Chest pain treatment plan generation method and system based on etiology adaptive migration LoRA

Abstract

The application discloses a chest pain treatment scheme generation method and system based on cause self-adaptive migration LoRA, relates to the technical field of artificial intelligence and intelligent medical treatment, and comprises the following steps: constructing a chest pain case data set, shielding explicit cause words in the chest pain case data set text, and obtaining a cause blind subset; constructing a basic large model and a LoRA fine-tuning framework, and training the basic large model and the LoRA fine-tuning framework through the cause blind subset; the cause blind subset is obtained by shielding the explicit cause words in the chest pain case text, and learning deviation caused by label leakage can be reduced; after pre-training through a whole-cause shared LoRA adapter, exclusive adapters are obtained through cause-specific migration training, overfitting during training of rare chest pain causes can be prevented; the cause probability output by a cause router and a pre-set reliability threshold are used to construct a first treatment model and a second treatment model, respectively, so that the basic expression capability of general chest pain cases and the specific diagnosis and treatment logic corresponding to rare causes can be taken into account.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and smart healthcare, and in particular to a method and system for generating chest pain treatment plans based on etiology adaptive migration LoRA. Background Technology

[0002] Chest pain is one of the most common symptoms in emergency and cardiovascular departments. It is a highly heterogeneous clinical scenario comprised of multiple etiologies, including unstable angina, acute ST-segment elevation myocardial infarction (STEMI), non-ST-segment elevation myocardial infarction (NSTEMI), heart failure, pulmonary embolism, aortic dissection, and cardiac neurosis. The treatment priorities, drug combinations, monitoring programs, and interventional pathways differ significantly depending on the etiology. For example, STEMI emphasizes emergency reperfusion therapy and continuous ECG monitoring; aortic dissection requires strict blood pressure control and priority surgical evaluation; while cardiac neurosis is mainly treated with psychological intervention and symptomatic relief.

[0003] With the application of large language models in the medical field, treatment plan generation methods based on efficient fine-tuning of LoRA (Low-Rank Adaptation) parameters have attracted widespread attention. Existing techniques typically employ the UnifiedLoRA method, which trains a unified adapter by mixing chest pain cases of all etiologies. This method can learn common expression patterns and medical order formats in treatment records.

[0004] However, in existing technologies, the unified LoRA architecture compresses data from multiple etiologies with significantly different diagnostic and treatment logics into the same parameter space during training. Common chest pain symptoms, which have a larger sample size, dominate model parameter updates, while the specific diagnostic and treatment features of high-risk diseases with small sample sizes, such as aortic dissection and pulmonary embolism, are easily covered by general templates. Furthermore, the model training phase does not mask the disease names directly labeled within the cases; the model directly matches the corresponding output template based on explicit etiology terms in the text, failing to derive treatment plans from basic clinical information such as symptoms and test indicators. Training independent adapters specifically for niche etiologies also leads to decreased generalization ability due to insufficient sample size. Therefore, existing technologies suffer from technical problems such as inability to balance general case representations with the specific diagnostic and treatment logic of niche etiologies, susceptibility to label leakage, and overfitting with small samples. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for generating chest pain treatment plans based on etiology adaptive migration LoRA, which solves the technical problems of existing technologies, such as the inability to take into account both general case expression and niche etiology-specific diagnostic and treatment logic, the tendency for label leakage, and overfitting in small samples.

[0006] The technical means employed in this invention are as follows: In a first aspect, the present invention provides a method for generating chest pain treatment plans based on etiological adaptive migration LoRA, including: Construct a chest pain case dataset, and mask the explicit etiological terms in the text of the chest pain case dataset to obtain an etiology-blinded subset; A basic large model and a LoRA fine-tuning framework are constructed. The basic large model and the LoRA fine-tuning framework are trained through the etiology-blinded subset to obtain a LoRA adapter that shares all etiologies. The etiology-blinded subset is divided into multiple chest pain factor sets according to etiology categories. The whole etiology-shared LoRA adapter is used as the initialization parameter, and transfer training is performed on each chest pain factor set to obtain multiple etiology-specific LoRA adapters. Input the clinical information of the patient to be treated into the etiology router to obtain the highest etiology probability value; Determine whether the highest probability value of the cause is not less than a preset confidence threshold; If the highest probability value of the cause is not less than the preset confidence threshold, the cause-specific LoRA adapter parameter corresponding to the highest probability value of the cause is fused with the weight of the basic large model to generate a first effective generation weight. A first treatment model is constructed based on the first effective generation weight, and a first treatment plan is generated based on the first treatment model. If the highest etiology probability value is less than the preset confidence threshold, the parameters of the all-etiology shared LoRA adapter are fused with the weights of the basic large model to generate a second effective generation weight. A second treatment model is constructed based on the second effective generation weight, and a second treatment plan is generated based on the second treatment model.

[0007] Furthermore, the masking of explicit etiological terms in the text of the chest pain case dataset includes: The target explicit etiological words in the chest pain case dataset are matched using regular expressions, and the target explicit etiological words are replaced with target placeholders. The rules for matching explicit cause words in the target context are as follows:

[0008] in, E represents the set of etiological words in the text that need to be blocked; E represents the set of causes of chest pain. The set of explicit terms corresponding to the cause e; x represents the input clinical text.

[0009] Furthermore, the step of training the base large model and the LoRA fine-tuning framework through the etiology-blinded subset to obtain a full etiology-shared LoRA adapter includes: The chest pain case texts from the etiology-blinded subset are input into the base model and the LoRA fine-tuning framework. The loss optimization objective is to minimize the cross-entropy loss between the treatment plan generation text and the real clinical record. When performing gradient backpropagation, the parameters of the base model are frozen, and only the parameters of the LoRA low-rank matrix are updated until the loss converges, thus obtaining the all-etiology shared LoRA adapter.

[0010] Furthermore, the formula for the transfer training of the chest pain lesion factor set is as follows:

[0011] in, The parameters of the LoRA adapter specific to the cause of the disease indicate that training is complete; This indicates a transfer training operation; This indicates that LoRA adapter parameters are shared across all etiologies; This represents the set of pathogenic factors corresponding to pathogenic factor e.

[0012] Furthermore, the step of inputting the clinical information of the patient to be processed into the etiology router to obtain the highest etiology probability value includes: The clinical information of the patient to be treated is input into the etiology router, and TF-IDF feature extraction is performed through the etiology router to obtain the target extraction feature. The target extraction feature is input into the logistic regression classifier to calculate the distribution probability of multiple chest pain etiologies. The highest etiology probability value is extracted from the distribution probability of multiple chest pain etiologies. The clinical information of the patients to be treated includes at least one of the following: the patient's chief complaint, present medical history, past medical history, vital signs, physical examination, laboratory tests, and imaging examinations.

[0013] Furthermore, the calculation formulas for the first effective generation weight and the second effective generation weight are as follows: when hour when hour in, Indicates the first effective generated weight; Indicates the second effective generation weight; Represents the weights of the basic large model; This indicates the highest probability cause predicted by the etiology router. Indicates the most probable cause of illness Corresponding LoRA adapter parameters specific to the underlying cause; This indicates that LoRA adapter parameters are shared across all etiologies; This represents the highest probability value of the cause of the disease output by the etiology router. This indicates a pre-set confidence threshold.

[0014] Furthermore, after generating the second treatment plan based on the second treatment model, the method further includes: The first treatment model and / or the second treatment model were evaluated based on the etiology consistency index and the quality index of the treatment text generation. The etiological consistency indicators include at least one of the following: label coverage rate, invalid label error accuracy rate, and Macro-F1. The quality indicators for the generated treatment text include at least one of ROUGE-1, ROUGE-2, ROUGE-L, keyword recall, and structural integrity.

[0015] Furthermore, when evaluating the first treatment model and / or the second treatment model using keyword recall metrics, the following are included: A treatment element categorization evaluation method is adopted, which divides treatment keywords into multiple target categories, calculates the keyword recall rate for each target category, and evaluates the first treatment model and / or the second treatment model based on the keyword recall rate; The target categories include interventional or surgical procedures, pharmacological treatments, examinations and monitoring, nursing and lifestyle care, and follow-up plans.

[0016] Secondly, the present invention also provides a chest pain treatment plan generation system based on etiology adaptive migration LoRA, comprising: The data preprocessing module is used to construct a chest pain case dataset, and to mask explicit etiological terms in the text of the chest pain case dataset to obtain an etiology-blinded subset. The shared adapter pre-training module is used to build a basic large model and a LoRA fine-tuning framework. The basic large model and the LoRA fine-tuning framework are trained through the etiology blinded subset to obtain a full etiology shared LoRA adapter. The dedicated adapter transfer training module is used to divide the etiology blinded subset into multiple chest pain factor sets according to etiology categories, and use the all-etiology shared LoRA adapter as the initialization parameter to perform transfer training on each of the chest pain factor sets to obtain multiple etiology-specific LoRA adapters. The etiology routing classification module is used to input the clinical information of the patients to be treated into the etiology router to obtain the highest etiology probability value; The confidence judgment module is used to determine whether the highest probability value of the cause of disease is not less than a preset confidence threshold. The first inference generation module is used to, if the highest etiology probability value is not less than the preset confidence threshold, fuse the etiology-specific LoRA adapter parameters corresponding to the highest etiology probability value with the weights of the basic large model to generate a first effective generation weight, construct a first treatment model based on the first effective generation weight, and generate a first treatment plan based on the first treatment model. The second inference generation module is used to fuse the parameters of the all-cause shared LoRA adapter with the weights of the basic large model if the highest etiology probability value is less than the preset confidence threshold, generate a second effective generation weight, construct a second treatment model based on the second effective generation weight, and generate a second treatment plan based on the second treatment model.

[0017] Compared with the prior art, the present invention has the following advantages: This invention obtains a blinded subset of causes by masking explicit etiological terms in chest pain case texts, which can reduce learning bias caused by label leakage; after pre-training with a LoRA adapter that shares all causes, a dedicated adapter is obtained through etiological transfer training, which can prevent overfitting during training for niche chest pain causes; based on the etiological router outputting etiological probabilities and pre-set confidence thresholds, a first treatment model and a second treatment model are constructed respectively, which can take into account both the basic expressive ability of general chest pain cases and the specific diagnostic and treatment logic corresponding to niche causes.

[0018] Based on the above reasons, this invention can be widely applied in fields such as artificial intelligence and smart healthcare. Attached Figure Description

[0019] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the method for generating a chest pain treatment plan based on etiology adaptive migration LoRA in an embodiment of the present invention. Figure 2 This is a comparison diagram of the mechanistic differences between the LoRA adapter shared by all etiologies and the LoRA adapter specific to the etiology in this embodiment of the invention; Figure 3 This is an indicator radar chart in an embodiment of the present invention; Figure 4 This is an approximate Bayesian forest graph for the categorized keyword recall of treatment elements in this embodiment of the invention; Figure 5 This is a confidence distribution diagram of etiology routing in an embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0024] Please see Figure 1 , Figure 1 This is a flowchart illustrating the method for generating a chest pain treatment plan based on etiology adaptive migration LoRA in an embodiment of the present invention.

[0025] This invention provides a method for generating chest pain treatment plans based on etiological adaptive migration LoRA, comprising the following steps: Step 101: Construct a chest pain case dataset, and mask the explicit etiological terms in the text of the chest pain case dataset to obtain an etiology-blinded subset.

[0026] Specifically, by constructing a chest pain case dataset, traversing all texts in the dataset, and identifying explicit etiological words in each text, a unified mask character is used to mask each identified explicit etiological word, resulting in an etiological blinding subset. By masking the identified explicit etiological words to obtain the etiological blinding subset, the interference of explicit etiological words in the text on the model's learning can be reduced. This ensures that the etiological blinding subset retains only neutral text content related to chest pain symptoms and the diagnosis and treatment process, alleviating the model's learning bias of directly matching conclusions based on explicit etiological words.

[0027] In some embodiments, masking explicit etiological terms in the chest pain case dataset text includes: matching target explicit etiological terms in the chest pain case dataset using regular expressions, and replacing the target explicit etiological terms with target placeholders. For example, the explicit etiological terms can be at least one of unstable angina, acute ST-segment elevation myocardial infarction, non-ST-segment elevation myocardial infarction, heart failure, pulmonary embolism, aortic dissection, and cardiac neurosis, and the target placeholder can be "etiological terms masked".

[0028] The rules for matching explicit cause words in the target context are as follows:

[0029] in, E represents the set of etiological words in the text that need to be blocked; E represents the set of causes of chest pain. The set of explicit terms corresponding to the cause e; x represents the input clinical text.

[0030] Please see Figure 2 , Figure 2 This is a comparison diagram of the mechanistic differences between the LoRA adapter shared by all etiologies and the LoRA adapter specific to the etiology in this embodiment of the invention.

[0031] Step 102: Construct a basic large model and a LoRA fine-tuning framework. Train the basic large model and the LoRA fine-tuning framework using a blinded subset of causes to obtain a LoRA adapter that shares all causes. The LoRA fine-tuning framework is a lightweight model fine-tuning architecture that is bound to the basic large model and equipped with a low-rank matrix adaptation module. By inputting the blinded subset of causes into the basic large model and the LoRA fine-tuning framework to perform model fine-tuning training, a LoRA adapter that shares all causes is obtained after iterative convergence. This allows the LoRA adapter to learn the correlation features between chest pain symptoms and various causes, reducing the extent to which the original parameters of the basic large model are modified.

[0032] In some embodiments, a shared LoRA adapter for all causes is obtained by training a base model and a LoRA fine-tuning framework using a blinded subset of causes. This includes: inputting chest pain case texts from the blinded subset of causes into the base model and the LoRA fine-tuning framework; minimizing the cross-entropy loss between the generated treatment plan text and the actual clinical record as the loss optimization objective; freezing the parameters of the base model during gradient backpropagation and updating only the parameters of the LoRA low-rank matrix until the loss converges, thus obtaining the shared LoRA adapter for all causes. Freezing the parameters of the base model and updating only the LoRA low-rank matrix reduces the number of parameters that need to be iteratively updated during training, thus reducing the computational requirements for training; and optimizing the model by minimizing the cross-entropy loss improves the fit between the LoRA adapter's output treatment plan text and the actual clinical record.

[0033] Step 103: Divide the etiology-blinded subset into multiple chest pain factor sets according to etiology categories. Using a shared LoRA adapter for all etiologies as initial parameters, perform transfer training on each chest pain factor set to obtain multiple etiology-specific LoRA adapters. By splitting the etiology-blinded subset according to etiology categories to obtain chest pain factor sets, the case texts corresponding to different pathogenic factors can be distinguished. By reusing the shared LoRA adapter for all etiologies as the initial parameters for transfer training, the parameter convergence period of single-category etiology adapters can be shortened. Through transfer training on each chest pain factor set, each etiology-specific LoRA adapter can be adapted to the feature patterns of the corresponding category of chest pain text, and can output the probability distribution of the corresponding etiology based on the patient's clinical input.

[0034] In some embodiments, the formula for chest pain lesion factor set transfer training is:

[0035] in, The parameters of the LoRA adapter specific to the cause of the disease indicate that training is complete; This indicates a transfer training operation; This indicates that LoRA adapter parameters are shared across all etiologies; This represents the set of pathogenic factors corresponding to pathogenic factor e.

[0036] Step 104: Input the clinical information of the patient to be treated into the etiology router to obtain the highest etiology probability value.

[0037] The etiology router can use TF-IDF feature extraction combined with a logistic regression classifier. The maximum number of TF-IDF features can be set to 10,000. The etiology router achieves a precision, recall, and F1 score of 1.000 for each class on the test set.

[0038] In some embodiments, inputting the clinical information of the patient to be treated into the etiology router to obtain the highest etiology probability value includes: inputting the clinical information of the patient to be treated into the etiology router, performing TF-IDF feature extraction through the etiology router to obtain target extraction features, inputting the target extraction features into a logistic regression classifier, calculating the distribution probability of multiple chest pain etiologies, and extracting the highest etiology probability value from the distribution probability of multiple chest pain etiologies; the clinical information of the patient to be treated includes at least one of the patient's chief complaint, present medical history, past medical history, vital signs, physical examination, laboratory tests, and imaging examinations.

[0039] TF-IDF was used to extract target features from the clinical information of patients to be treated, in order to distinguish text features related to the etiology of chest pain. The distribution probability of multiple chest pain etiologies was calculated by a logistic regression classifier, which can quantify the degree of matching between the clinical information of patients to be treated and the corresponding category of pathogenic factors.

[0040] Step 105: Determine if the highest etiology probability value is not less than a preset reliability threshold. If the highest etiology probability value is not less than the preset reliability threshold, fuse the etiology-specific LoRA adapter parameters corresponding to the highest etiology probability value with the weights of the base model to generate the first effective generation weights. Construct the first treatment model based on the first effective generation weights, and generate the first treatment plan based on the first treatment model. If the highest etiology probability value is less than the preset reliability threshold, fuse the all-etiology shared LoRA adapter parameters with the weights of the base model to generate the second effective generation weights. Construct the second treatment model based on the second effective generation weights, and generate the second treatment plan based on the second treatment model.

[0041] The preset reliability threshold can be 0.85. By comparing the highest probability value of the cause with the preset reliability threshold, the medical records of patients with clear causes or ambiguous causes can be distinguished. Then, the first or second treatment plan can be generated to generate a treatment plan that is appropriate for the patient's cause, thereby improving the matching effect of chest pain treatment plan.

[0042] In some embodiments, the formulas for calculating the first effective generation weight and the second effective generation weight are as follows: when hour when hour in, Indicates the first effective generated weight; Indicates the second effective generation weight; Represents the weights of the basic large model; This indicates the highest probability cause predicted by the etiology router. Indicates the most probable cause of illness Corresponding LoRA adapter parameters specific to the underlying cause; This indicates that LoRA adapter parameters are shared across all etiologies; This represents the highest probability value of the cause of the disease output by the etiology router. This indicates a pre-set confidence threshold.

[0043] In some embodiments, after generating a second treatment plan based on a second treatment model, the method further includes: evaluating the first treatment model and / or the second treatment model based on etiological consistency indicators and treatment text generation quality indicators; the etiological consistency indicators include at least one of tag coverage, invalid tag error accuracy, and Macro-F1; the treatment text generation quality indicators include at least one of ROUGE-1, ROUGE-2, ROUGE-L, keyword recall, and structural integrity. The treatment model is evaluated by calculating corresponding quantitative values ​​based on the etiological consistency indicators and treatment text generation quality indicators, comprehensively measuring the quality of the model's output content from two dimensions: etiological matching and text quality.

[0044] In some embodiments, when evaluating the first treatment model and / or the second treatment model using keyword recall metrics, the method includes: using a treatment element categorization evaluation approach to divide treatment keywords into multiple target categories, calculating the keyword recall rate for each target category, and evaluating the first treatment model and / or the second treatment model based on the keyword recall rate; the multiple target categories include interventional or surgical categories, drug treatment categories, examination and monitoring categories, nursing and lifestyle categories, and follow-up plan categories.

[0045] By differentiating target categories and then calculating the corresponding keyword recall rate, the performance differences between the first treatment model and / or the second treatment model in generating various treatment plans are evaluated, and the ability of the model to fully output various treatment contents is measured.

[0046] Please see Figure 3 , Figure 3 This is a radar chart illustrating the performance indicators in this embodiment of the invention. It is used to compare the performance of five models—Unified LoRA, Structured LoRA, EAT-LoRA True Rigorous, EAT-LoRA Backtracking, and EAT-LoRA Transfer—under five evaluation metrics. The radial axis of the radar chart represents the indicator values, ranging from 0.70 to 1.00. Higher values ​​indicate better performance in the corresponding dimension. The evaluation dimensions surrounding the axis are, in order, macro-average F1, accuracy, structural integrity, keyword recall, and text similarity. As can be seen from the coverage area, the EAT-LoRA Transfer model of this application outperforms the other comparative models in all five metrics. EAT-LoRA Backtracking and Unified LoRA are next, while Structured LoRA has the lowest overall indicator values ​​across all dimensions. This clearly demonstrates that the model proposed in this application has relatively superior overall performance in terms of etiology matching accuracy, treatment text completeness, diagnostic keyword coverage, and text generation similarity.

[0047] Please see Figure 4 , Figure 4This is an approximate Bayesian forest diagram for the recall of treatment element categorized keywords in this embodiment of the invention; it is used to compare the posterior mean difference between the EAT-LoRA true strict model and the unified LoRA model in the recall rate of five treatment keywords. The horizontal axis represents the difference in keyword recall rate between the two models (EAT-LoRA true strict result minus unified LoRA result), and the dashed line at 0.00 on the horizontal axis is the baseline of no difference; the vertical axis, from top to bottom, represents the five categories of diagnosis and treatment goals: intervention or surgery, drug treatment, examination and monitoring, follow-up plan, nursing and lifestyle. In the figure, the dots represent estimated differences in means between categories, and the horizontal lines represent confidence intervals for these differences. The p-value for the corresponding statistical test is indicated on the right side of each row. For the four categories—interventional or surgical, drug treatment, examination and monitoring, and follow-up plan—the confidence intervals are generally located to the right of the 0 baseline, and the p-values ​​are all significantly greater than 0.05. This indicates that the true strictness of EAT-LoRA outperforms the unified LoRA in recall for these four keyword categories, and the differences are not statistically significant. For the nursing and lifestyle category, the confidence interval crosses the 0 baseline, and the p-value is 0.045, which is less than 0.05, representing a statistically significant difference in recall performance between the two models in this dimension. In this category, the true strictness of EAT-LoRA has a lower mean recall than the unified LoRA. Overall, this proposed model demonstrates superior performance in recalling core chest pain diagnosis and treatment-related keywords.

[0048] Please see Figure 5 , Figure 5 This is a confidence distribution diagram of etiological routing in an embodiment of the present invention; it is used to show the confidence distribution of the etiological router output for seven causes of chest pain. The horizontal axis represents the confidence of etiological routing, with a value range of 0.65 to 0.95. The vertical axis, from top to bottom, represents the seven causes of chest pain: aortic dissection, cardiac neurosis, heart failure, non-ST-segment elevation myocardial infarction, unstable angina, ST-segment elevation myocardial infarction, and pulmonary embolism. The dots in the figure are divided into two categories: correct routing and incorrect routing. The outline reflects the numerical distribution density of the confidence of the corresponding etiological samples. The confidence levels of samples correctly identified by routing for various causes were mostly concentrated in the range of 0.85 to 0.95, indicating a generally high confidence level. Only a small number of samples with incorrect routing were distributed in the lower confidence range, with only a few low-confidence misjudgments observed for cardiac neurosis, heart failure, and pulmonary embolism. Among these, the confidence levels of correctly routed samples for unstable angina, ST-segment elevation myocardial infarction, and aortic dissection were concentrated in the range of 0.90 to 0.95, indicating that the causal router in this scheme can output matching results with high confidence for various causes of chest pain. The misjudged samples were mostly accompanied by lower confidence values, making it easier to screen reliable causal judgment results based on confidence thresholds.

[0049] This invention also provides a chest pain treatment plan generation system based on etiology-adaptive transfer LoRA, comprising: a data preprocessing module for constructing a chest pain case dataset and masking explicit etiology terms in the text of the chest pain case dataset to obtain an etiology-blinded subset; a shared adapter pre-training module for constructing a basic large model and a LoRA fine-tuning framework, training the basic large model and the LoRA fine-tuning framework through the etiology-blinded subset to obtain a full etiology-shared LoRA adapter; a dedicated adapter transfer training module for dividing the etiology-blinded subset into multiple chest pain factor sets according to etiology categories, using the full etiology-shared LoRA adapter as initialization parameters, and performing transfer training on each chest pain factor set to obtain multiple etiology-specific LoRA adapters; and an etiology routing classification module for classifying the clinical information of the patients to be processed. The system inputs a pathogenesis router to obtain the highest pathogenesis probability value; a confidence judgment module is used to determine whether the highest pathogenesis probability value is not less than a preset confidence threshold; a first inference generation module is used to, if the highest pathogenesis probability value is not less than the preset confidence threshold, fuse the pathogenesis-specific LoRA adapter parameters corresponding to the highest pathogenesis probability value with the weights of the basic large model to generate a first effective generation weight, construct a first treatment model based on the first effective generation weight, and generate a first treatment plan based on the first treatment model; a second inference generation module is used to, if the highest pathogenesis probability value is less than the preset confidence threshold, fuse the all-pathogenesis shared LoRA adapter parameters with the weights of the basic large model to generate a second effective generation weight, construct a second treatment model based on the second effective generation weight, and generate a second treatment plan based on the second treatment model.

[0050] The same or similar parts among the various embodiments in this specification can be referred to mutually, and will not be repeated here.

[0051] This invention obtains a blinded subset of causes by masking explicit etiological terms in chest pain case texts, which can reduce learning bias caused by label leakage; after pre-training with a LoRA adapter that shares all causes, a dedicated adapter is obtained through etiological transfer training, which can prevent overfitting during training for niche chest pain causes; based on the etiological router outputting etiological probabilities and pre-set confidence thresholds, a first treatment model and a second treatment model are constructed respectively, which can take into account both the basic expressive ability of general chest pain cases and the specific diagnostic and treatment logic corresponding to niche causes.

[0052] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention 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 or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating chest pain treatment plans based on etiological adaptive migration LoRA, characterized in that, include: Construct a chest pain case dataset, and mask the explicit etiological terms in the text of the chest pain case dataset to obtain an etiology-blinded subset; A basic large model and a LoRA fine-tuning framework are constructed. The basic large model and the LoRA fine-tuning framework are trained through the etiology-blinded subset to obtain a LoRA adapter that shares all etiologies. The etiology-blinded subset is divided into multiple chest pain factor sets according to etiology categories. The whole etiology-shared LoRA adapter is used as the initialization parameter, and transfer training is performed on each chest pain factor set to obtain multiple etiology-specific LoRA adapters. Input the clinical information of the patient to be treated into the etiology router to obtain the highest etiology probability value; Determine whether the highest probability value of the cause is not less than a preset confidence threshold; If the highest probability value of the cause is not less than the preset confidence threshold, the cause-specific LoRA adapter parameter corresponding to the highest probability value of the cause is fused with the weight of the basic large model to generate a first effective generation weight. A first treatment model is constructed based on the first effective generation weight, and a first treatment plan is generated based on the first treatment model. If the highest etiology probability value is less than the preset confidence threshold, the parameters of the all-etiology shared LoRA adapter are fused with the weights of the basic large model to generate a second effective generation weight. A second treatment model is constructed based on the second effective generation weight, and a second treatment plan is generated based on the second treatment model.

2. The method for generating a chest pain treatment plan based on etiological adaptive migration LoRA according to claim 1, characterized in that, The process of masking explicit etiological terms in the text of the chest pain case dataset includes: The target explicit etiological words in the chest pain case dataset are matched using regular expressions, and the target explicit etiological words are replaced with target placeholders. The rules for matching target explicit etiological terms are as follows: in, E represents the set of etiological words in the text that need to be blocked; E represents the set of causes of chest pain. The set of explicit terms corresponding to the cause e; x represents the input clinical text.

3. The method for generating a chest pain treatment plan based on etiological adaptive migration LoRA according to claim 1, characterized in that, The process of training the base model and the LoRA fine-tuning framework using the etiology-blinded subset to obtain a full etiology-shared LoRA adapter includes: The chest pain case texts from the etiology-blinded subset are input into the base model and the LoRA fine-tuning framework. The loss optimization objective is to minimize the cross-entropy loss between the treatment plan generation text and the real clinical record. When performing gradient backpropagation, the parameters of the base model are frozen, and only the parameters of the LoRA low-rank matrix are updated until the loss converges, thus obtaining the all-etiology shared LoRA adapter.

4. The method for generating a chest pain treatment plan based on etiological adaptive migration LoRA according to claim 1, characterized in that, The formula for the transfer training of the chest pain lesion factor set is as follows: in, The parameters of the LoRA adapter specific to the cause of the disease indicate that training is complete; This indicates a transfer training operation; This indicates that LoRA adapter parameters are shared across all etiologies; This represents the set of pathogenic factors corresponding to pathogenic factor e.

5. The method for generating a chest pain treatment plan based on etiological adaptive migration LoRA according to claim 1, characterized in that, The step of inputting the clinical information of the patient to be treated into the etiology router to obtain the highest etiology probability value includes: The clinical information of the patient to be treated is input into the etiology router, and TF-IDF feature extraction is performed through the etiology router to obtain the target extraction feature. The target extraction feature is input into the logistic regression classifier to calculate the distribution probability of multiple chest pain etiologies. The highest etiology probability value is extracted from the distribution probability of multiple chest pain etiologies. The clinical information of the patients to be treated includes at least one of the following: the patient's chief complaint, present medical history, past medical history, vital signs, physical examination, laboratory tests, and imaging examinations.

6. The method for generating a chest pain treatment plan based on etiological adaptive migration LoRA according to claim 1, characterized in that, The formulas for calculating the first effective generated weight and the second effective generated weight are as follows: when hour when hour in, Indicates the first effective generated weight; Indicates the second effective generation weight; Represents the weights of the basic large model; This indicates the highest probability cause predicted by the etiology router. Indicates the most probable cause of illness Corresponding LoRA adapter parameters specific to the underlying cause; This indicates that LoRA adapter parameters are shared across all etiologies; This represents the highest probability value of the cause of the disease output by the etiology router. This indicates a pre-set confidence threshold.

7. The method for generating chest pain treatment plans based on etiological adaptive migration LoRA according to claim 1, characterized in that, After generating the second treatment plan based on the second treatment model, the method further includes: The first treatment model and / or the second treatment model were evaluated based on the etiology consistency index and the quality index of the generated treatment text. The etiological consistency indicators include at least one of the following: label coverage rate, invalid label error accuracy rate, and Macro-F1. The quality indicators for the generated treatment text include at least one of ROUGE-1, ROUGE-2, ROUGE-L, keyword recall, and structural integrity.

8. The method for generating a chest pain treatment plan based on etiological adaptive migration LoRA according to claim 7, characterized in that, When using keyword recall metrics to evaluate the first treatment model and / or the second treatment model, the following are included: A treatment element categorization evaluation method is adopted, which divides treatment keywords into multiple target categories, calculates the keyword recall rate for each target category, and evaluates the first treatment model and / or the second treatment model based on the keyword recall rate; The target categories include interventional or surgical procedures, pharmacological treatments, examinations and monitoring, nursing and lifestyle care, and follow-up plans.

9. A chest pain treatment plan generation system based on etiological adaptive migration LoRA, characterized in that, include: The data preprocessing module is used to construct a chest pain case dataset, and to mask explicit etiological terms in the text of the chest pain case dataset to obtain an etiology-blinded subset. The shared adapter pre-training module is used to build a basic large model and a LoRA fine-tuning framework. The basic large model and the LoRA fine-tuning framework are trained through the etiology blinded subset to obtain a full etiology shared LoRA adapter. The dedicated adapter transfer training module is used to divide the etiology blinded subset into multiple chest pain factor sets according to etiology categories, and use the all-etiology shared LoRA adapter as the initialization parameter to perform transfer training on each of the chest pain factor sets to obtain multiple etiology-specific LoRA adapters. The etiology routing classification module is used to input the clinical information of the patients to be treated into the etiology router to obtain the highest etiology probability value; The confidence judgment module is used to determine whether the highest probability value of the cause of disease is not less than a preset confidence threshold. The first inference generation module is used to, if the highest etiology probability value is not less than the preset confidence threshold, fuse the etiology-specific LoRA adapter parameters corresponding to the highest etiology probability value with the weights of the basic large model to generate a first effective generation weight, construct a first treatment model based on the first effective generation weight, and generate a first treatment plan based on the first treatment model. The second inference generation module is used to fuse the parameters of the all-cause shared LoRA adapter with the weights of the basic large model if the highest etiology probability value is less than the preset confidence threshold, generate a second effective generation weight, construct a second treatment model based on the second effective generation weight, and generate a second treatment plan based on the second treatment model.

Images