A multi-path fusion learning method for medical record quality control of large medical models

Abstract

The present invention discloses a multi-path fusion learning method for medical record quality control of a large model. First, a reasonable prompt is designed to mine the thinking mode and step-by-step reasoning process of large models with large parameters of different series. Then, a multi-path fusion data construction strategy is adopted to guide the learning of large models with smaller parameters, so that they can finally give medical record quality control results with more comprehensive thinking angles and clearer reasoning logic. The present invention fully mines the thinking modes and thinking paths of large models of different series, and enhances the thinking and logical reasoning ability of large models with smaller parameters in a multi-path fusion learning manner, so that they can give more comprehensive, logical and more interpretable medical record quality control results.

Description

Technical Field

[0001] The invention relates to a method for medical record quality control using a large model, and belongs to the medical field. Background Art

[0002] As a core resource in the medical field, the complexity and diversity of electronic medical records have a profound impact on many subsequent medical tasks. Therefore, it is particularly important and challenging to implement strict quality management control (QC) of electronic medical records. With the rapid development and widespread application of large language models, they have also been introduced into the task of medical record QC. With its massive knowledge reserves and advanced underlying architecture, large models can conduct in-depth QC on various types of information in medical electronic medical records at the connotation level. However, in the field of medicine, which requires extremely high rigor and logic, the quality control results of large models often lack sufficient interpretability. On the one hand, its answers are difficult to present clearly and concisely in the form of medical diagnosis and treatment pathways; on the other hand, the purposeful approach is limited to the single-series, single-angle large model thinking mode, which is easy to fall into cognitive misunderstandings.

[0003] In the patent with application number CN202410219108.3 published on April 2, 2024, a medical record quality control method based on a large model is introduced: first, the medical record related information that needs quality control is structured and extracted, including current medical history, past history, allergy history and other diagnosis and treatment information; then it is sent to the standard treatment model, efficacy model, and node information model to obtain the parity plan information, efficacy information, and treatment course information, and the corresponding index is calculated by formula, and the final medical record quality control base is added. On the one hand, this method is relatively cumbersome and relies on professionals to formulate the calculation formula of the corresponding index. At the same time, the information output by the large model is not interpretable, that is, the thinking process of the plan, efficacy, and treatment course information given by the large model is not transparent enough, resulting in the final quality control results being not convincing enough.

[0004] The patent application number CN202410406561.5 published on May 7, 2024, involves a medical record quality control optimization method: first, construct defect instructions based on quality control rules and medical records to be quality controlled, and send them to the large model for training and learning, with the goal of generating defective medical record samples corresponding to the rules; then, the medical records to be quality controlled and defective medical record samples are combined to construct a defect identification model to determine whether the medical records have defects. The effect of this method is severely limited to the quality of defective samples constructed by a single large model, and the judgment results of the defect identification model lack a clear thinking process and are not interpretable. Summary of the invention

[0005] The problem to be solved by the present invention is that the current method of using large models for medical record quality control has defects such as weak interpretability and logic of the large model quality control answers, and the single thinking mode.

[0006] To solve the above technical problems, the technical solution of the present invention discloses a medical large model medical record quality control method based on multi-path fusion learning, which is characterized by including the following steps:

[0007] Step 1: Design a reasonable and general large model medical record quality control ability mining instruction template, which combines the original medical record text with medical record quality control rules to generate the final input prompt.

[0008] Step 2: Send the input prompt obtained in Step 2 into large parameter models of different series to obtain the thinking steps and step-by-step reasoning processes independently planned by all large parameter models.

[0009] Step 3: After collecting multi-path answers from large parameter models of different series, construct the final instruction fine-tuning training set through a multi-path fusion strategy.

[0010] Step 4: Send the instruction fine-tuning training set into a large model with smaller parameters to adjust the parameter weights of the large model with smaller parameters so that it fits the correct reasoning path.

[0011] Step 5: After parameter adjustment, the large model with smaller parameters, after absorbing the knowledge guidance of large parameter models of different series, can give a medical record quality control answer with a more comprehensive thinking angle, more rigorous logical reasoning, and stronger interpretability.

[0012] Preferably, in Step 1, the large model medical record quality control ability mining instruction template integrates the following four key elements:

[0013] Introduce the words and expressions of the medical record text to be quality controlled and the quality control rules;

[0014] Concatenate the original text of the medical record text to be quality controlled, the quality control rule description text, and the above-introduced words and expressions;

[0015] Specify the words and expressions that require the large model to first plan the thinking steps and then perform step-by-step logical reasoning;

[0016] Specify the language style and format for the large model to return the quality control result.

[0017] Preferably, in Step 2, in order to ensure that large parameter models of different series can accurately output in the required format, the Few-shot learning method is adopted. By pre-writing two correct <prompt, expected answer> pairs as historical conversation records and concatenating them with the current prompt and sending them into the large parameter model, for the same medical record and quality control rules, multiple answers with different thinking angles and reasoning logic paths can be obtained, and each answer includes the thinking steps and step-by-step reasoning process of the large parameter model for the current medical record quality control task.

[0018] Preferably, in step 3, the multi-path fusion strategy includes:

[0019] Step 301: Use regular expressions to clean the multi-path answers output by the large parameter model, including:

[0020] Use the re.sub() function in the regular expression to uniformly remove special characters;

[0021] Use the re.sub() function in regular expressions to filter out useless words;

[0022] Adopting the method of "first step +? second step +? third step +", the output of the large model is divided into the planning thinking steps, the process of step-by-step reasoning, and the final conclusion. If it is impossible to divide it into three steps, manual intervention and correction are performed to obtain a multi-path answer in a unified format.

[0023] Step 302: A similarity deduplication algorithm in the field of natural language processing is used to screen the cleaned data:

[0024] For each answer text obtained in step 301, pair it with all other answers in pairs to form multiple candidate pairs;

[0025] For each pair of candidates, the two answer texts obtained in the previous step are converted into vector form through the Word2vec model;

[0026] The cosine similarity is used to calculate the similarity between the two vectors: if the similarity calculation result of a pair of candidates exceeds the threshold of 0.95, it means that the two answers are highly similar in content. In order to maintain the diversity of the data, only one of the answers is retained and the other is considered redundant information and removed;

[0027] Step 303: After cleaning and deduplication, a multi-path answer set with high consistency and richness is finally obtained. Each answer is spliced ​​with its input prompt to finally obtain a set of question-answer pairs that can be used for training, that is, the instruction fine-tuning training set.

[0028] Preferably, in step 4, the instruction fine-tuning training set obtained in step 3 is randomly shuffled to form a multi-path fusion enhanced training set, and then based on the concept of transfer learning, a large model with fewer parameters is selected, and the multi-path fusion enhanced training set is used for fine-tuning training.

[0029] Preferably, during fine-tuning training:

[0030] The input text is converted into a machine-understandable vector form. These vectors are forward propagated according to the preset network architecture and calculation formula to obtain the probability prediction of each character at the current position. The large model compares these predicted probabilities with the correct labels and calculates the difference between the predicted value and the true value through the cross entropy loss function. The smaller the cross entropy loss value, the closer the prediction result of the large model is to the actual situation.

[0031] The goal of adjusting the model parameter weights is to increase the probability of the correct character appearing at the current position, that is, to minimize the value of the cross entropy loss function; by continuously repeating the training process, that is, letting the large model repeatedly learn the samples in the training set, we can eventually get an optimized large model enhanced by multi-path fusion.

[0032] The present invention discloses a multi-path fusion learning method for medical record quality control of a large model. First, a reasonable prompt is designed to mine the thinking mode and step-by-step reasoning process of large models with large parameters of different series. Then, a multi-path fusion data construction strategy is adopted to guide the learning of large models with smaller parameters, so that they can finally give medical record quality control results with more comprehensive thinking angles and clearer reasoning logic. The present invention fully mines the thinking modes and thinking paths of large models of different series, and enhances the thinking and logical reasoning ability of large models with smaller parameters in a multi-path fusion learning manner, so that they can give more comprehensive, logical and more interpretable medical record quality control results.

[0033] Compared with the prior art solutions, the present invention has the following beneficial effects:

[0034] 1. Unlike other inventions that only extract guidance data from a large model, the present invention deeply explores and obtains the patterns and paths of the large model's self-planning steps and step-by-step thinking for specific tasks. This process is more in-depth and replicable.

[0035] 2. In the medical field, which requires extremely high rigor and logic, this invention significantly improves the logical clarity and explainability of answers by deeply exploring the thinking patterns and paths of large models.

[0036] 3. Compared with those inventions that are limited to a single large model data or thinking mode, the present invention comprehensively learns from different series of large models to obtain more comprehensive and diverse answers from thinking angles. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0038] 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.

[0039] This paper proposes a medical record quality control method for large medical models based on multi-path fusion learning. The core of this method is to deeply explore the thinking patterns and paths of different series of large models, and use the data construction strategy of multi-path fusion to guide the training and learning of large models with relatively small parameters to achieve knowledge transfer.

[0040] Specifically, the method disclosed in the present invention first designs a set of reasonable and universal large-model medical record quality control capability mining instruction templates. The template combines the original medical record text with the medical record quality control rules to generate the final input prompt. Subsequently, this prompt is sent to different series of large parameter models to obtain their autonomously planned thinking steps and step-by-step reasoning processes.

[0041] After collecting multi-path answers from multiple series of large models, the final instruction fine-tuning training set is constructed through a multi-path fusion strategy. Then, this training set is sent to a large model with smaller parameters, and training methods such as Fine-tune or Lora are used to adjust the model's parameter weights so that it fits the correct reasoning path.

[0042] After parameter adjustment, the smaller parameter large model, after absorbing the knowledge guidance of multiple series of large models, can give medical record quality control answers with more comprehensive thinking angles, more rigorous logical reasoning, and stronger explanatory power.

[0043] (I) Constructing large-scale model medical record quality control capability mining instructions

[0044] In order to construct a reasonable and universal medical record quality control capability mining instruction, the present invention integrates the following four key elements into the instruction template:

[0045] (1) Introducing the medical record text to be quality controlled and the quality control rules, such as "The medical record text to be quality controlled is..." and "The current quality control rule description is...";

[0046] (2) Splice the original text of the medical record to be quality-controlled and the text describing the quality-control rules with the above-mentioned introduction words;

[0047] (3) Clearly require the big model to plan the thinking steps before performing step-by-step logical reasoning, such as “Step 1: Please plan the necessary steps to determine whether the current quality control rules have been passed; Step 2: Complete the reasoning step by step according to the set steps”;

[0048] (4) Specify the language style and format for the large model to return the quality control results, such as "Please output in JSON format, summarize, and give the final result of passing or failing."

[0049] By assembling the above elements in sequence, a common input prompt can be generated for multiple series of large parameter models.

[0050] (2) Obtain multi-path answers from multiple series of large parameter large models

[0051] To obtain multi-angle and comprehensive step-by-step reasoning answers, the present invention selects large parameter models of different base series and uses the above constructed prompt as input. To ensure that the large model without special training can accurately output in the required format. The present invention adopts the Few-shot learning method. By pre-writing two correct <prompt, expected answer> pairs as historical conversation records and splicing them with the current prompt and sending them into the model.

[0052] In this way, for the same medical record and quality control rules, answers with multiple thinking angles and reasoning logic paths can be obtained. Each answer contains the thinking steps and step-by-step reasoning process of the large model for the current medical record quality control task.

[0053] Specifically, the generative large model first converts each input character into a vector representation through the neural network vector layer. Then, according to the preset network architecture, the model performs numerical calculations on the vectors of the first N characters to predict the probability of each character appearing at the current N+1th position. Next, the model selects the character with the highest probability as the output at the current position. This process is repeated continuously until a specific termination symbol is encountered, thereby generating a complete output answer.

[0054] (3) Cleaning, filtering, and assembling of multi-path answers

[0055] To ensure the consistency of the final training set in terms of language style, output format, content, etc., the present invention performs a series of data processing on the obtained multi-path answers:

[0056] (1) Cleaning: To ensure the consistency of the large model's output in details such as text style and output format, the present invention uses regular expressions for cleaning;

[0057] (a) Use the re.sub() function in regular expressions to uniformly remove special characters such as "* - &", etc.;

[0058] (b) Use the re.sub() function in regular expressions to filter out useless words such as "Okay, I will follow your requirements...";

[0059] (c) Use regular expressions like “first step.+? second step.+? third step.+” to divide the output of the large model into the planning thinking steps, the step-by-step reasoning process, and the final conclusion. If it is impossible to divide it into three steps, manual intervention is required to correct it, and finally a multi-path answer in a unified format is obtained.

[0060] (2) Deduplication: In order to ensure the diversity and richness of the data, the present invention uses a similarity deduplication algorithm in the field of natural language processing for screening.

[0061] Specifically, for each answer text in the set, it will be paired with all other answers to form multiple candidate pairs. For each candidate pair, the two answer texts are converted into vector form through the Word2vec model. Then, the cosine similarity metric is used to calculate the similarity between the two vectors.

[0062] If the similarity calculation result of a pair of candidates exceeds the threshold of 0.95, it means that the two answers are highly similar in content. In order to maintain data diversity, only one of the answers is retained and the other is considered redundant information and eliminated.

[0063] (3) Assembly: After cleaning and deduplication, we can finally obtain a set of multi-path answers that are highly consistent and rich. Each answer is concatenated with its input prompt to finally obtain a set of question-answer pairs that can be used for training.

[0064] (IV) Multi-path fusion training and learning

[0065] The question-answering data set obtained through the above steps is randomly shuffled to form a multi-path fusion enhanced training set. Based on the concept of transfer learning, the present invention selects a large model with fewer parameters and uses the training set for fine-tuning training.

[0066] Taking into account that large models with large parameters have low efficiency in reasoning, the present invention selects a large model with relatively small parameters based on the idea of ​​transfer learning, and uses a multi-path fusion enhanced training set to fine-tune it.

[0067] During the specific training process, the input text will be converted into a vector form that the machine can understand. Then, these vectors will be forward propagated according to the preset network architecture and calculation formula to obtain the probability prediction of each character appearing at the current position. Then, the model compares these predicted probabilities with the correct labels and calculates the difference between the predicted value and the true value through the cross entropy loss function. The smaller the cross entropy loss value, the closer the model's prediction result is to the actual situation.

[0068] The goal of adjusting the model parameter weights is to increase the probability of the correct character appearing at the current position, that is, to minimize the value of the cross entropy loss function. By repeating the training process, that is, letting the model repeatedly learn the samples in the training set, we can eventually get an optimized large model enhanced by multi-path fusion. This model can show a more comprehensive thinking angle and stronger logic when answering questions.

[0069] (V) Predict and obtain the final large model step-by-step reasoning answer

[0070] Finally, in the actual reasoning stage, the actual medical record text and quality control rules to be controlled are converted into the autonomous decision-making guidance instructions P of the big model through step (1) and input into the big model M. The big model M that learns the thinking paths and derivation processes from multiple angles will give a more comprehensive and accurate step-by-step reasoning answer.

Claims

1. A medical record quality control method for a large medical model based on multi-path fusion learning, characterized in that: Including the following steps: Step 1: Design a reasonable and general instruction template for mining the medical record quality control ability of the large model. This instruction template combines the original medical record text with the medical record quality control rules to generate the final input prompt. Step 2: Feed the input prompt obtained in Step 2 into large parameter models of different series to obtain the thinking steps and step-by-step reasoning processes independently planned by all large parameter models. Step 3: After collecting the multi-path answers of large parameter models of different series, construct the final instruction fine-tuning training set through a multi-path fusion strategy. Step 4: Feed the instruction fine-tuning training set into a large model with relatively small parameters to adjust the parameter weights of the large model with relatively small parameters so that it fits the correct reasoning path. Step 5: After parameter adjustment, the large model with relatively small parameters, after absorbing the knowledge guidance of large parameter models of different series, can give a medical record quality control answer with a more comprehensive thinking angle, more rigorous logical reasoning, and stronger interpretability.

2. A medical record quality control method for a large medical model based on multi-path fusion learning as claimed in claim 1, characterized in that: In Step 1, the instruction template for mining the medical record quality control ability of the large model integrates the following four key elements: Introduce the words and expressions of the medical record to be quality controlled and the quality control rules. Concatenate the original text of the medical record to be quality controlled and the quality control rule description text with the above-introduced words and expressions. Specify the words and expressions that require the large model to first plan the thinking steps and then perform step-by-step logical reasoning. Specify the language style and format for the large model to return the quality control results.

3. A medical record quality control method for a large medical model based on multi-path fusion learning as claimed in claim 1, characterized in that: In Step 2, in order to ensure that large parameter models of different series can accurately output in the required format, the Few-shot learning method is adopted. By pre-writing two correct <prompt, expected answer> pairs as historical conversation records and concatenating them with the current prompt and feeding them into the large parameter model, for the same medical record and quality control rules, multiple answers with different thinking angles and reasoning logic paths can be obtained. Each answer contains the thinking steps and step-by-step reasoning process of the large parameter model for the current medical record quality control task.

4. The method for quality control of medical records of a large medical model based on multi-path fusion learning as claimed in claim 1, characterized in that: In Step 3, the multi-path fusion strategy includes: Step 301: Use regular expressions to clean the multi-path answers output by the large parameter model, including: Use the re.sub() function in regular expressions to uniformly remove special characters. Use the re.sub() function in regular expressions to filter out useless words and expressions. Use the method of "The first step...? The second step...? The third step..." to split the output of the large model into the planned thinking steps, the step-by-step reasoning process according to the steps, and the final conclusion. If it cannot be split into three steps, manual intervention is required for correction to obtain multi-path answers in a unified format. Step 302: Use a similarity deduplication algorithm in the field of natural language processing to screen the data after cleaning: For each answer text obtained through Step 301, pair it with all other answers pairwise to form multiple candidate pairs. For each pair of candidates, convert the two answer texts obtained in the previous step into vector form through the Word2vec model. The cosine similarity is used to calculate the similarity between the two vectors: if the similarity calculation result of a pair of candidates exceeds the threshold of 0.95, it means that the two answers are highly similar in content. In order to maintain the diversity of the data, only one of the answers is retained and the other is considered redundant information and removed; Step 303: After cleaning and deduplication, a multi-path answer set with high consistency and richness is finally obtained. Each answer is spliced ​​with its input prompt to finally obtain a set of question-answer pairs that can be used for training, that is, the instruction fine-tuning training set.

5. The method for quality control of medical records of a large medical model based on multi-path fusion learning as claimed in claim 1, characterized in that: In step 4, the instruction fine-tuning training set obtained in step 3 is randomly shuffled to form a multi-path fusion enhanced training set. Then, based on the concept of transfer learning, a large model with fewer parameters is selected, and the multi-path fusion enhanced training set is used for fine-tuning training.

6. A method for quality control of medical records of a large medical model based on multi-path fusion learning as claimed in claim 5, characterized in that: During fine-tuning training: The input text is converted into a machine-understandable vector form, which is forward propagated according to the preset network architecture and calculation formula to obtain the probability prediction of each character at the current position; The large model compares these predicted probabilities with the correct labels and calculates the difference between the predicted value and the true value through the cross entropy loss function. The smaller the cross entropy loss value, the closer the prediction result of the large model is to the actual situation. The goal of adjusting the model parameter weights is to increase the probability of the correct character appearing at the current position, that is, to minimize the value of the cross entropy loss function; by continuously repeating the training process, that is, letting the large model repeatedly learn the samples in the training set, we can eventually get an optimized large model enhanced by multi-path fusion.

Images