Large model reasoning ability enhancement method for intelligent triage of stomatology

Abstract

The invention discloses a large model reasoning ability enhancing method for intelligent triage of stomatology, which comprises the following steps: acquiring a plurality of medical record data, and selecting the medical record data meeting a preset condition as a medical record data set; converting the medical record data in the medical record data set into instruction data containing a reasoning process by adopting an open source reasoning large model, and screening the instruction data meeting a preset rule condition; and based on all instruction data, performing supervised fine tuning on the large model by adopting LoRA, taking the original large model as a reference model to maintain general knowledge, and training to obtain the oral medicine large model with enhanced reasoning ability. According to the scheme, the training data set is constructed based on the real medical record, and the general large model is finely adjusted by adopting technologies such as instruction fine adjustment and direct preference optimization, so that the reasoning ability in an intelligent triage scene is improved.

Description

Technical Field

[0001] The present invention relates to natural language processing technology, and particularly to a method for enhancing the inference ability of a large model for intelligent triage in stomatology. Background Art

[0002] In recent years, with the rapid development of artificial intelligence technology, the intelligent triage technology in stomatology has also been continuously evolving, and three main technical solutions have been formed. These solutions have their own characteristics, but at the same time, there are also certain limitations.

[0003] The first type of technical solution is a rule-based expert system. This type of system constructs a decision tree manually (for example, using the International Classification of Diseases standard ICD), establishes a mapping rule library between symptom keywords and disease types, and performs matching in a hard-coded manner. In structured interrogation scenarios, the rule-based expert system shows high stability and reliability, but its scalability and adaptability are poor, and it is difficult to meet the needs of modern medical intelligence.

[0004] The second type of technical solution is an intelligent triage method based on traditional machine learning models. This type of method mainly adopts the supervised learning paradigm, uses classic classification algorithms such as support vector machine (SVM), random forest, and XGBoost, and combines structured data (such as indicators like pain type and mouth opening degree) for classification. Some improved solutions also combine natural language processing technology to simply extract features from the patient's chief complaint text and add them to the input of the classification algorithm. Compared with the rule-based expert system, traditional machine learning methods have higher flexibility, but it is difficult to meet the requirements of actual triage scenarios.

[0005] The third type of technical solution is an intelligent triage method based on generative large models. In recent years, generative large models (such as GPT-4, DeepSeek, etc.) have made breakthroughs in the field of natural language processing, and their powerful language understanding and generation capabilities have provided new possibilities for intelligent triage. Existing solutions usually directly utilize general large models without fine-tuning, and only realize the generation of interrogation dialogues and the output of triage suggestions through prompt engineering. However, although general large models perform well in natural language understanding, there are still the following problems in the triage application in the field of stomatology:

[0006] 1. Insufficient domain expertise: The knowledge system of general large models mainly comes from publicly available Internet data, lacking in-depth professional knowledge in the field of stomatology, resulting in limited inference ability in specific medical scenarios.

[0007] 2. Difficult to meet the needs of evidence-based medicine: The inference logic of general large models is usually driven by the training of code and mathematical data. In the medical field, however, diagnostic and treatment decisions need to be based on strict evidence-based medical evidence, and general large models are difficult to fully meet this need.

[0008] 3. Lack of targeted optimization: Existing solutions usually only adjust the model output through prompt engineering, without performing domain-related fine-tuning on it, resulting in insufficient professional analysis ability of the model in oral medicine. Summary of the Invention

[0009] In view of the above deficiencies in the prior art, the method for enhancing the inference ability of the large model for oral medicine intelligent triage provided by the present invention solves the problem of insufficient inference ability in the intelligent triage scenario.

[0010] In order to achieve the above invention purpose, the technical solution adopted by the present invention is:

[0011] Provide a method for enhancing the inference ability of a large model for oral medicine intelligent triage, which includes the steps of:

[0012] S1. Obtain a number of medical record data, and select the medical record data that meets the preset conditions as the medical record data set;

[0013] S2. Use an open-source inference large model to convert the medical record data in the medical record data set into instruction data containing the inference process, and screen the instruction data that meets the preset rule conditions;

[0014] S3. Based on all the instruction data, use LoRA to perform supervised fine-tuning on the large model, and use the original large model as a reference model to maintain general knowledge, and train to obtain an oral medicine large model with enhanced inference ability.

[0015] Further, the method for enhancing the inference ability of the large model for oral medicine intelligent triage further includes step S4, using the direct preference optimization algorithm to perform alignment optimization on the output of the oral medicine large model to obtain the final oral medicine large model.

[0016] Further, the step S4 further includes:

[0017] S41. Obtain a number of medical record data, use the oral medicine large model to generate multiple different response responses for each medical record data, and calculate the semantic similarity between the predicted diagnosis and the true diagnosis in each response response;

[0018] S42. When the semantic similarity is greater than the first preset similarity, mark the response response as a positive sample, and when the semantic similarity is less than or equal to the second preset similarity, mark the response response as a negative sample;

[0019] S43. Screen the medical record data that simultaneously include positive samples and negative samples from multiple medical record data, and use all the screened medical record data to form a preference dataset;

[0020] S44. Apply the direct preference optimization algorithm to the preference dataset, adjust the model parameters of the oral medicine large model to optimize its generation strategy, and obtain the final oral medicine large model.

[0021] Further, when adjusting the oral medicine large model using the direct preference optimization algorithm, the expression of its optimization objective is:

[0022]

[0023] where, is the optimization objective; is the preference dataset; Δ(x, y1, y2) = logP θ (y1|x) ― logP θ (y2|x) is the logarithmic ratio of the generation probabilities of positive and negative samples; P θ (·) is the probability that the model generates a specific response y given the input x; y1 is the positive sample response; y2 is the negative sample response; x is the instruction input to the model; is the expected value on the preference dataset, which means that this calculation takes the average over the entire preference dataset; σ(·) is the sigmoid function.

[0024] Further, the step S1 further includes:

[0025] S11. Obtain a number of medical record data, input them into a pre-trained language model for quality scoring, and select the medical record data with the highest quality scores among the first preset number;

[0026] S12. Cluster all the medical record data, and select the medical record data with the highest quality scores and not selected before among the second preset number in each cluster;

[0027] S13. Merge all the selected medical record data as a high-quality and diverse medical record dataset.

[0028] Further, the method for clustering all the medical record data includes:

[0029] Initialize a candidate cluster number range k ∈ [k min , k max , where k min and k max are the preset minimum and maximum cluster numbers respectively;

[0030] Initialize each medical record data as a cluster and calculate the similarity between clusters. Iteratively merge the two most similar clusters until the number of clusters is reduced to k, and divide the medical record data into k clusters;

[0031] Calculate the silhouette coefficient of each cluster:

[0032] s k (i) = [b k (i) ― a k (i)] / max{a k (i), b k (i)}

[0033] where S(k) is the silhouette coefficient of the k-th cluster; N is the total number of medical record data in the k-th cluster; a k (i) is the average distance from the medical record data i in the k-th cluster to other medical record data in the same cluster; b k (i) is the average distance from the medical record data i in the k-th cluster to all samples in the nearest other cluster;

[0034] Traverse all candidate numbers of clusters k, and select the k that maximizes S(k) as the final number of clusters k opt = argmax k S(k), and divide all medical record data into k opt clusters.

[0035] Furthermore, the step S2 further includes:

[0036] S21. Input the medical record data in the medical record dataset into the open-source inference large model respectively, obtain the inference process and predicted diagnosis of the medical record data, and select the medical record data that meets the preset logical rules based on the inference process;

[0037] S22. Use the pre-trained word embedding model to transform the predicted diagnosis and true diagnosis corresponding to the medical record data selected in step S21 into high-dimensional semantic vectors;

[0038] S23. Calculate the semantic similarity between the high-dimensional semantic vectors of the predicted diagnosis and true diagnosis corresponding to the medical record data, and screen out the medical record data with a semantic similarity greater than the preset threshold;

[0039] S24. Use the present illness history and true diagnosis of the medical record data screened in step S23 as the input, and the inference process and predicted diagnosis corresponding to the medical record data as the output, and form instruction data with the input and output.

[0040] Furthermore, in step S3, the expression of the training objective for supervised fine-tuning is:

[0041]

[0042] Among them, is the training objective; is the average cross-entropy loss function; λ is a hyperparameter; is the KL divergence loss; T is the length of the instruction data; w t is the t-th word in the instruction data; w <t is the historical context of the first t - 1 words; P θ (w t |w <t ) is the predicted word probability distribution under the model parameter θ; P ref (w t |w <t ) is the probability distribution predicted by the reference model.

[0043] Furthermore, the calculation formula of the semantic similarity sim is:

[0044]

[0045] Among them, v true is the high-dimensional semantic vector corresponding to the true diagnosis; v pred is the high-dimensional semantic vector of the predicted diagnosis; ||·|| is the symbol for taking the modulus length.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] 1. In order to utilize medical record data to construct high-quality instruction data, this solution first generates instruction data of uncertain quality through an open-source large model, and then uses rule matching to screen out instruction data that meets the preset rule conditions, thereby screening out high-quality instruction data; this method efficiently converts the original medical record data into high-quality instruction data, providing a basis for the subsequent model fine-tuning stage.

[0048] 2. In order to improve the reasoning ability and output stability of the large model in the intelligent triage scenario, this solution proposes a two-stage fine-tuning method. By means of supervised fine-tuning and introducing a reference model, the reasoning ability of the model in the intelligent triage scenario is enhanced while maintaining the original knowledge of the model. Subsequently, the probability of the large model generating correct reasoning results is improved through direct preference optimization, thereby further enhancing the reliability of the triage results. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flowchart of a method for enhancing the reasoning ability of a large model for intelligent triage in stomatology.

[0050] Figure 2 is a detailed flowchart of the process for obtaining a medical record data set. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] The specific embodiments of the present invention will be described below to facilitate those skilled in the art of this technology to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art of this technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions created using the concept of the present invention are within the scope of protection.

[0052] Reference Figure 1 , Figure 1 shows a flowchart of a method for enhancing the inference ability of a large model for intelligent triage in stomatology; as Figure 1 shown, this method S includes steps S1 to S3.

[0053] In step S1, a number of medical record data are obtained, and the medical record data that meet the preset conditions are selected as the medical record data set; as Figure 2 shown, in implementation, this solution preferably further includes in step S1:

[0054] S11. Obtain a number of medical record data, input them into a pre-trained language model for quality scoring, and select the medical record data with the highest quality scores among the first preset number n1; the fine-tuning process of the pre-trained language model includes:

[0055] Professional doctors score the quality of a randomly sampled part of the medical record data; then, using a language model as the backbone of the scoring model, after replacing its last layer with a linear layer, using the medical record data as the input and the doctor's score as the output, fine-tune the language model to align with the expert preferences to obtain the pre-trained language model; use the pre-trained language model to score the quality of all the medical record data.

[0056] S12. Cluster all the medical record data, and select the medical record data with the highest quality scores and not selected before among the second preset number n2 in each cluster;

[0057] S13. Combine all the selected medical record data as a high-quality and diverse medical record data set.

[0058] In implementation, the method for clustering all the medical record data in this solution preferably includes the following steps:

[0059] Initialize a candidate number of clusters range k ∈ [k min , k max , where k min and k max are the preset minimum number of clusters and maximum number of clusters respectively;

[0060] Initialize each medical record data as a cluster and calculate the similarity between clusters, iteratively merge the two most similar clusters until the number of clusters is reduced to k, and divide the medical record data into k clusters;

[0061] Calculate the silhouette coefficient for each cluster:

[0062] s k (i) = [b k (i) - a k (i)] / max{a k (i), b k (i)}

[0063] Where S(k) is the silhouette coefficient of the k-th cluster; N is the total number of medical record data in the k-th cluster; a k (i) is the average distance from the medical record data i in the k-th cluster to other medical record data in the same cluster; b k (i) is the average distance from the medical record data i in the k-th cluster to all samples in the nearest other cluster;

[0064] Traverse all candidate cluster numbers k, and select the k that maximizes S(k) as the final cluster number k opt = argmax k S(k), and divide all medical record data into k opt clusters.

[0065] In step S2, use an open-source inference large model to convert the medical record data in the medical record dataset into instruction data containing the inference process, and filter the instruction data that meets the preset rule conditions;

[0066] In implementation, this solution preferably further includes the following steps in step S2:

[0067] S21. Input the medical record data in the medical record dataset into the open-source inference large model, such as DeepSeek-R1, to obtain the inference process and predicted diagnosis of the medical record data, and select the medical record data that meets the preset logical rules based on the inference process;

[0068] The preset logical rules specifically refer to defining a set of core medical keywords and phrases that must appear ("symptom analysis", "differential diagnosis", "exclusion basis"), and designing logical rules to ensure that the generated content conforms to the logical order, that is, first analyze the relevance between symptoms and diseases, then conduct a comparison of differential diagnoses, and finally list the basis for excluding other possibilities. Then, extract the relevant content in the generated text through regular expressions, and only when the text content contains all keywords and conforms to the logical order is it accepted, so as to ensure that the intermediate steps contain the required content and conform to the medical inference specification.

[0069] S22. Use a pre-trained word embedding model to convert the predicted diagnosis and true diagnosis corresponding to the medical record data selected in step S21 into high-dimensional semantic vectors;

[0070] S23. Calculate the semantic similarity between the high-dimensional semantic vectors corresponding to the predicted diagnosis and the true diagnosis based on the medical record data, and filter out the medical record data with a semantic similarity greater than a preset threshold. The calculation formula for the semantic similarity sim can be:

[0071]

[0072] where v true is the high-dimensional semantic vector corresponding to the true diagnosis; v pred is the high-dimensional semantic vector of the predicted diagnosis; ||·|| is the symbol for taking the modulus length.

[0073] S24. Use the present illness history and the true diagnosis of the medical record data screened in step S23 as the input, and the corresponding reasoning process and predicted diagnosis of the medical record data as the output, and form instruction data with the input and output.

[0074] In step S2, first generate instruction data of uncertain quality through an open-source large model, and then use rule matching and the cosine similarity distribution of word vectors to quantify the accuracy of the reasoning process and diagnosis results, so as to filter out high-quality instruction data. This method efficiently converts the original medical record data into high-quality instruction data, providing a basis for the subsequent model fine-tuning stage.

[0075] Since the real oral medicine medical records contain rich information about symptoms, diagnosis results, and the relationships between departments, they have great potential for improving the intelligent triage ability of the model. However, the medical record data lacks the intermediate reasoning process and cannot be directly used to fine-tune the reasoning model. Therefore, this solution filters from two aspects of data diversity and quality through steps S1 and S2 to improve the overall quality of the data and maintain diversity. Then, the filtered medical record data is sent into an open-source inference large model, and through the existing inference large model and the scoring algorithm based on semantic similarity, oral medicine instruction data that can be directly used for training is generated.

[0076] In step S3, based on all the instruction data, use LoRA to perform supervised fine-tuning on the large model, and use the original large model as a reference model to maintain general knowledge, and train an oral medicine large model with enhanced reasoning ability.

[0077] In step S3, LoRA simulates the effect of full-parameter updates by introducing low-rank decomposition matrices (i.e., adapters). During training, the main body weights of the model (i.e., the weights of the original model) remain frozen, and only the parameters of the adapters are updated. Since the main body weights of the model never change, during LoRA training, if the adapters are temporarily disabled (i.e., the parameters of the adapters are not used), then the output of the model is equivalent to the output of the original model. In this way, the prediction distribution of the reference model can be directly obtained without additionally loading a complete original model, thus avoiding an increase in video memory occupancy because only one set of model weights needs to be stored throughout the process.

[0078] In one embodiment of the present invention, the expression of the training objective for supervised fine-tuning in step S3 is:

[0079]

[0080] where, is the training objective; is the average cross-entropy loss function; λ is a hyperparameter; is the KL divergence loss; T is the length of the instruction data; w t is the t-th word in the instruction data; w <t is the historical context of the first t - 1 words; P θ (w t |w <t ) is the predicted word probability distribution under the model parameters θ; P ref (w t |w <t ) is the probability distribution predicted by the reference model.

[0081] By optimizing this objective, the performance of the model on basic medical inference tasks can be significantly improved. At the same time, in order to avoid overwriting the original general knowledge ability of the model, this solution uses the original model as the reference model. By adding the KL divergence between the probability distribution of the current model predicting the next word and the probability distribution of the reference model predicting the next word as a constraint, the inference ability of the model can be significantly improved.

[0082] In one embodiment of the present invention, the method for enhancing the inference ability of the large model of this solution further includes step S4, using the direct preference optimization algorithm to perform alignment optimization on the output of the oral medicine large model to obtain the final oral medicine large model; the detailed implementation process of this step includes:

[0083] S41. Obtain multiple medical record data, use the oral medicine large model to generate multiple different response responses for each medical record data, and calculate the semantic similarity between the predicted diagnosis and the true diagnosis in each response response;

[0084] S42. When the semantic similarity is greater than the first preset similarity, mark the response as a positive sample; when the semantic similarity is less than or equal to the second preset similarity, mark the response as a negative sample.

[0085] S43. Screen the medical record data that includes both positive samples and negative samples among multiple medical record data, and use all the screened medical record data to form a preference dataset to exclude samples where the model's multiple responses are all correct or all wrong.

[0086] S44. Apply the direct preference optimization algorithm on the preference dataset, adjust the model parameters of the oral medicine large model to optimize its generation strategy, and obtain the final oral medicine large model.

[0087] This solution applies the direct preference optimization algorithm on the preference dataset, adjusts the model parameters to optimize its generation strategy, and by maximizing the generation probability of positive samples and minimizing the generation probability of negative samples, it can improve the quality and stability of the model output.

[0088] When implemented, when this solution preferably uses the direct preference optimization algorithm to adjust the oral medicine large model, the expression of its optimization objective is:

[0089]

[0090] Where is the optimization objective; is the preference dataset; Δ(x, y1, y2) = logP θ (y1|x) ― logP θ (y2|x) is the logarithmic ratio of the generation probabilities of positive and negative samples; P θ (·) is the probability that the model generates a specific response y given the input x; y1 is the positive sample response; y2 is the negative sample response; x is the instruction input to the model. is the expected value on the preference dataset, which means this calculation takes the average over the entire preference dataset; σ(·) is the sigmoid function.

[0091] This solution can better align with user preferences and improve its stability and reliability in the medical scenario by minimizing The model can better align with user preferences and improve its stability and reliability in the medical scenario.

[0092] In summary, this solution constructs an instruction dataset through real medical records for model fine-tuning. Compared with manually constructing prompt words, it can better ensure that the output results of the oral medicine model conform to clinical practice; designing a two-stage fine-tuning method of supervised and direct preference optimization can effectively enhance the model's reasoning ability and output stability in the intelligent triage scenario, providing more reliable auxiliary support for medical decision-making.

Claims

1. A method for enhancing the reasoning ability of a large model for intelligent triage in stomatology, characterized in that: Includes steps: S1. Obtain a number of medical record data, and select the medical record data that meets the preset conditions as the medical record data set; S2. Use an open source reasoning model to convert the medical record data in the medical record data set into instruction data containing the reasoning process, and filter the instruction data that meets the preset rule conditions; S3. Based on all instruction data, LoRA is used to perform supervised fine-tuning on the large model. The original large model is used as a reference model to maintain common knowledge, and a large oral medicine model with enhanced reasoning ability is trained.

2. The method for enhancing the reasoning capability of a large model for intelligent triage in stomatology according to claim 1 is characterized in that: The method further includes step S4, wherein a direct preference optimization algorithm is used to perform alignment optimization on the output of the oral medicine large model to obtain a final oral medicine large model.

3. The method for enhancing the reasoning capability of a large model for intelligent triage in stomatology according to claim 2 is characterized in that: The step S4 further comprises: S41, obtaining multiple medical record data, using the oral medicine big model to generate multiple different responses for each medical record data, and calculating the semantic similarity between the predicted diagnosis and the actual diagnosis in each response; S42: when the semantic similarity is greater than a first preset similarity, mark the response as a positive sample; when the semantic similarity is less than or equal to a second preset similarity, mark the response as a negative sample; S43, screening medical record data including both positive samples and negative samples from multiple medical record data, and using all the screened medical record data to form a preference data set; S44. Apply the direct preference optimization algorithm on the preference data set, adjust the model parameters of the oral medicine large model to optimize its generation strategy, and obtain the final oral medicine large model.

4. The method for enhancing the reasoning capability of a large model for intelligent triage in stomatology according to claim 3, characterized in that: When the direct preference optimization algorithm is used to adjust the oral medicine large model, the expression of its optimization objective is: in, To optimize the goal; is the preference data set; Δ(x,y1,y2)=logP θ (y1|x)―logP θ (y2|x) is the logarithmic ratio of the probability of generating positive samples and negative samples; P θ (·) is the probability that the model generates a specific response y given an input x; y1 is a positive sample response; y2 is a negative sample response; x is the instruction input to the model; is the expected value on the preference data set; σ(·) is the sigmoid function.

5. The method for enhancing the reasoning capability of a large model for intelligent triage in stomatology according to claim 1, characterized in that: The step S1 further comprises: S11, obtaining a number of medical record data, and inputting the data into a pre-trained language model for quality scoring, and selecting a first preset number of medical record data with the highest quality score; S12, clustering all medical record data, and selecting a second preset number of medical record data that have not been selected and have the highest quality score in each cluster; S13. Merge all selected medical record data into a high-quality and diverse medical record dataset.

6. The method for enhancing the reasoning capability of a large model for intelligent triage in stomatology according to claim 5, characterized in that: Methods for clustering all medical record data include: Initialize a candidate cluster number range k∈[k min ,k max ], where k min and k max are the preset minimum and maximum number of clusters, respectively; Initialize each medical record data as a cluster and calculate the similarity between clusters. Iteratively merge the two most similar clusters until the number of clusters is reduced to k, and divide the medical record data into k clusters. Compute the silhouette coefficient for each cluster: s k (i)=[b k (i)―a k (i)] / max{a k (i),b k (i)} Where S(k) is the silhouette coefficient of the kth cluster; N is the total number of medical records in the kth cluster; a k (i) is the average distance between medical record data i in the kth cluster and other medical record data in the same cluster; b k (i) is the average distance from the medical record data i in the kth cluster to all samples in the nearest other clusters; Traverse all candidate cluster numbers k and select the k that maximizes S(k) as the final cluster number k opt = argmax k S(k), and divide all medical records into k opt Clusters.

7. The method for enhancing the reasoning capability of a large model for intelligent triage in stomatology according to claim 1, characterized in that: The step S2 further comprises: S21, inputting the medical record data in the medical record data set into the open source reasoning big model respectively, obtaining the reasoning process and predictive diagnosis of the medical record data, and selecting the medical record data that meets the preset logical rules based on the reasoning process; S22, using a pre-trained word embedding model to convert the predicted diagnosis and true diagnosis corresponding to the medical record data selected in step S21 into a high-dimensional semantic vector; S23, calculating the semantic similarity between the predicted diagnosis and the actual diagnosis according to the high-dimensional semantic vectors of the medical record data, and selecting the medical record data whose semantic similarity is greater than a preset threshold; S24. The current medical history and true diagnosis of the medical record data screened out in step S23 are taken as input, and the reasoning process and predicted diagnosis corresponding to the medical record data are taken as output, and the input and output are used to form instruction data.

8. The method for enhancing the reasoning capability of a large model for intelligent triage in stomatology according to claim 1, characterized in that: In step S3, the expression of the training objective for supervised fine-tuning is: in, For training objectives; is the average cross entropy loss function; λ is a hyperparameter; is the KL divergence loss; T is the length of the instruction data; w t is the tth word in the instruction data; w <t is the historical context of the previous t-1 words; P θ (w t |w <t ) is the predicted word probability distribution under the model parameter θ; P ref (w t |w <t ) is the probability distribution predicted by the reference model.

9. The method for enhancing the reasoning ability of a large model for intelligent triage in stomatology according to claim 2 or 7, characterized in that: The calculation formula of semantic similarity sim is: Among them, v true is the high-dimensional semantic vector corresponding to the true diagnosis; v pred is the high-dimensional semantic vector for predicting diagnosis; ||·|| is the modulus length symbol.

Images