Clinical test patient matching method and system based on large model

Abstract

On one hand, the invention provides a clinical test patient matching method based on a large model. The invention further discloses a clinical test patient matching system based on the large model. The clinical test patient matching system based on the large model is used for achieving the clinical test patient matching method based on the large model. According to the method, traditional patient containing and ranking are converted into case retrieval and problem-based filtering, a professional knowledge data set is constructed, a field large model is obtained based on large model fine tuning, and the cost can be greatly reduced and the accuracy can be improved by utilizing the understanding and processing capability of the large model for long texts. According to the method and the system, the answers of a plurality of containing and arranging standard questions can be more effectively integrated by fusing evaluation rules of various strategies, the medical semantic information is mined, the relationship between the patient and the clinical test is obtained, and the matching accuracy of the clinical test and the patient is effectively improved.

Description

Technical Field

[0001] The invention relates to a method and system for clinical trial patient matching based on a large model, and belongs to the technical field of large model digital medical treatment. Background Art

[0002] The main purpose of clinical trials is to evaluate the effectiveness of medical interventions and provide evidence for clinical practice. It is crucial to promote drug development and evidence-based medicine. Patient inclusion screening is the key to determining whether the target population can be included.

[0003] However, matching patients with appropriate clinical trials is challenging. Traditional patient exclusion technology usually matches patient medical records with exclusion criteria based on clinical trial exclusion criteria, and recommends appropriate clinical trials based on manual screening or traditional machine learning methods. In the face of a large amount of clinical diagnosis and treatment data, manual screening one by one is costly and inefficient. At the same time, when faced with redundant information, manual screening is complex and the evaluation accuracy is low. In some scenarios, the use of traditional machine learning methods cannot extract deeper medical semantic information and cannot match the corresponding relationships.

[0004] There are currently some patented methods based on big models or NLP technology to improve the accuracy of clinical trial patient matching: extract keywords from patient medical records through big models to form a disease-specific data set with medical entities and entity relationship structures, and match the corresponding clinical trials based on the knowledge graph. This method has poor generalization in scenarios where complex and long texts are processed, resulting in errors in the extracted entities and entity relationships. At the same time, the method based on rule-based keyword matching is not suitable for inclusion and exclusion criteria that include multiple ranges, and cannot accurately evaluate the degree of matching. Summary of the invention

[0005] The technical problem to be solved by the present invention is that the existing technical solutions for improving the accuracy of clinical trial patient matching based on large models or NLP technology have poor generalization in the scenario of processing complex and long texts, and are not suitable for inclusion and exclusion criteria covering multiple ranges.

[0006] In order to solve the above technical problems, one aspect of the present invention is to provide a method for clinical trial patient matching based on a large model, characterized in that it includes the following steps:

[0007] S1. Collect data, build instruction fine-tuning data sets, fine-tune pre-trained large models, and form a medical large model that can understand concepts related to clinical trials;

[0008] S2. Reconstruct the inclusion and exclusion criteria and decompose them into inclusion and exclusion criteria that are more conducive to retrieval, which specifically includes the following steps:

[0009] S2-1. Split and summarize the inclusion criteria of a single target clinical trial into multiple inclusion criteria {I 1 ,I 2 ,…,I N} and the exclusion criteria problem {E 1 ,E 2 ,…,E M}, where N represents the total number of included questions and M represents the total number of excluded questions;

[0010] S2-2. Define a single inclusion or exclusion problem as an inclusion or exclusion problem Q x , then the exclusion problem Q x Input the medical big model optimization output as multiple question query statements {Q x1 ,Q x2 ,…,Q xN}, where Q xN Representative Naqsh Problem Q x The Nth problem obtained by splitting;

[0011] S3. Combined with the inclusion and exclusion criteria, text search is performed on the patient's medical records, and the search results are organized and optimized, which specifically includes the following steps:

[0012] S3-1. Multiple query statements obtained in step S2-2 {Q x1 ,Q x2 ,…,Q xN}Extract semantic vector {S x1 ,S x2 ,…,S xN}, at the same time, the patient's medical records are divided into segments and semantic vectors are extracted {T 1 ,T 2 ,…,T M}, calculate the semantic vector {S x1 ,S x2 ,…,S xN} and the semantic vector {T 1 ,T 2 ,…,T M}, retrieve the text that matches the requirement;

[0013] S3-2. For a single standard problem Q x In the above step S3-1, a series of patient medical records that meet the search matching requirements are retrieved {Text x1 ,Text x2 ,…,Text xN}, where Text xN To represent the Nachrichten problem Q x The Nth text sentence retrieved;

[0014] S4. Combine the search results, integrate the information, construct the input, and output the answers to the exclusion questions one by one based on the medical big model reasoning;

[0015] S5. Aggregate the answers to the exclusion questions, establish an evaluation mechanism, calculate the degree of matching between clinical trials and patients, and form output results, which specifically includes the following steps:

[0016] S5-1. Use linear aggregation method to aggregate all the standard answers obtained in step S4 above, evaluate the matching degree between patients and clinical trials, and calculate the three percentage scores in the inclusion criteria: inclusion score I included , not included in the I uotincluded 、Irrelevant points I notrelevant And three percentage scores in the exclusion criteria: Exclusion score E excluded , not excluding E notexcluded , irrelevant fraction E notrelevant , the calculation formula is as follows:

[0017]

[0018]

[0019] Where N i-included Indicates that the answer is the number of inclusions, M i represents the total number of standard questions, N i-notincluded Indicates that the answer is the number not included, N i-notrelevant Indicates that the answer is an irrelevant number, N e-excluded Indicates that the answer is the number of exclusions, M e represents the total number of exclusion criteria questions, N e-notexcluded Indicates that the answer is a non-excluded number, N e-notrelevant Indicates that the answer is an irrelevant quantity;

[0020] Aggregate the six scores obtained above to get the linear aggregation score Score linear ;

[0021] S5-2. Use the large model aggregation method to aggregate all the exclusion criteria answers in a clinical trial, evaluate the matching degree between the patient and the clinical trial, and construct an aggregation question template. The model input is the exclusion criteria and all the exclusion criteria answers, and prompts to output a matching score Score I To reflect the degree of match between the patient and the clinical trial, the matching score Score I After normalization, we get Score LLM ;

[0022] S5-3. Combine the matching scores obtained by the two aggregation methods to obtain the patient matching clinical trial score Scoret :Score t =Score linear +Score LLM .

[0023] Preferably, the step S1 specifically includes the following steps:

[0024] S1-1. Collect clinical trial descriptions from the clinical trial registration database on the government website and diagnosis and treatment data from the patient medical record database;

[0025] S1-2. Merge and remove duplicate data sets and clean abnormal data;

[0026] S1-3. Construct a data set in the format of <patient-inclusion criteria-whether met>;

[0027] S1-4. Use the constructed structured dataset to fine-tune the large model to enable it to understand and process clinical trial exclusion data.

[0028] Preferably, in step S3-1, the cosine similarity is used to measure the degree of similarity between vectors, as shown in the following formula:

[0029]

[0030] In the formula, S xi represents the semantic vector of the i-th question sentence, T j Represents the semantic vector of the jth text segment extracted from the patient's medical records.

[0031] Preferably, the step S4 specifically includes the following steps:

[0032] S4-1. On the problem of quenching Q x and retrieved and reordered patient medical text {Text x1 ,Text x2 ,…,Text xN}Build model input;

[0033] S4-2. Reason based on the input based on the medical big model, output the answer, and optimize the answer to one of {compliant, inconsistent, irrelevant}.

[0034] Preferably, in step S5-1, the linear aggregation score Score linear The calculation formula is Score linear =I included -I notincluded -I notrelevant +E notexcluded -E excluded +E notrelevant .

[0035] Another aspect of the present invention is to disclose a system for clinical trial patient matching based on a large model, which is used to implement the above-mentioned clinical trial patient matching method based on a large model, characterized in that the system includes:

[0036] Model building module: used to collect data, build instruction fine-tuning data sets, fine-tune pre-trained large models, and form a medical large model that can understand concepts related to clinical trials;

[0037] Question construction module: used to reconstruct the inclusion and exclusion criteria and decompose them into inclusion and exclusion criteria questions that are more conducive to retrieval;

[0038] Information retrieval module: used to combine the inclusion and exclusion criteria to conduct text retrieval of patient medical records, and organize and optimize the retrieval results;

[0039] Model reasoning module: used to combine search results, integrate information, and construct input. Based on the medical big model reasoning, it outputs the answers to the exclusion questions one by one;

[0040] Aggregation evaluation module: used to aggregate answers to exclusion questions, establish an evaluation mechanism, calculate the degree of matching between clinical trials and patients, and form output results.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1. The present invention transforms the traditional patient inclusion into a case retrieval and question-based filtering process, in which a professional knowledge data set is constructed, and a domain big model is obtained based on big model fine-tuning. The big model's ability to understand and process long texts can greatly reduce costs while improving accuracy.

[0043] 2. The present invention can more effectively integrate answers to multiple exclusion criteria questions by integrating evaluation rules of multiple strategies, mine medical semantic information, obtain the relationship between patients and clinical trials, and effectively improve the accuracy of matching clinical trials with patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 The present invention discloses a general flow chart of a clinical trial patient matching method based on a large model;

[0045] Figure 2 A flowchart of the method for collecting data for the present invention, constructing a data set for fine-tuning instructions, fine-tuning a pre-trained large model, and forming a medical large model that can understand concepts related to clinical trials;

[0046] Figure 3 A flowchart of a method for disassembling and reconstructing the sorting criteria of the present invention into sorting criteria problems that are more conducive to retrieval;

[0047] Figure 4 The present invention combines the inclusion and exclusion criteria to conduct text retrieval of patient medical records, and organizes and optimizes the flow chart of the retrieval results;

[0048] Figure 5 This is a flowchart of a method for combining search results, integrating information, constructing input and then outputting answers to exclusion questions one by one based on medical big model reasoning;

[0049] Figure 6 A flowchart of the method for aggregating and sorting answers to questions, establishing an evaluation mechanism, calculating the degree of matching between clinical trials and patients, and forming output results for the present invention;

[0050] Figure 7 It is a schematic structural block diagram of the clinical trial patient matching system based on a large model of the present invention. DETAILED DESCRIPTION

[0051] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0052] One aspect of an embodiment of the present invention provides a method for clinical trial patient matching based on a large model, comprising the following steps:

[0053] S1. Collect data, build instruction fine-tuning data sets, fine-tune pre-trained large models, and form a medical large model that can understand concepts related to clinical trials.

[0054] In the embodiment of the present invention, step S1 specifically includes the following steps:

[0055] S1-1. Collect clinical trial descriptions from the clinical trial registration database on the government website and diagnosis and treatment data from the patient medical record database.

[0056] S1-2. Merge and remove duplicate data sets and clean abnormal data.

[0057] S1-3. Construct a data set in the format of <patient-inclusion criteria-whether met>.

[0058] S1-4. Use the constructed structured dataset to fine-tune the large model to enable it to understand and process clinical trial exclusion data.

[0059] S2. Reconstruct the inclusion and exclusion criteria and decompose them into inclusion and exclusion criteria problems that are more conducive to retrieval.

[0060] In the embodiment of the present invention, step S2 specifically includes the following steps:

[0061] S2-1. Split and summarize the inclusion criteria of a single target clinical trial into multiple inclusion criteria {I 1 ,I 2 ,…,I N} and the exclusion criteria problem {E 1 ,E 2 ,…,E M}, where N represents the total number of included questions and M represents the total number of excluded questions. In the embodiment of the present invention, the splitting algorithm used may be a natural language processing technology, such as sentence segmentation and named entity recognition assistance.

[0062] S2-2. Define a single inclusion or exclusion problem as an inclusion or exclusion problem Q x , then the exclusion problem Q x Input the medical big model optimization output as multiple question query statements {Q x1 ,Q x2 ,…,Q xN}, where Q xN Represents the Naqsh problem Q x The Nth question obtained by splitting. In the embodiment of the present invention, the medical large model prompt template for optimizing the output of the exclusion problem into multiple query statements can be constructed using the In-Context Learning training method.

[0063] S3. Combined with the inclusion and exclusion criteria, text searches are performed on the patient's medical records, and the search results are organized and optimized.

[0064] In the embodiment of the present invention, step S3 specifically includes the following steps:

[0065] S3-1. Multiple query statements obtained in step S2-2 above {Q x1 ,Q x2 ,…,Q xN}Extract semantic vector {S x1 ,S x2 ,…,S xN At the same time, the patient medical records are divided into segments and semantic vectors are extracted {T 1 ,T 2 ,…,T M}, use cosine similarity to measure the approximation between vectors, retrieve the text that matches the requirements, and the similarity calculation formula between all semantic vectors is as follows:

[0066]

[0067] In the formula, S xi represents the semantic vector of the i-th question sentence, T jRepresents the semantic vector of the jth text segment extracted from the patient's medical records.

[0068] In an embodiment of the present invention, a word vector model may be used to extract a semantic vector. Furthermore, the word vector model used is a BGE vector model.

[0069] S3-2. For a single standard problem Q x In the above step S3-1, a series of patient medical records that meet the search matching requirements are retrieved {Text x1 ,Text x2 ,…,Text xN}, where Text xN To represent the Nachrichten problem Q x The Nth text sentence retrieved. In order to highlight the more relevant and helpful texts to answer the question, and filter out low-quality fragments, the order of the texts is re-ranked based on the medical big model.

[0070] In the embodiment of the present invention, the medical large model prompt template for reordering the order of text can be constructed using the In-Context Learning training method.

[0071] S4. Combine the search results, integrate the information, and construct the input. Output the answers to the exclusion questions one by one based on the medical big model reasoning.

[0072] In the embodiment of the present invention, step S4 specifically includes the following steps:

[0073] S4-1. On the problem of quenching Q x and retrieved and reordered patient medical text {Text x1 ,Text x2 ,…,Text xN}Build model input.

[0074] S4-2. Reason based on the input based on the medical big model, output the answer, and optimize the answer to one of {compliant, inconsistent, irrelevant}.

[0075] S5. Aggregate the answers to the questions, establish an evaluation mechanism, calculate the degree of match between clinical trials and patients, and generate output results.

[0076] In the embodiment of the present invention, step S5 specifically includes the following steps:

[0077] S5-1. A linear aggregation method is used to aggregate all the standard answers obtained in step S4 to evaluate the matching degree between the patient and the clinical trial. The present invention calculates three percentage scores in the inclusion criteria: inclusion score I included , not included in the I notincluded、Irrelevant points I notrrlevant And three percentage scores in the exclusion criteria: Exclusion score E excluded , not excluding E notexcluded , irrelevant fraction E notrelevant , and its calculation formula is as follows:

[0078]

[0079] Where N i-included Indicates that the answer is the number of inclusions, M i represents the total number of standard questions, N i-notincluded Indicates that the answer is the number not included, N i-notrelevant Indicates that the answer is an irrelevant number, N e-excluded Indicates that the answer is the number of exclusions, M e represents the total number of exclusion criteria questions, N e-notexcluded Indicates that the answer is a non-excluded number, N e-notrelevant Indicates that the answer is an irrelevant quantity.

[0080] Aggregate the six scores obtained above to get the linear aggregation score Score linear , to assess the relevance of patients to clinical trials, the calculation formula is as follows:

[0081] Score linear =I included -I notincluded -I notrelevant +E notexcluded -E excluded

[0082] +E notrelevant

[0083] In the formula, Score linear The larger the value, the more patients meet the inclusion criteria and the higher the degree of matching.

[0084] S5-2. Use the large model aggregation method to aggregate all the exclusion criteria answers in a clinical trial to evaluate the matching degree between the patient and the clinical trial. Construct an aggregation question template, the model input is the exclusion criteria and all exclusion criteria answers, and prompts to output a matching score Score I To reflect the degree of match between the patient and the clinical trial, the score range is as follows:

[0085] 0 <Score I <100

[0086] Among them, Score I =0 means the patient is not related to this clinical trial. I =100, it is recommended that the patient be included in this clinical trial.

[0087] The matching score Score I After normalization, we get Score LLM .

[0088] S5-3. Combine the matching scores obtained by the two aggregation methods to obtain the patient matching clinical trial score Score t , and its calculation formula is as follows:

[0089] Score t =Score linear +Score LLM

[0090] The above formula combines two aggregation calculation methods to obtain the final score to evaluate the degree of match between patients and clinical trials. t The larger the value, the better the match between the patient and the clinical trial.

[0091] In the embodiment of the present invention, the question aggregation template can be constructed by the In-Context Learning training method.

[0092] Another aspect of the embodiments of the present invention is to provide a system for clinical trial patient matching based on a large model, which is used to implement the above method. The system includes the following modules:

[0093] Model building module: used to collect data, build instruction fine-tuning data sets, fine-tune pre-trained large models, and form a medical large model that can understand concepts related to clinical trials;

[0094] Question construction module: used to reconstruct the inclusion and exclusion criteria and decompose them into inclusion and exclusion criteria questions that are more conducive to retrieval;

[0095] Information retrieval module: used to combine the inclusion and exclusion criteria to conduct text retrieval of patient medical records, and organize and optimize the retrieval results;

[0096] Model reasoning module: used to combine search results, integrate information, and construct input. Based on the medical big model reasoning, it outputs the answers to the exclusion questions one by one;

[0097] Aggregation evaluation module: used to aggregate answers to exclusion questions, establish an evaluation mechanism, calculate the degree of matching between clinical trials and patients, and form output results.

Claims

1. A method for matching clinical trial patients based on a large model, characterized in that: The following steps are involved: S1. Collect data, build instruction fine-tuning data sets, fine-tune pre-trained large models, and form a medical large model that can understand concepts related to clinical trials; S2. Reconstruct the inclusion and exclusion criteria and decompose them into inclusion and exclusion criteria that are more conducive to retrieval, which specifically includes the following steps: S2-1. Split and summarize the inclusion criteria of a single target clinical trial into multiple inclusion criteria {I1,I2,…,I N } and the exclusion criteria problem {E1,E2,…,E M }, where N represents the total number of included questions and M represents the total number of excluded questions; S2-2. Define a single inclusion or exclusion problem as an inclusion or exclusion problem Q x , then the exclusion problem Q x Input the medical big model optimization output as multiple question query statements {Q x1 ,Q x2 ,…,Q xN }, where Q xN Representative Naqsh Problem Q x The Nth problem obtained by splitting; S3. Combined with the inclusion and exclusion criteria, text search is performed on the patient's medical records, and the search results are organized and optimized, which specifically includes the following steps: S3-1. Multiple query statements obtained in step S2-2 {Q x1 ,Q x2 ,…,Q xN }Extract semantic vector {S x1 ,S x2 ,…,S xN }, and at the same time, the patient's medical records are divided into sections and semantic vectors {T1, T2, …, T M }, calculate the semantic vector {S x1 ,S x2 ,…,S xN } and the semantic vector {T1,T2,…,T M }, retrieve the text that matches the requirement; S3-2. For a single standard problem Q x In the above step S3-1, a series of patient medical records that meet the search matching requirements are retrieved {Text x1 ,Text x2 ,…,Text xN }, where Text xN To represent the Nachrichten problem Q x The Nth text sentence retrieved; S4. Combine the search results, integrate the information, construct the input, and output the answers to the exclusion questions one by one based on the medical big model reasoning; S5. Aggregate the answers to the exclusion questions, establish an evaluation mechanism, calculate the degree of matching between clinical trials and patients, and form output results, which specifically includes the following steps: S5-1. Use linear aggregation method to aggregate all the standard answers obtained in step S4 above, evaluate the matching degree between patients and clinical trials, and calculate the three percentage scores in the inclusion criteria: inclusion score I included , not included in the I notincluded 、Irrelevant points I notrelevant And three percentage scores in the exclusion criteria: Exclusion score E excluded , not excluding E notexcluded , irrelevant fraction E notrelevant , the calculation formula is as follows: Where N i-included Indicates that the answer is the number of inclusions, M i represents the total number of standard questions, N i-notincluded Indicates that the answer is the number not included, N i-notrelevant Indicates that the answer is an irrelevant number, N e-excluded Indicates that the answer is the number of exclusions, M e represents the total number of exclusion criteria questions, N e-notexcluded Indicates that the answer is a non-excluded number, N e-notrelevant Indicates that the answer is an irrelevant quantity; Aggregate the six scores obtained above to get the linear aggregation score Score linear ; S5-2. Use the large model aggregation method to aggregate all the exclusion criteria answers in a clinical trial, evaluate the matching degree between the patient and the clinical trial, and construct an aggregation question template. The model input is the exclusion criteria and all the exclusion criteria answers, and prompts to output a matching score Score I To reflect the degree of match between the patient and the clinical trial, the matching score Score I After normalization, we get Score LLM ; S5-3. Combine the matching scores obtained by the two aggregation methods to obtain the patient matching clinical trial score Score t :Score t =Score linear +Score LLM .

2. A method for clinical trial patient matching based on a large model as claimed in claim 1, characterized in that: The step S1 specifically includes the following steps: S1-1. Collect clinical trial descriptions from the clinical trial registration database on the government website and diagnosis and treatment data from the patient medical record database; S1-2. Merge and remove duplicate data sets and clean abnormal data; S1-3. Construct a data set in the format of <patient-inclusion criteria-whether met>; S1-4. Use the constructed structured dataset to fine-tune the large model to enable it to understand and process clinical trial exclusion data.

3. A method for clinical trial patient matching based on a large model as claimed in claim 1, characterized in that: In step S3-1, the cosine similarity is used to measure the degree of similarity between vectors, as shown in the following formula: In the formula, S xi represents the semantic vector of the i-th question sentence, T j Represents the semantic vector of the jth text segment extracted from the patient's medical record.

4. A method for clinical trial patient matching based on a large model as claimed in claim 1, characterized in that: The step S4 specifically comprises the following steps: S4-1. On the problem of quenching Q x and retrieved and reordered patient medical text {Text x1 ,Text x2 ,…,Text xN }Build model input; S4-2. Reason based on the input based on the medical big model, output the answer, and optimize the answer to one of {compliant, inconsistent, irrelevant}.

5. A method for clinical trial patient matching based on a large model as claimed in claim 1, characterized in that: In step S5-1, the linear aggregation score Score linear The calculation formula is Score linear =I included -I notincluded -I notrelevant +E notexcluded -E excluded +E notrelevant .

6. A system for clinical trial patient matching based on a large model, used to implement the clinical trial patient matching method based on a large model as claimed in claim 1, characterized in that: The system comprises: Model building module: used to collect data, build instruction fine-tuning data sets, fine-tune pre-trained large models, and form a medical large model that can understand concepts related to clinical trials; Question construction module: used to reconstruct the inclusion and exclusion criteria and decompose them into inclusion and exclusion criteria questions that are more conducive to retrieval; Information retrieval module: used to combine the inclusion and exclusion criteria to conduct text retrieval of patient medical records, and organize and optimize the retrieval results; Model reasoning module: used to combine search results, integrate information, and construct input. Based on the medical big model reasoning, it outputs the answers to the exclusion questions one by one; Aggregation evaluation module: used to aggregate answers to exclusion questions, establish an evaluation mechanism, calculate the degree of matching between clinical trials and patients, and form output results.

Images