates the similarity of character shapes. Represents the contextual semantic probability of a large language model. This represents the first weighting coefficient. This represents the second weighting coefficient. This represents the third weighting coefficient.
7. The method for reviewing the content of a diagnostic report based on an intelligent agent according to claim 6, characterized in that, The phonological similarity is calculated in the following way: in, This represents the pinyin encoding sequence of candidate correction words. The pinyin encoding sequence representing the original observed word. This represents the edit distance of the Pinyin encoding sequence. This indicates the length of the pinyin encoding sequence of the candidate correction word. Indicates the length of the pinyin encoding sequence of the original observed word; in, Indicates the number of characters in the candidate correction words. This represents the a-th Chinese character in the candidate correction word. This represents the a-th Chinese character in the original observation word. The glyph embedding vector of the a-th Chinese character in the candidate correction word is generated by a pre-trained convolutional neural network. This shows the glyph embedding vector of the a-th Chinese character in the original observed word.
8. The method for reviewing the content of a diagnostic report based on an intelligent agent according to claim 5, characterized in that, The language standardization agent obtains the contextual semantic probability of candidate correction words through the following steps: The suspected typos in the text to be verified The original word element is replaced with a pre-defined special mask word element, forming a masked text sequence; The masked text sequence is input into the pre-trained medical large language model, and forward propagation computation is performed; wherein, the medical large language model is a bidirectional encoder model based on a transformer architecture, having multiple transformer coding layers; Extract the final output vector sequence of the coding layer of the multilayer transformer, and obtain the specific-dimensional dense vector corresponding to the position of the special masked word from the sequence as the context semantic feature representation vector. ; Obtain the candidate correction words Word embedding vectors in the word embedding matrix of the medical big language model ; Based on the contextual semantic feature representation vector and the word embedding vector, the contextual semantic probability of the candidate corrected word at the current position is calculated using the Softmax function.
9. The method for reviewing the content of a diagnostic report based on an intelligent agent according to claim 1, characterized in that, The arbitrary agent calculates the divergence of the confidence distribution in adjacent rounds and predicts the marginal payoff for continuing the game based on the historical divergence sequence. The game is terminated when it is determined that the confidence distribution has converged and there is no substantial room for improvement in continuing the game. This includes: After the t-th round of the game, the arbitrator agent collects the confidence score vectors of each agent for the same revised opinion. ; The confidence score vector is normalized into a probability distribution. ; Calculate the confidence distribution for round t-1. Compared with the confidence distribution of round t The Jansen-Shannon divergence between : in, , Indicates the Kulbeck-Leibler divergence; Arbitration agents based on historical divergence sequences Construct a linear prediction model to predict the expected divergence reduction if the next round of the game is played. : in, Indicates the time decay factor; The arbitral agent decides whether to terminate the game based on the following two conditions: when and When the arbitration agent determines that the confidence distribution has reached a stable state and there is no substantial room for improvement by continuing the game, it terminates the game process in advance and outputs the weighted median of the current confidence score vector as the final corrected confidence.
10. A diagnostic report content review system based on intelligent agents, characterized in that, The system includes: The vector extraction module is used to obtain the content of the diagnostic report to be reviewed, input the content of the diagnostic report into the pre-trained medical big language word segmentation model, and extract composite feature vectors. The composite feature vectors include: patient basic information embedding vector, chief complaint text embedding vector, auxiliary examination result embedding vector, preliminary diagnosis embedding vector, and medical order text embedding vector. A smart agent module is used to construct a review cluster consisting of multiple dedicated smart agents, the review cluster including: a language specification smart agent, a logical consistency smart agent, and an arbitration smart agent; The anomaly detection module is used to scan the composite feature vector set by the language specification agent and the logical consistency agent in the verification cluster to detect anomalies. The dialogue game module is used for each candidate revision suggestion. Under the coordination of the arbitration agent, each agent conducts multiple rounds of confidence game. In each round of game, each agent publishes a confidence score for the revision suggestion and reviews and discusses each other's opinions before adjusting their own scores. The convergence module is used to calculate the divergence of the confidence distribution in adjacent rounds through the arbitral agent, and predict the marginal payoff of continuing the game based on the historical divergence sequence. When it is determined that the confidence distribution has converged and there is no substantial room for improvement in continuing the game, the game is terminated, and the final corrected confidence and the corresponding correction suggestions are output.