Traditional Chinese medicine diagnosis and treatment system continuous learning method and system based on clinical feedback

Abstract

The application discloses a traditional Chinese medicine diagnosis and treatment system continuous learning method and system based on clinical feedback, and relates to the technical field of data processing. The method comprises the following steps: performing credibility evaluation on clinical feedback data to obtain a credibility score; retaining clinical feedback data with a credibility score greater than a first credibility score threshold as to-be-evaluated data; performing data value evaluation on the to-be-evaluated data to obtain a data value score; based on a data value score threshold, screening and classifying the to-be-evaluated data according to the credibility score and the data value score to obtain first-level, second-level and third-level learning data sets; performing supervised learning on the traditional Chinese medicine diagnosis and treatment system by using the first-level learning data set, performing reinforcement learning on the traditional Chinese medicine diagnosis and treatment system by using the second-level learning data set, and performing unsupervised learning on the traditional Chinese medicine diagnosis and treatment system by using the third-level learning data set, so that the continuous optimization of the traditional Chinese medicine diagnosis and treatment system is realized. The application realizes long-term self-adaptive evolution of the traditional Chinese medicine diagnosis and treatment system.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a continuous learning method and system for traditional Chinese medicine diagnosis and treatment systems based on clinical feedback. Background Technology

[0002] With the rapid development of TCM digitalization and medical artificial intelligence technologies, TCM diagnosis and treatment systems based on clinical data have gradually become widespread. Most existing TCM diagnosis and treatment systems use continuous learning methods that directly collect and use clinical feedback data to train models, thereby achieving continuous improvement in system performance through incremental learning.

[0003] However, existing technologies lack effective means of assessing the credibility of clinical feedback data. Mixed data quality can easily lead to low-quality data contaminating the training set and affecting model performance. Ignoring the differences in the intrinsic value of data, a large amount of repetitive and irrelevant data participates in learning, resulting in a waste of computing resources and low model optimization efficiency. The rigid learning strategy cannot be dynamically adjusted and is difficult to adapt to the performance changes of the TCM diagnosis and treatment system itself and the evolution of clinical practice, resulting in limited long-term adaptive evolution capabilities of the TCM diagnosis and treatment system. Summary of the Invention

[0004] This invention provides a method and system for continuous learning of traditional Chinese medicine diagnosis and treatment systems based on clinical feedback, aiming to solve the technical problem of limited long-term adaptive evolution capabilities of existing traditional Chinese medicine diagnosis and treatment systems.

[0005] In view of the above problems, the present invention provides a method and system for continuous learning of traditional Chinese medicine diagnosis and treatment system based on clinical feedback.

[0006] In a first aspect, the present invention provides a continuous learning method for a traditional Chinese medicine diagnosis and treatment system based on clinical feedback, including:

[0007] The collected clinical feedback data were evaluated for data credibility, and a credibility score was obtained for each clinical feedback data.

[0008] The clinical feedback data is filtered according to a preset first confidence score threshold, and clinical feedback data with a confidence score greater than the first confidence score threshold are retained as data to be evaluated.

[0009] Based on the novelty and relevance of the data, the data to be evaluated is evaluated to obtain a data value score for each data to be evaluated.

[0010] Based on the data value scoring threshold of the current dynamic adjustment of the performance of the traditional Chinese medicine diagnosis and treatment system, the data to be evaluated is screened and classified according to the credibility score and data value score to obtain a first-level learning data set, a second-level learning data set and a third-level learning data set;

[0011] The traditional Chinese medicine diagnosis and treatment system is subjected to supervised learning using the first-level learning dataset, reinforcement learning using the second-level learning dataset, and unsupervised learning using the third-level learning dataset, thereby achieving continuous optimization of the traditional Chinese medicine diagnosis and treatment system.

[0012] Secondly, this invention provides a continuous learning system for traditional Chinese medicine diagnosis and treatment based on clinical feedback, comprising:

[0013] The credibility assessment module is used to assess the credibility of the collected clinical feedback data and obtain a credibility score for each clinical feedback data.

[0014] The credibility screening module is used to screen the clinical feedback data according to a preset first credibility score threshold, and retain the clinical feedback data with a credibility score greater than the first credibility score threshold as data to be evaluated.

[0015] The value assessment module is used to assess the data value of the data to be assessed based on data novelty and task relevance, and to obtain a data value score for each data to be assessed.

[0016] The dynamic classification module is used to dynamically adjust the data value scoring threshold based on the current performance of the traditional Chinese medicine diagnosis and treatment system. Based on the credibility score and data value score, the data to be evaluated is screened and classified to obtain a first-level learning data set, a second-level learning data set, and a third-level learning data set.

[0017] The continuous optimization module is used to perform supervised learning on the traditional Chinese medicine diagnosis and treatment system using the first-level learning dataset, reinforcement learning on the traditional Chinese medicine diagnosis and treatment system using the second-level learning dataset, and unsupervised learning on the traditional Chinese medicine diagnosis and treatment system using the third-level learning dataset, so as to achieve continuous optimization of the traditional Chinese medicine diagnosis and treatment system.

[0018] One or more technical solutions provided in this invention have at least the following technical effects or advantages:

[0019] This invention provides a continuous learning method and system for a traditional Chinese medicine (TCM) diagnosis and treatment system based on clinical feedback. By assessing the credibility of clinical feedback data and using threshold screening, low-credibility data can be effectively eliminated, ensuring the basic quality of subsequent learning data. By combining data novelty and task relevance to conduct data value assessment, the learning value of the data to be evaluated can be accurately quantified. Furthermore, by dynamically adjusting the threshold based on the current performance of the TCM diagnosis and treatment system, the data to be evaluated can be classified and screened in a hierarchical manner, forming a differentiated set of learning data. Finally, the system is trained using supervised learning, reinforcement learning, and unsupervised learning methods for different levels of learning data, which can accurately match data characteristics and learning patterns, improve data utilization efficiency, continuously optimize the performance of the TCM diagnosis and treatment system, and enhance the learning effect and adaptive iteration capability of the TCM diagnosis and treatment system. Attached Figure Description

[0020] Figure 1 A flowchart illustrating the continuous learning method for a traditional Chinese medicine diagnosis and treatment system based on clinical feedback, provided in an embodiment of the present invention.

[0021] Figure 2 A system architecture diagram of the continuous learning method for a traditional Chinese medicine diagnosis and treatment system based on clinical feedback provided in an embodiment of the present invention;

[0022] Figure 3 A schematic diagram of the structure of a continuous learning system for traditional Chinese medicine diagnosis and treatment based on clinical feedback, provided in an embodiment of the present invention;

[0023] The components represented by each number in the attached diagram are explained below:

[0024] Credibility assessment module 11, credibility screening module 12, value assessment module 13, dynamic classification module 14, and continuous optimization module 15. Detailed Implementation

[0025] This invention provides a continuous learning method and system for traditional Chinese medicine diagnosis and treatment systems based on clinical feedback, which is used to address the technical problem of limited long-term adaptive evolution capabilities of existing traditional Chinese medicine diagnosis and treatment systems.

[0026] Example 1, as Figure 1 , Figure 2 As shown, this invention provides a continuous learning method for a traditional Chinese medicine diagnosis and treatment system based on clinical feedback, the method comprising:

[0027] S100: Conduct a data credibility assessment on the collected clinical feedback data and obtain a credibility score for each clinical feedback data.

[0028] In this embodiment of the invention, the collected clinical feedback data is evaluated for data credibility to obtain a credibility score for each piece of clinical feedback data. Clinical feedback data comes from diverse sources and varies in quality, exhibiting issues such as missing data records, contradictions in TCM diagnostic logic, and insufficient authority of data sources. Direct use of such data would reduce the diagnostic accuracy of the TCM treatment system. Therefore, a standardized credibility score is needed through multi-dimensional quantitative evaluation to provide a basis for subsequent data screening.

[0029] Step S100 in the method provided in this embodiment of the invention includes:

[0030] The clinical feedback data includes at least patient medical records, medication records, efficacy evaluation records, and physician feedback information.

[0031] First, clinical feedback data is collected. This clinical feedback data includes at least patient treatment records, medication records, efficacy evaluation records, and physician feedback information. Clinical feedback data refers to the collection of various data generated throughout the entire TCM treatment process, reflecting the treatment process and its effects; it serves as the data input for the continuous learning of the TCM treatment system. Patient treatment records include basic information about the patient at the time of consultation, chief complaints, medical history, TCM four diagnostic methods, preliminary diagnostic conclusions, and other treatment-related data. Medication records include TCM prescriptions issued after the patient's consultation, duration of medication, and any adjustments made to the medication. Efficacy evaluation records record the improvement of symptoms, changes in signs, and recovery progress after medication, used to evaluate the effectiveness of the treatment plan. Physician feedback information refers to the subjective evaluations and supplementary opinions given by physicians involved in the treatment based on their clinical experience regarding the patient's treatment process, the rationality of medication use, and the accuracy of the diagnosis.

[0032] Specifically, the scope of data collection should be determined, and the composition of clinical feedback data should be clarified to ensure that it covers at least four basic types of data: patient treatment records, medication records, efficacy evaluation records, and physician feedback information. Various types of clinical feedback data should be collected simultaneously through channels such as the hospital's TCM diagnosis and treatment system, physician mobile office terminals, and patient follow-up platforms. The collected data should be format-verified to ensure that the data format is uniform and free of garbled characters. Obviously invalid data should be initially removed, and data that conforms to basic standards should be retained for subsequent processing.

[0033] For example, taking a patient with insomnia due to liver stagnation and spleen deficiency as an example, the clinical feedback data collected is as follows: Patient's medical record: Basic information: name, gender, age; Chief complaint: insomnia for 1 month, accompanied by irritability, poor appetite, and fatigue; Medical history: no history of hypertension or diabetes; Four diagnostic methods: pale tongue with thin white coating, thready pulse; Preliminary diagnosis: liver stagnation and spleen deficiency syndrome. Medication record: Prescription: Xiaoyao San combined with Guipi Tang, modified, including Bupleurum chinense 10g, Angelica sinensis 12g, Atractylodes macrocephala 15g, etc.; Dosage and administration: 1 dose per day, decocted in water and taken warm in 2 divided doses; Duration of medication: 7 days. Efficacy evaluation record: Irritability was reduced after 3 days of medication, and normal sleep was possible after 7 days of medication, with improvement in symptoms of poor appetite and fatigue. Physician feedback: The patient's medical record is complete, the symptoms match the tongue and pulse, and the diagnosis is accurate. The medication regimen is consistent with the syndrome differentiation, and the efficacy is as expected.

[0034] Secondly, the credibility of the collected clinical feedback data was assessed.

[0035] This includes assessing the reliability of the collected clinical feedback data, including:

[0036] A comprehensive evaluation model is constructed, which receives multiple preset evaluation parameters as input. The evaluation parameters include at least two or more of the following: data source authority parameter, data record integrity parameter, internal logical consistency parameter, and provider historical accuracy parameter. The data source authority parameter is determined based on the data provider's registration qualification level or professional title level, and the data record integrity parameter is determined based on the filling ratio of preset mandatory data items in the clinical feedback data.

[0037] The comprehensive evaluation model integrates and calculates the current evaluation parameters of the input clinical feedback data, and outputs a quantitative score within a preset range as the credibility score of each clinical feedback data.

[0038] The methods for determining the internal logical consistency parameters include:

[0039] Extract symptom descriptions, tongue features, pulse characteristics, and preliminary diagnostic conclusions from clinical feedback data;

[0040] Based on a pre-defined knowledge base of TCM diagnostic rules, the medical compatibility among the symptom description text information, the tongue appearance feature information, the pulse appearance feature information, and the preliminary diagnosis conclusion information is determined, and a consistency judgment score is generated based on the compatibility judgment result. The higher the medical compatibility, the higher the consistency judgment score.

[0041] The consistency score is used as the internal logical consistency parameter.

[0042] First, extract symptom descriptions, tongue features, pulse characteristics, and preliminary diagnostic conclusions from the clinical feedback data. Symptom descriptions refer to the textual descriptions of the patient's subjective discomfort and objective signs during their visit. Tongue features refer to the color, coating, and shape of the patient's tongue recorded in the clinical feedback data, such as a pale tongue with a thin white coating or a red tongue with a yellow, greasy coating. Pulse characteristics refer to the pulse characteristics obtained by the physician during pulse diagnosis, such as a wiry and thready pulse or a deep and slow pulse. Preliminary diagnostic conclusions refer to the TCM syndrome differentiation diagnosis made by the physician based on the symptoms, tongue appearance, and pulse, such as liver qi stagnation and spleen deficiency syndrome or qi stagnation and blood stasis syndrome.

[0043] Specifically, the collected clinical feedback data is structured and analyzed to locate the fields for symptom descriptions, tongue images, pulse records, and diagnostic conclusions. Text extraction algorithms, such as keyword extraction and structured field reading, are used to accurately extract symptom description text information, tongue image feature information, pulse image feature information, and preliminary diagnostic conclusion information from the corresponding areas. The extracted information is then standardized in format, such as by unifying the text format and removing redundant expressions, to ensure the accuracy of subsequent compatibility assessments.

[0044] For example, taking the clinical feedback data of a patient with insomnia due to liver stagnation and spleen deficiency as an example: Analyzing the patient's clinical feedback data, the symptom description field is located as: insomnia for 1 month, accompanied by irritability, poor appetite, and fatigue; the tongue appearance field is: pale tongue with a thin white coating; the pulse appearance field is: wiry and thready pulse; and the preliminary diagnosis field is: liver stagnation and spleen deficiency syndrome. Key information is extracted as follows: symptom description text information: insomnia, irritability, poor appetite, fatigue; tongue appearance characteristics information: pale tongue with a thin white coating; pulse characteristics information: wiry and thready pulse; preliminary diagnosis conclusion information: liver stagnation and spleen deficiency syndrome.

[0045] Secondly, based on a pre-defined TCM diagnostic rules knowledge base, the medical compatibility among the symptom description text, tongue appearance features, pulse characteristics, and preliminary diagnostic conclusions is assessed. A consistency score is generated based on the compatibility assessment results, with higher medical compatibility resulting in a higher consistency score. The TCM diagnostic rules knowledge base is a pre-defined structured knowledge base containing recognized symptom, tongue, pulse, and syndrome correspondence logics, diagnostic norms, and medical compatibility assessment rules within the TCM field. Medical compatibility refers to the degree of matching between diagnostic information and TCM diagnostic logic; a higher degree of matching indicates stronger medical compatibility. The consistency score is a quantitative assessment of medical compatibility, with a pre-defined value range of 0-1; a higher score indicates greater consistency in the internal logic of the diagnostic information.

[0046] Specifically, the system invokes a pre-defined knowledge base of TCM diagnostic rules, using symptom descriptions, tongue features, and pulse characteristics as input to retrieve corresponding standard syndrome types from the knowledge base. The retrieved standard syndrome types are then compared with the preliminary diagnostic conclusions to determine their degree of matching: Complete match: 100% medical compatibility; Partial match: medical compatibility determined by the proportion of matching features; Complete mismatch: 0% medical compatibility. Following pre-defined scoring rules, the medical compatibility is converted into a consistency score within the range of 0-1, such as 1.0 for a complete match, 0.9 for a 90% match, and 0 for a complete mismatch.

[0047] For example, by calling the TCM diagnostic rules knowledge base and inputting the patient's symptom description text information, tongue appearance characteristics information, and pulse appearance characteristics information, the standard syndrome type is retrieved as liver stagnation and spleen deficiency syndrome; comparing the standard syndrome type with the preliminary diagnosis conclusion, the medical compatibility between the two is 95%; according to the scoring rules, a consistency judgment score of 0.95 is generated.

[0048] Next, the consistency judgment score is used as the internal logical consistency parameter. The internal logical consistency parameter is a quantitative indicator characterizing the internal logical rationality of the clinical feedback data's diagnostic and treatment information. It is one of the key parameters for data credibility assessment, and its value is completely consistent with the consistency judgment score. For example, a consistency judgment score of 0.95 is assigned as the internal logical consistency parameter of the patient's clinical feedback data. The parameter value 0.95 is then associated with the patient's symptom description text information, tongue appearance characteristics, pulse characteristics, and preliminary diagnostic conclusion information to complete the parameter determination.

[0049] Furthermore, a comprehensive evaluation model is constructed, which receives multiple preset evaluation parameters as input. The evaluation parameters include at least two or more of the following: data source authority parameter, data record integrity parameter, internal logical consistency parameter, and provider historical accuracy parameter. The data source authority parameter is determined based on the data provider's registration qualification level or professional title level, and the data record integrity parameter is determined based on the filling ratio of preset mandatory data items in the clinical feedback data.

[0050] The comprehensive evaluation model integrates multi-dimensional evaluation parameters to quantify the credibility of clinical feedback data, ultimately outputting a standardized credibility score. This can be achieved using conventional fusion algorithms such as weighted summation and neural networks. Evaluation parameters are multi-dimensional indicators used to measure the credibility of clinical feedback data, including at least two or more of the following: data source authority parameter, data record integrity parameter, internal logical consistency parameter, and provider historical accuracy parameter. Specifically, the data source authority parameter is determined based on the data provider's registration qualification level or physician professional title level, representing the reliability of the data source; the data record integrity parameter is determined based on the completion ratio of pre-defined mandatory data items in the clinical feedback data, representing the completeness of the data; the internal logical consistency parameter measures the degree of matching between symptom descriptions, tongue appearance, pulse appearance, and preliminary diagnostic conclusions in the clinical feedback data, conforming to the logic of traditional Chinese medicine diagnosis; and the provider historical accuracy parameter is determined based on the qualified proportion of previously submitted clinical feedback data by the data provider after verification, i.e., the proportion of reliable data in historical data, representing the quality of the provider's data submission.

[0051] Specifically, determine the combination of evaluation parameters. Select at least two or more parameters from the following parameters as inputs to the comprehensive evaluation model: data source authority, data record integrity, internal logical consistency, and provider historical accuracy. Clarify the rules for determining each evaluation parameter: Data source authority: Preset scores corresponding to professional titles / qualifications, such as 0.9 for chief physician, 0.8 for associate chief physician, and 0.7 for attending physician; 0.9 for tertiary TCM hospitals and 0.8 for secondary TCM hospitals. Determine the specific values ​​based on the actual situation of the data provider. Data record integrity: The calculation formula is: Data record integrity = Number of required fields filled / Preset total number of required fields, with a result range of 0-1. Provider historical accuracy: The calculation formula is: Provider historical accuracy = Number of reliable historical data from the provider / Total number of historical data submitted by the provider, with a result range of 0-1. Construct a comprehensive evaluation model: Determine the fusion algorithm for the comprehensive evaluation model. For example, if a weighted summation algorithm is used, the initial weights of each parameter can be preset to be equal, and can be fine-tuned later according to the actual evaluation effect. Clarify that the input of the comprehensive evaluation model is the selected evaluation parameters, and the output is a quantitative score in the range of 0-1, which is the reliability score.

[0052] For example, the following evaluation parameter combinations are selected: data source authority parameter, data record integrity parameter, internal logical consistency parameter, and provider historical accuracy parameter. The specific values ​​for each evaluation parameter are determined as follows: Data source authority parameter: The data provider is a chief physician, and the parameter value is set to 0.9 according to preset rules; Data record integrity parameter: There are 12 preset mandatory data items: 2 items of basic patient information, 4 items of four diagnostic methods, 1 item of diagnostic conclusion, 3 items of medication records, 1 item of efficacy evaluation, and 1 item of physician feedback. All patient data is filled in completely, so data record integrity = 1.0; Internal logical consistency parameter: 0.95; Provider historical accuracy parameter: Physician Li has submitted 100 clinical feedback data in the past, and after verification, 97 of them are reliable data, so the provider historical accuracy = 97 / 100 = 0.97. A comprehensive evaluation model is constructed: A weighted summation algorithm is used, with each parameter's weight preset to 0.25. The inputs to the comprehensive evaluation model are the above four parameters: 0.9, 1.0, 0.95, and 0.97, and the output is a reliability score.

[0053] Finally, the comprehensive evaluation model integrates and calculates the current evaluation parameters of the input clinical feedback data, outputting a quantitative score within a preset range as the credibility score for each piece of clinical feedback data. Integration calculation refers to the comprehensive evaluation model combining multiple input evaluation parameters through a preset algorithm to achieve comprehensive quantification of multi-dimensional indicators, ultimately outputting a single credibility score. The preset numerical range is 0-1, where 0 represents completely unreliable data and 1 represents completely reliable data; the higher the score, the higher the data credibility. The credibility score is the output of the comprehensive evaluation model and is a standardized quantitative indicator of the reliability of clinical feedback data, used for subsequent data screening.

[0054] For example, the parameters for data source authority (0.9), data record integrity (1.0), internal logical consistency (0.95), and provider historical accuracy (0.97) are input into the comprehensive evaluation model. A weighted summation algorithm is used, with each parameter having a weight of 0.25. The calculation process is: 0.9 × 0.25 + 1.0 × 0.25 + 0.95 × 0.25 + 0.97 × 0.25 ≈ 0.96. The output score is 0.96, meaning the credibility score of this clinical feedback data is 0.96, falling within the preset range of 0-1. This high score indicates extremely high data credibility.

[0055] In this embodiment of the invention, by standardizing the collection of clinical feedback data, the comprehensiveness and basic standardization of the data are ensured, providing qualified basic data for subsequent credibility assessment. By constructing a multi-dimensional comprehensive evaluation model and combining it with the knowledge base of TCM diagnostic rules to achieve internal logical consistency verification, the credibility of each piece of clinical feedback data is accurately quantified, effectively identifying low-credibility data. This ensures the data quality for the continuous learning of the TCM diagnosis and treatment system from the source, avoiding low-quality data from polluting the training set and affecting the system's training effect. At the same time, the standardized credibility scoring also provides a clear and quantifiable basis for subsequent data screening.

[0056] S200: The clinical feedback data is filtered according to a preset first confidence score threshold, and clinical feedback data with a confidence score greater than the first confidence score threshold are retained as data to be evaluated.

[0057] In this embodiment of the invention, the clinical feedback data is screened based on a preset first credibility score threshold, retaining clinical feedback data with a credibility score greater than the first credibility score threshold as data to be evaluated. After the data credibility assessment in S100, the clinical feedback data is assigned a credibility score in the range of 0-1. However, a large amount of low-credibility data still exists, such as data with low scores, contradictory medical logic, or unreliable sources. Even if low-credibility data enters the subsequent data value assessment stage, it is difficult for it to generate positive value for the continuous learning of the traditional Chinese medicine diagnosis and treatment system, and it will also consume computing resources and reduce learning efficiency. Therefore, it is necessary to preset a first credibility score threshold as a basic screening standard to filter out low-credibility data and retain only clinical feedback data with basic reliability as data to be evaluated, thus laying a solid data foundation for subsequent accurate value assessment.

[0058] The first credibility score threshold is a preset credibility score cutoff value, which is the standard for judging whether clinical feedback data has basic reliability. In this embodiment, the preset first credibility score threshold is 0.7. The data to be evaluated refers to the clinical feedback data with a credibility score greater than the first credibility score threshold after screening. This data is the sole source for subsequent data value assessment. All clinical feedback data with completed credibility scores are iterated through, and the credibility score of each data point is compared numerically with the preset first credibility score threshold. Clinical feedback data with credibility scores greater than the first credibility score threshold are retained, marked as data to be evaluated, and archived separately. Data with credibility scores less than or equal to the first credibility score threshold are removed to prevent them from entering subsequent stages. The selected data to be evaluated are batch-verified to confirm that there are no calculation errors in the scoring comparison process and that no data is omitted from the archiving process.

[0059] For example, the first confidence score threshold is preset to 0.7; the confidence scores of two sets of clinical feedback data are iterated: Clinical feedback data 1: clinical feedback data of patients with insomnia of liver stagnation and spleen deficiency type, with a confidence score of 0.96; Clinical feedback data 2: clinical feedback data of patients with wind-cold common cold, with a confidence score of 0.65; the confidence score of clinical feedback data 1 is 0.96 > 0.7, which meets the screening criteria and is retained as data to be evaluated; the confidence score of clinical feedback data 2 is 0.65 < 0.7, which does not meet the screening criteria and is removed; after confirming that the score calculation of the data to be evaluated in clinical feedback data 1 is correct and the archiving path is correct, the screening is completed.

[0060] In this embodiment of the invention, by using a hard screening method based on a first credibility scoring threshold, clinical feedback data with low credibility is directly eliminated. This prevents low-quality data with logical contradictions or unreliable sources from entering the subsequent value assessment stage, ensuring the basic reliability of the data to be assessed. It also reduces the processing workload of subsequent data value assessment, avoids wasting computational resources on low-quality data with no positive value, and improves the operational efficiency of the entire continuous learning process. Furthermore, it focuses subsequent assessment work on data with basic reliability, laying a data foundation for accurately assessing data value and achieving efficient learning.

[0061] S300: Based on the novelty and relevance of the data, the data to be evaluated is evaluated to obtain a data value score for each data to be evaluated.

[0062] In this embodiment of the invention, the data to be evaluated is assessed based on data novelty and task relevance to obtain a data value score for each piece of data. Although the data to be evaluated after the S200 screening possesses basic reliability, the value of different data for the continuous learning of the traditional Chinese medicine diagnosis and treatment system varies significantly: some data highly overlaps with historical training data, failing to provide new information for model optimization; some data has low relevance to the current performance optimization target, making it difficult to specifically improve the performance of the traditional Chinese medicine diagnosis and treatment system. Directly inputting all data to be evaluated into learning would result in wasted computational resources, low learning efficiency, and failure to achieve precise system optimization. Therefore, it is necessary to quantify data novelty and task relevance, dynamically allocate weights based on the current system performance, calculate data value scores, accurately distinguish the learning value of the data to be evaluated, and provide a quantitative basis for subsequent data classification and screening.

[0063] Step S300 in the method provided in this embodiment of the invention includes:

[0064] Calculate a first value coefficient to characterize the novelty of the data, wherein the first value coefficient reflects the overall degree of difference between the data to be evaluated and the data in the historical training data set of the traditional Chinese medicine diagnosis and treatment system, and the higher the overall degree of difference between the data to be evaluated and the historical training data set, the larger the value of the first value coefficient.

[0065] Calculate a second value coefficient to characterize task relevance, wherein the second value coefficient reflects the degree of correlation between the data to be evaluated and at least one performance optimization objective currently preset by the traditional Chinese medicine diagnosis and treatment system. The higher the degree of matching between the feature information of the data to be evaluated and the definition features of the current performance optimization objective, the larger the value of the second value coefficient.

[0066] Based on the current performance metric of the traditional Chinese medicine diagnosis and treatment system on the performance optimization target, a first weight is assigned to the first value coefficient, and a second weight is assigned to the second value coefficient;

[0067] Based on the first weight and the second weight, the first value coefficient and the second value coefficient are weighted and summed to obtain the data value score of the data to be evaluated.

[0068] First, a first value coefficient is calculated to characterize the novelty of the data. This first value coefficient reflects the overall degree of difference between the data to be evaluated and the data in the historical training data set of the traditional Chinese medicine diagnosis and treatment system. The higher the overall degree of difference between the data to be evaluated and the historical training data set, the larger the value of the first value coefficient. The first value coefficient is an indicator that quantifies the overall degree of difference between the data to be evaluated and the historical training data set of the traditional Chinese medicine diagnosis and treatment system. Its value ranges from 0 to 1. The higher the degree of difference, the larger the value of the first value coefficient; the lower the degree of difference, the smaller the value of the first value coefficient. The historical training data set refers to the collection of all reliable clinical feedback data that the traditional Chinese medicine diagnosis and treatment system has used for model training and optimization before this continuous learning, including diagnosis and treatment data of various TCM syndrome types and symptom combinations. The overall degree of difference refers to the average difference level between the features of the data to be evaluated, such as symptoms, tongue appearance, pulse appearance, and diagnostic conclusions, and all data features in the historical training data set, which can be quantified by a feature similarity algorithm.

[0069] Specifically, features consistent with the historical training data set are extracted from the data to be evaluated and standardized to ensure consistent feature dimensions. The archived historical training data set from the traditional Chinese medicine diagnosis and treatment system is obtained, and the core features of all data in the set are extracted to construct a historical feature library. A feature similarity algorithm, such as cosine similarity, is used to calculate the average similarity between the features of the data to be evaluated and all data features in the historical training data set; the lower the similarity, the higher the overall difference. According to a preset mapping rule, the overall difference is converted into a first value coefficient within the range of 0-1. The first value coefficient = 1 - average similarity; the lower the average similarity, the closer the first value coefficient is to 1; the higher the average similarity, the closer the first value coefficient is to 0.

[0070] For example, the features of the data to be evaluated are extracted; the historical training dataset is called, which contains 800 reliable clinical data sets, including 60 data sets related to insomnia due to liver stagnation and spleen deficiency. Using the cosine similarity algorithm, the average similarity between the data features of clinical feedback data 1 and the 60 similar historical data sets is calculated to be 0.08, which is extremely low, indicating that the overall difference is extremely high; according to the preset mapping rule, the first value coefficient = 1 - 0.08 = 0.92.

[0071] Secondly, a second value coefficient is calculated to characterize the relevance of the task. This second value coefficient reflects the degree of correlation between the data to be evaluated and at least one currently preset performance optimization goal of the traditional Chinese medicine (TCM) diagnosis and treatment system. The higher the matching degree between the feature information of the data to be evaluated and the defined features of the current performance optimization goal, the larger the value of the second value coefficient. The second value coefficient is an indicator that quantifies the degree of correlation between the data to be evaluated and at least one currently preset performance optimization goal of the TCM diagnosis and treatment system. Its value ranges from 0 to 1; the higher the correlation, the larger the value of the second value coefficient; the lower the correlation, the smaller the value of the second value coefficient. The current performance optimization goal refers to the optimization direction preset by the TCM diagnosis and treatment system in this continuous learning, such as improving the diagnostic accuracy of liver stagnation and spleen deficiency syndrome and qi stagnation and blood stasis syndrome, clarifying the syndrome types and corresponding features that need to be optimized. Feature matching degree refers to the matching ratio between the core features of the data to be evaluated and the core features defined by the current performance optimization goal.

[0072] Specifically, the performance optimization goal of the TCM diagnosis and treatment system for this continuous learning is determined, and the core features corresponding to the goal are identified; the core features of the data to be evaluated are compared with the core features of the current performance optimization goal, and the proportion of the number of matching features to the total number of core features of the optimization goal is calculated, i.e., the feature matching degree; the feature matching degree is directly mapped to a second value coefficient in the range of 0-1. The higher the matching degree, the larger the value of the second value coefficient. For example, a matching degree of 100% corresponds to a coefficient of 1.0, and a matching degree of 80% corresponds to a coefficient of 0.8.

[0073] For example, the optimization objective is to improve the diagnostic accuracy of liver stagnation and spleen deficiency syndrome and qi stagnation and blood stasis syndrome. The core features of liver stagnation and spleen deficiency syndrome are insomnia, irritability, poor appetite, thirst, pale tongue with thin white coating, wiry and thready pulse, and a diagnosis of liver stagnation and spleen deficiency syndrome, totaling 7 core features. The core features of the data to be evaluated are extracted as follows: insomnia, irritability, poor appetite, pale tongue with thin white coating, wiry and thready pulse, and a diagnosis of liver stagnation and spleen deficiency syndrome, totaling 6 core features. The feature matching degree is approximately 86% (6 / 7). According to the mapping rule, a matching degree of 86% corresponds to a second value coefficient of 0.86. Therefore, the second value coefficient of the data to be evaluated is determined to be 0.86.

[0074] Next, based on the current performance metric of the traditional Chinese medicine diagnosis and treatment system on the performance optimization target, a first weight is assigned to the first value coefficient, and a second weight is assigned to the second value coefficient.

[0075] Specifically, the process of assigning a first weight to the first value coefficient and a second weight to the second value coefficient based on the current performance metric of the traditional Chinese medicine diagnosis and treatment system on the performance optimization objective includes:

[0076] Obtain the prediction accuracy of the traditional Chinese medicine diagnosis and treatment system on the current performance optimization target, and use it as the current performance metric.

[0077] Set a weight allocation rule, wherein the second weight is negatively correlated with the current performance metric, the first weight is positively correlated with the current performance metric, and the sum of the first weight and the second weight is 1.

[0078] First, the predicted diagnosis accuracy rate of the traditional Chinese medicine (TCM) diagnosis and treatment system on the current performance optimization target is obtained as the current performance metric. The current performance metric refers to the predicted diagnosis accuracy rate of the TCM diagnosis and treatment system on the current performance optimization target, with a value ranging from 0 to 1. It is an indicator that measures the degree to which the system has achieved its current optimization target; the higher the accuracy rate, the better the performance of the TCM diagnosis and treatment system on that target. Through the performance testing module of the TCM diagnosis and treatment system, the predicted diagnosis accuracy rate of the TCM diagnosis and treatment system on the current performance optimization target is statistically analyzed and used as the current performance metric. For example, if the predicted diagnosis accuracy rate of the TCM diagnosis and treatment system on the current optimization target is 90%, then the current performance metric is 90%.

[0079] Secondly, a weighting rule is set, wherein the second weight is negatively correlated with the current performance metric, the first weight is positively correlated with the current performance metric, and the sum of the first weight and the second weight is 1. The weighting rule is: the higher the current performance metric, the lower the second weight; the first weight = 1 - the second weight, and the first weight + the second weight = 1. Here, the first weight is the weighted weight of the first value coefficient, which is positively correlated with the current performance metric; the higher the performance, the higher the first weight. The second weight is the weighted weight of the second value coefficient, which is negatively correlated with the current performance metric; the higher the performance, the lower the second weight. The second weight assignment rule is set as follows: second weight = 1 - current performance metric. For example, if the current performance metric is 90%, the second weight = 1 - 0.9 = 0.1, and the first weight = 1 - 0.1 = 0.9.

[0080] Finally, based on the first and second weights, the first and second value coefficients are weighted and summed to obtain the data value score of the data to be evaluated. Fusion calculation refers to the process of multiplying the first and second value coefficients by their corresponding weights and then summing the results to obtain the data value score. This ensures that the score comprehensively reflects the data's novelty, task relevance, and the current performance requirements of the traditional Chinese medicine (TCM) diagnosis and treatment system. The data value score is the final indicator that quantitatively represents the value of the data to be evaluated for the continuous learning of the TCM diagnosis and treatment system. Its value ranges from 0 to 1; the higher the score, the greater the value of the data for optimizing the TCM diagnosis and treatment system, and the more suitable it is for priority use in model training. Data value score = First value coefficient × First weight + Second value coefficient × Second weight. For example, if the first value coefficient = 0.92, the second value coefficient = 0.86, the first weight = 0.9, and the second weight = 0.1; the data value score = 0.92 × 0.9 + 0.86 × 0.1 ≈ 0.91.

[0081] The method provided in this embodiment of the invention further includes:

[0082] Add timestamp information to each of the data to be evaluated to record the time point when it was generated or collected;

[0083] Based on the time interval between the timestamp information and the current time, a time decay coefficient is calculated for each of the data to be evaluated, wherein the shorter the time interval, the larger the time decay coefficient.

[0084] In the data value assessment process, the time decay coefficient is introduced as a correction factor into the calculation of the data value score.

[0085] First, timestamp information is added to each of the data to be evaluated to record the time when it was generated or collected. Timestamp information refers to structured information used to accurately record the specific time when the data to be evaluated was generated or collected. For example, if the time when data 1 to be evaluated was generated is 2026-02-10 14:30:25, this timestamp is added; if the time when data 2 to be evaluated was generated is 2026-01-15 09:15:40, this timestamp is added.

[0086] Secondly, based on the time interval between the timestamp information and the current time, a time decay coefficient is calculated for each of the data to be evaluated. The shorter the time interval, the larger the time decay coefficient. The time interval refers to the difference between the timestamp time of the data to be evaluated and the current time when the system performs data value evaluation, calculated in days, and is a quantitative indicator of data timeliness. The time decay coefficient is a correction factor calculated based on the time interval, ranging from 0 to 1. The shorter the time interval, the stronger the timeliness, and the larger the time decay coefficient; the longer the time interval, the weaker the timeliness, and the smaller the time decay coefficient, used to adapt to clinical concept drift. Clinical concept drift refers to the gradual change over time in the TCM diagnosis and treatment scenario due to updates in medical concepts, changes in patient group characteristics, etc., leading to the correspondence between symptoms, syndrome types, and medications, and the evaluation standards for efficacy.

[0087] Specifically, the time interval is calculated by subtracting the timestamp from the current time for each data point, and then converting the result to days. The preset time decay coefficient mapping rules are: time interval ≤ 3 days: time decay coefficient = 0.98; 3 days < time interval ≤ 7 days: time decay coefficient = 0.95; 7 days < time interval ≤ 30 days: time decay coefficient = 0.90; 30 days < time interval ≤ 90 days: time decay coefficient = 0.80; time interval > 90 days: time decay coefficient = 0.70. For example, current time: 2026-02-11 10:00:00; calculated time interval: for data to be evaluated 1, the time interval ≈ 0.81 days ≤ 3 days, so the time decay coefficient = 0.98; for data to be evaluated 2, the time interval = 27 days, 7 days < 27 days ≤ 30 days, so the time decay coefficient = 0.90.

[0088] Therefore, in the data value assessment process, the time decay coefficient is introduced as a correction factor into the calculation of the data value score. The correction factor refers to the time decay coefficient as a weighting factor adjusting the original value score. It is incorporated into the score calculation through multiplication, reflecting the impact of timeliness on the actual value of the data. The original value score and corresponding time decay coefficient of each piece of data to be evaluated are retrieved; the corrected score = original value score × time decay coefficient. For example, the original score of data 1 to be evaluated is 0.91, and the time decay coefficient is 0.98, so the corrected data value score = 0.91 × 0.98 = 0.89; the original score of data 2 to be evaluated is 0.85, and the time decay coefficient is 0.90; the corrected data value score = 0.85 × 0.90 ≈ 0.77.

[0089] In this embodiment of the invention, by calculating data novelty and task relevance step by step, the learning value of the data to be evaluated is quantified in multiple dimensions, effectively distinguishing between high-value and low-value data. Based on the current performance metrics of the TCM diagnosis and treatment system, weights are dynamically allocated, conforming to the requirement that higher performance corresponds to lower task relevance weights. This achieves a precise match between value assessment and the optimization goals of the TCM diagnosis and treatment system. When performance is superior, novel data is prioritized; when performance is poor, relevant data is prioritized, improving the targeting of learning. This provides a standardized and quantifiable value basis for the subsequent hierarchical screening of data to be evaluated, allowing for the priority selection of high-value data for training, avoiding the waste of computational resources on low-value data, and ensuring that training data can specifically compensate for system performance shortcomings, promoting efficient and accurate optimization of the TCM diagnosis and treatment system. Simultaneously, by using a time decay coefficient to weaken the value weight of outdated data and strengthen the weight of highly timely data, the value score aligns with current diagnostic and treatment needs, avoiding misjudgments due to concept drift and improving assessment accuracy. The entire calculation process is traceable and repeatable, and the parameter settings conform to the TCM diagnosis and treatment rules and learning needs, ensuring the accuracy and reliability of data value scoring.

[0090] S400: Based on the data value scoring threshold dynamically adjusted according to the current performance of the traditional Chinese medicine diagnosis and treatment system, the data to be evaluated is screened and classified according to the credibility score and data value score to obtain a first-level learning data set, a second-level learning data set and a third-level learning data set.

[0091] In this embodiment of the invention, based on the data value scoring threshold dynamically adjusted according to the current performance of the traditional Chinese medicine diagnosis and treatment system, the data to be evaluated is screened and classified according to the credibility score and the data value score, resulting in a first-level learning dataset, a second-level learning dataset, and a third-level learning dataset. Although the data to be evaluated has undergone credibility and value assessment, if the data value scoring threshold is set fixedly, it cannot adapt to the real-time changing performance state of the traditional Chinese medicine diagnosis and treatment system, which can easily lead to a mismatch between data screening criteria and system optimization needs. At the same time, there are differences in the timeliness of data in clinical practice, and long-term outdated data can cause clinical concept drift, affecting the effectiveness of model learning. In addition, there are fundamental differences in the learning methods adapted to data with different credibility and value, and a single learning mode cannot maximize data value. Therefore, it is necessary to dynamically adjust the value scoring threshold through system performance, introduce a time decay coefficient to correct data value, combine credibility and value scores to complete the three-level data classification, and match corresponding learning modes for each level of data to achieve accurate adaptation between data and learning strategies, while adapting to clinical practice concept drift and ensuring the relevance and stability of the system's continuous learning.

[0092] Step S400 in the method provided in this embodiment of the invention includes:

[0093] Among them, based on the data value scoring threshold dynamically adjusted according to the current performance of the traditional Chinese medicine diagnosis and treatment system, the data to be evaluated is screened and classified according to the credibility score and data value score, including:

[0094] A basic data value scoring threshold is set in advance;

[0095] The ratio of the current performance metric to the preset baseline performance metric of the current performance optimization target is used as the threshold adjustment factor;

[0096] The basic data value scoring threshold is optimized and adjusted according to the threshold adjustment factor to obtain the data value scoring threshold actually used in the current period;

[0097] Data to be evaluated that have a data value score higher than the data value score threshold and a credibility score higher than the second credibility score threshold are included in the first-level learning data set, wherein the second credibility score threshold is greater than the first credibility score threshold.

[0098] First, a basic data value scoring threshold is pre-set. This threshold serves as an initial benchmark for data value screening and forms the basis for subsequent dynamic adjustments. Its value ranges from 0 to 1 and can be preset based on the initial performance of the TCM diagnosis and treatment system and the baseline quality of clinical data. It does not change in real-time with system performance. The range of the basic data value scoring threshold is determined by considering the application scenario of the TCM diagnosis and treatment system, the overall quality level of clinical data, and the initial performance optimization requirements of the system. A specific basic threshold value is preset to ensure that it can initially screen out data with certain learning value while also reserving reasonable space for subsequent dynamic adjustments. This avoids overly high values ​​leading to the omission of high-value data or overly low values ​​leading to misjudgment of low-value data. For example, considering the initial performance of the TCM diagnosis and treatment system and the value distribution of reliable clinical data, the preset basic data value scoring threshold is 0.7.

[0099] Secondly, the ratio of the current performance metric to the preset baseline performance metric of the current performance optimization target is used as the threshold adjustment factor. The preset baseline performance metric is a fixed constant set for the current performance optimization target, ranging from 0 to 1. It is preset by technical personnel based on clinical diagnostic and treatment needs and system design goals, representing the ideal basic performance level of the system at that optimization target. The threshold adjustment factor = current performance metric / preset baseline performance metric. The threshold adjustment factor is a quantification coefficient used to dynamically adjust the basic data value scoring threshold. It is positively correlated with the current performance metric; the higher the current performance, the larger the adjustment factor; the lower the current performance, the smaller the adjustment factor. For example, if the current performance metric is 90% and the preset baseline performance metric is 80%, the threshold adjustment factor = current performance metric / preset baseline performance metric = 90% / 80% = 1.125.

[0100] Next, the basic data value scoring threshold is optimized and adjusted according to the threshold adjustment factor to obtain the actual data value scoring threshold used in the current period. Optimization adjustment refers to the process of adjusting the basic data value scoring threshold by multiplication using the threshold adjustment factor, so that the adjusted threshold is adapted to the current system performance. The higher the system performance, the higher the current threshold; the lower the system performance, the lower the current threshold. The actual data value scoring threshold used in the current period refers to the actual critical value used to screen high-value data for evaluation within this continuous learning period after dynamic adjustment. Its value ranges from 0 to 1 and is dynamically updated according to changes in system performance, and may differ across different learning periods. The optimization adjustment is performed using multiplication, and the adjustment formula is: Actual data value scoring threshold used in the current period = Basic data value scoring threshold × Threshold adjustment factor. For example, if the basic data value scoring threshold = 0.7 and the threshold adjustment factor = 1.125, the actual data value scoring threshold used in the current period = 0.7 × 1.125 ≈ 0.79.

[0101] Furthermore, data to be evaluated that have a data value score higher than the data value score threshold and a credibility score higher than the second credibility score threshold are included in the first-level learning data set. The second credibility score threshold is greater than the first credibility score threshold. The second credibility score threshold is a critical value used to distinguish between high-credibility and medium-to-low-credibility data, ranging from 0 to 1. It is the core credibility standard for this first-level data screening, explicitly requiring its value to be greater than the preset first credibility score threshold in S200 to ensure the high credibility of the first-level data. The first-level learning data set refers to the high-value, high-credibility, high-quality data set obtained in this screening. This set will subsequently be used as high-confidence labeled data for supervised learning of the traditional Chinese medicine diagnosis and treatment system, directly and accurately updating model parameters.

[0102] For example, a second credibility score threshold is preset: combining the high credibility requirement of the first-level data, the second credibility threshold is set to 0.85, confirming that 0.85 > the first credibility score threshold of 0.7. The parameters of the data to be evaluated are retrieved: data value score = 0.89, credibility score = 0.96. A comparison of the two thresholds shows that the data value score of 0.89 > the current data value score threshold of 0.79, meeting the value requirement; the credibility score of 0.96 > the second credibility score threshold of 0.85, meeting the credibility requirement. This data meets both conditions, is marked as first-level learning data, included in the first-level learning data set, archived separately, and used for subsequent supervised learning.

[0103] The screening and classification of the data to be evaluated also includes:

[0104] Data to be evaluated that have a data value score higher than the data value score threshold and a credibility score lower than or equal to the second credibility score threshold are included in the secondary learning data set.

[0105] A preset secondary data value scoring threshold is established, wherein the secondary data value scoring threshold is lower than the data value scoring threshold.

[0106] Data to be evaluated that has a data value score lower than the data value score threshold but higher than or equal to the secondary data value score threshold, and a credibility score higher than the second credibility score threshold, will be classified into the third-level learning data set.

[0107] First, data to be evaluated that have a data value score higher than the data value score threshold and a credibility score lower than or equal to the second credibility score threshold are included in the secondary learning data set. The secondary learning data set consists of high-value, medium-to-low-credibility data to be evaluated, which will subsequently be used as reinforcement learning data to construct a simulated learning environment and train the system to optimize diagnostic and treatment strategies under uncertainty. The previously determined current data value score threshold and second credibility score threshold are retrieved; all data to be evaluated not yet included in the primary learning data set are iterated through; the two scores of each data point are compared to determine whether the data value score is higher than the current data value score threshold and whether the credibility score is lower than or equal to the second credibility score threshold; data that simultaneously meets both conditions are marked as secondary learning data and included in the secondary learning data set for reinforcement learning.

[0108] For example, the current data value score threshold is 0.79, and the second credibility score threshold is 0.85. Iterate through the data to be evaluated that are not included in the first-level set, and select one data point: data value score = 0.85, credibility score = 0.8. Conditional judgment: if the data value score 0.85 > 0.79 and the credibility score 0.80 ≤ 0.85, both conditions are met; mark this data as second-level learning data, include it in the second-level learning data set, archive it separately, and label it for reinforcement learning to construct a simulated learning scenario.

[0109] Secondly, a secondary data value scoring threshold is preset, wherein the secondary data value scoring threshold is lower than the primary data value scoring threshold. The secondary data value scoring threshold is a critical value used to distinguish between medium-value and low-value data to be evaluated, ranging from 0 to 1. It must be lower than the data value scoring threshold actually used in the current period and serves as the value standard for screening the subsequent three-level learning data set. Medium-value data refers to data whose data value score is lower than the current high-value threshold but higher than or equal to the secondary threshold; it possesses some learning value and does not need to be discarded. Low-value data refers to data whose data value score is lower than the secondary data value scoring threshold; it has extremely low learning value and will be directly removed from subsequent system learning.

[0110] Specifically, based on the value score distribution of the data to be evaluated, the needs of unsupervised learning in the TCM diagnosis and treatment system, and the current data value score threshold, the range of the secondary data value score threshold is determined. This threshold must be lower than the current data value score threshold and higher than the value baseline corresponding to the first confidence threshold to avoid low-value data being mixed in. The specific value of the secondary data value score threshold is preset to ensure it is strictly lower than the current data value score threshold, while reserving a reasonable range to effectively distinguish between medium-value and low-value data. The preset basis and specific value of the secondary data value score threshold are recorded and archived in the parameter record of this learning cycle, stored in association with the basic threshold, current threshold, and second confidence threshold for easy subsequent traceability and verification. It is confirmed that the secondary threshold is < the current data value score threshold and falls within the range of 0-1 to avoid values ​​that are too high, leading to the omission of medium-value data, or values ​​that are too low, leading to misjudgment of low-value data.

[0111] For example, the data value scores to be evaluated are distributed between 0.70 and 0.95. The current data value score threshold is 0.79. Based on the basic requirements of unsupervised learning for data value, the preset secondary threshold should be lower than 0.79 and higher than 0.70. The secondary data value score threshold is determined to be 0.75. It is confirmed that 0.75 < 0.79 and is within the range of 0-1. The threshold preset is reasonable and can be used for subsequent three-level data screening.

[0112] Furthermore, data to be evaluated that have a data value score lower than the data value score threshold but higher than or equal to the secondary data value score threshold, and a credibility score higher than the second credibility score threshold, are categorized into the tertiary learning dataset. The tertiary learning dataset refers to a dataset of data to be evaluated with moderate data value and high credibility. This dataset will be subsequently used in the unsupervised learning process to uncover potential, unlabeled clinical patterns through data analysis, or for model consolidation training to maintain the stability of existing knowledge.

[0113] The system retrieves the previously determined thresholds for current data value scoring, secondary data value scoring, and second credibility scoring. It then iterates through all data not yet included in the first or second-level learning datasets. Each data point's score is compared to the second-level threshold to determine if it simultaneously meets both of the following conditions: if the data value score is less than the current data value threshold and greater than or equal to the secondary data value threshold, it is classified as medium-value data; if the credibility score is greater than the second-level credibility threshold, it is classified as high-credibility data. Data meeting both conditions is marked as third-level learning data, included in the third-level learning dataset, and archived separately for unsupervised learning. Data not included in the first, second, or third-level datasets is directly removed and not used in subsequent system learning to avoid wasting computational resources.

[0114] For example, the current data value score threshold is 0.79, the secondary data value score threshold is 0.75, and the second confidence score threshold is 0.85. Iterating through the data not yet included in the first or second-level sets, one data point is selected: data value score = 0.77, confidence score = 0.88. Conditional judgment: data value score 0.77 < 0.79 and ≥ 0.75, and confidence score 0.88 > 0.85, both conditions are met. This data is marked as third-level learning data, included in the third-level learning data set, archived separately, and labeled for unsupervised learning, mining potential clinical patterns, or model consolidation training.

[0115] In this embodiment of the invention, a dynamic adjustment mechanism that is positively correlated with system performance through a threshold adjustment factor allows the data value screening criteria to be adapted to the real-time performance of the traditional Chinese medicine diagnosis and treatment system; a time decay coefficient is introduced to correct data value, effectively adapting to the concept drift in clinical practice and avoiding interference from outdated data in model learning; a three-level data classification is completed by combining credibility and value scores, achieving precise matching between data quality and learning strategies; first-level data is used to accurately update model parameters through supervised learning, second-level data is used to improve decision-making ability in uncertain scenarios through reinforcement learning, and third-level data is used to mine potential clinical patterns and consolidate model knowledge through unsupervised learning, comprehensively improving the efficiency, relevance, and long-term stability of the system's continuous learning.

[0116] S500: Supervised learning is performed on the traditional Chinese medicine diagnosis and treatment system using the first-level learning dataset, reinforcement learning is performed on the traditional Chinese medicine diagnosis and treatment system using the second-level learning dataset, and unsupervised learning is performed on the traditional Chinese medicine diagnosis and treatment system using the third-level learning dataset, thereby achieving continuous optimization of the traditional Chinese medicine diagnosis and treatment system.

[0117] In this embodiment of the invention, the first-level learning dataset is used for supervised learning of the traditional Chinese medicine (TCM) diagnosis and treatment system, the second-level learning dataset is used for reinforcement learning, and the third-level learning dataset is used for unsupervised learning, thereby achieving continuous optimization of the TCM diagnosis and treatment system. After the three-level learning data is classified in S400, datasets with different characteristics are adapted to different machine learning paradigms: first-level data has high credibility and high value, serving as high-quality labeled data for accurately updating the core model; second-level data has high value but limited credibility, suitable for exploring diagnosis and treatment decisions in uncertain scenarios through reinforcement learning; third-level data has high credibility but moderate value, and can be used to uncover potential clinical patterns or consolidate model knowledge through unsupervised learning. If a single learning mode is used, it is impossible to maximize the learning value of various types of data, nor can it be used to specifically optimize the system's core diagnostic accuracy, decision-making ability in complex scenarios, knowledge stability, and other performance dimensions. Therefore, it is necessary to match a dedicated learning paradigm to the third-level dataset, achieving comprehensive and refined continuous optimization of the TCM diagnosis and treatment system through a combination of supervised, reinforcement, and unsupervised learning.

[0118] First, the primary learning dataset is used to perform supervised learning on the traditional Chinese medicine (TCM) diagnosis and treatment system. The primary learning dataset is then converted into high-confidence labeled data for supervised learning of the diagnostic model of the TCM diagnosis and treatment system, enabling direct and accurate updating of model parameters. Supervised learning is a machine learning paradigm based on labeled data. By minimizing the error between the model's predictions and the labeled results, iterative backpropagation updates the model parameters, achieving precise parameter optimization.

[0119] Specifically, the core input features of each primary data point are extracted: symptom description, tongue appearance features, and pulse appearance features. Labels are added: preliminary diagnostic conclusions and clinically validated optimal treatment plans. Features are standardized, such as by unifying encoding formats and normalizing numerical ranges, to ensure the input format is compatible with the diagnostic model. Training / validation subsets are divided: the primary learning data set is split into training and validation sets in a 7:3 ratio to avoid model overfitting. The training set is input into the diagnostic model, using the labeled values ​​as prediction targets. The prediction error, such as cross-entropy loss, is calculated using backpropagation. The model's weights, biases, and other core parameters are iteratively updated to gradually reduce the deviation between predicted and labeled values. The validation set is used to evaluate the model's training effect. If the accuracy does not meet expectations, hyperparameters such as the learning rate, number of iterations, and batch size are adjusted, and retraining is performed until performance meets the target. The optimized diagnostic model parameters are saved and used to replace the original parameters, completing the accurate update of the diagnostic model.

[0120] Secondly, the traditional Chinese medicine diagnosis and treatment system is subjected to reinforcement learning using the aforementioned secondary learning dataset. The secondary learning dataset and its corresponding clinical information are used to construct a reinforcement learning simulation environment to train the decision-making model and explore optimized diagnosis and treatment strategies under uncertainty.

[0121] Specifically, each data point in the secondary learning dataset is encapsulated as an independent simulated learning scenario. Each scenario contains complete information on the initial patient state, initial treatment plan, and clinical outcome. The initial patient state information in each data point is extracted and transformed into a standardized vector, such as one-hot encoding or numerical normalization, and defined as the state S of the reinforcement learning environment. The initial treatment plan generated by the traditional Chinese medicine treatment system based on state S is defined as the action A that the reinforcement learning agent can choose in this state. A preset reward rule is used to calculate the total reward value: efficacy weight (0.6): significant effect = 0.6, effective = 0.3, ineffective = -0.3; safety weight (0.4): none. Adverse reaction = 0.4, mild adverse reaction = -0.2, severe adverse reaction = -0.7; total reward value = efficacy score + safety score; the reinforcement learning agent sequentially executes the process of perceiving state S, selecting action A, receiving reward value, and updating policy in multiple simulated scenarios, and iteratively optimizes the policy network parameters through the temporal difference (TD) algorithm, with the goal of maximizing long-term cumulative reward; the above interaction process is repeated in all secondary simulated scenarios until the agent's cumulative reward value tends to stabilize, indicating that the decision policy has been optimized; the trained decision model is integrated into the traditional Chinese medicine diagnosis and treatment system to replace the original decision module and improve the system's decision performance in complex and uncertain scenarios.

[0122] Finally, the three-level learning dataset is used to perform unsupervised learning on the traditional Chinese medicine diagnosis and treatment system. The three-level learning dataset is used in the unsupervised learning process to uncover potential, unlabeled clinical patterns through data analysis, or it is used for model consolidation training to maintain the stability of existing knowledge.

[0123] Specifically, the core features of each tertiary data point are extracted, redundant information is removed, missing feature values ​​are corrected, and the feature space is unified to ensure data quality is suitable for unsupervised learning. The K-means clustering algorithm is used to cluster the tertiary data, with the number of clusters preset based on TCM clinical knowledge, such as 8 clusters according to TCM syndrome types. The clustering results are analyzed, common features within each cluster are extracted, and potential associations not explicitly labeled are discovered. TCM experts are invited to verify the discovery results, and patterns consistent with clinical logic are added to the TCM diagnostic rules knowledge base. The discovered potential clinical patterns are added to the system knowledge base, and the model parameters after consolidation training are solidified, completing the supplementation and stabilization of system knowledge.

[0124] Through the specific implementation methods described above, the embodiments of the present invention achieve the following technical effects:

[0125] This invention provides a continuous learning method and system for a traditional Chinese medicine (TCM) diagnosis and treatment system based on clinical feedback. Through multi-dimensional data credibility assessment, it quantifies the reliability of clinical feedback data, eliminating low-quality data with logical contradictions and unreliable sources at the source, ensuring the basic quality of the data. By screening through a first credibility threshold, the data is further refined to obtain data with basic reliability for evaluation, laying a solid foundation for subsequent value assessment. The data value is quantified by combining data novelty and task relevance, and weights are dynamically allocated based on the system's current performance to achieve precise matching between value assessment and system optimization goals. The value scoring threshold is dynamically adjusted based on system performance, and clinical concept drift is corrected using a time decay coefficient, completing the precise classification of the three-level learning data set and achieving scientific adaptation between data and learning models. Supervised, reinforcement, and unsupervised combined learning paradigms are adopted for the three levels of data respectively. First-level data is used to accurately update model parameters and solidify diagnostic accuracy; second-level data is used to improve the system's decision-making ability in complex scenarios; and third-level data is used to mine potential clinical patterns and consolidate model knowledge, maximizing the value of clinical feedback data. Ultimately, this achieves a comprehensive improvement in the diagnostic accuracy, decision-making performance, and knowledge stability of the TCM diagnosis and treatment system, enabling adaptive and sustainable optimization of the TCM diagnosis and treatment system.

[0126] Example 2, as Figure 3 As shown, this invention provides a continuous learning system for traditional Chinese medicine diagnosis and treatment based on clinical feedback, the system comprising:

[0127] The credibility assessment module 11 is used to assess the credibility of the collected clinical feedback data and obtain a credibility score for each clinical feedback data.

[0128] The credibility screening module 12 is used to screen the clinical feedback data according to a preset first credibility score threshold, and retain the clinical feedback data with a credibility score greater than the first credibility score threshold as data to be evaluated.

[0129] Value assessment module 13 is used to assess the data value of the data to be assessed based on data novelty and task relevance, and obtain a data value score for each data to be assessed.

[0130] The dynamic classification module 14 is used to dynamically adjust the data value scoring threshold based on the current performance of the traditional Chinese medicine diagnosis and treatment system, and to filter and classify the data to be evaluated according to the credibility score and the data value score, so as to obtain a first-level learning data set, a second-level learning data set and a third-level learning data set.

[0131] The continuous optimization module 15 is used to perform supervised learning on the traditional Chinese medicine diagnosis and treatment system using the first-level learning data set, reinforcement learning on the traditional Chinese medicine diagnosis and treatment system using the second-level learning data set, and unsupervised learning on the traditional Chinese medicine diagnosis and treatment system using the third-level learning data set, so as to achieve continuous optimization of the traditional Chinese medicine diagnosis and treatment system.

[0132] In one embodiment, the credibility assessment module 11 is further configured to:

[0133] The clinical feedback data includes at least patient medical records, medication records, efficacy evaluation records, and physician feedback information.

[0134] This includes assessing the reliability of the collected clinical feedback data, including:

[0135] A comprehensive evaluation model is constructed, which receives multiple preset evaluation parameters as input. The evaluation parameters include at least two or more of the following: data source authority parameter, data record integrity parameter, internal logical consistency parameter, and provider historical accuracy parameter. The data source authority parameter is determined based on the data provider's registration qualification level or professional title level, and the data record integrity parameter is determined based on the filling ratio of preset mandatory data items in the clinical feedback data.

[0136] The comprehensive evaluation model integrates and calculates the current evaluation parameters of the input clinical feedback data, and outputs a quantitative score within a preset range as the credibility score of each clinical feedback data.

[0137] The methods for determining the internal logical consistency parameters include:

[0138] Extract symptom descriptions, tongue features, pulse characteristics, and preliminary diagnostic conclusions from clinical feedback data;

[0139] Based on a pre-defined knowledge base of TCM diagnostic rules, the medical compatibility among the symptom description text information, the tongue appearance feature information, the pulse appearance feature information, and the preliminary diagnosis conclusion information is determined, and a consistency judgment score is generated based on the compatibility judgment result. The higher the medical compatibility, the higher the consistency judgment score.

[0140] The consistency score is used as the internal logical consistency parameter.

[0141] In one embodiment, the value assessment module 13 is further used for:

[0142] Calculate a first value coefficient to characterize the novelty of the data, wherein the first value coefficient reflects the overall degree of difference between the data to be evaluated and the data in the historical training data set of the traditional Chinese medicine diagnosis and treatment system, and the higher the overall degree of difference between the data to be evaluated and the historical training data set, the larger the value of the first value coefficient.

[0143] Calculate a second value coefficient to characterize task relevance, wherein the second value coefficient reflects the degree of correlation between the data to be evaluated and at least one performance optimization objective currently preset by the traditional Chinese medicine diagnosis and treatment system. The higher the degree of matching between the feature information of the data to be evaluated and the definition features of the current performance optimization objective, the larger the value of the second value coefficient.

[0144] Based on the current performance metric of the traditional Chinese medicine diagnosis and treatment system on the performance optimization target, a first weight is assigned to the first value coefficient, and a second weight is assigned to the second value coefficient;

[0145] Based on the first weight and the second weight, the first value coefficient and the second value coefficient are weighted and summed to obtain the data value score of the data to be evaluated.

[0146] Specifically, the process of assigning a first weight to the first value coefficient and a second weight to the second value coefficient based on the current performance metric of the traditional Chinese medicine diagnosis and treatment system on the performance optimization objective includes:

[0147] Obtain the prediction accuracy of the traditional Chinese medicine diagnosis and treatment system on the current performance optimization target, and use it as the current performance metric.

[0148] Set a weight allocation rule, wherein the second weight is negatively correlated with the current performance metric, the first weight is positively correlated with the current performance metric, and the sum of the first weight and the second weight is 1.

[0149] In one embodiment, the value assessment module 13 is further used for:

[0150] Add timestamp information to each of the data to be evaluated to record the time point when it was generated or collected;

[0151] Based on the time interval between the timestamp information and the current time, a time decay coefficient is calculated for each of the data to be evaluated, wherein the shorter the time interval, the larger the time decay coefficient.

[0152] In the data value assessment process, the time decay coefficient is introduced as a correction factor into the calculation of the data value score.

[0153] First, timestamp information is added to each of the data to be evaluated to record the time when it was generated or collected.

[0154] Secondly, based on the time interval between the timestamp information and the current time, a time decay coefficient is calculated for each of the data to be evaluated, wherein the shorter the time interval, the larger the time decay coefficient.

[0155] Therefore, in the data value assessment process, the time decay coefficient is introduced as a correction factor into the calculation of the data value score.

[0156] In one embodiment, the dynamic classification module 14 is further configured to:

[0157] Among them, based on the data value scoring threshold dynamically adjusted according to the current performance of the traditional Chinese medicine diagnosis and treatment system, the data to be evaluated is screened and classified according to the credibility score and data value score, including:

[0158] A basic data value scoring threshold is set in advance;

[0159] The ratio of the current performance metric to the preset baseline performance metric of the current performance optimization target is used as the threshold adjustment factor;

[0160] The basic data value scoring threshold is optimized and adjusted according to the threshold adjustment factor to obtain the data value scoring threshold actually used in the current period;

[0161] Data to be evaluated that have a data value score higher than the data value score threshold and a credibility score higher than the second credibility score threshold are included in the first-level learning data set, wherein the second credibility score threshold is greater than the first credibility score threshold.

[0162] The screening and classification of the data to be evaluated also includes:

[0163] Data to be evaluated that have a data value score higher than the data value score threshold and a credibility score lower than or equal to the second credibility score threshold are included in the secondary learning data set.

[0164] A preset secondary data value scoring threshold is established, wherein the secondary data value scoring threshold is lower than the data value scoring threshold.

[0165] Data to be evaluated that has a data value score lower than the data value score threshold but higher than or equal to the secondary data value score threshold, and a credibility score higher than the second credibility score threshold, will be classified into the third-level learning data set.

[0166] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A continuous learning method for a traditional Chinese medicine diagnosis and treatment system based on clinical feedback, characterized in that, The method includes: The collected clinical feedback data were evaluated for data reliability, and a reliability score was obtained for each piece of clinical feedback data, including: A comprehensive evaluation model is constructed, which receives multiple preset evaluation parameters as input. The evaluation parameters include at least two or more of the following: data source authority parameter, data record integrity parameter, internal logical consistency parameter, and provider historical accuracy parameter. The data source authority parameter is determined based on the data provider's registration qualification level or professional title level, and the data record integrity parameter is determined based on the filling ratio of preset mandatory data items in the clinical feedback data. The comprehensive evaluation model integrates and calculates the current evaluation parameters of the input clinical feedback data, and outputs a quantitative score within a preset range as the credibility score of each clinical feedback data. The clinical feedback data is filtered according to a preset first confidence score threshold, and clinical feedback data with a confidence score greater than the first confidence score threshold are retained as data to be evaluated. The data value of the data to be evaluated is assessed based on data novelty and task relevance, resulting in a data value score for each piece of data, including: Calculate a first value coefficient to characterize the novelty of the data, wherein the first value coefficient reflects the overall degree of difference between the data to be evaluated and the data in the historical training data set of the traditional Chinese medicine diagnosis and treatment system, and the higher the overall degree of difference between the data to be evaluated and the historical training data set, the larger the value of the first value coefficient. Calculate a second value coefficient to characterize task relevance, wherein the second value coefficient reflects the degree of correlation between the data to be evaluated and at least one performance optimization objective currently preset by the traditional Chinese medicine diagnosis and treatment system. The higher the degree of matching between the feature information of the data to be evaluated and the definition features of the current performance optimization objective, the larger the value of the second value coefficient. Based on the current performance metric of the traditional Chinese medicine diagnosis and treatment system on the performance optimization objective, a first weight is assigned to the first value coefficient, and a second weight is assigned to the second value coefficient, including: Obtain the predictive diagnosis and treatment accuracy of the traditional Chinese medicine diagnosis and treatment system on the current performance optimization target, and use it as the current performance metric. Set a weight allocation rule, wherein the second weight is negatively correlated with the current performance metric, the first weight is positively correlated with the current performance metric, and the sum of the first weight and the second weight is 1; Based on the first weight and the second weight, the first value coefficient and the second value coefficient are weighted and summed to obtain the data value score of the data to be evaluated; Based on the data value scoring threshold of the current dynamic adjustment of the performance of the traditional Chinese medicine diagnosis and treatment system, the data to be evaluated is screened and classified according to the credibility score and data value score to obtain a first-level learning data set, a second-level learning data set and a third-level learning data set; The traditional Chinese medicine diagnosis and treatment system is subjected to supervised learning using the first-level learning dataset, reinforcement learning using the second-level learning dataset, and unsupervised learning using the third-level learning dataset, thereby achieving continuous optimization of the traditional Chinese medicine diagnosis and treatment system. The determination method for the internal logical consistency parameter includes: Extract symptom descriptions, tongue features, pulse characteristics, and preliminary diagnostic conclusions from clinical feedback data; Based on a pre-defined knowledge base of TCM diagnostic rules, the medical compatibility among the symptom description text information, the tongue appearance feature information, the pulse appearance feature information, and the preliminary diagnosis conclusion information is determined, and a consistency judgment score is generated based on the compatibility judgment result. The higher the medical compatibility, the higher the consistency judgment score. The consistency score is used as the internal logical consistency parameter.

2. The continuous learning method for a traditional Chinese medicine diagnosis and treatment system based on clinical feedback according to claim 1, characterized in that, The clinical feedback data includes at least patient medical records, medication records, efficacy evaluation records, and physician feedback information.

3. The continuous learning method for a traditional Chinese medicine diagnosis and treatment system based on clinical feedback according to claim 1, characterized in that, Based on the data value scoring threshold dynamically adjusted according to the current performance of the traditional Chinese medicine diagnosis and treatment system, the data to be evaluated is screened and classified according to the credibility score and data value score, including: Pre-set a basic data value scoring threshold; The ratio of the current performance metric to the preset baseline performance metric of the current performance optimization target is used as the threshold adjustment factor; The basic data value scoring threshold is optimized and adjusted according to the threshold adjustment factor to obtain the data value scoring threshold actually used in the current period; Data to be evaluated that have a data value score higher than the data value score threshold and a credibility score higher than the second credibility score threshold are included in the first-level learning data set, wherein the second credibility score threshold is greater than the first credibility score threshold.

4. The continuous learning method for a traditional Chinese medicine diagnosis and treatment system based on clinical feedback according to claim 3, characterized in that, The screening and classification of the data to be evaluated also includes: Data to be evaluated that have a data value score higher than the data value score threshold and a credibility score lower than or equal to the second credibility score threshold are included in the secondary learning data set. A preset secondary data value scoring threshold is established, wherein the secondary data value scoring threshold is lower than the data value scoring threshold. Data to be evaluated that has a data value score lower than the data value score threshold but higher than or equal to the secondary data value score threshold, and a credibility score higher than the second credibility score threshold, will be classified into the third-level learning data set.

5. The continuous learning method for a traditional Chinese medicine diagnosis and treatment system based on clinical feedback according to claim 1, characterized in that, The method also includes: Add timestamp information to each of the data to be evaluated to record the time point when it was generated or collected; Based on the time interval between the timestamp information and the current time, a time decay coefficient is calculated for each of the data to be evaluated, wherein the shorter the time interval, the larger the time decay coefficient. In the data value assessment process, the time decay coefficient is introduced as a correction factor into the calculation of the data value score.

6. A continuous learning system for TCM diagnosis and treatment based on clinical feedback, characterized in that: A method for implementing the continuous learning method of a traditional Chinese medicine diagnosis and treatment system based on clinical feedback as described in any one of claims 1-5, the system comprising: The credibility assessment module is used to assess the credibility of the collected clinical feedback data and obtain a credibility score for each clinical feedback data. The credibility screening module is used to screen the clinical feedback data according to a preset first credibility score threshold, and retain the clinical feedback data with a credibility score greater than the first credibility score threshold as data to be evaluated. The value assessment module is used to assess the data value of the data to be assessed based on data novelty and task relevance, and to obtain a data value score for each data to be assessed. The dynamic classification module is used to dynamically adjust the data value scoring threshold based on the current performance of the traditional Chinese medicine diagnosis and treatment system. Based on the credibility score and data value score, the data to be evaluated is screened and classified to obtain a first-level learning data set, a second-level learning data set, and a third-level learning data set. The continuous optimization module is used to perform supervised learning on the traditional Chinese medicine diagnosis and treatment system using the first-level learning dataset, reinforcement learning on the traditional Chinese medicine diagnosis and treatment system using the second-level learning dataset, and unsupervised learning on the traditional Chinese medicine diagnosis and treatment system using the third-level learning dataset, so as to achieve continuous optimization of the traditional Chinese medicine diagnosis and treatment system.

Images