Traditional Chinese medicine clinical decision support system for primary osteoporosis

Abstract

The invention discloses a primary osteoporosis traditional Chinese medicine clinical decision support system, and relates to the field of traditional Chinese medicine artificial intelligence, and the system comprises a data processing module which is used for carrying out the term standardization processing and syndrome element extraction of original traditional Chinese medicine literature data, and constructing a structured traditional Chinese medicine syndrome and treatment database; the dialectical reasoning module is constructed based on a machine learning model and is used for outputting a traditional Chinese medicine syndrome element diagnosis result according to the input four diagnosis information; and the prescription recommendation module is constructed based on a knowledge graph and is used for recommending corresponding treatment methods and rules and traditional Chinese medicine prescriptions according to the traditional Chinese medicine syndrome element diagnosis results and / or the specific clinical symptoms. According to the scheme, objectivity and consistency of syndrome element differentiation are remarkably enhanced, the defect that a traditional method is insufficient in composite syndrome type processing capacity is overcome, and the self-adaptive capacity of prescription recommendation along with dynamic evolution of clinical practice is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of traditional Chinese medicine artificial intelligence, and in particular to a traditional Chinese medicine clinical decision support system for primary osteoporosis. BACKGROUND

[0002] Primary osteoporosis is a degenerative disease of the skeletal system associated with aging, and its prevention and treatment is a global public health challenge. Traditional Chinese medicine has accumulated rich clinical experience in the prevention and treatment of primary osteoporosis, and has formed an individualized diagnosis and treatment system centered on syndrome differentiation. This system relies on the collection and analysis of four diagnostic information by physicians to determine the TCM syndrome type and formulate prescriptions accordingly. However, the process of TCM syndrome differentiation is highly nonlinear and subjective, and its inheritance and promotion are heavily dependent on the personal experience of physicians, which limits the accessibility and consistency of high-quality diagnosis and treatment resources. In recent years, although some studies have attempted to apply computer technology to TCM auxiliary diagnosis, most systems only implement simple knowledge base queries or logical judgments based on single rules, failing to simulate the complex TCM differentiation thinking and flexible prescription decision-making process. Therefore, how to build a clinical decision support system that effectively integrates TCM theoretical knowledge, massive clinical data, and modern artificial intelligence technology, thus achieving precise, interpretable, and highly compliant clinical decision support, has become a key technical problem in the field of TCM informatization and intelligentization.

[0003] The existing TCM auxiliary decision-making system mainly has the following deficiencies: First, in the syndrome differentiation model, most systems are based on fixed, pre-defined syndrome type rules for judgment, lack the ability to effectively analyze and identify complex syndrome types, and cannot handle the common syndrome element combination problem in clinical practice. The model has poor flexibility and adaptability, and is difficult to continuously learn and optimize from new data. Second, in the prescription recommendation link, traditional systems mostly use simple frequency statistics or fixed formula matching, and fail to build deep and networked associations between drugs, syndrome elements, and symptoms, resulting in single and lack of individualized adjustment dimensions in the recommended results, and unable to simulate the clinical thinking of famous doctors "adding and subtracting according to the syndrome". Finally, from the system architecture level, the syndrome differentiation and prescription modules are often isolated from each other, and cannot form an end-to-end intelligent pipeline from "four diagnostic information" to "syndrome element diagnosis" to "prescription generation". The entire decision-making process is like a "black box", lacking the necessary explainability, making it difficult for clinicians to understand and trust the recommended results of the system, thus severely restricting its application and promotion in real clinical scenarios. Therefore, based on the above problems, the present application proposes a traditional Chinese medicine clinical decision support system for primary osteoporosis. SUMMARY

[0004] To address the aforementioned problems, the present invention aims to provide a TCM clinical decision support system for primary osteoporosis. This system addresses the issues of inconsistent diagnostic standards, highly subjective prescription decisions, and difficulty in replicating and promoting high-quality medical resources in traditional TCM diagnosis and treatment due to its heavy reliance on physicians' personal experience. The goal is to standardize, intelligentize, and personalize the TCM diagnosis and treatment process for primary osteoporosis, thereby improving the accuracy, consistency, and accessibility of clinical decisions.

[0005] To achieve the above objectives, this invention provides a TCM clinical decision support system for primary osteoporosis. The system first constructs a structured TCM syndrome and treatment database through terminology standardization and syndrome element extraction. Then, it uses an integrated machine learning model that integrates multi-dimensional feature screening to achieve objective identification of TCM syndrome elements. Finally, it generates personalized prescriptions based on a knowledge graph with a dynamic weight evolution mechanism. The system's performance is then verified and continuously optimized through a human-machine collaborative evaluation module, ultimately forming a complete TCM clinical intelligent decision support closed loop.

[0006] In a first aspect, the present invention provides a TCM clinical decision support system for primary osteoporosis, comprising: The data processing module is used to standardize the terminology and extract the syndrome elements from the original TCM literature data in order to solve the problem of ambiguity in TCM terminology and to build a structured TCM syndrome and treatment database suitable for computer processing. The diagnostic reasoning module, built upon the integration of multiple machine learning models, is used to output TCM syndrome element diagnostic results based on the input information from the four diagnostic methods, thereby achieving accurate analysis and judgment of complex syndromes and improving the objectivity and consistency of syndrome differentiation. The prescription recommendation module is built on a knowledge graph with weighted relationships. It is used to make personalized recommendations of treatment principles and Chinese medicine prescriptions based on the diagnosis results of the TCM syndrome elements and / or specific clinical symptoms, thereby realizing flexible combination and priority ranking of prescription drugs. The dialectical reasoning module and the prescription recommendation module are sequentially connected to form an end-to-end clinical decision support architecture, completing an automated and intelligent closed loop from symptom input to prescription recommendation.

[0007] Furthermore, the data processing module maps non-standard TCM diagnostic information, syndrome types, and Chinese herbal medicine names to a standardized terminology system based on preset TCM terminology standards. It also decomposes complex TCM syndrome types into combinations of single disease location syndrome elements and disease nature syndrome elements based on the syndrome element differentiation theory. This effectively reduces the complexity of syndrome types and provides machine learning models with more granular and easier-to-learn classification targets.

[0008] Furthermore, the diagnostic reasoning module uses a sparse penalized regression model to select a subset of features with significant predictive ability for the target syndrome from all the four diagnostic methods. It also integrates multiple classification algorithms to construct an independent binary classification prediction model for each TCM syndrome and adaptively selects the optimal prediction model for each syndrome based on a preset comprehensive performance evaluation index. This approach overcomes the limitations of a single algorithm and ensures that each syndrome can obtain the most robust prediction effect.

[0009] Furthermore, the classification algorithm includes at least the following three categories: Improve decision tree algorithms, neural network algorithms, support vector machine algorithms, and random forest algorithms.

[0010] Furthermore, the dialectical reasoning module further performs feature screening based on a multi-dimensional feature contribution scoring function. This function integrates the statistical significance of the feature, its redundancy with other features, and its clinical prior knowledge weight in traditional Chinese medicine theory, in order to screen out a subset of features with high discriminative power for the target syndrome.

[0011] Furthermore, the prescription recommendation module constructs a graph data structure with syndrome elements, treatment principles, Chinese medicines, and clinical symptoms as entities, and the treatment and correlation relationships between them as edges. Based on the historical usage frequency of Chinese medicines under the corresponding syndrome elements, quantitative weights are assigned to the syndrome element-Chinese medicine relationship, and the recommended drugs are sorted and screened according to the weights. This makes the recommendation results both conform to the statistical laws of traditional Chinese medicine theory and have adjustable flexibility.

[0012] Furthermore, the weights are dynamic weights, and the system also includes a weight evolution module, which periodically updates the dynamic weights based on the historical usage frequency, time decay factor, and clinical feedback signals of entity relationships in the knowledge graph through a weight evolution function.

[0013] Furthermore, the prescription recommendation module is further configured as follows: Supports independent and combined searches of three core relationships: syndrome element-treatment principle, syndrome element-traditional Chinese medicine, and symptom-traditional Chinese medicine; It supports dynamically adjusting the number and range of recommended drugs based on user-defined weight thresholds.

[0014] Furthermore, the system also includes a human-machine collaborative evaluation module, which simulates the Turing test environment, generates virtual medical records, and allows the system and human physicians to independently diagnose and prescribe the patients. The similarity and acceptability of the prescriptions are compared through double-blind evaluation to assess the clinical intelligence level of the system.

[0015] Furthermore, the system also includes a multimodal input interface for receiving and structuring diagnostic information input from text, voice, or structured forms.

[0016] Furthermore, the output of the dialectical reasoning module includes an interpretability output unit, which is used to show the user the key diagnostic information that leads to the judgment of a specific syndrome element and its contribution.

[0017] In a second aspect, the present invention also provides a non-transitory computer-readable medium storing computer-executable instructions, which, when executed by a processor, control the operation of the system described in the first aspect; the medium includes, but is not limited to, solid-state memory, optical storage devices, or cloud storage systems, ensuring that the system can be stably implemented on a variety of hardware platforms.

[0018] This invention provides a TCM clinical decision support system for primary osteoporosis. First, the system transforms heterogeneous TCM literature and clinical data into a standardized knowledge base that can be processed by a computer through a structured data construction method based on the syndrome differentiation theory. Then, it employs a multimodal feature contribution weighted machine learning model to achieve accurate and interpretable automated syndrome differentiation from the four diagnostic methods to TCM syndrome elements. Finally, based on a knowledge graph with a dynamic weight evolution mechanism, it generates personalized prescriptions according to the syndrome differentiation results and clinical symptoms. The overall performance of the system is then validated and continuously optimized through a human-machine collaborative evaluation system.

[0019] This system not only significantly enhances the objectivity and consistency of syndrome differentiation, overcoming the shortcomings of traditional methods in handling complex syndromes, but also achieves adaptive capabilities in prescription recommendations that dynamically evolve with clinical practice, effectively simulating the diagnostic thinking of renowned physicians who "add or subtract ingredients according to the syndrome." Ultimately, by providing a transparent and interpretable reasoning process and rigorously evaluated reliable outputs, the system greatly enhances clinicians' trust and willingness to use it, promoting the transformation of TCM diagnostic and treatment experience from reliance on personal inheritance to a standardized service model that is calculable, evolvable, and scalable.

[0020] Beneficial effects By implementing the TCM clinical decision support system for primary osteoporosis provided by the present invention, the following technical effects are achieved: (1) By systematically decomposing complex TCM syndromes into combinations of single disease locations and disease-related elements, and establishing a standardized terminology system, this method provides a data foundation with appropriate granularity and clear semantics for subsequent computational analysis. Compared with methods that directly use original syndrome types or unstructured terminology, this method effectively solves the inherent heterogeneity and nonlinearity problems of TCM data, provides high-quality, computable learning objectives for machine learning models, and is a key prerequisite for the intelligentization of the entire system.

[0021] (2) By designing an evaluation process that simulates a double-blind testing environment, the output of the artificial intelligence system is objectively compared with the decisions of human experts under the same conditions. This provides a more convincing verification method that goes beyond conventional performance indicators for evaluating the clinical intelligence level of TCM AI systems. It can comprehensively evaluate the rationality of the system's syndrome differentiation and treatment logic and the acceptability of prescriptions in real clinical scenarios, greatly enhancing the credibility and clinical translation value of the system's results.

[0022] (3) By introducing a contribution scoring function that integrates statistical significance, feature redundancy, and prior knowledge of TCM clinical practice, the machine learning diagnostic process is fundamentally optimized. Compared with traditional single feature selection methods, this model can more accurately screen out the four diagnostic methods that have high discriminative power for specific syndrome elements, effectively reduce collinearity interference between features, thereby significantly improving the prediction accuracy and generalization ability of the syndrome element diagnostic model, and enhancing the transparency and interpretability of the diagnostic decision-making process.

[0023] (4) By constructing a weighted evolution function that integrates time decay and clinical feedback signals, the self-updating and optimization of knowledge graph relationships are realized. Compared with prescription recommendation methods based on static historical frequency, this mechanism enables the system to adaptively track the evolution of clinical medication trends and respond to actual user feedback, thereby completely changing the rigid nature of traditional knowledge graphs and significantly improving the timeliness, personalization, and clinical compliance of prescription recommendations. Attached Figure Description

[0024] To make the above-described TCM clinical decision support system for primary osteoporosis of the present invention more apparent and understandable, the accompanying drawings used in the specific embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram illustrating the basic framework of the TCM clinical decision support system proposed in this application. Figure 2 This is a flowchart illustrating the literature retrieval and screening process for this application. Figure 3 A schematic diagram illustrating the characteristic variables of the "kidney" syndrome element in the pathogenesis. Figure 4 A schematic diagram representing the characteristic variables of the "liver" syndrome element in the pathogenesis analysis; Figure 5 A schematic diagram representing the characteristic variables of the "spleen" syndrome element in the location of the disease; Figure 6 A schematic diagram representing the characteristic variables of the disease location syndrome element "essence deficiency"; Figure 7A schematic diagram representing the characteristic variables of "Yang deficiency" in the pathogenesis; Figure 8 A schematic diagram representing the characteristic variables of "Yin deficiency" in the pathogenesis; Figure 9 A schematic diagram representing the characteristic variables of "Qi deficiency" in the pathogenesis; Figure 10 A schematic diagram representing the characteristic variables of the disease location syndrome element "blood deficiency"; Figure 11 A schematic diagram representing the characteristic variables of the pathogenesis factor "blood stasis"; Figure 12 A schematic diagram representing the characteristic variables of "qi stagnation" in the pathogenesis; Figure 13 A schematic diagram representing the characteristic variables of the "wind" syndrome element in the pathogenesis. Figure 14 A schematic diagram representing the characteristic variables of the "cold" syndrome element in the pathogenesis; Figure 15 A schematic diagram representing the characteristic variables of the pathogenesis factor "phlegm and dampness"; Figure 16 A schematic diagram of the interface for searching the relationship between syndrome elements and treatment principles; Figure 17 This is a schematic diagram of the many-to-many search interface for the relationship between syndrome elements and traditional Chinese medicine. Figure 18 This is a schematic diagram of the symptom-traditional Chinese medicine relationship search interface. Detailed Implementation

[0026] Example 1: This embodiment relates to a method for constructing a TCM syndrome differentiation and treatment database for a TCM clinical decision support system for primary osteoporosis. The basic framework of the system is as follows: Figure 1 As shown. The construction of this database began with the systematic retrieval and organization of multi-source literature. The literature sources cover Chinese and English databases such as CNKI, Wanfang Data Knowledge Service Platform, VIP Chinese Journal Service Platform, Biomedical Literature Database, and Ancient and Modern Medical Case Cloud Platform. The literature retrieval and screening process is as follows: Figure 2As shown. To comprehensively cover relevant data on oral administration of traditional Chinese medicine for the treatment of primary osteoporosis, a combined search strategy was adopted. Chinese search terms included "primary osteoporosis," "perimenopausal osteoporosis," "postmenopausal osteoporosis," "senile osteoporosis," "idiopathic osteoporosis," "osteoporosis," "osteoporosis," "osteoporosis pain," "osteoporosis," "traditional Chinese medicine," "Chinese herbal medicine," "medical records," and "famous doctor's experience," while English search terms included "Primary Osteoporosis," "Postmenopausal Osteoporosis," "Senile Osteoporosis," "Idiopathic Osteoporosis," "Osteoporosis," "Traditional Chinese Medicine," "Chinese Herbal Medicine," "Osteoporotic Pain," "Osteopenia," and "Bone loss," etc. Corresponding search expressions were constructed according to the rules of each database. In addition, modern books were manually consulted to supplement the literature and ensure the comprehensiveness and representativeness of the data.

[0027] The literature screening process strictly followed pre-defined inclusion and exclusion criteria. Inclusion criteria included that the literature source be formal journal articles, covering clinical research, renowned physician experience, medical records, or expert consensus and guidelines; the research subjects be patients with primary osteoporosis or related osteopenia; the treatment method be oral administration of traditional Chinese medicine compound prescriptions; and the literature must provide detailed information on at least one of the following: TCM syndrome differentiation, four diagnostic methods, treatment principles, or prescription information. Exclusion criteria included duplicate publications, multilingual publications, basic research, reviews, secondary research, statistical data, and purely theoretical research; low-quality literature containing obvious errors, falsified data, or serious discrepancies between syndrome differentiation and prescription; and literature involving patients with other diseases or fragility fractures. After a step-by-step screening, 3163 articles were ultimately included from 29789 articles retrieved from online databases and supplemented by 167 books for subsequent data extraction.

[0028] Data extraction was completed independently by two people using two machines, with cross-verification ensuring accuracy. The extracted content mainly included information from the four diagnostic methods (inspection, auscultation and olfaction), TCM syndrome types, TCM treatment principles, and prescription medications. Based on this, TCM terminology was standardized. The standardization of the four diagnostic methods, syndrome types, and treatment principles primarily referenced authoritative works such as *Internal Medicine of Traditional Chinese Medicine*, *Gynecology of Traditional Chinese Medicine*, *Diagnostics of Traditional Chinese Medicine*, *Standardized Terminology for Common Symptoms in TCM Clinical Practice*, and *Differential Diagnosis of Symptoms in Traditional Chinese Medicine*, combined with the characteristics of primary osteoporosis and clinical expert advice, to determine representative terminology. For example, "lower back pain / lumbar pain / lower leg pain" was standardized as "lower back and lower limb pain," and "lethargy / dejection / exhaustion / fatigue" was standardized as "fatigue and weakness." For TCM syndrome differentiation, to address the common problem of complex syndromes in clinical practice, the syndrome element differentiation theory is used to break down complex syndromes into single syndrome elements. For example, "Qi and Blood Deficiency Syndrome" is broken down into "Qi Deficiency Syndrome + Blood Deficiency Syndrome," and "Kidney Deficiency and Liver Qi Stagnation Syndrome" is broken down into "Kidney Essence Deficiency Syndrome + Liver Qi Stagnation Syndrome." For syndromes that are difficult to clearly break down, such as "Heart and Spleen Deficiency Syndrome" and "Heart and Kidney Disharmony Syndrome," the original terminology is retained. The standardization of TCM drug names mainly refers to "Traditional Chinese Materia Medica." Except for drugs whose medicinal properties and functions are significantly different before and after processing, only the names of the raw medicinal slices are recorded without distinguishing the processing methods. To further reduce data dispersion, "Deer Antler," "Deer Antler Glue," "Deer Antler Powder," and "Deer Velvet" are classified under the "Deer Antler" category, and "Leopard Bone," "Dog Bone," "Monkey Bone," "Tiger Bone," "Chicken Bone," "Deer Bone," "Red Deer Bone," "Ox Bone," "Sheep Bone," "Wild Boar Bone," and "Cuttlefish Bone" are classified under the "Animal Bone" category. Through standardized processing, 262 items of information from the four diagnostic methods, 23 basic TCM syndrome types, and 352 kinds of Chinese herbal medicine decoction pieces were finally summarized.

[0029] Based on the standardization of syndrome types, further extraction of TCM syndrome elements was conducted. Syndrome elements are the basic components of a syndrome type, including the location and nature of the disease. This method, based on "Syndrome Element Differentiation" and combined with the actual situation of the included literature, formulated a TCM syndrome element extraction scheme for primary osteoporosis. For example, "Kidney Deficiency Syndrome" was extracted as the location syndrome element "Kidney" and the nature syndrome element "Essence Deficiency," while "Liver Qi Stagnation Syndrome" was extracted as the location syndrome element "Liver" and the nature syndrome element "Qi Stagnation." The syndrome element extraction process was completed independently by two people and involved multiple rounds of interactive discussions. When disagreements arose, they were arbitrated by experts with senior professional titles in the fields of TCM internal medicine and TCM diagnostics. Finally, seven location syndrome elements (Kidney, Liver, Spleen, Stomach, Heart, Lung, and Gallbladder) and twelve nature syndrome elements (Essence Deficiency, Qi Deficiency, Yin Deficiency, Yang Deficiency, Blood Deficiency, Blood Stasis, Qi Stagnation, Qi Reversal, Wind, Cold, Phlegm-Dampness, and Heat) were extracted. Among them, "Phlegm" and "Dampness" were combined into the "Phlegm-Dampness" syndrome element due to the limited number of literatures and their close relationship.

[0030] After completing terminology standardization and syndrome element extraction, Microsoft Excel 2023 software was used for data entry and management to establish a TCM syndrome differentiation and treatment database for primary osteoporosis. This database consists of two parts: a TCM syndrome differentiation information database and a TCM prescription information database. The TCM syndrome differentiation information database contains information on the four diagnostic methods (inspection, auscultation and olfaction, inquiry, and palpation), TCM syndrome types, and TCM syndrome elements. The TCM prescription information database contains TCM syndrome elements, treatment principles, and TCM prescription information. The data is stored in text format for easy subsequent analysis and retrieval.

[0031] Data cleaning is a crucial step in database construction, improving data quality by identifying and correcting errors, inconsistencies, and missing values. In this embodiment, data cleaning includes data encoding, duplicate value handling, missing value handling, and outlier handling. During the data encoding stage, the four diagnostic methods, syndrome elements, and prescription information are binarized, with "present" set to "1" and "absent" set to "0," transforming the data into a computer-readable format. Duplicate value handling involves comparing and deleting duplicate records row by row using Excel software. Missing value handling, based on the characteristics of text data, considers information not recorded as equivalent to "absent," and therefore does not require special processing. Outlier handling fully respects the personalized characteristics of TCM diagnosis and treatment, deleting only outliers that are clearly inconsistent with TCM theory.

[0032] Through the above steps, a standardized and structured TCM syndrome differentiation and treatment database for primary osteoporosis was constructed. This database not only encompasses rich clinical data but also achieves deep data integration through terminology standardization and syndrome element extraction, providing a high-quality data foundation for subsequent TCM syndrome differentiation decision support and prescription recommendation modules. The database construction method demonstrates creativity in the standardization of TCM data processing, the decomposition of complex syndrome types, and syndrome element extraction, laying a solid foundation for the development of TCM clinical decision support systems.

[0033] Example 2: This embodiment relates to a method for constructing a TCM syndrome differentiation decision support system for a TCM clinical decision support system for primary osteoporosis. Based on the aforementioned constructed TCM syndrome and treatment database, the core of this system lies in using various machine learning algorithms to construct a computer program capable of automatically determining the TCM syndrome elements based on the patient's four diagnostic methods, thereby achieving intelligent and standardized TCM syndrome differentiation.

[0034] The first step in system construction is feature selection. The input features are 262 standardized TCM diagnostic methods (four diagnostic methods), while the output consists of 13 extracted major TCM syndrome elements. Considering that some syndrome elements often co-occur in clinical practice, potentially leading to feature confusion, this system employs an innovative feature selection strategy combining algorithms and expert knowledge. First, the data is randomly divided into training and testing sets in an 8:2 ratio, and preliminary feature selection is performed based on the minimum absolute contraction and selection operator. The minimum absolute contraction and selection operator uses L1 regularization to compress the coefficients of unimportant features to zero, thereby selecting feature variables that significantly contribute to the prediction of each syndrome element. The results show that 166 of the 262 features have non-zero coefficients. Building on this, a clinical expert intervention mechanism is further introduced. Experts in TCM internal medicine and TCM diagnostics manually review the features selected by the minimum absolute contraction and selection operator based on the basic principles of TCM and actual clinical situations, eliminating variables that clearly do not conform to TCM theory. For example, for syndromes that may coexist, such as "kidney" and "blood stasis," "spleen" and "kidney," or "yin deficiency" and "yang deficiency," experts carefully evaluate and adjust the associated features based on their inherent pathogenesis. Through this dual screening process that integrates data-driven and knowledge-driven approaches, 157 key diagnostic information features for predicting 13 TCM syndromes were ultimately identified. Based on the characteristics of TCM diagnostic information collection, these features were systematically categorized into inquiry, inspection, and palpation. Figures 3-15 The top 15 characteristic variables of importance for each TCM syndrome element are shown.

[0035] After feature selection, the system proceeds to the construction and optimization phase of the machine learning model. This system employs 12 mainstream machine learning algorithms, building independent prediction models for each TCM syndrome element based on the selected feature subset. These algorithms include adaptive boosting, artificial neural networks, Naive Bayes, categorical feature boosting, decision trees, resilient regression networks, K-nearest neighbors, gradient boosting, gradient boosting decision trees, lightweight gradient boosting machines, support vector machines, and random forests. The development environment is based on PyTorch 2.0, using PyCharm 2024.2.2 as the development platform, and primarily employing the Scikit-Learn 1.3.1 framework. To improve the model's predictive performance, hyperparameters were tuned to suit the characteristics of different algorithms, and the F1 score was used as the core metric to retain the optimal hyperparameter combination for the final model construction.

[0036] Evaluating the predictive efficacy of models is crucial to ensuring their clinical application value. The system employs multiple indicators, including accuracy, precision, recall, F1 score, and area under the receiver operating characteristic (AUC) curve, to comprehensively evaluate all models corresponding to each syndrome element and select the model with the best predictive efficacy. Evaluation results show that different syndrome elements have their most suitable machine learning models. For example, for the syndrome element "kidney," the CatBoost model performed best, with an accuracy of 0.9247, an F1 score of 0.9529, and an AUROC as high as 0.9616; for the syndrome element "liver," the XGBoost model was optimal, with an F1 score of 0.7379 and an AUROC of 0.9160; for the syndrome element "essence deficiency," the LGBM model was optimal, with an F1 score of 0.9362; and for the syndrome element "yang deficiency," the CatBoost model was optimal, with an F1 score of 0.8133. This result confirms the necessity and superiority of selecting differentiated modeling strategies for different diagnostic objectives.

[0037] Finally, by writing Python scripts, the optimal models corresponding to the 13 TCM syndrome elements obtained above were integrated to construct a unified computer program capable of simultaneously predicting multiple major TCM syndrome elements of primary osteoporosis. This program can receive structured patient diagnostic information from the four diagnostic methods (inspection, auscultation, palpation, and olfaction) as input and output a binary judgment result regarding the presence or absence of each TCM syndrome element. Clinicians can then flexibly combine the syndrome element information to determine the complete TCM syndrome type. For example, when a patient's diagnostic information includes "lower back and limb pain, fixed location of pain, fluctuating pain intensity, aggravated by exertion, weakness and fatigue in muscles and bones, difficulty in lifting heavy objects, soreness and weakness in the lower back and knees, fatigue and weakness, numbness in the limbs, pale or lackluster complexion, poor sleep or insomnia, excessive dreaming, forgetfulness and confusion, tinnitus, hot flashes and sweating, dry mouth and throat with a desire to drink, red tongue, little coating, thready pulse, rapid pulse, weak pulse, deep pulse," the program can automatically determine the location of the disease, "kidney," and the nature of the disease, "essence deficiency" and "yin deficiency," as "yes," while the remaining factors are "no." Based on this, physicians can easily combine and determine the patient's TCM syndrome type as "kidney yin deficiency syndrome." This system provides a visual Python application interface that intuitively displays the factors with positive predictions, greatly assisting clinicians in making rapid and standardized diagnoses and effectively reducing subjective bias caused by differences in experience.

[0038] In summary, the TCM syndrome differentiation decision support system constructed in this embodiment successfully achieves automated and high-precision identification of TCM syndrome elements in primary osteoporosis by integrating minimum absolute contraction and selection operators with clinical expert knowledge for dual feature screening, comparison and optimization of multiple machine learning algorithms, and optimal model integration based on performance indicators. This method not only improves the objectivity and repeatability of TCM syndrome differentiation but also provides accurate syndrome element diagnostic basis for subsequent intelligent prescription recommendations, representing a significant breakthrough in the field of intelligent TCM diagnosis and treatment.

[0039] Example 3: This embodiment relates to a method for constructing a traditional Chinese medicine (TCM) prescription recommendation system for a TCM clinical decision support system for primary osteoporosis. The core of this system lies in utilizing knowledge graph technology to construct a system capable of intelligently recommending treatment principles and prescriptions based on the patient's TCM syndrome elements and specific clinical symptoms, thereby simulating the prescribing thinking of TCM experts and achieving precise medication recommendations.

[0040] The construction of this system began with in-depth mining and structured processing of three core relationships in the aforementioned TCM syndrome and treatment database: syndrome element-treatment principle, syndrome element-Chinese herbal medicine, and symptom-Chinese herbal medicine. The development environment was based on PyTorch 2.2 and PyCharm 2024.2.2, using Networkx 3.2.1 as the main framework for building the knowledge graph, and Neo4j 5.14.0, a high-performance graph database, was selected as the data storage and management system to achieve effective management and efficient querying of complex entity relationships.

[0041] In the data preprocessing stage, the relationship between evidence elements and governance principles was first systematically constructed. The evidence element-governance principle relationship retrieval interface is as follows: Figure 16 As shown, by summarizing literature data and combining it with clinical practice experience, relatively fixed TCM treatment principles were established for each TCM syndrome element. For example, the syndrome element of "essence deficiency" corresponds to "replenishing essence," "yang deficiency" corresponds to "warming yang," and "qi deficiency" corresponds to "tonifying qi"; the syndrome element of "kidney" corresponds to "tonifying the kidney," and "spleen" corresponds to "strengthening the spleen." It is worth noting that for the syndrome element of "liver," based on its pathogenesis characteristics, two treatment principles, "soothing the liver" or "nourishing the liver," are associated, reflecting the flexibility of syndrome differentiation.

[0042] Secondly, a core correlation between syndrome elements and traditional Chinese medicine (TCM) was constructed, and a many-to-many search interface for syndrome element-TCM relationships was established, such as... Figure 17 As shown, this is a many-to-many relationship, meaning one syndrome element corresponds to multiple Chinese herbal medicines (TCMs), and one TCM herb can act on multiple syndrome elements. Based on the medication data recorded in the literature, each syndrome element-TCM pair was statistically summarized, and its frequency of occurrence was quantified as a correlation weight. This weight represents the clinical frequency of the TCM herb under the corresponding syndrome element and is an important basis for subsequent quantitative recommendations. For example, for the syndrome element "kidney" (location of disease), Rehmannia glutinosa (64.44%), Eucommia ulmoides (48.63%), and Angelica sinensis (44.97%) had the highest weights; for the syndrome element "blood stasis" (pathogenesis of disease), Angelica sinensis (55.21%), Rehmannia glutinosa (51.45%), and Eucommia ulmoides (43.76%) were the core drugs. This frequency-based weighting method makes the recommendation results more clinically statistically grounded.

[0043] Third, a symptom-TCM relationship was constructed, and the symptom-TCM relationship retrieval interface is shown below.Figure 18 As shown, by summarizing empirical medications explicitly mentioned in the literature for specific clinical symptoms, 56 key symptom-TCM association pairs were formed. For example, when a patient experiences "lower back and lower limb pain," the system associates herbs such as Eucommia ulmoides, Cibotium barometz, Taxillus chinensis, and Achyranthes bidentata; when experiencing "anxiety and depression," it associates herbs such as Bupleurum chinense, Cyperus rotundus, Curcuma longa, and Triticum aestivum; and when experiencing "poor appetite," it associates herbs such as Crataegus pinnatifida, Massa fermentata, and chicken gizzard lining to promote digestion and relieve stagnation. This design allows the system to not only recommend main prescriptions based on core pathogenesis but also to flexibly add or subtract prescriptions according to the patient's prominent symptoms, making it more closely aligned with clinical practice.

[0044] After data preprocessing, entity recognition and relationship extraction were performed using Python scripts to construct a knowledge graph from the three types of relationships. In the knowledge graph, nodes represent four types of entities: "syndrome elements," "treatment principles," "traditional Chinese medicine," and "symptoms," while edges represent the relationships between them. A "weight" attribute was added to the syndrome element-traditional Chinese medicine edge, and a "specific drug" identifier was added to the symptom-traditional Chinese medicine edge. The completed knowledge graph data was then imported in batches into a Neo4j database for persistent storage.

[0045] Based on this knowledge graph, four core search functions were developed, forming the core logic of prescription recommendation. First, inputting one or more "syndrome elements" returns their corresponding "treatment principles." Second, inputting a single "syndrome element" returns all associated Chinese herbal medicines and their weight scores, sorted in descending order of score. Third, inputting multiple "syndrome elements" simultaneously, the system iterates through the neighboring nodes of each syndrome element, sums the weights of the same Chinese herbal medicine under different syndrome elements as its comprehensive quantitative score, and then sorts and returns the results, thus enabling compound prescription recommendations for multiple syndrome element combinations. Users can set thresholds in the program to control the number of drugs output. Fourth, inputting a specific "symptom" returns commonly used "specialized drugs" associated with it, used for prescription additions and subtractions.

[0046] The application of this module is demonstrated through a specific case: When the system inputs the syndrome elements "Kidney," "Liver," "Essence Deficiency," and "Yin Deficiency," it first retrieves the corresponding treatment principles as "Tonifying the Kidney, Nourishing the Liver, Replenishing Essence, and Nourishing Yin." Subsequently, the system calculates a comprehensive list of recommended medications based on the weight of each syndrome element. If the patient also experiences symptoms such as "irritability or hot flashes in the palms and soles" and "fatigue and weakness," the system further utilizes the "symptom-medicine" relationship, recommending herbs like "Lycium chinense root bark" and "Anemarrhena asphodeloides" for clearing deficiency heat, and "Ginseng" and "Astragalus membranaceus" for tonifying Qi. Finally, through Neo4j's visualization interface, all relevant syndrome elements, symptoms, treatment principles, and medicine nodes, along with their relationships, are clearly displayed graphically, allowing users to intuitively understand the prescription generation logic. By setting a weight threshold (e.g., ≥600), core medications such as "Rehmannia glutinosa," "Eucommia ulmoides," "Phellodendron chinense," and "Anemarrhena asphodeloides" can be selected from the recommendation list to form the final personalized prescription.

[0047] In summary, the knowledge graph-based prescription recommendation system constructed in this embodiment achieves comprehensive and structured prescription reasoning from pathogenesis to treatment principles, and from primary symptoms to secondary symptoms, by establishing a multi-dimensional, quantitative association network of syndrome elements, treatment principles, traditional Chinese medicine, and symptoms. This method not only transforms the rich clinical experience and flexible diagnostic and therapeutic concepts of traditional Chinese medicine into a computable and searchable knowledge system, but also enhances the interpretability of the system through visualization technology, providing clinicians with precise and personalized intelligent prescription assistance. It is a model of the deep integration of traditional Chinese medicine knowledge engineering and artificial intelligence. Example 4: This embodiment involves a multi-dimensional and comprehensive validation experiment on the aforementioned TCM clinical decision support system for primary osteoporosis. The aim is to comprehensively evaluate the reliability, accuracy, and clinical applicability of the TCM syndrome differentiation decision support system and the TCM prescription recommendation system through quantitative indicators and intelligence level tests. This validation system overcomes the limitations of traditional model evaluation and constructs a comprehensive validation scheme that integrates internal performance evaluation and external Turing tests.

[0048] First, a rigorous predictive performance evaluation was conducted on the machine learning-based TCM syndrome differentiation decision support module. This module is a binary classification prediction model, therefore, five indicators were used for quantitative evaluation: accuracy, precision, recall, F1 score, and area under the receiver operating characteristic (AUC) curve. Independent models of 13 major TCM syndrome elements were tested. The results showed that the AUROC values ​​of all optimal models were significantly higher than the random guessing baseline (0.5). Key syndrome elements such as "Kidney" (CatBoost model, AUROC=0.9616), "Essence Deficiency" (LGBM model, AUROC=0.9313), and "Blood Stasis" (LGBM model, AUROC=0.9467) demonstrated excellent discriminative ability, with most models maintaining F1 scores above 0.8, confirming the module's high accuracy and robustness in syndrome element identification.

[0049] Secondly, the medication compatibility rate of the knowledge graph-based prescription recommendation module was evaluated. One hundred real medical cases were randomly selected from the constructed database of TCM syndrome differentiation and treatment information for primary osteoporosis as the test set. The TCM syndrome elements and specific clinical symptoms recorded in each medical case were used as input data. The prescription recommendation module was run to obtain a list of recommended medications, which was then compared with the original prescriptions actually written by clinicians in the medical cases. The compatibility rate was calculated as follows: Medication compatibility rate = (Number of medications recommended by the model that are the same as the original prescription / Total number of medications in the original prescription) × 100%. Statistical analysis showed that the average medication compatibility rate of this model reached 76.64% ± 7.21%, indicating that this knowledge graph recommendation system can highly cover the prescription and medication patterns of human physicians and has significant guiding value for clinical practice.

[0050] Finally, to assess the approximation of the system's diagnostic and treatment results to the level of human experts, a model intelligence evaluation method based on the Turing test was introduced. The specific implementation plan is as follows: Based on the system database, 40 virtual medical cases with standardized structure and terminology were constructed according to the input features selected by the diagnostic module. Subjects were divided into a "computer group" and a "human group." The computer group was operated by a computer professional who obtained the system's syndrome element diagnosis, treatment principles, and medication recommendations based on the four diagnostic methods of the medical cases. Subsequently, four master's students in Traditional Chinese Medicine (TCM) internal medicine, without altering the system's core conclusions, standardized the expression to complete the diagnostic and treatment process. The human group consisted of four TCM internal medicine physicians with intermediate or higher professional titles who conducted completely autonomous diagnostic and treatment analysis on the same batch of medical cases. All subjects had a time limit of 8 minutes to consider each medical case. The outputs of both groups were anonymized and standardized by a specialist to eliminate judgment bias caused by language style. Subsequently, five unsuspecting TCM internal medicine physicians were invited as judges to determine whether the results of each medical case came from a computer or a human within the time limit. The core evaluation metric is the classification error rate (number of incorrect judgments / total number of medical cases × 100%). Experimental results show that the average classification error rate of the judges reached the preset threshold of 50% or higher. This means that the process of judges distinguishing between human and computer prescriptions is close to random guessing, thus proving that the intelligent TCM prescriptions generated by this system have initially passed the Turing test and possess a high level of intelligence and human-like diagnostic and treatment capabilities.

[0051] In summary, through rigorous testing across three dimensions—internal predictive efficacy evaluation, prescription compliance rate verification, and external Turing test—the TCM clinical decision support system constructed in this invention not only possesses excellent computational performance but also approaches the level of human experts in clinical logic and intelligence, providing a solid and reliable experimental basis for its widespread application in the real world. Example 5: Building upon the aforementioned embodiments, a multimodal feature contribution weighting mechanism is introduced to optimize the machine learning differentiation process of TCM syndrome elements. Traditional feature selection methods rely solely on statistical significance, neglecting the differences in clinical importance of different features within TCM theory and the interaction effects between features. This model constructs a feature contribution scoring function that integrates the statistical significance of features, the weight of prior TCM clinical knowledge, and the strength of nonlinear interactions between features, thereby dynamically adjusting the weight of each feature during model training. This scoring function, based on an extended form of information entropy and covariance matrix, quantifies the contribution of features to the discrimination of specific syndrome elements and reduces redundancy issues caused by highly correlated features, ultimately improving the accuracy and robustness of the syndrome element differentiation model. Extract all features from the four diagnostic methods (inspection, auscultation and olfaction) of a structured TCM diagnostic database and perform binary encoding. Let the total number of features be... Each feature is represented as (in ).

[0052] For each element, define a feature contribution score. as follows: In the formula, Features The covariance of the evidence elements reflects statistical correlation; Features The sum of the absolute correlation coefficients with all other features is used to penalize redundant features; The redundancy penalty coefficient is preferably 0.1; Features Information entropy is used to measure the degree of dispersion of feature distribution; It is a very small constant to prevent division by zero errors; This is a priori weight adjustment factor, preferably 0.5, used to balance data-driven approaches and prior knowledge; The prior weights are assigned based on the knowledge of TCM experts.

[0053] For each element, only the score is retained. The top m features form a subset of evidence-specific features.

[0054] Using the selected feature subset, multiple machine learning models are trained, and the ensemble weights of the models are dynamically adjusted based on the importance of the evidence elements. The final output is the evidence element diagnostic result.

[0055] This model significantly improves the accuracy and interpretability of syndrome differentiation by introducing a feature contribution scoring function. On the test set, the model improves the F1 score of common syndrome elements by an average of approximately 5.2% and the AUROC score by an average of approximately 4.1%. Furthermore, feature redundancy is reduced by 30%, and model training time is reduced by approximately 15% because it avoids interference from irrelevant features. Compared with existing syndrome differentiation models based on single statistical feature selection methods, the syndrome differentiation optimization model based on multimodal feature contribution weighting proposed in this embodiment demonstrates significant comprehensive advantages in simulation experiments. The results show that this model achieves more accurate screening of key syndrome differentiation features by integrating statistical correlation, feature redundancy, and prior clinical knowledge in traditional Chinese medicine. Compared with the limited generalization ability of traditional methods due to neglecting the interaction effects and clinical importance between features, this model effectively improves the accuracy and robustness of syndrome differentiation decisions. More importantly, this model endows the prediction results with interpretability, clearly tracing the four diagnostic methods that contribute the most to the diagnostic conclusion, thereby significantly enhancing clinicians' trust in the AI-assisted diagnostic process and the system's practicality.

[0056] Example 6: Building upon the aforementioned embodiments, a dynamic weight evolution mechanism is designed to enable the entity relationship weights in the knowledge graph to adaptively update with the influx of new data, thereby simulating the accumulation and optimization process of TCM prescription experience. Traditional knowledge graph weights are typically calculated based on static frequency, failing to reflect dynamic changes and contextual dependencies in clinical practice. This mechanism constructs a weight evolution function that integrates historical usage frequency, recent data trends, and clinical feedback signals to dynamically adjust the weights of syndrome element-TCM herbal relationships. This function, based on the ideas of exponential smoothing and feedback reinforcement learning, introduces a time decay factor and a feedback gain term to ensure that the knowledge graph maintains the stability of historical experience while rapidly adapting to new clinical models, ultimately improving the personalization and timeliness of prescription recommendations.

[0057] Historical usage frequencies of the "syndrome element-Chinese herbal medicine" relationship were extracted from the TCM syndrome differentiation and treatment database, and initial weights were calculated. ,in To prove With Chinese medicine The frequency of co-occurrence.

[0058] As new data flows in, the weights are iteratively updated. The weight evolution formula is as follows: In the formula, For time Time Evidence With Chinese medicine The weights; This represents the rate of change of the frequency of this relationship within the current time period; This is a time decay factor for the relationship since the last update, used to reduce the impact of outdated data; The attenuation coefficient is preferably 0.05; This serves as a clinical feedback signal, calculated based on the user's adoption rate of the recommended prescription; , and Let be the weighting coefficient, satisfying Preferred , , Each of these factors controls the balance between historical weights, the impact of new data, and the intensity of feedback.

[0059] The knowledge graph is reconstructed regularly using updated weights, and the graph retrieval algorithm is optimized to prioritize the recommendation of high-weight traditional Chinese medicines.

[0060] Based on the syndrome elements and symptoms input by the user, the system retrieves the highest-weighted combination of traditional Chinese medicine from the knowledge graph and generates a prescription.

[0061] This mechanism, through dynamic weight evolution, makes prescription recommendations more aligned with changes in clinical practice. In a test with 100 new medical cases, the medication compliance rate of recommended prescriptions increased to 84.3%, and the user feedback adoption rate increased by 12.5%. Furthermore, the knowledge graph update cycle was shortened by 50% due to the reduced need for manual intervention caused by automated weight adjustments. Compared to existing knowledge graph recommendation methods that use static historical frequency-based weight calculations, the knowledge graph prescription recommendation mechanism based on dynamic weight evolution constructed in this embodiment has achieved fundamental progress in simulation experiments. The results show that this mechanism, by introducing a time decay factor and clinical feedback signals, allows the prescription logic of the knowledge graph to continuously evolve, rather than adhering to outdated data patterns. Compared to the rigidity and lag exhibited by existing static methods when facing new clinical practices and personalized needs, this mechanism can adaptively capture micro-changes in medication trends and incorporate real physician feedback. This not only significantly improves the degree of alignment and personalization between prescription recommendations and actual clinical needs but also enables the system to self-optimize through continuous interaction, thus achieving a leap from a static knowledge base to a dynamic intelligent agent.

Claims

1. A TCM clinical decision support system for primary osteoporosis, characterized in that, include: The data processing module is used to standardize terminology and extract syndrome elements from raw TCM literature data, and to build a structured TCM syndrome and treatment database. The dialectical reasoning module, built on a machine learning model, is used to output TCM syndrome element diagnostic results based on the input four diagnostic methods. The prescription recommendation module, built on a knowledge graph, is used to recommend corresponding treatment principles and Chinese herbal prescriptions based on the diagnostic results of the TCM syndrome elements and / or specific clinical symptoms.

2. The system according to claim 1, characterized in that: The data processing module maps non-standard TCM diagnostic information, syndrome types, and Chinese herbal medicine names to a standardized terminology system, and breaks down complex TCM syndrome types into combinations of single disease location syndrome elements and disease nature syndrome elements.

3. The system according to claim 1, characterized in that: The diagnostic reasoning module uses a sparse penalized regression model to select a subset of features with significant predictive ability for the target syndrome from all the four diagnostic information, and integrates multiple classification algorithms to construct an independent binary classification prediction model for each TCM syndrome. The system selects the optimal prediction model for each element and integrates them according to the preset comprehensive performance evaluation index.

4. The system according to claim 3, characterized in that, The classification algorithms include at least the following three categories: Improve decision tree algorithms, neural network algorithms, support vector machine algorithms, and random forest algorithms.

5. The system according to claim 3, characterized in that: The dialectical reasoning module further performs feature screening based on a multi-dimensional feature contribution scoring function. This function integrates the statistical significance of the feature, its redundancy with other features, and its clinical prior knowledge weight in traditional Chinese medicine theory, in order to screen out a subset of features with high discriminative power for the target syndrome.

6. The system according to claim 1, characterized in that: The prescription recommendation module constructs a graph data structure with syndrome elements, treatment principles, traditional Chinese medicines, and clinical symptoms as entities, and the treatment and correlation relationships between them as edges. It assigns weights to the syndrome element-traditional Chinese medicine relationships based on the frequency of use of traditional Chinese medicines under the corresponding syndrome elements, and sorts and filters the recommended drugs according to the weights.

7. The system according to claim 6, characterized in that, The prescription recommendation module is further configured as follows: Supports independent and combined searches of three core relationships: syndrome element-treatment principle, syndrome element-traditional Chinese medicine, and symptom-traditional Chinese medicine; It supports dynamically adjusting the number and range of recommended drugs based on user-defined weight thresholds.

8. The system according to claim 1, characterized in that: The system also includes a human-machine collaborative evaluation module, which simulates the Turing test environment, generates virtual medical records, and allows the system and human physicians to independently diagnose and prescribe the patients. The similarity and acceptability of the prescriptions are compared through double-blind evaluation to assess the clinical intelligence level of the system.

9. The system according to claim 1, characterized in that: The system also includes a multimodal input interface for receiving and structuring diagnostic information input from text, voice, or structured forms.

10. The system according to claim 1, characterized in that: The output of the dialectical reasoning module includes an interpretable output unit, which displays to the user the key diagnostic information that leads to the judgment of a specific syndrome element and its contribution.

Images