Multi-role configuration and effective judgment method based on semantic arbitration

Abstract

The invention discloses a multi-role configuration and effective judgment method based on semantic arbitration, and belongs to the technical field of artificial intelligence, and the method comprises the steps: setting a multi-organization role and an arbitration role; receiving and vectorizing a test description text, distributing the test description text as a first clinical test context vector to multiple organizations, randomly assigning a first speaking party and setting an empty discussion pool; the first speaking party generates speaking content based on the clinical test context vector, vectorizes the speaking content and writes the speaking content into a discussion pool, calculates the similarity between a speaking content vector and the clinical test context vector, and calculates an arbitration score in combination with multi-organization attributes; the arbitration role executes three-state judgment according to the arbitration score and the parameters, an arbitration decision is generated, and an effective speaking content vector is determined; fusing, updating and writing the effective vector and the clinical test context vector into a discussion pool; discussion is terminated according to preset conditions, and based on the final vector and the discussion pool, effective judgment is output and a log is recorded; according to the method, a multi-organization discussion scene can be effectively simulated, and the clinical test discussion efficiency and accuracy are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence, and in particular to a multi-role configuration and effective determination method based on semantic arbitration. BACKGROUND

[0002] In professional discussions in the field of clinical trials, it is a common working mode for multiple experts to discuss and give trial suggestions based on trial drugs or instruments. However, the existing discussion process lacks effective management and determination mechanism, and the speeches of experts may deviate from the theme, provide invalid or incorrect information, etc., resulting in low efficiency of the discussion and difficulty in quickly and accurately obtaining effective treatment plans. At the same time, the existing natural language processing technology cannot accurately determine and arbitrate the effectiveness of expert speeches in the multi-role discussion scene according to the professional needs of the clinical trial field. Therefore, there is an urgent need for a method that can perform semantic arbitration on multi-role discussion content, reasonably configure roles, and accurately determine the effectiveness of speeches, in order to improve the quality and efficiency of medical discussions. SUMMARY

[0003] In view of the deficiencies of the prior art, the present application provides a multi-role configuration and effective determination method based on semantic arbitration, which sets multiple organization roles and arbitration roles; receives a trial description text and vectorizes it as a first clinical trial context vector distributed to multiple organizations, randomly assigns a first speaker and sets an empty discussion pool; the first speaker generates speech content based on the clinical trial context vector and vectorizes it and writes it into the discussion pool, calculates the similarity of the speech content vector and the clinical trial context vector, and calculates the arbitration score in combination with the multiple organization attributes; the arbitration role performs three-state determination according to the arbitration score and parameters, generates an arbitration decision and determines the effective speech content vector; the effective vector is fused and updated with the clinical trial context vector and written into the discussion pool; the discussion is terminated according to the preset conditions, the effective determination is output based on the final vector and the discussion pool, and the log is recorded; the method can effectively simulate a multi-organization discussion scene and improve the efficiency and accuracy of clinical trial discussions.

[0004] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0005] The multi-role configuration and effective determination method based on semantic arbitration comprises:

[0006] A plurality of organization roles and arbitration roles are set, and attribute tuples are configured for each organization role, and arbitration parameter sets are configured for the arbitration roles;

[0007] A trial description text is received, and after being vectorized by a preset semantic embedding model, a clinical trial description vector is obtained, which is distributed to all configured organization roles as a first discussion clinical trial context vector, and a first speaker is randomly assigned, and an empty discussion pool is set;

[0008] The first speaker generates a speech content based on a clinical trial context vector, writes the speech content vector into a discussion pool after vectorization, calculates the similarity between the speech content vector and the current clinical trial context vector, and calculates an arbitration score in combination with the organizational role attribute information;

[0009] The arbitration role performs a three-state judgment according to the arbitration score and arbitration parameters, generates an arbitration decision and broadcasts it to all organizational roles, and determines an effective speech content vector according to the arbitration decision;

[0010] The effective speech content vector is fused with the current clinical trial context vector to update the clinical trial context vector and write it into the discussion pool;

[0011] Terminate the discussion according to the preset end condition, output the effective decision based on the final clinical trial context vector and the discussion pool, and record the whole process log.

[0012] Specifically, the process of setting multiple organizational roles includes:

[0013] Load a multi-organizational information data table from an information system and perform entity alignment on the multi-organizational information data table;

[0014] Map the multi-organizational professional field to a semantic vector through a pre-trained medical knowledge graph to obtain a professional vector of the multi-organization, and set an experience weight;

[0015] Collect historical clinical trial discussion records of the multi-organization, extract keyword distribution using TF-IDF, and generate a style vector after normalization;

[0016] Concatenate the professional vector of the multi-organization, the experience weight, and the style vector into a text description, and input it into a contrast learning model to generate a role configuration file;

[0017] Configure an attention bias parameter for each role, model the collaboration relationship between the multi-organizations through a graph neural network, and generate a role association matrix;

[0018] According to the role association matrix, in combination with a preset role template and a role configuration file, each organizational role is instantiated.

[0019] Specifically, the process of configuring an arbitration parameter set for the arbitration role includes:

[0020] Set an arbitration threshold range , wherein and represent the lowest effective threshold and the highest conflict threshold, respectively;

[0021] By analyzing historical data of arbitration decisions in historical clinical trial discussions, machine learning algorithms are used to train arbitration weight coefficients corresponding to different professional fields;

[0022] In combination with the pre-constructed clinical trial complexity evaluation index system, the difficulty level of the clinical trial is classified, and based on the difficulty level of the clinical trial, arbitration parameters associated with the difficulty level of the clinical trial are generated; the arbitration parameters include arbitration threshold, disagreement threshold, speaking order rule, and discussion termination condition parameter.

[0023] Specifically, the implementation process of the clinical trial description vector includes:

[0024] The input trial description text is preprocessed, the deep semantic features of the preprocessed trial description text are extracted using the BioBERT model, and different token semantic representation vectors are obtained;

[0025] The token semantic representation vectors are weighted and integrated using a multi-head attention mechanism to generate an initial clinical trial description vector;

[0026] The historical clinical trial vector library is called, the semantic distance between the current initial clinical trial description vector and the historical case vectors in the historical clinical trial vector library is calculated using cosine similarity, and the historical case vectors with the highest similarity are selected as reference vectors according to the semantic distance sorting;

[0027] The initial clinical trial description vector and the reference vector are weighted and summed according to the similarity weight to generate an extended initial vector containing historical references, i.e., the clinical trial description vector is obtained.

[0028] Specifically, the method of randomly assigning a first speaker includes:

[0029] Based on the generated role association matrix, the degree of each multi-organization node in the role association matrix is calculated to obtain the centrality index of the multi-organization node, and the centrality index is normalized; the degree of the multi-organization node in the role association matrix is the number of other multi-organization nodes directly connected to the multi-organization node; the other multi-organization nodes are multi-organization nodes other than itself;

[0030] The similarity between the professional vector of each organization role and the clinical trial description vector is calculated, the organization roles are sorted according to the similarity value, and an initial priority queue is generated;

[0031] The normalized centrality index and the sorting position in the initial priority queue are respectively assigned weights, the comprehensive score of each organization role is calculated, and the organization roles are re-ordered according to the comprehensive score to obtain a candidate multi-organization set;

[0032] Check if there are organization roles with the same comprehensive score in the candidate multi-organization set, if there are organization roles with the same comprehensive score, introduce a random disturbance factor to obtain a candidate multi-organization sorting result;

[0033] Set a main role rotation period parameter, record the identification of the current discussion period, if the main role rotation period is reached, select the next organizational role as the new first speaker candidate from the candidate multi-organizational ranking result in order, if the current discussion period does not reach the main role rotation period, select the first ranked organizational role as the first speaker candidate from the candidate multi-organizational ranking result.

[0034] Specifically, the process that the current first speaker generates the speech content based on the clinical trial context vector includes:

[0035] The first speaker inputs the clinical trial context vector into the pre-trained LSTM-based medical dialogue generation model to extract clinical trial time sequence features;

[0036] Map the attribute tuple to an attribute embedding vector, and perform weighted fusion with the clinical trial time sequence features to obtain a fusion feature representation constrained by the attribute tuple;

[0037] Input the fusion feature representation into the decoder part of the LSTM model, and at each time step, the decoder generates a word probability distribution according to the current state and the fusion feature representation;

[0038] At each time step, a word is selected for generation according to the generated word probability distribution using a reinforcement learning strategy, and the corresponding reward is calculated according to the generated word sequence, then the parameters of the medical dialogue generation model are adjusted according to the reward value, and through iterative training, a text sequence optimized by the reinforcement learning strategy is generated;

[0039] Post-process the text sequence optimized by the reinforcement learning strategy to convert it into a structured text containing diagnostic opinions, differential diagnosis and treatment suggestions to generate the speech content.

[0040] Specifically, the arbitration role performs three-state judgment according to the arbitration score, which includes comparing the arbitration score with the lowest effective threshold and the highest conflict threshold;

[0041] When the arbitration score is lower than the lowest effective threshold, it is determined as invalid speech;

[0042] When the arbitration score is higher than the highest conflict threshold, it is determined as valid and unique speech;

[0043] When the arbitration score is within the arbitration threshold range, the arbitration decision is generated by a decision tree model combined with the arbitration weight coefficient and the arbitration parameter.

[0044] Specifically, the arbitration decision determines the valid speech content vector, which includes:

[0045] If the arbitration decision is invalid speech, directly set the valid speech content vector to empty;

[0046] If the arbitration decision is valid and unique speech, it includes:

[0047] The speech content vector determined to be valid and unique speech is denoised to obtain a denoised speech content vector; by calculating the inner product and the module length between vectors, the orthogonal projection component of the speech content vector in the direction of the clinical trial context vector is obtained; a gating mechanism is introduced to dynamically control the information fusion strength according to the size of the orthogonal projection component; the speech content vectors adopted in multiple rounds are time-decay weighted to generate an effective vector set with confidence markers;

[0048] If the arbitration decision is multiple speech comprehensive validity, multiple speech content vectors are weighted and fused according to the attributes of the organization roles to generate an effective speech content vector.

[0049] Specifically, the process of outputting the effective determination based on the final clinical trial context vector and the discussion pool includes:

[0050] The final clinical trial context vector is input into a clinical trial conclusion prediction model based on a convolutional neural network, and the output result of the clinical trial conclusion prediction model is corrected in combination with the statistical information of the valid speech content vector in the discussion pool to generate a final effective determination result.

[0051] Compared with the prior art, the beneficial effects of the present application are:

[0052] 1. The present application proposes a multi-role configuration and effective determination method based on semantic arbitration, by setting multiple organization roles and arbitration roles, and configuring exclusive attribute tuples and arbitration parameter sets for different roles, a simulation framework close to the real medical discussion scene is constructed, after receiving the trial description text, the vectorization processing is carried out and the distribution is carried out, the first speaker is randomly specified to start the discussion, this process design makes the clinical trial discussion can be carried out in an orderly and efficient way, lays a good foundation for the subsequent speech generation and arbitration decision, and helps to more comprehensively and deeply discuss the clinical trial.

[0053] 2. The present application proposes a multi-role configuration and effective determination method based on semantic arbitration, in the discussion process, the first speaker generates speech content based on the clinical trial context vector and calculates the arbitration score, the arbitration role determines the valid speech content vector according to the score and the parameter, and then updates the valid vector and the clinical trial context vector, this series of operations not only realizes the scientific evaluation and screening of the speech content, but also constantly enriches the clinical trial context information and improves the discussion quality, finally terminates the discussion according to the preset condition and outputs the effective determination and records the log, which provides strong support for medical decision-making, and is convenient for subsequent review analysis. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 A schematic diagram of the multi-role configuration and effective determination method based on semantic arbitration of the present application;

[0055] Figure 2 A principle flowchart of the multi-role configuration and effective determination method based on semantic arbitration of the present application;

[0056] Figure 3 A first speaker determination flowchart of the multi-role configuration and effective determination method based on semantic arbitration of the present application. DETAILED DESCRIPTION

[0057] Embodiment 1

[0058] Referring to Figure 1 and Figure 2 , the present application provides an embodiment: a multi-role configuration and effective determination method based on semantic arbitration, the method comprising steps S1-S7, comprising the following steps:

[0059] S1: setting multiple organization roles and an arbitration role, configuring an attribute tuple for each organization role, and configuring an arbitration parameter set for the arbitration role;

[0060] In the present application, the multiple organization roles refer to clinical researchers, contract research organizations, independent expert committees, regulatory agencies, and patient organizations; and the arbitration role refers to a sponsor.

[0061] Further, the process of configuring an attribute tuple for each organization role comprises:

[0062] (1) extracting a professional vector, an experience weight, and a style vector of the organization role from a role configuration file;

[0063] (2) associating the professional vector with a preset professional domain knowledge base to obtain detailed knowledge mastery indicators of the organization role in various professional domains as professional attributes;

[0064] (3) calculating an experience value attribute based on the experience weight in combination with work experience, the number of clinical trial participations, and other data in a multi-organization information data table;

[0065] (4) performing cluster analysis on the style vector to determine the speaking style type of the organization role, such as rigorous type, lively type, and concise type, as a style attribute;

[0066] (5) combining the professional attribute, the experience value attribute, and the style attribute into an attribute tuple, and assigning a unique identifier to each attribute tuple.

[0067] S2: receiving the trial description text, obtaining the clinical trial description vector after vectorization, and distributing the clinical trial description vector to all configured organization roles as the first discussed clinical trial context vector, randomly assigning the first speaker, and setting an empty discussion pool at the same time;

[0068] Further, the process of distributing the clinical trial description vector to all configured organization roles as the first discussed clinical trial context vector comprises:

[0069] (1) According to the identifier of the organization role, a mapping relationship with the clinical trial description vector is established;

[0070] (2) Through a message queue or a distributed communication mechanism, the clinical trial description vector is accurately and correctly distributed to the computing node corresponding to each organization role;

[0071] (3) During the distribution process, the distribution time, receiving node and other information are recorded to ensure the traceability of data transmission;

[0072] (4) After each organization role receives the clinical trial description vector, it is stored as the first discussed clinical trial context vector.

[0073] Further, after S2, it further comprises: performing anomaly detection on the clinical trial description vector, judging whether there is data noise or abnormal value by calculating the norm and dimension distribution of the vector; if an anomaly is detected, the abnormal value is corrected by using the median absolute deviation method to ensure the accuracy of the clinical trial context vector input to the organization role, wherein the median absolute deviation method is a prior art content in the art and is not the inventive scheme of the present application, and will not be described here.

[0074] S3: The first speaker generates speech content based on the clinical trial context, vectorizes the speech content and writes it into the discussion pool, calculates the similarity between the speech content vector and the current clinical trial context vector, and calculates the arbitration score combined with the organization role attribute information;

[0075] Further, the method of vectorizing the speech content comprises:

[0076] (1) Obtaining the speech content text to be vectorized;

[0077] (2) The layered attention mechanism is used to process the speech content text before vectorization, and a text representation processed by the layered attention is obtained, wherein the layered attention mechanism usually includes two levels of word level and sentence level. At the word level, attention weight is allocated to each word in a sentence to highlight important words. At the sentence level, attention weight is allocated to each sentence to focus on key sentences. Through this layered processing, key information in long text can be better captured, and important information can be avoided from being submerged due to long text;

[0078] (3) A pre-defined medical term dictionary and a corresponding standard vector space are used to identify key medical terms from the text representation processed by the layered attention. Then, a linear transformation method is used to align and correct the vector representation of the key medical terms with the pre-defined standard vector space, so that the vector of the key medical term has a more accurate semantic representation in the standard vector space, reduces the semantic deviation caused by term ambiguity or different expressions, and obtains a text representation after completion of vector space alignment and correction;

[0079] (4) The text representation after completion of vector space alignment and correction is input into a pre-trained neural network model, and the text is further feature extracted and encoded through the encoding layer of the model. Then, a fixed dimension vector is extracted from the output of the model as a speech semantic vector representation;

[0080] (5) The KL divergence between the speech semantic vector and the clinical trial context vector is calculated, wherein the KL divergence is an index for measuring the difference between two probability distributions, which can be used here to measure the semantic deviation between the speech semantic vector and the clinical trial context vector. The smaller the KL divergence value, the closer the speech semantic and the clinical trial context. The larger the KL divergence value, the higher the semantic deviation, wherein the KL divergence calculation formula is prior art content in the art, and is not the inventive scheme of the present application, which will not be described here;

[0081] (6) While storing the speech semantic vector, the timestamp and role identification of the speech are recorded, and the speech semantic vector representation with timestamp and role identification metadata is output, wherein the timestamp is the time sequence information of the speech in the discussion process, and the role identification is the role information of the speech organization, such as internal medicine organization and surgery organization. These metadata are of great significance for subsequent analysis of the time distribution of the speech, the speech characteristics of different role organizations, etc.

[0082] Further, the arbitration score is calculated as follows:

[0083] (1) The vectorized speech content vector and the current clinical trial context vector are standardized to eliminate the influence of vector dimension difference;

[0084] (2) Calculate the basic similarity between the vectorized speech content vector and the clinical trial context vector using the cosine similarity algorithm. By calculating the dot product of the two vectors, an initial value reflecting the semantic association degree is obtained. The value ranges from -1 to 1. The closer the value is to 1, the higher the semantic matching degree, and the closer the value is to -1, the greater the semantic difference;

[0085] (3) Extract three core dimensions of professional matching degree, experience level, and historical speech credibility from the organizational role attribute information. Among them, the professional matching degree is determined by the cosine similarity between the professional field vector of the organizational role and the clinical trial field vector; the experience level is divided into 1-5 levels according to the length of service of the organizational role, the number of participating clinical trial projects and other indicators, corresponding to the quantitative scores of 0.2-1.0; the historical speech credibility is calculated based on the frequency of past speeches being adopted, the formula is the number of historical valid speeches divided by the total number of speeches, obtaining a quantitative value between 0-1;

[0086] (4) According to the type and stage of the clinical trial, preset the weight coefficients of each attribute dimension, among which the professional matching degree weight is usually set to 0.4-0.6, the experience level weight is set to 0.2-0.3, and the historical speech credibility weight is set to 0.2-0.3, and the sum of the three weight coefficients is 1. Multiply the quantitative value of each attribute by the corresponding weight coefficient to obtain the weighted score of each attribute, and then add the three weighted scores to obtain the comprehensive weighted score of the organizational role attribute, which ranges from 0 to 1;

[0087] (5) Assign weights to the basic semantic similarity and the comprehensive weighted score of the organizational role attribute, respectively, among which the semantic similarity weight is set to 0.5-0.7, and the attribute comprehensive weighted score weight is set to 0.3-0.5, and the sum of the two weights is 1. Multiply the basic semantic similarity value by its corresponding weight, and add the comprehensive weighted score of the organizational role attribute multiplied by its corresponding weight to obtain the final arbitration score.

[0088] S4: The arbitration role performs three-state judgment according to the arbitration score and arbitration parameters, generates an arbitration decision and broadcasts it to all organizational roles, and at the same time, determines the effective speech content vector according to the arbitration decision;

[0089] S5: Fuse the effective speech content vector with the current clinical trial context vector, update the clinical trial context vector and write it into the discussion pool;

[0090] Further, the fusing of the effective speech content vector and the current clinical trial context vector is realized by a gated recurrent unit structure, taking the effective speech content vector and the clinical trial context vector as inputs, and realizing the updating of the clinical trial context vector, wherein the gated recurrent unit structure is a prior art in the field and is not the inventive scheme of the present application, and will not be described here.

[0091] S6: terminating the discussion according to a preset ending condition, outputting an effective judgment based on the final clinical trial context vector and the discussion pool, and recording a whole-process log.

[0092] Further, the terminating the discussion according to the preset ending condition comprises:

[0093] setting a maximum discussion round threshold, terminating the discussion when the number of discussion rounds reaches the threshold; calculating the difference degree of the current clinical trial context vector and the last round clinical trial context vector, and terminating the discussion when the difference degree is less than a preset convergence threshold; detecting whether there is continuous multiple rounds of invalid speech in the discussion pool, and terminating the discussion if there is.

[0094] Further, the recording the whole-process log comprises: recording the organization role, speech content, speech time and arbitration score after each multi-organization speech; recording the arbitration decision and effective speech content vector after the arbitration of the arbitration role; recording the updated clinical trial context vector in each round during the discussion; storing all the recorded information in a distributed log system in chronological order, and performing encryption processing.

[0095] Embodiment 2

[0096] Please refer to Figure 3 The process of setting multiple organization roles in the embodiment comprises:

[0097] loading a multi-organization information data table from an information system, and performing entity alignment on the multi-organization information data table;

[0098] mapping the multi-organization professional field into a semantic vector through a pre-trained medical knowledge graph to obtain a professional vector of the multi-organization, and setting an experience weight;

[0099] collecting historical clinical trial discussion records of the multi-organization, extracting keyword distribution using TF-IDF, and generating a style vector after normalization, wherein TF-IDF is a prior art in the field and is not the inventive scheme of the present application, and will not be described here.

[0100] splicing the professional vector of the multi-organization, the experience weight and the style vector into a text description, and inputting the text description into a contrast learning model to generate a role configuration file;

[0101] An attention bias parameter is configured for each role, a collaboration relationship between multiple organizations is modeled by a graph neural network to generate a role correlation matrix, wherein the graph neural network is prior art in the field and is not the creative scheme of the present application, and will not be described here;

[0102] According to the role correlation matrix, each organization role is instantiated by combining a preset role template and a role configuration file.

[0103] Further, the specific process of setting the experience weight includes:

[0104] First, the historical experience data of multiple organizations is collected and cleaned, and core indicators such as participation time and the number of successful projects are screened out; then the indicators are quantified and graded, and the weights are allocated according to the importance and the single experience score is calculated; then the single experience score is added to obtain an initial comprehensive score, and the dimension difference is eliminated by normalization processing; then the experience weight in the interval of 0-1 is finally obtained by dynamic correction through a time decay factor.

[0105] Further, the single experience score is obtained by multiplying the single indicator quantified value of each organization by the corresponding weight.

[0106] Further, the introduction of the time decay factor to correct the normalized score includes: giving 1.2 times weight to the experience data in the recent 3 years, 1.0 times weight to the experience data in 3-5 years, and 0.8 times weight to the experience data more than 5 years, in order to highlight the reference value of recent experience, and obtain the final experience weight.

[0107] Further, the acquisition approach of the preset role template includes:

[0108] First, the responsibility definition and behavior standard of different posts are extracted from medical guidelines, hospital regulations, and industry standards as the basic framework of the template; then the actual performance data of each role in the historical case discussion is analyzed, and the high-frequency behavior mode and effective interaction mode are summarized and supplemented to the template; finally, medical field experts are invited to evaluate, correct and improve the preliminary template to ensure that the template meets the needs of clinical practice and professional standards, and the role template is obtained.

[0109] The process of configuring the arbitration parameter set for the arbitration role includes:

[0110] Setting an arbitration threshold range , wherein, and represent the minimum effective threshold and the maximum conflict threshold, respectively;

[0111] The historical data of arbitration decisions in the discussion of historical clinical trials is analyzed, and arbitration weight coefficients corresponding to different professional fields are trained by using a machine learning algorithm. The machine learning algorithm is prior art in the field and is not the creative scheme of the present application, and thus is not described herein.

[0112] In combination with a pre-constructed clinical trial complexity evaluation index system, the clinical trial is classified according to difficulty levels, and arbitration parameters associated with the difficulty levels of the clinical trial are generated based on the difficulty levels of the clinical trial. The arbitration parameters include an arbitration threshold, a disagreement threshold, a speaking order rule, and a discussion termination condition parameter.

[0113] The arbitration threshold is a critical value for judging the effectiveness of speaking, including a minimum effective threshold and a maximum conflict threshold. The disagreement threshold is a critical value for judging whether there is a substantial conflict in multi-role speaking. If the value is below the threshold, it is considered that the opinions are basically consistent, and if the value is above the threshold, in-depth discussion needs to be started. The speaking order rule is used to regulate the sequence of multi-role speaking and ensure that key roles participate in the discussion at the appropriate stage. The present application presets the speaking order according to the importance of the roles. The discussion termination condition parameter is used to determine when to end the discussion, including the maximum number of rounds, the convergence similarity threshold, and the core opinion achievement rate.

[0114] Further, the generation process of the arbitration threshold includes:

[0115] The effective speaking score range of different difficulty level tests is extracted from historical clinical trial data. For example, the effective speaking score of a simple test is concentrated in 0.6-0.9, and the base interval is set accordingly. The initial value of the minimum effective threshold for a simple test is 0.6, and the initial value of the maximum conflict threshold is 0.85. The initial value of the minimum effective threshold for a complex test is 0.7, and the initial value of the maximum conflict threshold is 0.9, thereby obtaining the arbitration threshold.

[0116] Further, the generation process of the disagreement threshold includes:

[0117] The speaking similarity distribution under different difficulty levels in historical tests is analyzed, and the critical similarity value for reaching an agreement or generating a disagreement in opinions in tests of the same difficulty level is calculated. For example, the disagreement critical value in a simple test is 0.6, and the disagreement critical value in a complex test is 0.5.

[0118] Further, in the present application, the discussion termination condition is set as:

[0119] The maximum number of rounds for a simple test is set to 3 rounds. The convergence similarity threshold is set to 0.8, and if the speaking similarity is greater than or equal to 0.8 for two consecutive rounds, the discussion is terminated. The core opinion achievement rate is greater than or equal to 80%. The maximum number of rounds for a complex test is set to 5 rounds. The convergence similarity threshold is set to 0.7. The core opinion achievement rate is greater than or equal to 60%.

[0120] Further, the process of clinical trial difficulty level classification includes:

[0121] (1) Construct a multi-dimensional and multi-level clinical trial complexity evaluation index system, including symptom manifestation dimension, symptom severity, symptom duration, examination result dimension;

[0122] (2) Collect clinical trial data and pre-process them;

[0123] (3) Determine the weight of each index by using the analytic hierarchy process, including:

[0124] Taking the clinical trial complexity evaluation as the target layer, the symptom manifestation, examination result, disease diagnosis, and treatment response as the criterion layer, and the specific indexes under each criterion layer as the scheme layer;

[0125] Compare the importance of each index, construct a judgment matrix, for example, for the number of symptoms, symptom severity, and symptom duration under the symptom manifestation dimension, the experts think that the symptom severity is slightly more important than the number of symptoms, and the symptom duration is obviously more important than the number of symptoms;

[0126] Obtain the weight vector of each index by solving the maximum eigenvalue and the corresponding eigenvector of the judgment matrix, and perform consistency check on the judgment matrix to ensure the rationality of the weight;

[0127] (4) Calculate the complexity score of each clinical trial according to the weight of each index and the pre-processed clinical trial data;

[0128] (5) According to the complexity score of the clinical trial, use clustering analysis to classify the clinical trial into different difficulty levels. For example, the clinical trial is classified into three levels of simple, medium, and complex, wherein the clustering analysis is the prior art content in the field and is not the creative scheme of the present application, and will not be repeated here.

[0129] The implementation process of the clinical trial description vector includes:

[0130] Pre-process the input trial description text, extract the deep semantic features of the pre-processed trial description text using the BioBERT model, and obtain different token semantic representations, wherein the BioBERT model is the prior art content in the field and is not the creative scheme of the present application, and will not be repeated here;

[0131] Use the multi-head attention mechanism to weight and integrate the token semantic representations to generate an initial clinical trial description vector;

[0132] Further, the multi-head attention mechanism is used to integrate the token semantic representations to generate an initial clinical trial description vector, including:

[0133] (1) After extracting the deep semantic features, a multi-head attention mechanism is used to process the word vectors, wherein the multi-head attention mechanism allows the model to focus on different parts of the text from multiple different representation subspaces, thereby better capturing long-distance dependencies and important semantic information in the text;

[0134] (2) The word vector sequence output by the BioBERT model is used as the input of the multi-head attention mechanism, wherein the multi-head attention mechanism performs weighted summation on each word vector, and the weight is determined by the semantic similarity between the word vectors. Through parallel calculation of multiple attention heads, multiple different attention weight distributions are obtained, and then the attention weight distributions are spliced and linearly transformed to obtain a comprehensive vector representation;

[0135] (3) The vectors processed by the multi-head attention mechanism are further subjected to feature fusion and dimension reduction operations, and finally an initial clinical trial description vector of 768 dimensions is generated. The initial clinical trial description vector of 768 dimensions contains the main semantic information of the trial description text and can better represent the overall characteristics of the clinical trial.

[0136] The historical clinical trial vector library is called, the cosine similarity is used to calculate the semantic distance between the current initial clinical trial description vector and the historical case vectors in the historical clinical trial vector library, and the historical case vector with the highest similarity is selected as the reference vector according to the semantic distance;

[0137] It should be explained that the cosine similarity measures the similarity between two vectors by calculating the cosine of the included angle. For the current initial clinical trial description vector and the historical case vectors in the historical clinical trial vector library, the dot product and the modulus of each are calculated, and then the similarity value is calculated according to the cosine similarity formula. The similarity value ranges from -1 to 1, and the closer the value is to 1, the more similar the semantics of the two vectors, and the closer the value is to -1, the less similar the semantics, and the value of 0 indicates that the two vectors are orthogonal, i.e. no semantic association.

[0138] The initial clinical trial description vector and the reference vector are weighted and summed according to the similarity weight to generate an expanded initial vector containing historical references, i.e. a clinical trial description vector is obtained.

[0139] The method for randomly assigning a first speaker includes:

[0140] Based on the generated role association matrix, the degree of each multi-organization node in the role association matrix is calculated to obtain a centrality index of the multi-organization node, and the centrality index is normalized; the degree of the multi-organization node in the role association matrix is the number of other multi-organization nodes directly connected to the multi-organization node; the other multi-organization nodes are multi-organization nodes other than the multi-organization node itself.

[0141] calculate the similarity between the professional vector of each organization role and the clinical trial description vector, sort the organization roles according to the similarity value, and generate an initial priority queue;

[0142] In the present application, the cosine similarity is used to calculate the similarity between the professional vector of each organization role and the clinical trial description vector, and the cosine similarity is a prior art content in the field and is not the creative scheme of the present application, which will not be described here;

[0143] The normalized centrality index and the sorting position in the initial priority queue are respectively assigned weights, the comprehensive score of each organization role is calculated, and the organization roles are re-ordered according to the comprehensive score to obtain a candidate multi-organization set;

[0144] Further, the comprehensive score is to convert the sorting position of the organization role in the initial priority queue into a corresponding score value, that is, the higher the score, the earlier the sorting, and then the normalized centrality index of the organization role, that is, the centrality value after eliminating the dimension effect and the value range is unified, and the corresponding score value of the role are taken as two core evaluation dimensions. Finally, the normalized centrality index of each organization role is multiplied by its corresponding weight, and then the sorting score of the role in the initial priority queue is added to the corresponding weight, and the sum of the two product results is the comprehensive score of the organization role.

[0145] Check whether there is an organization role with the same comprehensive score in the candidate multi-organization set, if there is an organization role with the same comprehensive score, introduce a random disturbance factor to obtain a candidate multi-organization ranking result; the range of the random disturbance factor is ;

[0146] Set a main role rotation period parameter, record the identifier of the current discussion period, if the main role rotation period is reached, select the next organization role as a new first speaker candidate from the candidate multi-organization ranking result in order, if the current discussion period does not reach the main role rotation period, select the first ranked organization role as the first speaker candidate from the candidate multi-organization ranking result.

[0147] The process of generating the speaking content by the current first speaker based on the clinical trial context vector includes:

[0148] The first speaker inputs the clinical trial context vector into a pre-trained LSTM-based medical dialogue generation model to extract the clinical trial time sequence feature, wherein the Long Short-Term Memory (LSTM) is a prior art content in the field and is not the creative scheme of the present application, which will not be described here;

[0149] The LSTM model can effectively capture the time sequence dependency in the clinical trial context vector by virtue of its internal memory unit and gating mechanism, perform layer-by-layer processing and feature extraction on the input vector, and output the time sequence feature representation of the clinical trial. The LSTM model is prior art in the field and is not the inventive solution of the present application, and thus will not be described here.

[0150] The attribute tuple is mapped to an attribute embedding vector, which is fused with the clinical trial time sequence feature to obtain a fused feature representation constrained by the attribute tuple;

[0151] The attribute embedding vector is fused with the clinical trial time sequence feature, and this fusion method enables the model to consider both the time sequence information of the clinical trial and the subjective attribute information of the multi-organization when generating text.

[0152] For example, if the professional inclination of the multi-organization is in the field of cardiovascular diseases, the probability of cardiovascular disease-related words will be increased when generating disease-related words; if the experience inclination of the multi-organization is inclined to conservative treatment, the probability of aggressive treatment plan words will be reduced when generating treatment plans; if the multi-organization has advanced diagnostic technology capabilities, the probability of words referring to advanced diagnostic technology will be increased when generating diagnostic opinions. In this way, the generated speech content conforms to the professional inclination, experience inclination, and capability characteristics of the multi-organization.

[0153] The fused feature representation is input to the decoder part of the LSTM model. At each time step, the decoder generates a word probability distribution based on the current state and the fused feature representation;

[0154] At each time step, a word is selected for generation based on the generated word probability distribution using a reinforcement learning strategy, and the corresponding reward is calculated based on the generated word sequence. Then, the parameters of the medical dialogue generation model are adjusted based on the reward value, and through iterative training, a text sequence optimized by the reinforcement learning strategy is generated;

[0155] The reward function is designed based on the quality, accuracy, completeness, and compliance with medical standards of the generated speech content. For example, if the generated speech content contains accurate diagnostic opinions, reasonable differential diagnosis, and effective treatment suggestions, a higher reward is given; if the speech content contains errors or is incomplete, a lower reward or penalty is given. Then, the reinforcement learning algorithm, such as the policy gradient algorithm, is used to adjust the parameters of the medical dialogue generation model, so that the model is more inclined to select words that can obtain high rewards. The policy gradient algorithm is prior art in the field and is not the inventive solution of the present application, and thus will not be described here.

[0156] The text sequence optimized by the reinforcement learning strategy is post-processed to convert it into structured text containing diagnosis opinions, differential diagnosis and treatment suggestions, and the speech content is generated.

[0157] Further, the optimized text sequence is post-processed to convert it into structured text containing diagnosis opinions, differential diagnosis and treatment suggestions, and the specific steps include:

[0158] (1) The relationship extraction method is used to extract key information from the text sequence, such as disease name, symptom description, diagnosis basis, differential diagnosis disease list and corresponding treatment suggestions, wherein the relationship extraction method is prior art in the art and is not the inventive scheme of the present application, and will not be repeated here;

[0159] (2) The extracted key information is organized, for example, the diagnosis opinion is listed separately, the differential diagnosis is sorted according to the possibility, the treatment suggestion is divided into drug treatment, surgical treatment and other different categories, and the structured text is formed.

[0160] The arbitration role performs a three-state determination according to the arbitration score, including comparing the arbitration score with the lowest effective threshold and the highest conflict threshold;

[0161] When the arbitration score is lower than the lowest effective threshold, it is determined as invalid speech;

[0162] When the arbitration score is higher than the highest conflict threshold, it is determined as valid and unique speech;

[0163] When the arbitration score is within the arbitration threshold range, the arbitration decision is generated by a decision tree model combined with the arbitration weight coefficient and the arbitration parameter.

[0164] The valid speech content vector is determined according to the arbitration decision, including:

[0165] If the arbitration decision is invalid speech, the valid speech content vector is directly set to empty;

[0166] If the arbitration decision is valid and unique speech, it includes:

[0167] The speech content vector determined as valid and unique speech is denoised to obtain a denoised speech content vector; the inner product and the module length between the vectors are calculated to obtain the orthogonal projection component of the speech content vector in the direction of the clinical trial context vector; a gating mechanism is introduced to dynamically control the information fusion strength according to the size of the orthogonal projection component; the time decay weighting is performed on the speech content vectors adopted in multiple rounds to generate a valid vector set with confidence markers;

[0168] If the arbitration decision is a multi-speech comprehensive effective, the multiple speech content vectors are weighted and fused according to the attribute of the organization role to generate an effective speech content vector.

[0169] The process of outputting the effective decision based on the final clinical trial context vector and the discussion pool includes:

[0170] The final clinical trial context vector is input into a clinical trial conclusion prediction model based on a convolutional neural network, and the output result of the clinical trial conclusion prediction model is corrected in combination with the effective speech content vector in the discussion pool to generate a final effective decision result, wherein the convolutional neural network is prior art content in the art and is not the inventive scheme of the present application, and is not described here.

[0171] The embodiments of the application are described above in combination with the drawings, but the application is not limited to the above specific embodiments, and the above specific embodiments are only illustrative and not restrictive, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments without departing from the purpose of the application and the scope of protection, and these are all within the protection of the application.

[0172] If the technical solution of the present disclosure involves personal information, the product applying the technical solution of the present disclosure has been explicitly informed of the personal information processing rules before processing the personal information and has obtained the personal independent consent. If the technical solution of the present disclosure involves sensitive personal information, the product applying the technical solution of the present disclosure has obtained the personal independent consent before processing the sensitive personal information and at the same time meets the requirement of "explicit consent". For example, at the personal information collection device such as camera, a clear and prominent mark is set to inform that the personal information collection range has been entered and the personal information will be collected. If the individual voluntarily enters the collection range, it is considered to agree to collect the personal information. Or, on the device for processing personal information, the personal information processing rules are informed by using obvious marks / information, and the personal authorization is obtained by pop-up information or asking the individual to upload the personal information. The personal information processing rules can include personal information processor, personal information processing purpose, processing method and personal information type, etc.

Claims

1. A multi-role configuration and effective determination method based on semantic arbitration, characterized in that: include: Set up multiple organizational roles and arbitration roles, configure attribute tuples for each organizational role, and configure arbitration parameter sets for arbitration roles; Receive the trial description text, vectorize it using the preset semantic embedding model, and obtain the clinical trial description vector. This vector is then distributed to all configured organizational roles as the first clinical trial context vector for discussion. The first speaker is randomly assigned, and an empty discussion pool is set. The current first speaker generates a speech based on the clinical trial context vector, vectorizes the speech content, and writes it into the discussion pool. The similarity between the speech content vector and the current clinical trial context vector is calculated, and the arbitration score is calculated based on the organizational role attribute information. The arbitration role performs a three-state judgment based on the arbitration score and arbitration parameters, generates an arbitration decision, and broadcasts it to all organizational roles. At the same time, the effective speech content vector is determined based on the arbitration decision. Fuse the effective speech content vector with the current clinical trial context vector, update the clinical trial context vector and write it into the discussion pool; Terminate the discussion according to the preset end conditions, output a valid judgment based on the final clinical trial context vector and discussion pool, and record the entire process log.

2. The multi-role configuration and validity determination method based on semantic arbitration according to claim 1, characterized in that: The process of setting up multiple organizational roles includes: Loading multi-organization information data tables from the information system and performing entity alignment on the multi-organization information data tables; Through pre-trained medical knowledge graphs, the professional fields of multiple organizations are mapped into semantic vectors, and professional vectors of multiple organizations are obtained, and experience weights are set; Collect historical clinical trial discussion records from multiple organizations, use TF-IDF to extract keyword distribution, and generate style vectors after normalization; The professional vectors, experience weights, and style vectors of multiple organizations are concatenated into text descriptions and input into a contrastive learning model to generate role profiles. Configure attention bias parameters for each role, model the collaborative relationship between multiple organizations through graph neural networks, and generate a role association matrix; Each organizational role is instantiated based on the role association matrix, combined with the preset role template and role profile.

3. The multi-role configuration and validity determination method based on semantic arbitration according to claim 2, characterized in that: The process of configuring the arbitration parameter set for the arbitration role includes: Setting the Quorum Threshold Range ,in, and Represent the lowest valid threshold and the highest conflict threshold respectively; By analyzing historical data on arbitration decisions in clinical trial discussions, machine learning algorithms are used to train arbitration weight coefficients corresponding to different professional fields. Combined with the pre-built clinical trial complexity assessment index system, clinical trials are divided into difficulty levels, and based on the difficulty level of the clinical trial, arbitration parameters associated with the difficulty level of the clinical trial are generated; the arbitration parameters include arbitration threshold, disagreement threshold, speaking order rules, and discussion termination condition parameters.

4. The multi-role configuration and validity determination method based on semantic arbitration according to claim 3, characterized in that: The implementation process of the clinical trial description vector includes: Preprocess the input test description text, use the BioBERT model to extract the deep semantic features of the preprocessed test description text, and obtain different token semantic representation vectors; A multi-head attention mechanism is used to perform weighted integration on the token semantic representation vectors to generate the initial clinical trial description vector; Call the historical clinical trial vector library, use cosine similarity to calculate the semantic distance between the current initial clinical trial description vector and the historical case vectors in the historical clinical trial vector library, sort them according to the semantic distance, and select the historical case vector with the highest similarity as the reference vector; The initial clinical trial description vector and the reference vector are weighted and summed according to the similarity weight to generate an extended initial vector containing historical references, that is, the clinical trial description vector.

5. The multi-role configuration and validity determination method based on semantic arbitration according to claim 4, characterized in that: The method for randomly designating a first speaker includes: Based on the generated role association matrix, the degree of each multi-organization node in the role association matrix is ​​calculated to obtain the centrality index of the multi-organization node, and the centrality index is normalized; the degree of the multi-organization node in the role association matrix is ​​the number of other multi-organization nodes directly connected to the multi-organization node; the other multi-organization nodes are multi-organization nodes other than the multi-organization node itself; Calculate the similarity between the professional vector of each organizational role and the clinical trial description vector, sort the organizational roles according to the similarity value, and generate an initial priority queue; Assign weights to the normalized centrality index and the ranking position in the initial priority queue, calculate the comprehensive score for each organizational role, and re-rank the organizational roles based on the comprehensive score to obtain a candidate multi-organization set; Check whether there are organizational roles with the same comprehensive score in the candidate multi-organization set. If there are organizational roles with the same comprehensive score, introduce a random perturbation factor to obtain the candidate multi-organization ranking result; Set a protagonist rotation cycle parameter and record the identifier of the current discussion cycle. If the protagonist rotation cycle is reached, the next organizational role is selected from the candidate multi-organization ranking results in order as the new first speaker candidate. If the current discussion cycle has not reached the protagonist rotation cycle, the organizational role ranked first from the candidate multi-organization ranking results is selected as the first speaker candidate.

6. The multi-role configuration and validity determination method based on semantic arbitration according to claim 5, characterized in that: The process of the current first speaker generating speech content based on the clinical trial context vector includes: The first speaker inputs the clinical trial context vector into a pre-trained LSTM-based medical dialogue generation model to extract clinical trial temporal features. Mapping the attribute tuple into an attribute embedding vector and performing weighted fusion with the clinical trial time series feature to obtain a fused feature representation constrained by the attribute tuple; The fused feature representation is input into the decoder part of the LSTM model. At each time step, the decoder generates a word probability distribution based on the current state and the fused feature representation. At each time step, a reinforcement learning strategy is used to select a word for generation based on the probability distribution of generated words. The corresponding reward is calculated based on the generated word sequence. The parameters of the medical dialogue generation model are then adjusted based on the reward value. Through iterative training, a text sequence optimized by the reinforcement learning strategy is generated. The text sequence optimized by the reinforcement learning strategy is post-processed and converted into structured text containing diagnostic opinions, differential diagnoses and treatment recommendations to generate speech content.

7. The multi-role configuration and validity determination method based on semantic arbitration according to claim 6, characterized in that: The arbitration role performs the three-state determination according to the arbitration score, including: comparing the arbitration score with the minimum valid threshold and the maximum conflict threshold; When the arbitration score is lower than the minimum valid threshold, the speech is deemed invalid; When the arbitration score is higher than the highest conflict threshold, it is judged as valid and the only statement; When the arbitration score is within the arbitration threshold range, an arbitration decision is generated through a decision tree model in combination with the arbitration weight coefficient and the arbitration parameter.

8. The multi-role configuration and validity determination method based on semantic arbitration according to claim 7, characterized in that: Determining the effective speech content vector according to the arbitration decision includes: If the arbitration decision is an invalid speech, the valid speech content vector is directly set to empty; If the arbitration decision is valid and the only decision made, it includes: The speech content vectors determined to be valid and unique are subjected to denoising to obtain the denoised speech content vectors. The orthogonal projection components of the speech content vectors in the direction of the clinical trial context vectors are obtained by calculating the inner product and modulus between the vectors. A gating mechanism is introduced to dynamically control the information fusion strength based on the size of the orthogonal projection components. The speech content vectors adopted in multiple rounds are weighted by time decay to generate a set of valid vectors with confidence marks. If the arbitration decision is that multiple speeches are comprehensive and valid, then the multiple speech content vectors are weighted and fused according to the attributes of the organizational role to generate a valid speech content vector.

9. The multi-role configuration and validity determination method based on semantic arbitration according to claim 8, characterized in that: The process of outputting a valid judgment based on the final clinical trial context vector and the discussion pool includes: The final clinical trial context vector is input into the clinical trial conclusion prediction model based on convolutional neural network. The output result of the clinical trial conclusion prediction model is corrected by combining the statistical information of the effective speech content vector in the discussion pool to generate the final effective judgment result.

Images