Intelligent mining method and system for potential subjects in clinical research based on cooperation of large and small models

Abstract

The invention provides a large and small model collaborative intelligent mining method and system for potential subjects in clinical research, and relates to the technical field of data analysis. The large-scale patient data is preliminarily screened and quickly matched by using efficient BM25 retrieval and a lightweight small model, so that the overall calculation cost is remarkably reduced, and the response time is shortened. Meanwhile, for boundary patients in a fuzzy zone in a screening result, a large model is called for deep semantic analysis and re-checking, the accuracy of a key decision is ensured, and the limitation of a lightweight model on complex semantic understanding is avoided. According to the hierarchical collaborative working mode, massive irrelevant data is quickly filtered through a low-cost model, and a high-computing-power model is only introduced into a key decision node, so that the optimal balance among the screening precision, the computing efficiency and the cost in large-scale real-world patient data screening is realized. The core contradiction that precision and cost are difficult to consider in the prior art is effectively solved through a large and small model cooperative working mode.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and more specifically, to a method and system for intelligently identifying potential subjects in clinical research using a combination of large and small models. Background Technology

[0002] In intelligent patient screening scenarios for clinical research, existing technical solutions generally employ a single-model architecture, with a single model handling the entire process from patient candidate recall and semantic understanding of inclusion and exclusion to final inclusion and exclusion judgment. This architecture presents significant performance trade-offs in practical implementation: if a large-scale model is used to infer for each candidate patient individually, the model call volume increases linearly with the patient scale, leading to a substantial increase in computational costs and excessively high response latency, making it difficult to meet the core business requirements of high concurrency and real-time performance. Conversely, if rule-based matching or lightweight models are used, the model's limitations result in insufficient performance in key areas such as complex inclusion and exclusion semantic parsing and implicit inclusion and exclusion condition inference, making it prone to overfitting to specific data distributions and exhibiting weak model generalization ability, ultimately leading to large fluctuations in patient screening accuracy. Therefore, existing single-model architecture screening solutions consistently face a core contradiction in large-scale patient screening tasks: the inability to coordinate and optimize accuracy and cost.

[0003] Meanwhile, in the actual operation of patient recruitment for clinical research, multiple operational constraints such as target sample size and sub-center quotas need to be taken into account. However, existing intelligent screening solutions mostly focus only on calculating the matching degree between patients and inclusion / exclusion criteria, or simply output a list of matched patients. They do not integrate actual operational needs such as recruitment quota limits and dynamic allocation of sub-center quotas into the core screening logic, resulting in screening results that cannot be directly connected to the operational requirements of actual recruitment. Summary of the Invention

[0004] The problem that this invention aims to solve is that existing single-model screening schemes always face a core contradiction in large-scale patient screening tasks, where it is difficult to coordinate and optimize accuracy and cost.

[0005] To address the aforementioned problems, in a first aspect, this invention provides a method for intelligently identifying potential clinical research subjects using a combination of large and small models, comprising: Preprocess the patient data to obtain patient description text and patient token; Clinical trial data is preprocessed to obtain standard trial texts; Input the experimental names, abstracts, and prompt word templates from the clinical trial data into the large model to obtain a clinical query word bag; Based on the patient token documents and clinical query bag, a BM25 search was performed to obtain a set of candidate patients; Input the patient description text and standard trial text into the small model to obtain the first matching probability and first matching label of the candidate patient; Based on the first matching probability, a candidate patient queue is obtained, and the candidate patients are numbered in order; When the number is less than the lower boundary of the review interval or greater than the upper boundary, the candidate patient corresponding to the number is recorded as a clear patient; When the number is within the review interval, the candidate patient corresponding to the number is recorded as a border patient, and the patient description text and standard trial text of the border patient are input into the large model to obtain the second matching probability and the second matching label; based on the first matching probability and the second matching probability of the border patient, as well as the first matching label and the second matching label, the comprehensive matching probability and the comprehensive matching label are obtained. Based on the first-match probability of clearly identified patients and the combined match probability of borderline patients, combined with the trial recruitment quota, a recommended list of subjects and corresponding matching labels were determined.

[0006] Secondly, the present invention also provides an intelligent system for identifying potential clinical trial subjects using a combination of large and small models, comprising: The patient data processing module is used to preprocess patient data to obtain patient description text and patient token; The clinical data processing module is used to preprocess clinical trial data to obtain standard trial texts; The clinical term extraction module is used to input the experimental names, abstracts, and prompt word templates from clinical trial data into the large model to obtain a clinical query term bag. The candidate patient screening module is used to perform BM25 retrieval based on patient token documents and clinical query bag-of-terms to obtain a set of candidate patients; The matching analysis module is used to input patient description text and standard trial text into a small model to obtain the first matching probability and first matching label of the candidate patients; it is also used to sort according to the first matching probability to obtain a candidate patient queue and number the candidate patients in order. The matching verification module is used to mark the candidate patient corresponding to the number as a clear patient when the number is less than the lower boundary or greater than the upper boundary of the verification interval; it is also used to mark the candidate patient corresponding to the number as a boundary patient when the number is within the verification interval, and input the patient description text and standard trial text of the boundary patient into the large model to obtain the second matching probability and the second matching label; based on the first matching probability and the second matching probability of the boundary patient, as well as the first matching label and the second matching label, the comprehensive matching probability and the comprehensive matching label are obtained. The subject recommendation module is used to determine the subject recommendation list and corresponding matching tags based on the first matching probability of clear patients and the comprehensive matching probability of boundary patients, combined with the trial recruitment quota.

[0007] This invention provides a method and system for intelligently identifying potential subjects in clinical research using a combination of large and small model approaches. Compared with existing technologies, it has the following advantages: By leveraging efficient BM25 retrieval and lightweight small models for initial screening and rapid matching of large-scale patient data, overall computational costs and response time were significantly reduced. Simultaneously, for borderline patients in the ambiguous areas of the screening results, a large model was invoked for deep semantic analysis and verification, ensuring the accuracy of critical decisions and avoiding the limitations of lightweight models in complex semantic understanding. This hierarchical and collaborative working model—BM25 coarse screening + small model fine screening + large model on-demand verification—quickly filters massive amounts of irrelevant data using low-cost models, while introducing high-performance models only at critical decision nodes. This achieves an optimal balance between screening accuracy, computational efficiency, and cost in large-scale real-world patient data screening. The collaborative working model effectively resolves the core contradiction in existing technologies where accuracy and cost are difficult to balance. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 A flowchart illustrating a method for intelligently identifying potential clinical trial subjects using a combination of large and small models, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a clinical research potential subject intelligent mining system that uses a combination of large and small models, as provided in an embodiment of the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application are described clearly and completely. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0011] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0012] like Figure 1 As shown in the embodiments of this application, a method for intelligently identifying potential clinical research subjects using a combination of large and small models is provided, comprising: S1: Preprocess the patient data to obtain the patient description text and the patient token.

[0013] S2: Preprocess the clinical trial data to obtain the standard trial text.

[0014] S3: Input the trial names, abstracts, and prompt word templates from the clinical trial data into the large model to obtain the clinical query bag-of-words. The large model refers to a model with billions or more parameters, capable of binary classification using a classification head / prompt paradigm and keyword extraction using an extractive / generative approach. Suitable models include the Qwen3-32B model, the Qwen series full-scale models, the ChatGLM3-6B model, or the Baichuan2-7B model.

[0015] S4: Perform a BM25 search based on the patient token document and the clinical query bag to obtain a candidate patient set.

[0016] S5: Input the patient description text and standard test text into the small model to obtain the first matching probability and first matching label of the candidate patient. The small model refers to a model with hundreds of millions or less of parameters, which is lightweight, easy to deploy, and focuses on text understanding / classification tasks. Small models such as the Chinese-Roberta-WWM-Ext model, TinyBERT / MediumBERT model, or Chinese-MacBERT model can be selected.

[0017] S6: Sort the candidates according to the first matching probability to obtain the candidate patient queue, and number the candidates in order.

[0018] S7: When the number is less than the lower boundary of the review interval or greater than the upper boundary, the candidate patient corresponding to the number is recorded as a clear patient.

[0019] S8: When the number is in the review interval, the candidate patient corresponding to the number is recorded as a border patient, and the patient description text and standard trial text of the border patient are input into the large model to obtain the second matching probability and the second matching label; based on the first matching probability and the second matching probability of the border patient, as well as the first matching label and the second matching label, the comprehensive matching probability and the comprehensive matching label are obtained.

[0020] S9: Based on the first matching probability of clear patients and the comprehensive matching probability of borderline patients, combined with the trial recruitment quota, determine the recommended list of subjects and their corresponding matching labels.

[0021] In this optional embodiment, efficient BM25 retrieval and lightweight small models are first used to perform preliminary screening and rapid matching of large-scale patient data, significantly reducing overall computational costs and response time. Simultaneously, for borderline patients in the ambiguous areas of the screening results, a large model is invoked for deep semantic analysis and verification, ensuring the accuracy of key decisions and avoiding the limitations of lightweight models in complex semantic understanding. This hierarchical and collaborative working mode—BM25 coarse screening + small model fine screening + large model on-demand verification—quickly filters massive amounts of irrelevant data using low-cost models, while introducing high-computing-power models only at key decision nodes. This achieves an optimal balance between screening accuracy, computational efficiency, and cost in large-scale real-world patient data screening. The collaborative working mode of large and small models effectively solves the core contradiction in existing technologies where accuracy and cost are difficult to balance.

[0022] The following is a detailed description of each step.

[0023] S1: Preprocess the patient data to obtain the patient description text and the patient token.

[0024] Read the clinical trial table and the patient table. Let's assume the i-th patient... Structured and unstructured text information are used as the raw input. The structured information includes four fields: age, gender, BMI, and blood pressure. The unstructured information includes chief complaint, present illness, diagnosis, medication history, past medical history, arcimname (examination item name), examSee (lab report), examConclusion (lab conclusion), examDiagnose (lab diagnosis), and orderSubTypeName (examination category / subtype). The above information is first summarized into a patient feature set: in, Age text, Text indicating gender, Represents BMI text, This text represents blood pressure. This indicates the main complaint text. This indicates the present medical history text. This indicates the diagnostic text. This indicates the text of medication history. Indicates historical texts. This represents the text "arcimname". This indicates the text for examSee. This represents the exam conclusion text. This indicates the text "examDiagnose". This indicates the text "orderSubTypeName".

[0025] To accommodate the different input requirements of the large and small models for subsequent sparse retrieval and deep semantics, patient data is processed in a dual-path manner.

[0026] S11: Integrate the structured and unstructured information in the patient data according to the medical narrative logic, retain field boundaries and semantic tags, and obtain the patient description text. The structured information includes age, gender, BMI and blood pressure, and the unstructured information includes chief complaint, present illness, diagnosis, medication history, past medical history, arcimname, examSee, examConclusion, examDiagnose and orderSubTypeName.

[0027] Specifically, the patient's various fields are textually integrated according to medical narrative logic to form a well-structured and coherent sequence of complete patient descriptions. This sequence explicitly preserves field boundaries and semantic labels, enabling the model to understand patient information within a unified natural language context. (Patient description text) as follows: ; .

[0028] S12: Perform word segmentation on the chief complaint text, present medical history text, diagnosis text, and test text in the patient data, and perform word frequency weighting on key fields to obtain the patient token.

[0029] Specifically, a patient-weighted token document is constructed for BM25 sparse retrieval. To improve the matching capability of key medical information in the initial screening and recall stage, this invention further constructs a sparse retrieval document from four fields: "diagnosis, examination, present medical history, and chief complaint," and applies term frequency weighting to the key fields to enhance their contribution to BM25 term frequency statistics. The patient tokens are as follows: ; in, This represents the patient's token. This indicates word segmentation processing. Indicates the main complaint text, Indicates the present medical history text, This indicates the diagnostic text. This represents the test text, where n1 and n2 represent the number of times the corresponding token is repeated. These are used to increase the frequency weight of high-value fields, thereby improving the relevance and recall quality with the test search terms. This indicates a sequential concatenation operation, which combines multiple token sequences to form a longer token document.

[0030] In the data preprocessing stage, a weighted token strategy is employed to repeatedly encode and strengthen structured fields with significant clinical meaning in patient medical records. This approach enhances the sensitivity of sparse retrieval to core medical entities, ensuring that key features have a higher weight in word frequency statistics, thereby improving the hit rate for potential subjects in the initial screening stage.

[0031] S2: Preprocess the clinical trial data to obtain the standard trial text.

[0032] Specifically, for the trial side, the j-th clinical trial It includes six core information categories: trial name, related diseases, trial summary, inclusion criteria, exclusion criteria, and recruitment quantity. To support subsequent semantic screening and large-scale model inference, the trial-side fields are concatenated into a standard trial text. : .

[0033] Through the above preprocessing, the patient side obtains data for retrieval. With used for semantic judgment The test side obtained a unified test text. This provides standardized input for subsequent stratified screening processes.

[0034] S3: Input the experimental name, abstract, and prompt word template from the clinical trial data into the large model to obtain the clinical query word bag.

[0035] Specifically, the goal of the BM25 initial screening phase is to rapidly recall a set of candidate patients potentially relevant to the current trial from the full patient pool, thereby reducing the computational cost of subsequent small and large model inference. Instead of directly using the complete trial text as the search query in the initial screening phase, semantic keywords are first generated from the trial name and abstract using a large model. Then, a sparse bag-of-words search term is constructed based on these keywords to perform the BM25 recall.

[0036] For clinical trials, semantic extraction is performed from the trial name and trial summary, and the prompt word template P is sent to the Qwen3-32B large model to form the input of the large model: The large model extracts a set of semantic keywords from clinical trials based on the input content. : in, This indicates large model processing. Represents a set of keywords The kth keyword in the text.

[0037] Subsequently, the keywords were segmented and cleaned using regular expressions to construct a sparse clinical query term bag. as follows: in, Represents a set of keywords The mth keyword in the text This indicates word segmentation processing.

[0038] Using a more focused, noise-reduced set of keywords during the retrieval phase will improve retrieval hit rate and recall quality, and reduce unnecessary recall scope.

[0039] S4: Based on the patient token documents and the bag-of-clinical-queries, perform a BM25 search to obtain a candidate patient set. This step specifically includes the following:

[0040] S41: Perform a BM25 search based on the patient token and clinical query bag to obtain a preliminary screening score for each patient.

[0041] Specifically, for any patient token, the BM25 scoring function (coarse screening score) is defined as: in, t ( , ) represents the word frequency of keyword q in the patient token, | | represents the length of the patient's token. The average length of the corpus, 1 is the word frequency saturation adjustment parameter. Adjust the parameters for document length normalization. The total number of documents in the corpus. This represents the number of documents containing the keyword q.

[0042] S42: Based on the preliminary screening score, select a specified number of patients with the highest preliminary screening score to obtain a candidate patient set.

[0043] Specifically, all patients are ranked according to their initial screening scores, and the top k patients with the highest scores are selected as the candidate set. This candidate patient set is denoted as: in, Indicating clinical trials The corresponding candidate patient set, This indicates that the top k patients in the sequence are selected as the candidate patient set. This indicates the patient's code.

[0044] In the initial screening stage, this application does not directly use the original text of clinical trials for matching. Instead, it utilizes a large-scale model to perform deep semantic understanding of the trial names and abstracts, generating a keyword set containing synonyms, hyponyms, and related medical terms. This strategy effectively bridges the semantic gap between standardized trial inclusion and exclusion criteria and unstructured, conversational electronic medical record text, significantly improving the recall rate for the target population.

[0045] S5: Input the patient description text and standard trial text into the small model to obtain the first matching probability and first matching label of the candidate patient.

[0046] Specifically, the goal of the deep semantic screening stage is to perform more granular semantic discrimination on patients in the candidate patient set, thereby outputting the matching probability between patients and trials, and providing a basis for ranking and constructing uncertainty intervals (verification intervals). A small model is obtained by training the chinese-roberta-wwm-ext model, enabling the trained model to perform binary classification of trial text and patient text pairs.

[0047] S51: Concatenate and fuse the patient description text and the standard test text to obtain the fused input sequence.

[0048] Specifically, regarding the standard test text Patient description text Construct the input sequence: Where [CLS] refers to the classification start marker of the Transformer input sequence, and [SEP] refers to the separator marker in the Transformer input. This indicates a sequential splicing operation, splicing standard trial text and patient description text to form an input sequence.

[0049] S52: The fused input sequence is processed by the word segmenter to obtain word embedding, sentence embedding and position embedding.

[0050] Specifically, after the experimental text and patient text are concatenated to form the input sequence, the tokenizer processes it to obtain three types of indexes: the first is a token ID that maps each token to a number in the vocabulary, used to look up the corresponding lexical embedding. Secondly, it distinguishes between the test segment and the patient's segment ID, which is used to look up the segment embedding in the table. Thirdly, it represents the sequential position ID of each token in the sequence, which is used to look up the position embedding in a table. .

[0051] S53: Add word embeddings, sentence embeddings, and position embeddings element by element to obtain the input embedding representation.

[0052] Specifically, the representation vector at each position in the fused input sequence is obtained by adding the three types of embeddings element-wise, thus forming the final input embedding representation: S54: Input the input embedding representation into the small model to obtain the first matching probability and the first matching label.

[0053] Specifically, the input embedding representation is used as the input to the Transformer network, and the output feature sequence is obtained through multi-layer encoding of the Transformer: Take the hidden layer vector at the first position [ ] As a global semantic representation, its dimension is 768. To transform the 768-dimensional semantic vector into a scalar score usable for binary classification, the model uses a linear classifier head at the top level, outputting an unnormalized matching score. Then through the sigmoid function Map the score to a probability value in the interval [0,1]: During the model training phase, the true labels of each training sample are used... With model output probability Calculate the binary cross-entropy loss: The small model is trained based on the aforementioned loss, and after multiple iterations of optimization, a small model with convergent loss is obtained. This small model is used for fine screening to provide candidate patients with a first-order matching probability that can be ranked. Matching the first tag This also provides a basis for judgment when calling larger models later.

[0054] Before using the first matching probability output by the small model to screen patients who need to be reviewed, it is necessary to construct a dynamic review interval based on the trial recruitment quota and proportion coefficient.

[0055] Specifically, to further achieve a balance between accuracy and cost, a review interval construction mechanism based on trial recruitment quotas is introduced in the inference stage: first, the candidate patient set is sorted according to the first matching probability in the small model's fine screening results, and then combined with the trial recruitment quotas. By constructing a verification interval, the large model verification is triggered only for difficult samples (boundary samples) within the interval, thereby significantly reducing the number of large model calls while ensuring high-quality recommendations.

[0056] The lower boundary of the review interval is ; The upper boundary of the verification interval is ; in, Indicates the number of positions available for the trial. This indicates the total number of patients in the candidate patient set. Indicates rounding down. This indicates rounding up to the nearest integer. This represents a proportional coefficient indicating the width of the control interval, used to control the scope of large-scale model verification, that is, to balance cost and effectiveness. It can be understood as: [regarding the quota]. Fluctuating up or down by a certain percentage yields an interval. The lower bound is... Expand the interval to the left of the quota, with the upper bound using Extend the interval to the right of the quota. For example, if =100, =0.3, then the theoretical interval is At this point, the large model only reviews patients ranked around 70 to 130, because these patients are close to the "recruitment boundary" and are most prone to misjudgment, thus having the highest review value. The larger the value, the wider the uncertainty interval, the more patients are validated by the large model, and the higher the accuracy is usually, but the cost is also higher. The smaller the size, the narrower the scope of review, resulting in lower costs but a higher risk of boundary misjudgment.

[0057] By incorporating trial recruitment quotas into the final subject recommendation logic, the screening results can directly meet the recruitment needs of actual operations, avoiding the problem of screening results being out of sync with actual recruitment goals in traditional schemes. This improves the overall efficiency and practicality of intelligent discovery of potential subjects for clinical research.

[0058] S6: Sort the candidates according to the first matching probability to obtain the candidate patient queue, and number the candidates in order.

[0059] Specifically, the candidates are sorted in descending order based on the first matching probability to obtain a queue of candidates, and then numbered sequentially.

[0060] S7: When the number is less than the lower boundary of the review interval or greater than the upper boundary, the candidate patient corresponding to the number is recorded as a clear patient.

[0061] Specifically, a differentiated processing strategy is adopted for different regions: for those ranked in Previous patients were directly considered highly matched patients, and their first match probability and first match label, output by the fine screening small model, were directly used to enter the recommendation priority queue; for those ranked in... Subsequent patients are considered low-matched and do not need to proceed to the next review process; they can directly use the first match probability and first match label output by the fine screening model. Clearly matched patients include both high-matched and low-matched patients.

[0062] S8: When the number is in the review interval, the candidate patient corresponding to the number is recorded as a border patient, and the patient description text and standard trial text of the border patient are input into the large model to obtain the second matching probability and the second matching label; based on the first matching probability and the second matching probability of the border patient, as well as the first matching label and the second matching label, the comprehensive matching probability and the comprehensive matching label are obtained.

[0063] Specifically, for those in For patients within a given range, the Qwen3-32B large-scale model was used to perform deep inference and verification on patient description texts and standard trial texts. The output of the large-scale model was: in, Indicates the second matching probability. Indicates the second matching tag. This indicates the large-scale model processing. The aforementioned differentiated processing strategy allows the large model to focus on serving boundary samples, improving the accuracy of the final list, while avoiding high-cost inference on all candidates.

[0064] This application does not perform large-scale model inference on all samples. Instead, after a small model provides matching probabilities and ranks candidate patients, a dynamic verification interval is constructed for a sample set near the ranking boundary. Only difficult samples falling within this interval and at the decision boundary are called to perform deep inference using the large model, achieving on-demand allocation of computing resources and significantly reducing computational overhead while maintaining high accuracy.

[0065] In the fusion and discrimination stage, the results of the small model and the large model are weighted and fused to obtain the final matching score (i.e., the comprehensive matching probability). A consistency-first strategy is then used to determine the final label to improve decision stability and interpretability. The comprehensive matching probability is... in, Indicates the overall matching probability. Indicates the probability of the first match. Indicates the second matching probability. This represents the fusion weight, used to control the contribution ratio of the small model and the large model to the final score.

[0066] The overall matching label adopts a consistency-first strategy, which selects the first matching label output by the smaller model. The second matching label output by the large model The system outputs a binary conclusion determining whether a patient meets the inclusion requirements, improving the eligibility and interpretability of the recommended list. The comprehensive matching labels are... in, This indicates a comprehensive matching of tags. Indicates the first matching tag. This indicates the second matching tag, where 1 indicates a high degree of matching and 0 indicates a low degree of matching.

[0067] S9: Based on the first matching probability of clear patients and the comprehensive matching probability of borderline patients, combined with the trial recruitment quota, determine the recommended list of subjects and their corresponding matching labels.

[0068] Specifically, patients are reordered based on the first matching probability of highly matched and poorly matched patients and the combined matching probability of borderline patients, and the top-ranked patients are truncated. [Number] patients were included in the recommended enrollment list (i.e., the subject recommendation list): in, Indicating clinical trials The corresponding recommended list of subjects, Indicating clinical trials With patients The overall matching probability obtained during pairing. This indicates extracting the first part of the sequence. A candidate patient set of [number] patients is selected. Simultaneously, corresponding matching labels indicating "recommendation / eligibility" are output. This set is written into a standard format file and provided to the clinical team for actual patient recruitment, facilitating selection. Through comprehensive judgment using multiple models and dimensions, a highly confident and interpretable recommended enrollment list can be generated, providing accurate and reliable subject recruitment support for clinical research teams.

[0069] like Figure 2 As shown in the embodiments of this application, a clinical research potential subject intelligent mining system based on large and small model collaboration includes: The patient data processing module 10 is used to preprocess patient data to obtain patient description text and patient token.

[0070] The clinical data processing module 20 is used to preprocess clinical trial data to obtain standard trial texts.

[0071] The clinical term bag extraction module 30 is used to input the experimental name, abstract and prompt word template from the clinical trial data into the large model to obtain the clinical query term bag.

[0072] The candidate patient screening module 40 is used to perform BM25 retrieval based on the patient token document and the clinical query bag to obtain a candidate patient set.

[0073] The matching analysis module 50 is used to input patient description text and standard trial text into the small model to obtain the first matching probability and first matching label of the candidate patients; it is also used to sort according to the first matching probability to obtain a candidate patient queue and number the candidate patients in order.

[0074] The matching verification module 60 is used to record the candidate patient corresponding to the number as a clear patient when the number is less than the lower boundary or greater than the upper boundary of the verification interval; it is also used to record the candidate patient corresponding to the number as a boundary patient when the number is in the verification interval, and input the patient description text and standard trial text of the boundary patient into the large model to obtain the second matching probability and the second matching label; based on the first matching probability and the second matching probability of the boundary patient, as well as the first matching label and the second matching label, the comprehensive matching probability and the comprehensive matching label are obtained.

[0075] The subject recommendation module 70 is used to determine the subject recommendation list and corresponding matching tags based on the first matching probability of clear patients and the comprehensive matching probability of boundary patients, combined with the trial recruitment quota.

[0076] In this optional embodiment, the matching analysis module 50 first uses a small model to quickly match and rank a large number of candidate patients, significantly reducing the number of patients requiring processing by a large model. This effectively avoids the linear increase in computing power costs and excessive response latency caused by a single full-scale large model architecture. The matching verification module 60 only calls the large model for deep semantic analysis on boundary patients in the verification interval during ranking, making accurate inferences on complex inclusion and exclusion conditions and implicit conditions, overcoming the limitation of weak generalization ability of lightweight models in key links. The subject recommendation module 70 combines the first matching probability of clear patients with the comprehensive matching probability of boundary patients, and dynamically determines the final recommendation list strictly according to the trial recruitment quota, ensuring that the screening results can be directly adapted to the target sample size and sub-center quotas and other actual operational constraints. This modular collaborative design achieves optimal allocation of computing resources through a hierarchical processing mechanism, keeping the number of large model calls to a minimum while maintaining high accuracy. At the same time, it embeds operational requirements into the core screening logic, fundamentally eliminating the contradiction between accuracy and cost in existing technologies, and solving the pain point that screening results cannot be matched with actual recruitment requirements.

[0077] This application utilizes a tiered, combined large and small model approach for clinical trial subject screening. Through a processing flow of "BM25 high-efficiency coarse screening + small model in-depth fine screening + large model on-demand verification," it automatically identifies and screens candidate subjects who highly match the inclusion and exclusion criteria of the target clinical trial. Compared to existing approaches that rely on human experience for patient screening, rely entirely on standardized or lightweight models, or rely entirely on large models, this application offers the following advantages: 1. Significantly reduce inference costs and system latency while ensuring the accuracy of screening.

[0078] This application adopts a collaborative architecture of "BM25 coarse screening—RoBERTa / BERT fine screening—large model on-demand verification," and through a confidence-based adaptive invocation strategy, triggers large model inference only for difficult samples that are difficult for the small model to determine, thereby significantly reducing the number of large model calls. This effectively controls computational power consumption and inference latency in scenarios involving massive patient data screening, improving system throughput and scalability, making the method more suitable for high-concurrency, real-time or near-real-time clinical recruitment needs.

[0079] 2. Significantly improves robustness and cross-institutional adaptability to complex texts of real clinical medical records.

[0080] To address the common characteristics of clinical medical records, such as long texts, noisy information, mixed Chinese and English, and a high density of Latin abbreviations and technical terms, this application reduces the risk of feature fragmentation due to word segmentation errors by employing domain-specific medical terminology normalization rules and a regularization engine. Simultaneously, it utilizes structured serialization technology to reorganize scattered heterogeneous fields into a coherent natural language context according to clinical logic, enhancing the model's ability to understand negative expressions, threshold conditions, and contextual dependencies. This improves the stability, consistency, and transferability of screening results, reduces dependence on specific hospital field systems, and lowers the adaptation costs for cross-center deployment.

[0081] 3. Form a verifiable, measurable, and iterative engineering closed loop to improve clinical operational availability.

[0082] This application provides detailed and fused results of multi-model scores and labels at the output level, making the screening process traceable and reviewable, which is convenient for researchers to review and control quality. At the same time, it directly incorporates operational constraints such as actual recruitment quotas into the result truncation and output logic, so that the algorithm output is naturally aligned with the clinical recruitment plan, improving the feasibility and efficiency of clinical operations.

[0083] An electronic device provided in this application includes a memory and a processor; the memory is used to store a computer program; the processor is used to implement the above-described method for intelligently identifying potential clinical research subjects using a combination of large and small models when executing the computer program.

[0084] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the above-described method for intelligently identifying potential clinical research subjects through a combination of large and small models.

[0085] In this embodiment, the beneficial effects of electronic devices and computer-readable storage media are similar to those of the above-described method for intelligently identifying potential clinical trial subjects in a synergistic manner with size models, and will not be repeated here.

[0086] The present invention describes electronic devices that can serve as servers or clients of this application, which are examples of hardware devices that can be applied to various aspects of this application. Electronic devices are intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices can also represent various forms of mobile devices, such as personal digital assistant devices, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the application described and / or claimed herein.

[0087] Electronic devices include a computing unit that can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). The RAM can also store various programs and data required for device operation. The computing unit, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0088] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. In this application, the separately described modules may or may not be physically separate. Some or all of the modules can be selected to achieve the purpose of the embodiments of this application according to actual needs. Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0089] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0090] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for intelligently identifying potential subjects in clinical research using a combination of large and small models, characterized in that, include: Preprocess the patient data to obtain patient description text and patient token; Clinical trial data is preprocessed to obtain standard trial texts; Input the experimental names, abstracts, and prompt word templates from the clinical trial data into the large model to obtain a clinical query word bag; Based on the patient token documents and clinical query bag, a BM25 search was performed to obtain a set of candidate patients; Input the patient description text and standard trial text into the small model to obtain the first matching probability and first matching label of the candidate patient; Based on the first matching probability, a candidate patient queue is obtained, and the candidate patients are numbered in order; When the number is less than the lower boundary of the review interval or greater than the upper boundary, the candidate patient corresponding to the number is recorded as a clear patient; When the number is within the review interval, the candidate patient corresponding to the number is recorded as a border patient, and the patient description text and standard trial text of the border patient are input into the large model to obtain the second matching probability and the second matching label; based on the first matching probability and the second matching probability of the border patient, as well as the first matching label and the second matching label, the comprehensive matching probability and the comprehensive matching label are obtained. Based on the first-match probability of clearly identified patients and the combined match probability of borderline patients, combined with the trial recruitment quota, a recommended list of subjects and corresponding matching labels were determined.

2. The method for intelligently identifying potential clinical research subjects using a combination of large and small models as described in claim 1, characterized in that, The process involves performing a BM25 search based on the patient token and clinical query bag-of-terms to obtain a candidate patient set, including: Based on the patient token and clinical query bag, a BM25 search was performed to obtain a preliminary screening score for each patient; Based on the preliminary screening score, a specified number of patients with the highest preliminary screening score are selected to obtain the candidate patient set.

3. The method for intelligently identifying potential clinical research subjects using a combination of large and small models as described in claim 1, characterized in that... The process of inputting patient description text and standard trial text into a small model to obtain the first matching probability and first matching label of the candidate patient includes: The patient description text and the standard trial text are concatenated and fused to obtain the fused input sequence; The fused input sequence is processed by the word segmenter to obtain word embeddings, sentence embeddings, and position embeddings; The word embedding, sentence embedding, and position embedding are added element by element to obtain the input embedding representation; The input embedding representation is fed into the small model to obtain the first matching probability and the first matching label.

4. The method for intelligently identifying potential clinical research subjects using a combination of large and small models as described in claim 1, characterized in that, Before recording the candidate patient corresponding to the number as a clear patient when the number is less than the lower boundary or greater than the upper boundary of the review interval, the method further includes: Based on the number of recruits and the proportional coefficient for the trial, a dynamic review interval is obtained.

5. The method for intelligently identifying potential clinical research subjects using a combination of large and small models as described in claim 4, characterized in that... The lower boundary of the review interval is ; The upper boundary of the verification interval is ; in, Indicates the number of positions available for the trial. This indicates the total number of patients in the candidate patient set. The proportional coefficient representing the width of the control interval. Indicates rounding down. This indicates rounding up to the nearest integer.

6. The method for intelligently identifying potential clinical research subjects using a combination of large and small models as described in claim 1, characterized in that, The overall matching probability is ； in, Indicates the overall matching probability. Indicates the probability of the first match. Indicates the second matching probability. This indicates the fusion weight.

7. The method for intelligently identifying potential clinical research subjects using a combination of large and small models as described in claim 1, characterized in that, The comprehensive matching tags are: ； in, This indicates a comprehensive matching of tags. Indicates the first matching tag. This indicates the second matching tag, where 1 indicates a high degree of matching and 0 indicates a low degree of matching.

8. The method for intelligently identifying potential clinical trial subjects using a combination of large and small models as described in claim 1, characterized in that, The preprocessing of patient data to obtain patient description text and patient token includes: The structured and unstructured information in the patient data are integrated into the text according to the medical narrative logic, and field boundaries and semantic tags are preserved to obtain the patient description text. The structured information includes age, gender, BMI and blood pressure, and the unstructured information includes chief complaint, present illness, diagnosis, medication history, past medical history, arcimname, examSee, examConclusion, examDiagnose and orderSubTypeName. The chief complaint text, present medical history text, diagnosis text, and test text in the patient data are segmented into words, and the key fields are weighted by word frequency to obtain the patient token.

9. The method for intelligently identifying potential clinical research subjects using a combination of large and small models as described in claim 8, characterized in that, The patient's token is ； in, This represents the patient's token. This indicates word segmentation processing. Indicates the main complaint text, Indicates the present medical history text, This indicates the diagnostic text. This represents the verification text, where n1 and n2 represent the number of times the corresponding token is repeated. This indicates a sequential splicing operation.

10. A system for intelligently identifying potential clinical trial participants using a combined large and small model approach, characterized in that, include: The patient data processing module is used to preprocess patient data to obtain patient description text and patient token; The clinical data processing module is used to preprocess clinical trial data to obtain standard trial texts; The clinical term extraction module is used to input the experimental names, abstracts, and prompt word templates from clinical trial data into the large model to obtain a clinical query term bag. The candidate patient screening module is used to perform BM25 retrieval based on patient token documents and clinical query bag-of-terms to obtain a set of candidate patients; The matching analysis module is used to input patient description text and standard trial text into a small model to obtain the first matching probability and first matching label of the candidate patients; it is also used to sort according to the first matching probability to obtain a candidate patient queue and number the candidate patients in order. The matching verification module is used to mark the candidate patient corresponding to the number as a clear patient when the number is less than the lower boundary or greater than the upper boundary of the verification interval; it is also used to mark the candidate patient corresponding to the number as a boundary patient when the number is within the verification interval, and input the patient description text and standard trial text of the boundary patient into the large model to obtain the second matching probability and the second matching label; based on the first matching probability and the second matching probability of the boundary patient, as well as the first matching label and the second matching label, the comprehensive matching probability and the comprehensive matching label are obtained. The subject recommendation module is used to determine the subject recommendation list and corresponding matching tags based on the first matching probability of clear patients and the comprehensive matching probability of boundary patients, combined with the trial recruitment quota.

Images