Intelligent diagnosis and dynamic optimization method combining traditional Chinese medicine and western medicine

Abstract

The invention relates to the technical field of artificial intelligence and medical diagnosis, in particular to an intelligent diagnosis and dynamic optimization method combining traditional Chinese medicine and western medicine. The intelligent diagnosis and dynamic optimization method based on combination of traditional Chinese medicine and western medicine includes the steps that multi-modal data such as traditional Chinese medicine tongue picture images are collected and preprocessed, traditional Chinese medicine syndrome differentiation and western medicine disease models are input respectively, a supervision model group is distributed or a collaborative diagnosis mechanism is started according to model confidence, and a structured diagnosis report is generated through double verification. And obtaining a compliance diagnosis result through data processing, and dynamically adjusting a knowledge calling priority according to user feedback to obtain a calling result. According to the method, a complete closed loop of data acquisition, intelligent diagnosis, safety supervision and feedback optimization is constructed, the system is endowed with self-learning and self-adaptive capabilities while the precision, the safety and the compliance of combined diagnosis of traditional Chinese medicine and western medicine are improved, and an efficient and reliable technical path is provided for intelligent diagnosis and treatment of complex diseases.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and medical diagnosis technology, and in particular to an intelligent diagnosis and dynamic optimization method combining traditional Chinese and Western medicine. Background Art

[0002] In the field of medical diagnosis, integrated Chinese and Western medicine diagnosis can combine the advantages of traditional Chinese medicine syndrome differentiation and treatment with Western medicine precision testing, providing a more comprehensive solution for the diagnosis and treatment of complex diseases. However, existing technologies in integrated Chinese and Western medicine diagnosis have the following key issues that need to be addressed:

[0003] Low efficiency of multimodal data fusion: Traditional Chinese medicine data (such as tongue images, pulse diagnosis, and symptom texts) and Western medicine data (such as medical images, test reports, and electronic medical records) are heterogeneous. Traditional methods lack a unified preprocessing and feature fusion mechanism, resulting in the inadequate release of data value.

[0004] Insufficient diagnostic reliability of a single model: When the TCM syndrome differentiation model and the Western medicine disease model operate independently, they are easily affected by data noise, model bias, etc., resulting in substandard confidence or conflicting conclusions, and lack of an effective collaborative decision-making mechanism to integrate multi-source evidence.

[0005] Lack of diagnostic security and compliance: The existing system lacks multi-level verification of diagnostic results, making it difficult to avoid logical errors such as medication contraindications and diagnostic contradictions. At the same time, medical data faces the risk of privacy leakage during the sharing and aggregation process, and traditional methods do not meet strict compliance requirements (such as data de-identification and encrypted aggregation).

[0006] Poor model adaptability: Model updates rely on manual parameter adjustment or fixed training data, and are unable to dynamically optimize knowledge call priority and model capabilities based on user feedback, making it difficult to adapt to the complexity and diversity of clinical scenarios.

[0007] In summary, the existing integrated Chinese and Western medicine diagnostic technology has significant deficiencies in data processing, decision-making mechanisms, security compliance, and model evolution. There is an urgent need for an intelligent diagnostic method that can integrate multimodal data, improve diagnostic reliability, ensure data security, and support dynamic optimization. Summary of the Invention

[0008] In view of this, the purpose of the present invention is to provide an intelligent diagnosis and dynamic optimization method combining traditional Chinese and Western medicine to solve the problems of low efficiency of multimodal data fusion, insufficient reliability of single model diagnosis, lack of diagnostic safety and compliance, and poor model adaptability in the existing traditional Chinese and Western medicine combined diagnosis technology in the prior art.

[0009] According to a first aspect of an embodiment of the present invention, a method for intelligent diagnosis and dynamic optimization combining traditional Chinese and Western medicine is provided, the method comprising:

[0010] Collecting multimodal data; the multimodal data includes: traditional Chinese medicine data and Western medicine data;

[0011] The TCM data includes: tongue image, symptom text, and pulse diagnosis signal;

[0012] The Western medicine data include: electronic medical records, medical images, and test reports;

[0013] Preprocessing the multimodal data, inputting the preprocessed TCM data into a preset TCM syndrome differentiation model, and outputting TCM syndrome classification;

[0014] Inputting the pre-processed Western medicine data into a preset Western medicine disease model and outputting a disease prediction result;

[0015] Based on the TCM syndrome classification and disease prediction results, determine whether the confidence level of the single model meets the standards;

[0016] If the criteria are met, the user is assigned to a supervision model group based on the user's risk level according to the TCM syndrome classification and disease prediction results;

[0017] If the standards are not met, the collaborative diagnosis mechanism will be activated to obtain a comprehensive decision-making model vote. After the results are integrated based on the comprehensive decision-making model vote, the supervision model group will be assigned according to the user's risk level;

[0018] Using a preset double verification mechanism to verify the supervision model group allocated according to user risk level, if the verification is passed, a structured diagnosis report is generated;

[0019] Otherwise, the conflicting suggestion is automatically rejected;

[0020] Using a preset federated learning framework to de-identify the terminal layer data of the structured diagnostic report, generate encrypted feature vectors, and homomorphically encrypt the aggregation parameters of the central layer of the structured diagnostic report through secure multi-party computation, and add noise in combination with differential privacy to obtain compliance diagnosis results;

[0021] User feedback is collected for the obtained compliance diagnosis results, and the knowledge call priority is dynamically adjusted according to the user feedback to obtain the call results.

[0022] Furthermore, the preprocessing of the multimodal data includes:

[0023] unifying the perspective of the acquired tongue image through image registration, and extracting texture features of the tongue image;

[0024] Encoding the symptom text using a preset TCM terminology library code;

[0025] Convert the preset pulse diagnosis signal into time-frequency domain features;

[0026] Extracting the features of the lesion area of the medical image through a preset convolutional neural network;

[0027] Parsing the inspection report into a numerical vector;

[0028] The electronic medical records are structured using NLP technology.

[0029] Furthermore, judging whether the confidence of a single model meets the standard based on the TCM syndrome classification and disease prediction results includes:

[0030] According to the TCM syndrome classification and disease prediction results, confidence thresholds are set for the TCM syndrome differentiation model and the Western medicine disease model respectively;

[0031] Calculate the confidence of the TCM syndrome classification output by the TCM syndrome differentiation model, and the confidence of the disease prediction results output by the Western medicine disease model;

[0032] If the confidence of the TCM syndrome differentiation model is greater than or equal to its corresponding confidence threshold, and the confidence of the Western medicine disease model is greater than or equal to its corresponding confidence threshold, then the confidence of the single model is judged to meet the standard; otherwise, the confidence of the single model is judged to fail to meet the standard.

[0033] Furthermore, the collaborative diagnosis mechanism is activated to obtain a comprehensive decision-making model vote, including:

[0034] Extract the evidence related to Chinese and Western medicine related to the current case from the historical medical record database and build an evidence chain;

[0035] Organize a small model group of multiple experts to generate pro and con arguments for the current case and conduct adversarial debates;

[0036] The comprehensive decision-making model is used to vote on the results after the adversarial debate to obtain the voting results.

[0037] Furthermore, the allocation of supervision model groups according to user risk levels specifically includes:

[0038] Determine the user's risk level based on the user's age, disease severity, and past medical history; preset corresponding supervision model groups for different risk levels;

[0039] According to the determined user risk level, the corresponding supervision model group is assigned to supervise the diagnosis results.

[0040] Furthermore, the preset double verification mechanism includes:

[0041] Rule engine verification: Using the preset rule engine, the diagnostic results after the supervision model group is assigned according to the user's risk level are logically verified to check for logical errors such as medication contraindications and diagnostic contradictions;

[0042] Semantic similarity verification: The semantic similarity between diagnostic recommendations and taboo rules is calculated using the twin-tower model. If the semantic similarity exceeds a preset threshold, a conflict is determined.

[0043] Only when both the rule engine verification and the semantic similarity verification pass, can it be considered to have passed the double verification.

[0044] Furthermore, the method of using a preset federated learning framework to de-identify the terminal layer data of the structured diagnostic report and generate an encrypted feature vector specifically includes:

[0045] De-identify the structured diagnostic report data at the terminal layer to remove any information that can directly or indirectly identify the user;

[0046] The de-identified data is encrypted using an encryption algorithm to generate an encrypted feature vector.

[0047] Furthermore, the central layer of the structured diagnostic report is homomorphically encrypted with aggregated parameters through secure multi-party computation, and noise is added in combination with differential privacy, including:

[0048] At the center layer, secure multi-party computation is used to perform homomorphic encryption aggregation on the encrypted feature vectors uploaded by each terminal layer to obtain the aggregation parameters;

[0049] Noise that meets the requirements of differential privacy is added to the aggregation parameters to protect the privacy of the data.

[0050] Furthermore, dynamically adjusting the knowledge call priority according to user feedback and obtaining the call result includes:

[0051] Establish a knowledge call priority database to record the call priority of different knowledge;

[0052] Dynamically adjust the knowledge call priority in the knowledge call priority database according to user feedback on the diagnosis results;

[0053] Perform knowledge call according to the adjusted knowledge call priority and obtain the call result.

[0054] Furthermore, the method further includes: calculating the merit points of the TCM syndrome differentiation model and the Western medicine disease model based on user feedback and the accuracy of the diagnosis results;

[0055] The models are divided into expert and novice groups according to their merit points;

[0056] The expert group's model was trained with more complex case data to improve its diagnostic capabilities in complex scenarios.

[0057] For the model of the novice group, a knowledge distillation method is used to learn knowledge from the model of the expert group to quickly improve its diagnostic ability;

[0058] Regularly update the model's merit points, dynamically adjust the group to which the model belongs, and achieve dynamic optimization of the model.

[0059] The intelligent diagnosis and dynamic optimization method of integrated traditional Chinese and Western medicine provided by the present invention effectively solves the shortcomings of the existing technology through multi-dimensional technological innovation and has the following significant beneficial effects:

[0060] 1. Deep integration of multimodal data to improve the quality of basic diagnosis

[0061] A standardized preprocessing process (such as image registration, time-frequency domain conversion, and NLP structured processing) has been designed for heterogeneous data such as traditional Chinese medicine tongue images, pulse diagnosis, symptom texts and Western medicine images and test reports. This process converts unstructured data into computable feature vectors, providing high-quality input for subsequent model reasoning, ensuring the semantic consistency and complementarity of traditional Chinese and Western medicine data, and significantly improving the standardization and effectiveness of data preprocessing.

[0062] 2. Collaborative diagnosis and double verification to ensure diagnostic accuracy and safety

[0063] Collaborative diagnosis mechanism: When the confidence level of a single model does not meet the standard, a historical medical case evidence chain is constructed to conduct a debate with multiple expert models to generate comprehensive decision-making voting results, significantly improving the diagnostic accuracy of complex cases and effectively solving the reliability issues of a single model caused by data noise or bias.

[0064] Double verification system: The rule engine verification quickly intercepts logical conflicts such as medication contraindications and diagnostic contradictions, and the semantic similarity verification identifies implicit semantic risks, forming a multi-level security barrier of "rule logic + semantic understanding" to ensure that treatment recommendations comply with the clinical standards of traditional Chinese and Western medicine, and build a diagnostic safety protection system covering clear contraindications and ambiguous semantic risks.

[0065] 3. Federated learning and privacy protection technology to meet medical data compliance requirements

[0066] The terminal layer removes sensitive information through de-identification and hashing. The central layer utilizes secure multi-party computation (MPC) for homomorphic encryption aggregation and incorporates differential privacy to add noise, enabling secure sharing of data without leaving the domain. These technical processes ensure data privacy while maintaining high model performance, fully complying with medical data security standards and providing a compliant technical path for cross-institutional data collaboration.

[0067] 4. Dynamic optimization and self-evolution capabilities to adapt to changes in clinical scenarios

[0068] Model merit is calculated based on user feedback. Through knowledge distillation and complex case training, a self-evolving closed loop of "capability stratification and targeted optimization" is formed. Furthermore, knowledge call priority is dynamically adjusted based on diagnostic results, enabling the system to continuously improve its adaptability in complex disease diagnoses, achieving an upgrade from a "static model" to an intelligent diagnostic system with self-learning capabilities.

[0069] 5. Cross-modal feature fusion and quantitative decision-making to enhance the depth of integration of Chinese and Western medicine

[0070] By calculating bidirectional attention weights for both traditional Chinese and Western medicine data through a dynamic spatiotemporal attention mechanism, we achieve the organic integration of cross-modal features, breaking down the technical barriers between traditional Chinese and Western medicine diagnosis. Methods for quantifying evidence weights and model contributions make diagnostic evidence traceable, providing transparent support for clinical decision-making and promoting deep collaboration and complementarity between traditional Chinese and Western medicine diagnostic elements.

[0071] In summary, the present invention constructs a complete closed loop of "data collection-intelligent diagnosis-safety supervision-feedback optimization". While improving the diagnostic accuracy, safety and compliance of integrated traditional Chinese and Western medicine, it gives the system self-learning and self-adaptation capabilities, providing an efficient and reliable technical path for the intelligent diagnosis and treatment of complex diseases.

[0072] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0074] Figure 1 It is a schematic diagram of an intelligent diagnosis and dynamic optimization method combining traditional Chinese and Western medicine according to an exemplary embodiment. DETAILED DESCRIPTION

[0075] Exemplary embodiments will be described in detail herein, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, like numbers in different figures represent like or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all possible embodiments consistent with the present invention. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present invention, as detailed in the appended claims.

[0076] Example 1

[0077] See also Figure 1 , Figure 1 FIG1 is a schematic diagram of an intelligent diagnosis and dynamic optimization method combining traditional Chinese and Western medicine according to an exemplary embodiment, the method comprising:

[0078] S1. Collecting multimodal data; the multimodal data includes: traditional Chinese medicine data, Western medicine data;

[0079] The TCM data includes: tongue image, symptom text, and pulse diagnosis signal;

[0080] The Western medicine data include: electronic medical records, medical images, and test reports;

[0081] S2. Preprocessing the multimodal data, inputting the preprocessed TCM data into a preset TCM syndrome differentiation model, and outputting a TCM syndrome classification;

[0082] S3. Input the pre-processed Western medicine data into a preset Western medicine disease model and output the disease prediction results;

[0083] S4. Based on the TCM syndrome classification and disease prediction results, determine whether the confidence level of the single model meets the requirements;

[0084] S5. If the target is met, the user is assigned to a supervision model group according to the TCM syndrome classification and disease prediction results according to the user's risk level;

[0085] If the standards are not met, the collaborative diagnosis mechanism is activated to obtain a comprehensive decision model vote. After integrating the results based on the comprehensive decision model vote, the supervision model group is assigned according to the user's risk level.

[0086] S7. Using a preset double verification mechanism to verify the user risk level distribution supervision model group, if the verification is passed, a structured diagnostic report is generated;

[0087] S8. Otherwise, the conflicting suggestion is automatically rejected;

[0088] S9. De-identify the terminal-layer data of the structured diagnostic report using a pre-defined federated learning framework to generate encrypted feature vectors, homomorphically encrypt aggregated parameters of the central layer of the structured diagnostic report using secure multi-party computation, and add noise using differential privacy to obtain a compliance diagnosis result.

[0089] S10. Collect user feedback on the obtained compliance diagnosis results, dynamically adjust the knowledge call priority according to the user feedback, and obtain the call results.

[0090] Furthermore, the present invention achieves deep fusion and semantic unification of heterogeneous data of Chinese and Western medicine through standardized multimodal data preprocessing, improves the diagnostic accuracy of complex cases and builds a multi-level security barrier with the help of collaborative diagnosis mechanism and double verification system, and uses federated learning and privacy protection technology to achieve safe and compliant sharing of medical data without "data leaving the domain". It dynamically adjusts the knowledge call priority based on user feedback and forms a self-evolutionary closed loop through merit point model grouping and knowledge distillation. At the same time, it promotes deep collaboration of Chinese and Western medicine diagnostic elements through cross-modal feature fusion and quantitative decision-making, and finally constructs a complete closed loop covering data collection, intelligent diagnosis, safety supervision to feedback optimization, providing an efficient, accurate, safe and adaptive intelligent Chinese and Western medicine combined diagnosis technology solution for complex diseases.

[0091] This embodiment takes the diagnosis of chronic gastritis patients as an example to specifically illustrate the execution process and key technical implementation of the method:

[0092] 1. Multimodal Data Acquisition and Preprocessing

[0093] 1. Data Collection

[0094] The patients' traditional Chinese medicine data were collected: tongue images (RGB format, resolution 640×480), symptom text ("burning pain in the stomach, dry mouth and bitter taste"), and pulse diagnosis signals (radial artery pulse wave, sampling frequency 100Hz); Western medicine data: gastroscopy images (DICOM format), test reports (including CRP=15mg / L, Hp antibody positive), and electronic medical records (previous diagnosis of "chronic superficial gastritis").

[0095] 2. Standardization

[0096] Traditional Chinese Medicine data preprocessing:

[0097] Tongue images are unified in shooting perspective (eliminating light and angle deviation) through image registration technology, and texture features are extracted using discrete cosine transform to obtain a dimension of N×D TCM The feature matrix X TCM ;

[0098] For example, the low-frequency, medium-frequency, and high-frequency texture coefficients of each region are extracted by discrete cosine transform (DCT). If the number of coefficients is D TCM , then the dimension of each feature vector is D TCM .

[0099] In a specific implementation, N represents the number of extracted features (or the number of feature vectors).

[0100] For example, if the tongue image is divided into N local areas (such as block processing), or N texture feature points are extracted, then N is the number of areas / points.

[0101] DTCM : Represents the dimension (length) of each feature vector, that is, the amount of attribute information contained in a single feature (related to the description granularity of TCM texture features).

[0102] Symptom text is encoded using the TCM terminology database, mapped to BERT word vectors (dimension 768), and combined with TF-IDF to generate symptom semantic features;

[0103] The pulse diagnosis signal is converted into time-frequency domain features through wavelet transform, and the energy spectrum of each frequency band is calculated by the following formula:

[0104] E(f)=∣∫x(t)ψ*(t,f)dt∣2 (1);

[0105] Specifically, (ψ is the Morlet wavelet basis function), x(t) is the time domain pulse wave signal (unit: mmHg or pressure value) collected by pulse diagnosis, and the sampling frequency is usually 50-200Hz; ψ(t,f) is the Morlet wavelet, and the expression is Where f is the center frequency and ψ(t,f) is its complex conjugate (used to extract signal phase information).

[0106] Western medicine data preprocessing:

[0107] Gastroscopy images are extracted through CNN (ResNet-50) to extract the features of the lesion area, and the output dimension is M×D WM The feature matrix X WM ;

[0108] The test report is parsed into a numerical vector: CRP = 15, Hp antibody = 1;

[0109] Electronic medical records use NLP technology to perform entity recognition and relationship extraction, extracting key information such as "chronic superficial gastritis" and "burning pain in the stomach".

[0110] 2. Dual-model independent reasoning and confidence judgment

[0111] Model input and inference

[0112] Traditional Chinese Medicine Syndrome Differentiation Model (M1): Based on the Transformer Encoder architecture, it inputs tongue texture features, symptom word vectors, and pulse diagnosis time-frequency domain features, and outputs the TCM syndrome classification as "liver and stomach stagnation heat syndrome" with a confidence level of 0.82;

[0113] Western medicine disease model (M2): Based on the multimodal Transformer, it inputs gastroscopy image features, test indicators and medical record text, and outputs the disease prediction result as "acute exacerbation of chronic gastritis" with a confidence level of 0.75.

[0114] Confidence Judgment

[0115] The preset confidence thresholds for the TCM model are 0.8 and 0.8 for the Western medicine model. The current confidence level of the TCM model (0.82) meets the standard, while that of the Western medicine model (0.75) does not, triggering the collaborative diagnosis mechanism.

[0116] 3. Activation of the collaborative diagnosis mechanism and comprehensive decision-making

[0117] 1. Evidence chain generation and adversarial debate

[0118] 200 cases related to "liver and stomach stagnation and heat syndrome + chronic gastritis" were retrieved from the historical medical record database, and evidence of the correlation between traditional Chinese medicine and Western medicine was extracted (e.g., the correlation between "liver and stomach stagnation and heat syndrome" and the degree of gastric mucosal inflammation was r = 0.78);

[0119] A multi-expert small model group (including TCM syndrome models and Western medicine inflammation models) generated pro and con arguments:

[0120] Proponent: The red tongue with yellow fur (supporting liver and stomach heat syndrome) is logically consistent with the gastroscopic mucosal congestion (supporting acute inflammation);

[0121] Counterargument: There is insufficient evidence to directly link a stringy and rapid pulse (corresponding to liver depression) with Hp infection (the cause in Western medicine).

[0122] 2. Comprehensive decision-making voting

[0123] The “maximum model” that integrates five expert-level small models uses a weighted voting mechanism, and the weight of evidence is calculated using the following formula:

[0124]

[0125] Among them, e i represents the i-th piece of evidence (diagnostic basis from the expert model, such as tongue features, imaging results, etc.); Sim(e i ,query)evidence i Correlation with the current case (query) (0-1, the higher the correlation); Conf(e i ) Evidence i Confidence level; w i Evidence i The normalized weight of .

[0126] The final comprehensive diagnosis was "liver and stomach heat syndrome (confidence 0.85) combined with acute exacerbation of chronic gastritis (confidence 0.88)".

[0127] In specific implementation, in the diagnosis of integrated traditional Chinese and Western medicine, "the 'largest model' integrating 5 expert-level small models adopts a weighted voting mechanism" means generating the final diagnosis result by integrating the outputs of multiple small models focusing on different fields (such as traditional Chinese medicine syndromes, Western medicine imaging, and test index analysis) and combining them with evidence weight calculation.

[0128] The five small models focus on different diagnostic dimensions:

[0129] Small model 1: TCM tongue image-syndrome correlation analysis;

[0130] Small Model 2: Traditional Chinese Medicine Pulse Diagnosis-Syndrome Classification;

[0131] Small model 3: Western medicine gastroscopy imaging-lesion identification;

[0132] Small model 4: Test report - Inflammation index analysis;

[0133] Small model 5: Electronic medical record-medical history logical reasoning;

[0134] Each small model outputs an independent diagnostic result (e.g., small model 1 supports "liver and stomach heat syndrome", small model 3 supports "acute inflammation of gastric mucosa"), but there may be conflicting conclusions or insufficient confidence.

[0135] More specifically, the weights are normalized using the above-mentioned evidence weight formula so that the contribution ratio of each piece of evidence is clear and traceable.

[0136] This mechanism achieves multi-model "intelligent aggregation" through "evidence quality quantification-weighted decision-making", integrating multi-dimensional evidence from traditional Chinese and Western medicine and quantifying uncertainty through confidence. In the case of chronic gastritis, high confidence levels of 0.85 and 0.88 verified its reliability, providing clinical collaborative diagnostic support of "1+1>2", and is the core technical path to improve decision-making accuracy and safety in complex medical scenarios. 4. Dynamic Supervision and Privacy Protection

[0137] 1. Double verification mechanism

[0138] Rule engine verification: Checks whether there are any medication contraindications in the treatment recommendations (e.g., the warming tonic Huangqi should be avoided for "liver and stomach heat syndrome," but there is no conflict in the current recommendations);

[0139] In one embodiment, a rule engine verification mechanism is utilized to quickly intercept logical conflicts, specifically based on preset medical rules, to quickly identify obvious diagnostic or medication contraindication conflicts.

[0140] Clear contraindications for both Traditional Chinese Medicine (TCM) and Western Medicine diagnosis and treatment are built into the system. For example, TCM contraindications include: "Liver and Stomach Stagnation Heat Syndrome" should not be treated with warming and tonic herbs (such as Astragalus and Codonopsis) because their properties conflict with the "stagnant heat" of the syndrome; Western Medicine contraindications include: Penicillins conflict with medications for patients with a history of penicillin allergy; and TCM and Western Medicine compatibility contraindications include: The combined use of licorice and aspirin may increase the risk of gastrointestinal irritation. Rules are structured using an "IF-THEN" structure (e.g., "IF the syndrome is Liver and Stomach Stagnation Heat Syndrome AND the prescription contains Astragalus THEN trigger the contraindication").

[0141] When the system generates treatment recommendations (such as traditional Chinese medicine prescriptions or Western medicine combinations), the rule engine automatically scans the drug / syndrome combinations in the recommendations; it matches the preset rules line by line. If it is found that the prescription corresponding to "liver and stomach heat syndrome" contains astragalus, it is directly marked as "conflict" and the recommendation is automatically rejected. For example: in the case of chronic gastritis, the recommended prescription is "Zuojin Wan combined with Chaihu Shugan San" (excluding astragalus). The rule engine verifies that there is no conflict and proceeds to the next step.

[0142] Semantic similarity verification: The cosine similarity between the diagnostic suggestion and the taboo rule is calculated using the twin-tower model, using the following formula:

[0143]

[0144] The calculation result is 0.32 (<threshold 0.6), which means there is no conflict and the verification passes.

[0145] Specifically, it handles fuzzy semantic conflicts (such as synonyms and implicit associations) that cannot be covered by the rule engine, and calculates the semantic matching degree between diagnostic suggestions and taboo rules through the twin-tower model.

[0146] Specifically, at the encoding layer:

[0147] Input 1: Diagnostic suggestion (e.g., “It is recommended to take Chinese medicine with heat-clearing effect”), which is encoded into a vector v through an NLP model such as BERT 建议 ;

[0148] Input 2: Taboo rules (e.g., "Hot-heat drugs should not be used for liver and stomach stagnation syndrome"), encoded as vector v 规则 .

[0149] Similarity calculation: Using the cosine similarity formula:

[0150]

[0151] The result range is [-1, 1], with higher values indicating more semantic similarity (0.6 is the preset threshold, which can be adjusted according to clinical needs).

[0152] The judgment logic is as follows:

[0153] If Sim ≥ threshold (such as 0.6):

[0154] If the semantics of the diagnostic suggestion and the contraindication rule are highly similar, it is judged as a "potential conflict" (for example, the suggestion mentions "use of warming yang drugs", and the similarity with the rule "Yure syndrome should avoid warming heat" is 0.7, triggering rejection);

[0155] If Sim < threshold:

[0156] The semantic differences are large and are judged as "no conflict" (for example, in the case where the similarity is 0.32 < 0.6, it passes the verification).

[0157] In specific implementation, the rule engine and semantic similarity verification form a complementary security protection system:

[0158] Rule engine: It has built-in clear rules for Chinese and Western medicine diagnosis and medication contraindications (such as "liver and stomach heat syndrome should not use the warming tonic drug Astragalus"). It uses "IF-THEN" logic to quickly scan treatment suggestions and accurately intercept "hard conflicts" that directly conflict with preset rules, ensuring that diagnostic suggestions that are obviously in violation of regulations are immediately excluded.

[0159] Semantic similarity verification: With the help of the dual-tower model, the diagnostic recommendations and taboo rules are semantically encoded and cosine similarity is calculated to deeply identify "soft conflicts" not covered by the rule engine (such as the implicit semantic association between "warming tonic drugs" and "warming yang drugs"), avoiding potential risks caused by differences in natural language expressions.

[0160] The two work together to form a double insurance: the rule engine quickly filters and clarifies taboos, and the semantic verification captures ambiguous semantic risks, jointly ensuring that treatment recommendations comply with traditional Chinese and Western medical standards, and building a multi-level diagnostic safety barrier.

[0161] Terminal layer processing: De-identify sensitive information such as patient ID and address, and generate irreversible feature summaries (such as hash values);

[0162] During implementation, direct identification information (PII) such as patient ID, address, and contact information will be deleted or replaced. For example:

[0163] Replace "patient 12345" with the anonymous ID "P_20250423_001";

[0164] Generalize the address (for example, only retain the city level and hide the street number).

[0165] Hash the indirect identification information in medical data (such as unique symptom combinations in medical records) to generate irreversible feature summaries:

[0166] Using encryption hash algorithms such as SHA-256, "red tongue with yellow fur + rapid and tense pulse + burning pain in the stomach" is converted into a hash value of fixed length (such as a 64-bit string) to ensure that the original data cannot be inferred from the summary.

[0167] Central layer aggregation: The parameters of each terminal model are aggregated through homomorphic encryption through secure multi-party computing (MPC). The formula is:

[0168] E(∑x i )=∑E(x i )mod n (5);

[0169] And add differential privacy noise (Laplace mechanism, privacy budget ∈=1) to ensure that "data does not leave the domain".

[0170] Specifically, without decrypting the original data, the model parameters of each terminal are securely aggregated to achieve global model updates. Multiple terminals (such as different hospitals) can collaboratively calculate the aggregated results of model parameters (such as average values, gradient updates) without sharing the original data.

[0171] Furthermore, each terminal encrypts the locally trained model parameters (such as the Transformer weights of the TCM main model) (e.g., using the RSA homomorphic encryption algorithm);

[0172] The central server receives the encrypted parameters and performs aggregation operations such as addition / multiplication (such as calculating the average of all terminal parameters). The result is still in ciphertext form.

[0173] Formula example (additive homomorphic encryption): E(x)+E(y)=E(x+y);

[0174] That is, the sum of the ciphertexts is equal to the ciphertext after the addition of the plaintexts, ensuring that the intermediate result cannot be decrypted.

[0175] More specifically, random noise that conforms to the Laplace distribution is injected into the aggregated parameters to blur the influence of individual terminal data. The formula is:

[0176]

[0177] Where Δf is the sensitivity (maximum variation), ∈=1 is the privacy budget (the smaller the value, the stronger the privacy protection). In practice, even if an attacker obtains the aggregated parameters, he cannot infer the individual patient

[0178] Or the specific data of the terminal, meeting the "differential privacy" standard (such as changing a single data point does not affect the probability distribution of the final result).

[0179] In its specific implementation, the federated learning privacy protection mechanism builds a complete chain of "data localization processing-parameter encryption transmission-global security aggregation" through terminal-layer de-identification, central-layer security aggregation and noise addition, ensuring that the integrated Chinese and Western medicine diagnosis system strictly abides by privacy protection specifications during data sharing and model training, providing key technical support for the compliant implementation of medical AI.

[0180] 5. User Feedback and Model Self-Evolution

[0181] 1. Structured diagnostic report generation

[0182] The output report contains:

[0183] TCM Syndrome: Liver and stomach heat stagnation syndrome, recommended prescription "Zuojin Wan combined with Chaihu Shugan San";

[0184] Western medicine diagnosis: acute exacerbation of chronic gastritis, recommended combined medication "omeprazole + amoxicillin";

[0185] Diagnostic basis: red tongue with yellow fur (texture feature matching degree 0.91) and endoscopic mucosal congestion (lesion feature weight 0.85).

[0186] 2. Dynamic optimization and knowledge updating

[0187] Merit calculation: The contribution of TCM model features is calculated using the following formula

[0188] The average attention weight is 0.78, and the Shapley value of the Western medicine model is Comprehensive merit points, as follows

[0189] In specific implementation, regarding the feature contribution (TCM model):

[0190] The contribution of the model to the effective features in the feature extraction stage is calculated based on the attention mechanism, specifically the mean of the feature attention weights in all time steps (range 0-1, the higher the value, the more critical the features extracted by the model).

[0191] In specific implementation, when the traditional Chinese medicine model processes tongue images, the average attention weight of the "red tongue coating with yellow fur" texture feature is 0.78, indicating that the model can effectively focus on core visual features (such as the typical tongue image of "liver and stomach heat syndrome") in syndrome classification.

[0192] In specific implementation, the Shapley value (Western medicine model)

[0193] Technical meaning: Derived from game theory, it measures the contribution of a single model in multi-model collaborative decision-making (ranging from 0 to 1, with higher values indicating a greater impact on the final diagnostic result).

[0194] Calculation logic: For the Western medicine model, a Shapley value of 0.35 means that in the diagnosis of "acute exacerbation of chronic gastritis", the decision contribution of the model accounts for 35% of the total decision (for example, the proportion of support for the final conclusion of the gastroscopic image features it outputs).

[0195] In specific implementation, the comprehensive merit score = w1*feature contribution + w2*Shapley value (w1+w2=1) is used; among them, w1 (such as 0.6): the traditional Chinese medicine model relies more on feature extraction capabilities (unstructured data such as tongue image and pulse diagnosis require strong feature screening), so it is given a higher weight;

[0196] w2 (e.g., 0.4): The Western medicine model focuses on multimodal decision fusion (imaging, tests, and medical records require logical integration), and the Shapley value can better reflect its collaborative value.

[0197] Through this quantification mechanism, the system achieves a self-evolutionary closed loop of "continuously strengthening superior models and targeted optimization of inferior models." This is a key technical support for upgrading the integrated Chinese and Western medicine diagnosis system from a "static model" to an "intelligent agent." Model grouping: Traditional Chinese Medicine models with a merit score greater than 0.7 enter the expert group, responsible for reasoning complex cases (such as chronic gastritis with gastric ulcers); Western Medicine models with a merit score of 0.65 enter the novice group, learning the attention matrix of the expert group through knowledge distillation.

[0198] In specific implementation, when the student model (novice group) is trained, it not only learns the true label (such as "acute exacerbation of chronic gastritis"), but also uses the cross-entropy loss function as follows:

[0199]

[0200] Fit the soft labels of the teacher model (expert group). Among them, p teacher (c) is the soft probability of the teacher model (including implicit associations between syndromes, such as the similarity between “liver and stomach stagnation and heat syndrome” and “damp-heat syndrome”), which helps the student model capture complex dialectical logic.

[0201] Final loss function:

[0202] here is the knowledge distillation loss (learning the knowledge of the teacher model), is the cross entropy loss (learning the true label), λ = 0.5 means giving By balancing the contributions of the two, the model can absorb the knowledge of the teacher model and fit the real data distribution during training.

[0203] Knowledge base update: User feedback "symptoms relieved after 2 weeks of treatment" was marked as a "victory case", and the call priority of the corresponding "liver and stomach heat syndrome-chronic gastritis" related knowledge was increased (updated the weight of the fourth-level knowledge base).

[0204] In summary, this content achieves capability stratification and collaboration through model grouping, uses knowledge distillation and combined loss functions to promote knowledge transfer between models, and combines the knowledge base update mechanism to form a closed loop of "learning-optimization-application", thereby improving the system's diagnostic capabilities and adaptability.

[0205] The Western medicine model enters the novice group due to its merit score of 0.65, and learns the attention matrix of the expert group through knowledge distillation. Furthermore, the model uses cross-modal feature fusion technology to process Chinese and Western medicine data, dynamic spatiotemporal

[0206] The dynamic spatiotemporal attention mechanism calculates the bidirectional attention weights as follows:

[0207]

[0208] Fusion Features Strengthen the semantic association between tongue texture and gastric mucosal lesions. This allows the novice model to more effectively learn the cross-modal association knowledge from the expert model, improving diagnostic capabilities.

[0209] 2. Merit Quantification and Grouping:

[0210] Based on the historical 30-day data, the model's average merit score Pi = 0.68, standard deviation σi = 0.09, threshold Ti = Pi + 1.5σi = 0.815, and the dynamic baseline method is used to avoid manual parameter adjustment deviation.

[0211] This example verifies the following technical points through real cases:

[0212] Multimodal preprocessing: tongue image registration error <5%, CNN image feature extraction accuracy >92%;

[0213] Collaborative diagnosis effect: The misdiagnosis rate of a single model was 15%, which was reduced to 7% after collaboration, and the proportion of cases with an evidence chain matching degree greater than 0.8 increased by 40%;

[0214] Privacy protection performance: After differential privacy processing, the model accuracy drops by less than 1.5%, which complies with the medical data security standard (NIST SP 800-53).

[0215] Through the above implementation method, the system realizes the closed loop of "data collection-inference diagnosis-supervision protection-feedback optimization", which is suitable for the diagnosis of complex cases combining traditional Chinese and Western medicine, especially in scenarios such as chronic gastritis and diabetic complications, showing high diagnostic accuracy and self-evolution capabilities.

[0216] It can be understood that the same or similar parts of the above embodiments can be referenced to each other, and the contents not described in detail in some embodiments can refer to the same or similar contents in other embodiments.

[0217] It should be noted that, in the description of the present invention, the terms "first", "second", etc. are used for descriptive purposes only and should not be understood as indicating or implying relative importance. In addition, in the description of the present invention, unless otherwise specified, the meaning of "plurality" is at least two.

[0218] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may be performed out of the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention pertain.

[0219] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0220] Those skilled in the art will understand that all or part of the steps in the method of the above embodiment can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiment.

[0221] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing module, or each unit may exist physically separately, or two or more units may be integrated into a single module. The aforementioned integrated modules may be implemented in the form of hardware or in the form of software functional modules. If the integrated modules are implemented in the form of software functional modules and sold or used as independent products, they may also be stored in a computer-readable storage medium.

[0222] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.

[0223] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0224] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. An intelligent diagnosis and dynamic optimization method combining traditional Chinese and Western medicine, characterized in that: The method comprises: Collecting multimodal data; the multimodal data includes: traditional Chinese medicine data and Western medicine data; The TCM data includes: tongue image, symptom text, and pulse diagnosis signal; The Western medicine data include: electronic medical records, medical images, and test reports; Preprocessing the multimodal data, inputting the preprocessed TCM data into a preset TCM syndrome differentiation model, and outputting TCM syndrome classification; Inputting the pre-processed Western medicine data into a preset Western medicine disease model and outputting a disease prediction result; Based on the TCM syndrome classification and disease prediction results, determine whether the confidence level of the single model meets the standards; If the criteria are met, the user is assigned to a supervision model group based on the user's risk level according to the TCM syndrome classification and disease prediction results; If the standards are not met, the collaborative diagnosis mechanism will be activated to obtain a comprehensive decision-making model vote. After the results are integrated based on the comprehensive decision-making model vote, the supervision model group will be assigned according to the user's risk level; Using a preset double verification mechanism to verify the supervision model group allocated according to user risk level, if the verification is passed, a structured diagnosis report is generated; Otherwise, the conflicting suggestion is automatically rejected; Using a preset federated learning framework to de-identify the terminal layer data of the structured diagnostic report, generate encrypted feature vectors, and homomorphically encrypt the aggregation parameters of the central layer of the structured diagnostic report through secure multi-party computation, and add noise in combination with differential privacy to obtain compliance diagnosis results; User feedback is collected for the obtained compliance diagnosis results, and the knowledge call priority is dynamically adjusted according to the user feedback to obtain the call results.

2. The method according to claim 1, characterized in that The preprocessing of the multimodal data includes: unifying the perspective of the acquired tongue image through image registration, and extracting texture features of the tongue image; Encoding the symptom text using a preset TCM terminology library code; Convert the preset pulse diagnosis signal into time-frequency domain features; Extracting the features of the lesion area of the medical image through a preset convolutional neural network; Parsing the inspection report into a numerical vector; The electronic medical records are structured using NLP technology.

3. The method according to claim 1, characterized in that The determination of whether the confidence level of a single model meets the standards based on the TCM syndrome classification and disease prediction results includes: According to the TCM syndrome classification and disease prediction results, confidence thresholds are set for the TCM syndrome differentiation model and the Western medicine disease model respectively; Calculate the confidence of the TCM syndrome classification output by the TCM syndrome differentiation model, and the confidence of the disease prediction results output by the Western medicine disease model; If the confidence of the TCM syndrome differentiation model is greater than or equal to its corresponding confidence threshold, and the confidence of the Western medicine disease model is greater than or equal to its corresponding confidence threshold, then the confidence of the single model is judged to meet the standard; otherwise, the confidence of the single model is judged to fail to meet the standard.

4. The method according to claim 1, wherein The collaborative diagnosis mechanism is initiated to obtain a comprehensive decision-making model vote, including: Extract the evidence related to Chinese and Western medicine related to the current case from the historical medical record database and build an evidence chain; Organize a small model group of multiple experts to generate pro and con arguments for the current case and conduct adversarial debates; The comprehensive decision-making model is used to vote on the results after the adversarial debate to obtain the voting results.

5. The method according to claim 1, wherein The allocation of supervision model groups according to user risk levels specifically includes: Determine the user's risk level based on the user's age, disease severity, and past medical history; preset corresponding supervision model groups for different risk levels; According to the determined user risk level, the corresponding supervision model group is assigned to supervise the diagnosis results.

6. The method according to claim 1, characterized in that The preset double verification mechanism includes: Rule engine verification: Using the preset rule engine, the diagnostic results after the supervision model group is assigned according to the user's risk level are logically verified to check for logical errors such as medication contraindications and diagnostic contradictions; Semantic similarity verification: The semantic similarity between diagnostic recommendations and taboo rules is calculated using the twin-tower model. If the semantic similarity exceeds a preset threshold, a conflict is determined. Only when both the rule engine verification and the semantic similarity verification pass, can it be considered to have passed the double verification.

7. The method according to claim 1, characterized in that The method of using a preset federated learning framework to de-identify the terminal layer data of the structured diagnostic report and generate an encrypted feature vector specifically includes: De-identify the structured diagnostic report data at the terminal layer to remove any information that can directly or indirectly identify the user; The de-identified data is encrypted using an encryption algorithm to generate an encrypted feature vector.

8. The method according to claim 1, characterized in that The central layer of the structured diagnostic report is homomorphically encrypted with aggregated parameters through secure multi-party computation, and noise is added in combination with differential privacy, including: At the center layer, secure multi-party computation is used to perform homomorphic encryption aggregation on the encrypted feature vectors uploaded by each terminal layer to obtain the aggregation parameters; Noise that meets the requirements of differential privacy is added to the aggregation parameters to protect the privacy of the data.

9. The method according to claim 1, characterized in that The dynamically adjusting the knowledge call priority according to user feedback and obtaining the call result includes: Establish a knowledge call priority database to record the call priority of different knowledge; Dynamically adjust the knowledge call priority in the knowledge call priority database according to user feedback on the diagnosis results; Perform knowledge call according to the adjusted knowledge call priority and obtain the call result.

10. The method according to claim 1, characterized in that The method further includes: calculating the merit points of the TCM syndrome differentiation model and the Western medicine disease model based on user feedback and the accuracy of the diagnosis results; The models are divided into expert and novice groups according to their merit points; The expert group's model was trained with more complex case data to improve its diagnostic capabilities in complex scenarios. For the model of the novice group, a knowledge distillation method is used to learn knowledge from the model of the expert group to quickly improve its diagnostic ability; Regularly update the model's merit points, dynamically adjust the group to which the model belongs, and achieve dynamic optimization of the model.

Images