Multi-modal data dynamic treatment and double-engine intelligent decision-making method for traditional Chinese medicine biased physique

Abstract

A multi-modal data dynamic treatment and double-engine intelligent decision-making method for traditional Chinese medicine biased physique belongs to the technical field of real world research, and comprises the following steps: step 1, accessing multi-source data; step 2, performing cleaning treatment on the access data; step 3, performing intelligent analysis on the data after cleaning treatment; 4, clinical suggestions are provided through double-engine clinical decisions. According to the invention, a traditional Chinese medicine biased physique intelligent platform based on dynamic data management is constructed, and multi-modal data aggregation, precise NLP processing and time sequence physique evolution modeling are realized; deployment is achieved through an EBM + RWE (evidence-based medicine + real world) dual-drive aided decision engine (CDSS) system, good universality and generalization performance of the system are guaranteed, and disease prevention of a'preventive disease 'level can be better and accurately carried out on individuals.

Description

Technical Field

[0001] The present invention belongs to the technical field of real-world research, and specifically relates to a method for dynamic governance of multi-modal data of traditional Chinese medicine biased constitutions and dual-engine intelligent decision-making. Background Art

[0002] Based on the theory of "preventive treatment of disease" in "Plain Questions", personal constitutions are classified into balanced constitution, qi-deficiency constitution, etc. through the nine-constitution classification method, which is the core basis for constitution identification. This traditional traditional Chinese medicine health management system relies on manual constitution determination, has a low level of data processing automation, and lacks dynamic prediction ability, resulting in insufficient generalization ability of the traditional system, limited application scenarios, and relatively high costs of various types.

[0003] Real World Study (RWS) is a research method for collecting and analyzing data in real clinical, community or home environments. It aims to evaluate the actual impact of a certain treatment measure or research factor on the health of patients in the real world. Different from traditional randomized controlled trials (RCTs), real world research pays more attention to the health status and treatment effects of patients in the actual diagnosis and treatment process, and is not restricted by strict inclusion criteria and intervention measures. The advantage of this research method lies in the authenticity and wide range of its data sources, which can include data from various channels such as outpatient clinics, hospitals, examinations, surgeries, pharmacies, wearable devices, and social media. By analyzing these data, researchers can obtain information on treatment effects and safety closer to the real medical environment, thus providing stronger support for clinical decision-making and the formulation of medical policies.

[0004] Both the US Food and Drug Administration (FDA) and the China Food and Drug Administration (CFDA) have promoted real world research (RWS), and the construction of traditional Chinese medicine disease-specific databases has gradually emerged. There has been a batch of various modified generative artificial intelligences based on large language models such as ChatGPT as the baseline model, through fine-tuning, prompt engineering frameworks, external Chinese medicine expert knowledge bases and disease-specific knowledge bases, and real world testing.

[0005] The "Construction and Application of Disease-Specific Databases Based on Big Data Research Platforms" mentioned in the prior art literature [Zhao Qianqian. Construction and Application of Disease-Specific Databases Based on Big Data Research Platforms [J]. China Digital Medicine, 2020, 15(12): 89-92] proposed a general architecture for disease-specific databases, but did not solve the problems of traditional Chinese medicine text NLP and dynamic constitution modeling.

[0006] The prior art document [Guo Zhuang, Yu Lili, Zhang Juan, et al. Literature Research on the Auxiliary Constitution Identification of Traditional Chinese Medicine Diagnosis Equipment [J]. World Chinese Medicine, 2024, 19(18): 2790-2794] mentioned that "Literature Research on the Auxiliary Constitution Identification of Traditional Chinese Medicine Diagnosis Equipment" used a body mass scale to evaluate static constitution and ignored the correlation of time series data.

[0007] In summary, the prior art has the following defects:

[0008] 1. Poor data integration ability: The data between hospitals and communities has strong heterogeneity. Multiple different clinical standards are mostly adopted, with low collection efficiency, diverse data types, a mixture of continuous variables and categorical variables. Currently, the available OGG / ETL (database synchronization technology) tools only support general data types and lack mapping rules for traditional Chinese medicine fields.

[0009] 2. Defects in NLP word segmentation: Unregistered words in traditional Chinese medicine (such as "liver depression transforming into fire") cannot be recognized by general word segmentation tools, resulting in a high error rate and deviation in subsequent curative effect analysis.

[0010] 3. Lack of dynamic analysis ability: A person's constitution fluctuates dynamically according to individual activities and seasonal changes. The static constitution classification method cannot guide the adjustment of personalized conditioning plans, and data capture and collection must follow individual fluctuations. Summary of the Invention

[0011] The present invention discloses a multi-modal data dynamic governance and dual-engine intelligent decision-making method for the biased constitution of traditional Chinese medicine, constructs an intelligent platform for the biased constitution of traditional Chinese medicine based on dynamic data governance, and realizes multi-modal data aggregation, precise NLP processing, and time-series constitution evolution modeling; it is deployed through an EBM+RWE (Evidence-Based Medicine+Real-World Evidence) dual-driven auxiliary decision-making engine (CDSS) system to ensure its good universality and generalization performance, and better accurately target individuals for disease prevention at the level of "preventing disease before its onset".

[0012] To achieve the above object, the technical solution of the present invention is:

[0013] The multi-modal data dynamic governance and dual-engine intelligent decision-making method for the biased constitution of traditional Chinese medicine includes the following steps:

[0014] Step 1. Access multi-source data;

[0015] Step 2. Clean and govern the accessed data;

[0016] Step 3. Conduct intelligent analysis on the data after cleaning and governance;

[0017] Step 4. Provide clinical suggestions through a dual-engine clinical decision-making.

[0018] Preferably, the step 1 includes: connecting to the hospital HIS system through OGG / CDC technology, using ETL tools to structure the community questionnaire survey data, and generating a standardized TCM disease-specific data set.

[0019] Preferably, the step 3 comprises:

[0020] Step 31: Use the open domain Chinese medicine dictionary for preliminary word segmentation;

[0021] Step 32: Automatically identify new terms by calculating the combination probability of adjacent words based on the contextual meaning of the text in which the new terms are located;

[0022] Step 33: Identify unregistered words based on the boundary entropy clustering algorithm. The formula is as follows:

[0023] Boundary entropy (s) = -Σp(left neighbor)log p (left neighbor)-Σp(right neighbor)log p (right adjacent character).

[0024] Preferably, the step 4 comprises:

[0025] Step 41: Use the Transformer model to fuse the time series data and output the physical state transition probability matrix, as shown in the following formula:

[0026] P(t i |t i-1 )=Softmax(W·Concat[H i-1 , T season ]);

[0027] Step 42: construct an evidence-based rule base based on the rules of traditional Chinese medicine classics and clinical guidelines, and make decisions based on the processed data according to the static rules of the evidence-based rule base;

[0028] Step 43: Based on the decision-making judgment, provide clinical recommendations according to the dynamic optimization of the real world engine (RWE) improved by TCM characteristic data mining.

[0029] Preferably, in step 41, the Transformer model fuses the time series data into medical records and seasonal variables.

[0030] The beneficial effects of the TCM biased constitution multimodal data dynamic management and dual-engine intelligent decision-making method of the present invention are:

[0031] 1. Data integration efficiency: Through the adaptive field mapping model developed by this technology based on the Transformer architecture, the multi-source data cleaning efficiency has been improved by 60% (the comparison literature is [Zhao Qianqian. Construction and Application of Special Disease Databases Based on Big Data Research Platforms [J]. Chinese Digital Medicine, 2020, 15(12): 89-92]).

[0032] 2. Word segmentation accuracy: The F1 value of unregistered word recognition in Traditional Chinese Medicine NLP reaches 93.2% (the traditional tool <75%).

[0033] 3. Dynamic prediction ability: The prediction error of physical constitution state migration is reduced to 8.7%, which can support the adjustment of quarterly intervention plans.

[0034] 4. Clinical adaptability: The compliance rate of the recommended solutions of the auxiliary decision-making system with expert consensus exceeds 90% (the single EBM model is only 72%). Brief Description of the Drawings

[0035] Figure 1 This is the technical roadmap of the implementation plan of the present invention. Detailed Implementation Modes

[0036] The following is only the preferred embodiment of the present invention and is not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0037] The following embodiments can be understood as separately expressing a part of the local structure or method of the present invention, or can also be understood as the combination of embodiments explaining the connotation of a larger range of the structure or method of the present invention.

[0038] Example 1

[0039] The dynamic governance method of multi-modal data of Traditional Chinese Medicine's biased physical constitution and the dual-engine intelligent decision-making method include the following steps:

[0040] Step 1: Access multi-source data;

[0041] Step 2: Clean and manage the accessed data;

[0042] Step 3: Conduct intelligent analysis on the data after cleaning and management;

[0043] Step 4: Provide clinical suggestions through dual-engine clinical decision-making.

[0044] Example 2

[0045] Based on Example 1, this example discloses:

[0046] The step 1 includes: connecting the hospital HIS system through OGG / CDC technology, and processing the community questionnaire survey data with ETL tools to generate a standardized TCM disease dataset. By standardizing the coding, a data dictionary module is established. Clinical data and community data from multiple hospitals are included to establish a high-quality dataset so that it can be used for subsequent model training after cleaning.

[0047] Example 3

[0048] Based on Example 1, this embodiment discloses:

[0049] The step 3 comprises:

[0050] Step 31: Use an open domain TCM dictionary (such as TCM Clinical Terminology) to perform preliminary word segmentation; for example:

[0051] Original sentence: "The patient's pulse is stringy and thin, considering liver depression and spleen deficiency" → word segmentation result: patient / pulse / stringy and thin / consider / liver depression and spleen deficiency;

[0052] Step 32: Automatically identify new terms based on the context of the text in which the new terms are located by calculating the combination probability of adjacent words; for example:

[0053] When "肝郁脾虚" appears for the first time, "肝郁" and "脾虚" are displayed separately → after analyzing the context, it is determined to be an integral term, that is, the literal meaning of the context includes both liver depression and body deficiency. Therefore, the probability of "肝郁" and "脾虚" forming a new term together is high. Then, by analyzing the relationship between the words adjacent to the front of "肝郁" and "肝郁" and the meaning of the context, and combining the relationship between the words adjacent to the back of "脾虚" and "脾虚" and the meaning of the context, the probability of "肝郁" and the words adjacent to the front forming a new term and the probability of "脾虚" and the words adjacent to the back forming a new term are analyzed. The probability of "liver depression and spleen deficiency" was finally determined as a new term; the above "liver depression" can be replaced by word A, and "spleen deficiency" can be replaced by word B; for example, the word adjacent to "liver depression" is "perennial", combined with the context "chest and flank pain, depression, loss of appetite, intestinal rumbling and diarrhea, fatigue", it clearly expresses the connotation of "liver depression" and "spleen deficiency", but does not emphasize the relationship between "perennial" and "liver depression", so it is determined that the probability of "liver depression" and "spleen deficiency" combining is greater than the probability of "perennial" and "liver depression" combining into a new term, and the probability analysis of "spleen deficiency" combining with the adjacent words at the back is the same as above.

[0054] Step 33: Identify unregistered words (such as "Yin deficiency and Yang hyperactivity") based on the boundary entropy clustering algorithm, the formula is as follows:

[0055] Boundary entropy (s) = -Σp(left neighbor)log p (left neighbor)-Σp(right neighbor)log p (right adjacent character).

[0056] Example 4

[0057] Based on Example 1, this example discloses:

[0058] Step 4 described above includes:

[0059] Step 41: Use the Transformer model to fuse time-series data (medical records, seasonal variables) and output the physical state transition probability matrix, as shown in the following formula:

[0060] P(t i |t i-1 ) = Softmax(W·Concat[H i-1 , T season )

[0061] For example:

[0062] Example input data: Physical fitness scores, medication records (such as "Liujunzi Decoction"), and seasonal factors (spring / summer / autumn / winter encoded as 1-4) of a patient within 3 years;

[0063] Example output: Current physical state: Phlegm-dampness constitution (score 80) → Predicted to turn into a balanced constitution in autumn (score 55);

[0064] Step 42: Construct an evidence-based rule base (EBM) based on the rule content of traditional Chinese medicine classics and clinical guidelines, and make decision judgments based on the static rules of the evidence-based rule base according to the processed data; for example, the following Python pseudocode:

[0065] IF physical constitution

[0066] # Medication rules for qi-deficiency constitution

[0067] if physical constitution == "qi-deficiency constitution" and symptoms contain ("fatigue", "spontaneous sweating"):

[0068] recommended prescriptions.append("Sijunzi Decoction")

[0069] if concurrent symptoms == "palpitation":

[0070] recommended prescriptions.append("Guipi Decoction")

[0071] Step 43: Based on the decision judgment, provide clinical suggestions through the dynamic optimization of the real-world engine (RWE) improved by data mining with traditional Chinese medicine characteristics. By analyzing real-world data (Real-World Data, RWD) such as historical electronic medical records, medication records, and follow-up data, dynamically optimize the original rule base.

[0072] Existing modern medicine RWE systems (such as IBM Watson) rely on SNOMED CT standard terminology, while traditional Chinese medicine needs to use its own "Traditional Chinese Medicine Clinical Terminology Set". Design new composite evaluation indicators, such as "constitution deviation degree + symptom remission period".

[0073] The present invention conducts cluster analysis on indicators such as "constitution score change rate" and "symptom remission period" of patients, discovers potential curative effect rules; mines the association rules between prescriptions and curative effects based on the Apriori algorithm, and adds them to the recommendation system after expert review. And new cases are updated to the model in real time through the online learning update route. Application examples are as follows:

[0074] Analysis of historical cases reveals that the combination of "Sijunzi Decoction + Astragalus membranaceus" increases the effective rate for qi-deficiency constitution by 15% (p < 0.05), and the recommended plan is automatically updated.

[0075] Example of association rule: {constitution = phlegm-dampness constitution, medication = Erchen Decoction, season = summer} → {significant curative effect ↑} (support = 0.35, confidence = 0.82).

[0076] The working principle of the present invention:

[0077] 1. Multi-source heterogeneous data intelligent aggregation method: Dynamically generate a Chinese medicine field mapping rule library to support automatic conversion of JSON / XML / relational databases.

[0078] 2. NLP word segmentation algorithm based on EM and entropy optimization: Solve the problem of out-of-vocabulary word recognition in the description of traditional Chinese medicine curative effects.

[0079] 3. Temporal constitution dynamic scoring model: A Transformer architecture that integrates seasonal variables and treatment intervention events.

[0080] 4. EBM + RWE dual-driven decision engine: An evidence-based rule and dynamic curative effect network collaborative reasoning method.

[0081] As shown in the following table, the advantages of the present invention compared with the existing methods are given:

[0082]

[0083] It should be noted that the method given in the present invention can also be replaced with the following forms, but these replacement forms are all inferior to the above-mentioned scheme disclosed in the present invention. Specifically as follows:

[0084] Replacement of data synchronization technology: The OGG technology can be replaced with a flink stream message queue to achieve real-time data stream collection, but it will greatly increase the hardware cost.

[0085] NLP Word Segmentation Alternative: Use a pre-trained traditional Chinese BERT model for entity recognition (the fine-tuning strategy of the model needs to be adjusted. The performance of the BERT model is usually positively correlated with the amount of high-quality training data, with high time, manpower, and computing costs and requirements for equipment).

[0086] Model Architecture Alternative: Replace Transformer with LSTM + Attention (Long Short-Term Memory + Attention Mechanism), but a seasonal feature time-domain encoding module needs to be added. It is not as good as directly using Transformer in the present invention. The architecture of Transformer contains a multi-head self-attention mechanism and can be applied only through fine-tuning training without adding new modules.

Claims

1. Dynamic management of multimodal data of TCM biased constitution and dual-engine intelligent decision-making method, which is characterized by: It includes the following steps: Step 1: Multi-source data access; Step 2: Clean and manage the accessed data; Step 3: Intelligently analyze the data after cleaning and management; Step 4: Provide clinical suggestions through a dual-engine clinical decision-making system.

2. The dynamic governance method for multi-modal data of traditional Chinese medicine's biased constitutions and the dual-engine intelligent decision-making method according to claim 1, characterized in that, The said Step 1 includes: Connecting to the hospital HIS system through OGG / CDC technology, and using an ETL tool to structurally process the community questionnaire data to generate a standardized traditional Chinese medicine (TCM) disease-specific dataset.

3. The dynamic governance method for multi-modal data of traditional Chinese medicine's biased constitutions and the dual-engine intelligent decision-making method according to claim 1, characterized in that, The said Step 3 includes: Step 31: Perform preliminary word segmentation using an open-domain TCM dictionary; Step 32: Automatically identify new terms based on the context of the text where the new terms are located by calculating the combination probability of adjacent characters; Step 33: Identify out-of-vocabulary words based on the boundary entropy clustering algorithm, and the formula is as follows: Boundary entropy (s) = -Σp(left neighboring word) log p (left neighboring word) - Σ p(right neighboring word) log p (right neighboring word).

4. The dynamic governance method for multi-modal data of traditional Chinese medicine's biased constitutions and the dual-engine intelligent decision-making method according to any one of claims 1-3, characterized in that, The said Step 4 includes: Step 41: Use a Transformer model to fuse time-series data and output a physical state transition probability matrix, as shown in the following formula: P(t i |t i-1 ) = Softmax(W·Concat[H i-1 , T season ); Step 42: Construct an evidence-based rule library based on the rule content of TCM classics and clinical guidelines, and make decision judgments based on the static rules of the evidence-based rule library according to the processed data; Step 43: On the basis of the decision judgment, provide clinical suggestions based on the dynamic optimization of the real-world engine (RWE) improved by TCM characteristic data mining.

5. The dynamic governance method for multi-modal data of traditional Chinese medicine's biased constitutions and the dual-engine intelligent decision-making method according to claim 4, characterized in that, In the said Step 41, the time-series data fused by the Transformer model are medical records and seasonal variables.

Images