A method for reviewing and correcting traditional Chinese medicine AI diagnosis results based on human-computer collaborative decision-making

Abstract

The application discloses a Chinese medicine AI diagnosis result review and correction method based on man-machine collaborative decision-making, belongs to the technical field of Chinese medicine auxiliary diagnosis, and constructs a multi-modal four-examination data hierarchical fusion framework and a time sequence cause-effect syndrome differentiation reasoning engine to generate an interpretable syndrome differentiation result containing a syndrome type evolution process; a syndrome differentiation confidence-surgeon experience matching degree two-dimensional dynamic weight distribution model is introduced to realize deep collaborative decision-making of AI and doctors; a fine-grained conflict detection and correction guiding mechanism based on syndrome element orthogonal decomposition is proposed; and a correction strength-doctor grade two-factor weighted incremental knowledge updating closed loop is established. The Chinese medicine AI diagnosis result review and correction method based on man-machine collaborative decision-making can realize deep collaboration of AI and doctors in the whole diagnosis process, and significantly improve the accuracy, interpretability and clinical practicability of Chinese medicine AI diagnosis.

Description

Technical Field

[0001] This invention relates to the field of TCM auxiliary diagnostic technology, and in particular to a method for reviewing and correcting TCM AI diagnostic results based on human-machine collaborative decision-making. Background Technology

[0002] With the rapid development of artificial intelligence technology, TCM AI diagnostic systems are being used more and more widely in clinical practice. These systems learn from a large amount of TCM clinical case data, and can automatically analyze patients' four diagnostic methods and provide syndrome differentiation results, which improves the efficiency and standardization of TCM diagnosis to a certain extent.

[0003] However, existing TCM AI diagnostic systems still have the following key problems: (1) Strong black box nature, and the dialectical logic is untraceable: Most existing systems use deep learning models for end-to-end training, which can only give the final dialectical result and cannot explain its reasoning process. Physicians cannot understand why the AI ​​comes to such a conclusion, and it is difficult to effectively evaluate and verify its results, which seriously affects clinicians' trust in the AI ​​system.

[0004] (2) Ignoring the dynamic evolution of syndrome types: Most existing causal reasoning models are based on static knowledge graphs, focusing only on the patient's current symptoms and ignoring the dynamic evolution of TCM syndrome types and the impact of the temporal sequence of symptom appearance on the diagnosis results. The core of TCM syndrome differentiation and treatment is to grasp the development and change trend of the disease. This deficiency leads to a significant decrease in the accuracy of existing systems when dealing with diseases with long course and complex and changeable syndrome types.

[0005] (3) Human-machine collaboration is merely a formality: Most existing "AI + physician" models are sequential, with AI providing the results first and physicians then reviewing them, rather than true collaborative decision-making. The AI ​​system does not fully utilize physicians' clinical experience, and physicians do not receive effective support from the AI ​​system, resulting in the AI ​​system's advantages not being fully realized.

[0006] (4) Low accuracy in diagnosing complex diseases: Traditional Chinese medicine emphasizes individualization and holistic treatment. For complex and difficult diseases, a single AI model often fails to make accurate diagnoses. Existing systems perform reasonably well when dealing with simple and common diseases, but their accuracy drops significantly when faced with complex diseases. Summary of the Invention

[0007] The purpose of this invention is to provide a method for reviewing and correcting TCM AI diagnostic results based on human-machine collaborative decision-making. This method can achieve deep collaboration between AI and TCM doctors throughout the entire diagnostic process, significantly improving the accuracy, interpretability, and clinical applicability of TCM AI diagnosis.

[0008] To achieve the above objectives, this invention provides a method for reviewing and correcting TCM AI diagnostic results based on human-machine collaborative decision-making, comprising the following steps: S1. Multimodal four diagnostic methods data acquisition and standardized preprocessing; S2. Construct a hierarchical and integrated TCM feature extraction network to perform multi-scale feature extraction and fusion on the standardized four diagnostic methods data to generate a fusion feature vector of the four diagnostic methods. S3. Based on the TCM knowledge graph and the temporal causal reasoning engine, generate preliminary AI diagnosis results, and output the diagnosis reasoning path including the evolution of the syndrome and the confidence level of each diagnosis element. S4. Construct a dynamic weight allocation model with two dimensions: diagnostic confidence and doctor experience matching. The decision weights of AI and doctor are automatically allocated based on the confidence of the AI ​​diagnostic results and the experience level of the doctor currently treating the patient. S5. Human-machine collaborative diagnosis and decision-making: AI shows physicians the diagnosis and reasoning path, key evidence support and similar case references. At the same time, it performs fine-grained conflict detection and correction guidance based on the orthogonal decomposition of diagnosis elements. Physicians adjust the AI ​​results based on clinical experience, and the system calculates the final diagnosis result according to dynamic weights. S6. Define a deep verification mechanism for conflict detection. When there is a significant conflict between the AI ​​and the doctor's diagnosis, the multi-expert consultation process will be automatically triggered and a conflict analysis report will be generated. S7. Incremental knowledge updates and model self-optimization: The correction results of physicians and the conclusions of expert consultations are evaluated and weighted, and used as new training samples to perform weighted incremental learning on the AI ​​model, continuously improving the accuracy of diagnosis.

[0009] Preferably, S1 specifically includes the following steps: S11. Data collection for visual diagnosis includes tongue images, facial images, and body posture images. Standardized lighting conditions and shooting angles are used, and quantitative features of the tongue body, tongue coating, and facial color are extracted through image segmentation algorithms. S12. Auscultation data acquisition includes speech signals and breath sound signals. Speech features are extracted using Mel frequency cepstral coefficients, and breath sound features are extracted using wavelet transform. S13. The data collection of consultations adopts a structured questionnaire combined with natural language processing technology to convert patients' chief complaints and medical history descriptions into standardized TCM symptom terms. S14. Palpation data acquisition uses a digital pulse diagnostic instrument to obtain pulse wave signals and extract the frequency, rhythm, intensity, and morphological quantitative characteristics of the pulse. S15. Map all the data from the four diagnostic methods to a unified TCM symptom ontology database to generate a standardized feature matrix of the four diagnostic methods.

[0010] Preferably, in S2, the hierarchical fusion TCM feature extraction network specifically includes: The bottom feature extraction layer uses a convolutional neural network to process the raw data of inspection, auscultation and palpation, and a bidirectional long short-term memory network to process the consultation text data. The middle-layer feature fusion layer uses an attention mechanism to weightedly fuse features from different modalities to generate a single-modality feature vector; The high-level semantic fusion layer constructs a semantic fusion network based on the Five Elements theory and the organ differentiation system of traditional Chinese medicine, mapping the single-modal feature vector to the semantic space of traditional Chinese medicine differentiation, and generating the four diagnostic features vectors.

[0011] Preferably, in S3, the process of generating preliminary AI diagnostic results based on the TCM knowledge graph and temporal causal reasoning engine specifically includes the following steps: S31. Construct a TCM knowledge graph containing at least 2 million combinations of TCM symptoms and tongue and pulse signs, 4,000 syndrome types and corresponding TCM prescriptions, and define the causal and evolutionary transfer relationships between symptoms-syndromes, syndrome types-diseases, syndrome types-prescriptions, and syndrome types. S32. Input the four diagnostic and treatment fusion feature vectors and the patient's historical diagnosis and treatment time series data into the causal reasoning engine. First, use the temporal causal graph convolutional network to extract the dynamic evolution features of symptoms and syndrome types, calculate the evolution and transition probability of each syndrome type, and then combine the Bayesian network to calculate the comprehensive posterior probability of each syndrome type at the current time. S33. Generate a dialectical reasoning path that includes the evolution of syndrome types, and show the complete reasoning process from the appearance of symptoms to the current syndrome type, including the contribution of each symptom to each syndrome type at different time points and the transition probability of syndrome type evolution. S34. Calculate the overall confidence level of the dialectical results and the individual confidence levels of each dialectical element.

[0012] Preferably, in S4, the constructed dual-dimensional dynamic weight allocation model of dialectical confidence and physician experience matching specifically includes: The confidence level (C) of the AI-generated identification results is divided into three levels: high confidence, medium confidence, and low confidence. Based on the physician's professional title, years of practice, specialty, and historical diagnostic accuracy, the physician's experience level E is calculated and classified into junior physicians, intermediate physicians, and senior physicians. The decision weight W of the AI ​​is calculated using the following formula. AI The decision weight of the physician W doctor : W AI =α×C+β×(1-E); W doctor =1-W AI ; Where α and β are adjustment coefficients, α+β=1, and the optimal values ​​are determined through training with clinical data; When the AI's assessment result is low confidence and the physician is a junior physician, the process of review by a senior physician is automatically triggered.

[0013] Preferably, step S5 specifically includes the following steps: S51. AI presents physicians with a visualized diagnostic reasoning path diagram, marking the top 5 symptoms with the highest support and the 3 most likely syndrome types, while also displaying a prediction of the patient's syndrome type evolution trend. S52. AI automatically retrieves historical cases with similar symptoms and disease progression to the current patient, and displays the diagnosis results, treatment plans and prognoses of similar cases; S53. AI detects fine-grained conflicts between physician adjustments and AI preliminary results in real time based on orthogonal decomposition of dialectical elements, and pushes targeted evidence and correction guidance for dialectical elements with significant conflicts. S54. Physicians modify the AI's diagnostic results based on the actual diagnosis, including adjusting the syndrome type, adding or deleting diagnostic elements, and correcting the evolution trend of the syndrome type. S55. The AI ​​calculates the final diagnosis result based on dynamic weights. When there is a difference between the doctor's adjustment and the AI's suggestion, the difference is automatically recorded and an adjustment explanation is generated. S56. Generate a complete diagnostic report that includes preliminary AI results, physician adjustments, final diagnosis, and prognostic recommendations.

[0014] Preferably, in S53, the AI's preliminary diagnostic results and the physician's real-time adjustments are decomposed into three independent orthogonal diagnostic elements: location of disease, nature of disease, and disease progression. The degree of conflict between the AI ​​and the physician on each element is calculated. For diagnostic elements with a degree of conflict exceeding a threshold, key supporting evidence, counter-evidence cases, and authoritative expert consensus corresponding to that element are automatically pushed to the physician to guide the physician to make precise element-level corrections rather than overall syndrome adjustments.

[0015] Preferably, the orthogonal decomposition of diagnostic elements and the calculation of conflict degree specifically include: constructing three independent semantic spaces for disease location, disease nature, and disease progression; mapping all syndrome names and symptom terms to the three semantic spaces based on the TCM symptom ontology and knowledge graph, obtaining the normalized vector representation of each syndrome and symptom in the three spaces; calculating the cosine distance between the AI ​​diagnostic result vector and the physician adjustment content vector in each semantic space as the conflict degree of the element; when the conflict degree of an element is greater than 0.6, it is determined that the element has a significant conflict, and the targeted correction guidance process is automatically triggered.

[0016] Preferably, step S6 specifically includes the following steps: S61. Define the conflict threshold: when there are significant differences between the diagnosis results of AI and physician in terms of disease location, disease nature or disease severity, and the confidence level of both is higher than 0.7, it is judged as a significant conflict. S62. Automatically generate conflict analysis reports, providing detailed comparisons of the reasoning paths, evidence, syndrome evolution analysis, and similar case support between AI and physicians; S63. Push conflict cases to at least two senior physicians for consultation. The system automatically summarizes the consultation opinions and calculates the degree of consensus. S64. The consultation conclusion shall be used as the final diagnosis, and the case shall be marked as a difficult case and stored in the knowledge base.

[0017] Preferably, S7 specifically includes the following steps: S71. Establish a case quality assessment model to evaluate the quality of physicians' correction results and expert consultation conclusions, and select high-quality training samples. Calculate the learning weight of each training sample based on the two factors of correction intensity and physician level. The correction intensity is quantified according to the degree of difference between the AI ​​and the physician's diagnosis results, and the physician level weight is determined according to the physician's experience level. The correction sample weight of senior physicians is 2-3 times that of junior physicians. S72. The AI ​​model is updated using a weighted incremental learning algorithm. The parameter update magnitude is adjusted according to the sample learning weight. High-value correction samples are learned first, and model parameters are adjusted only using new high-quality samples to avoid catastrophic forgetting. S73. Regularly update the TCM knowledge graph, incorporating newly discovered symptom-syndrome relationships, syndrome evolution patterns, and clinical experience into the knowledge graph; S74. Generate a model optimization report to show the changes in diagnostic accuracy for different syndrome types and the model convergence speed.

[0018] Therefore, the beneficial effects of the above-mentioned method for reviewing and correcting TCM AI diagnostic results based on human-machine collaborative decision-making in this invention are as follows: (1) Combination of high interpretability and dynamism: Based on the TCM knowledge graph and the temporal causal reasoning engine, this invention constructs a syndrome differentiation model, which can not only generate a complete syndrome differentiation reasoning path and clearly show the reasoning process from symptoms to syndrome type, but also capture the dynamic evolution law of syndrome type, show the patient's syndrome type change process from onset to the present and future evolution trend, solve the black box problem and static reasoning defects of traditional AI system, and significantly improve clinicians' trust in AI diagnosis results.

[0019] (2) Deep human-machine collaboration: This invention innovatively proposes a two-dimensional dynamic weight allocation model of "diagnosis confidence degree - doctor's experience matching degree", which realizes deep collaboration between AI and doctors in the entire diagnostic process. The AI ​​system can automatically adjust the decision weight according to its own confidence degree and the doctor's experience level, giving full play to their respective advantages.

[0020] (3) High accuracy in diagnosing complex diseases: This invention significantly improves the diagnostic accuracy of complex diseases and difficult diseases by hierarchical fusion of multimodal four diagnostic data, temporal causal reasoning and identification of combined diseases, combined with a multi-expert consultation mechanism.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 This is a flowchart of an embodiment of a method for reviewing and correcting TCM AI diagnostic results based on human-machine collaborative decision-making according to the present invention. Detailed Implementation

[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0025] Example 1: like Figure 1 As shown, this embodiment provides a method for reviewing and correcting TCM AI diagnostic results based on human-machine collaborative decision-making, including the following steps: S1. Multimodal Four Diagnostic Methods Data Acquisition and Standardized Preprocessing: Data collection for visual diagnosis: Tongue images of patients were collected using a standardized tongue imager. The tongue body and tongue coating regions were extracted using the Deeplabv3+ image segmentation algorithm, and 16-dimensional quantitative features such as tongue color (pale white, light red, red, crimson, purple), tongue shape (large, thin, teeth marks, cracks), coating color (white, yellow, gray, black), and coating texture (thin, thick, greasy, dry) were calculated. Facial images of patients were collected using facial image acquisition equipment, and 6-dimensional quantitative features such as facial color (pale, sallow, reddish, bluish-purple, dark) and luster (glossy, dull) were extracted.

[0026] Auscultation data acquisition: The patient's speech and respiratory sound signals were acquired using a microphone at a sampling rate of 44.1 kHz. The speech signal was analyzed by extracting 13-dimensional Mel-frequency cepstral coefficients (MFCCs) and their first and second-order differences, resulting in 39 features. The respiratory sound signal was analyzed by performing a 5-level wavelet decomposition, extracting the energy and entropy values ​​of each level, resulting in 20 features.

[0027] Consultation data collection: A structured electronic medical record system, combined with a BERT pre-trained model, is used to perform natural language processing on patients' chief complaints and medical history descriptions, converting unstructured text into standardized TCM symptom terms. The TCM symptom ontology used in this system contains more than 120,000 TCM symptom terms, covering all common symptoms in "Traditional Chinese Medicine Diagnostics".

[0028] Palpation data acquisition: A digital pulse diagnostic instrument was used to collect pulse wave signals from the six positions (cun, guan, chi) on both hands of the patient, with a sampling rate of 1000Hz. 24-dimensional quantitative characteristics of the pulse, including frequency, rhythm, intensity, and morphology, were extracted, encompassing quantitative indicators of common pulse types such as superficial, deep, slow, rapid, weak, strong, slippery, and hesitant.

[0029] Data standardization: Map all diagnostic data to a unified TCM symptom ontology database to generate a standardized four diagnostic feature matrix with dimensions N×D, where N is the number of symptoms and D is the quantitative feature dimension of each symptom.

[0030] S2. Layered fusion of TCM feature extraction: Constructing a three-layer fusion network for extracting TCM features: The bottom feature extraction layer uses a ResNet50 convolutional neural network to process visual image data and outputs a 2048-dimensional image feature vector; it uses a one-dimensional convolutional neural network to process the time-series data of auscultation and palpation and outputs 512-dimensional and 512-dimensional feature vectors respectively; and it uses a BiLSTM network to process the text data of medical consultation and outputs a 1024-dimensional text feature vector.

[0031] The middle layer feature fusion layer uses a multi-head attention mechanism to weightedly fuse features from different modalities. The attention mechanism can automatically learn the importance of different modal features to the diagnostic results. For example, for patients with liver disease, tongue features in inspection and pulse features in palpation may have higher weights. After fusion, a 4096-dimensional single-modal feature vector is generated.

[0032] High-level semantic fusion layer: Based on the Five Elements theory and the organ differentiation system of traditional Chinese medicine, a semantic fusion network is constructed. This network maps the single-modal feature vector to the semantic space of traditional Chinese medicine differentiation, takes into account the generating and restraining relationships between organs and the correlation between symptoms, and finally generates a 2048-dimensional fusion feature vector of the four diagnostic methods.

[0033] S3. Explainable temporal causal dialectical reasoning: Constructing a Traditional Chinese Medicine (TCM) Knowledge Graph: This system constructs a TCM knowledge graph containing 2 million combinations of TCM symptoms with tongue and pulse signs, 4,000 syndrome types, and corresponding TCM formulas, as well as the causal and evolutionary relationships between them. Relationships within the knowledge graph include symptom-syndrome (e.g., a red tongue with little coating leading to Yin deficiency syndrome), syndrome-disease (e.g., liver and kidney Yin deficiency syndrome leading to hypertension), syndrome-formula (e.g., liver and kidney Yin deficiency syndrome corresponding to Liuwei Dihuang Wan), and syndrome-syndrome (e.g., wind-cold common cold syndrome can evolve into wind-heat common cold syndrome or phlegm-dampness accumulation in the lungs syndrome), etc.

[0034] Causal reasoning: The fusion feature vector of the four diagnostic methods and the patient's historical treatment time-series data over the past 6 months are input into the causal reasoning engine. First, a temporal causal graph convolutional network is used to extract the dynamic evolution features of symptoms and syndrome types, and the probability of the patient's transition from wind-cold common cold syndrome to phlegm-dampness accumulation in the lungs syndrome is calculated to be 0.87. Then, a Bayesian network is used to calculate the comprehensive posterior probability of each syndrome type at the current time, where the posterior probability of phlegm-dampness accumulation in the lungs syndrome is 0.89, and the posterior probability of lung yin deficiency syndrome is 0.58.

[0035] Reasoning Path Generation: The system generates a visualized dialectical reasoning path diagram, showing the complete reasoning process from the onset of symptoms to the current syndrome type. For example, the reasoning path diagram shows that the patient developed symptoms of chills and fever, nasal congestion and runny nose 3 months ago (contribution of 0.85 to the wind-cold common cold syndrome), and then gradually developed symptoms of cough with phlegm, chest tightness and epigastric fullness (contribution of 0.82 to the phlegm-dampness accumulation in the lungs syndrome), accompanied by symptoms of "white and greasy tongue coating and slippery pulse" (contribution of 0.78 to the phlegm-dampness accumulation in the lungs syndrome), with a syndrome type evolution and transition probability of 0.87.

[0036] Confidence Calculation: The system calculates the overall confidence level of the diagnostic results and the individual confidence levels of each diagnostic element. For example, the overall confidence level of the phlegm-dampness accumulation in the lung syndrome is 0.89, the confidence level of the location of the disease (lung) is 0.92, the confidence level of the nature of the disease (phlegm-dampness) is 0.88, and the confidence level of the disease progression (chronic protracted stage) is 0.85.

[0037] S4. Dynamic weight allocation: AI confidence assessment: The confidence level (C) of the AI's analytical results is divided into three levels: high confidence (C≥0.85), medium confidence (0.6≤C<0.85), and low confidence (C<0.6). In this embodiment, the confidence level of the AI's analytical results is 0.89, which falls under the high confidence level.

[0038] Physician experience assessment: The physician's experience level E is calculated based on their professional title, years of practice, specialty, and historical diagnostic accuracy rate. For example, a TCM internal medicine attending physician with 10 years of experience has a historical diagnostic accuracy rate of 0.88 and an experience level E of 0.72, classifying them as a senior physician.

[0039] Weight calculation: According to formula W AIThe decision weights of AI are calculated as α×C+β×(1-E). In this embodiment, α=0.7, β=0.3, therefore W AI =0.7×0.89+0.3×(1-0.72)=0.623+0.084= 0.707, the physician's decision weight W doctor =1-0.707=0.293.

[0040] S5. Human-machine collaborative dialectical decision-making and fine-grained correction guidance: The AI ​​system displays to physicians: the patient's standardized four diagnostic methods data and raw data, a visualized path of syndrome differentiation and reasoning and the evolution of syndrome types, the top 5 symptoms with the highest support and their contribution, the 3 most likely syndrome types and their probabilities, 10 historical cases with similar symptoms and disease progression to the current patient and their diagnostic results, treatment plans and prognoses.

[0041] Fine-grained conflict detection and correction guidance: When the physician begins to input adjustments, the system automatically decomposes the preliminary AI results and the physician's real-time input into three orthogonal elements: location, nature, and progression. In this embodiment, the location vector of the preliminary AI results is [0.92, 0.05, 0.03] (lung, spleen, others), the nature vector is [0.88, 0.07, 0.05] (phlegm-dampness, spleen deficiency, others), and the progression vector is [0.85, 0.12, 0.03] (chronic persistent phase, acute exacerbation phase, recovery phase); the location vector corresponding to the physician's input of phlegm-dampness accumulating in the lungs and spleen deficiency is [0.65, 0.32, 0.03], the nature vector is [0.58, 0.37, 0.05], and the progression vector is consistent with the AI. The system calculated that the cosine distance of the location element was 0.52 and the cosine distance of the nature element was 0.48, neither of which reached the significant conflict threshold of 0.6. However, the system still proactively pushed supporting evidence (contribution of 0.75) for the symptoms of loss of appetite and loose stools to the physician regarding the nature of spleen deficiency, as well as three authoritative cases of similar phlegm-dampness accumulation in the lungs combined with spleen deficiency, to help the physician confirm the rationality of the correction.

[0042] The physician reviewed the AI's diagnostic results, reasoning process, and correction guidance information, and combined this with their own clinical experience to evaluate and confirm that adding the diagnostic element of concurrent spleen deficiency was correct.

[0043] The system calculates the final diagnosis based on dynamic weights: the AI-suggested weight for phlegm-dampness accumulation in the lungs is 0.707, while the physician-suggested weight for phlegm-dampness accumulation in the lungs combined with spleen deficiency is 0.293. After comprehensive consideration, the system's final diagnosis is phlegm-dampness accumulation in the lungs combined with spleen deficiency.

[0044] The system generates a complete diagnostic report, including preliminary AI results, physician adjustments, final diagnosis, recommended treatment plan, and prognostic suggestions.

[0045] S6. Conflict Detection and Deep Review: In this embodiment, there is no significant conflict between the AI's diagnosis and the physician's diagnosis, so there is no need to trigger a multi-expert consultation process. If the AI's diagnosis is Lung Yin Deficiency Syndrome (confidence level 0.88), while the physician's diagnosis is Phlegm-Dampness Accumulation in the Lung Syndrome (confidence level 0.92), the system will determine it as a significant conflict, automatically generate a conflict analysis report, and compare in detail the reasoning paths, evidence, syndrome evolution analysis, and similar case support of the two, and then push the case to two senior physicians for consultation.

[0046] S7, Weighted Incremental Model Update: Case quality assessment: The system assesses the quality of the physician's corrections. In this embodiment, the physician's adjustments are reasonable, so the case is marked as a high-quality training sample.

[0047] Sample weight calculation: The sample learning weight is calculated based on a two-factor model of correction intensity and physician level. In this embodiment, the physician is a senior physician (level weight 2.2), and the correction intensity is 0.31 (the degree of difference between the AI ​​and the physician's diagnosis results). Therefore, the sample learning weight is 2.2 × 0.31 + 1.0 = 1.682, which is higher than the average weight of 1.0 for ordinary samples.

[0048] Weighted Incremental Learning: An improved version of the Elastic Weights Consolidation (EWC) algorithm is used for weighted incremental learning, adjusting the parameter update magnitude based on the sample learning weights. The update magnitude of the model parameters for this sample is 1.682 times that of ordinary samples, enabling the model to quickly learn the key points of diagnosis for phlegm-dampness accumulation in the lungs complicated by spleen deficiency.

[0049] Knowledge graph update: The contribution of poor appetite and loose stools to the syndrome of phlegm-dampness accumulation in the lungs and spleen deficiency was adjusted from 0.62 to 0.75. At the same time, the evolution and transfer probability of phlegm-dampness accumulation in the lungs to the syndrome of phlegm-dampness accumulation in the lungs and spleen deficiency was updated.

[0050] Model Optimization Report: The system generates a monthly model optimization report, showing that the diagnostic accuracy of phlegm-dampness accumulation in the lungs syndrome has increased from 85.3% to 89.7%, the identification accuracy of spleen deficiency syndrome has increased from 72.6% to 81.2%, and the model convergence speed has increased by 42% compared to the previous month.

[0051] Example 2: Based on the method of this invention, this embodiment is applied in a primary healthcare institution. A junior physician (experience level E=0.35) at this institution treated a patient who complained of "recurrent cough for 3 months." After analyzing the patient's four diagnostic methods data and combining it with the patient's historical medical records, the AI ​​used a temporal causal graph convolutional network to find that the patient initially had a wind-cold type of common cold, which gradually progressed to internal heat. The current diagnosis is phlegm-heat stagnation in the lungs, with a confidence level C=0.58, which is considered low confidence.

[0052] According to the dynamic weight allocation model, the decision weight W of AI is... AI =0.7×0.58+0.3×(1-0.35)=0.406+0.195=0.601, the physician's decision weight W doctor =0.399. Meanwhile, because the AI's assessment result was of low confidence and the physician was a junior physician, the system automatically triggered a review process by a senior physician.

[0053] The system demonstrated the AI's diagnostic reasoning path, syndrome evolution process, and similar case references to the junior physician. After reviewing the information, the junior physician determined that the patient's symptoms were more consistent with the phlegm-dampness accumulating in the lungs syndrome and began inputting adjustments. The system immediately performed fine-grained conflict detection, calculating the conflict level of the pathological elements to be 0.72 (exceeding the threshold of 0.6). It automatically pushed to the junior physician the core differentiating points between phlegm-heat stagnation in the lungs syndrome and phlegm-dampness accumulating in the lungs syndrome in terms of tongue coating and pulse, as well as a comparison of 5 easily confused similar cases. After reviewing the guidance information, the junior physician further confirmed their judgment and added the key symptom of a white, greasy tongue coating as a basis for correction.

[0054] The system forwarded the case to a senior physician (experience level E=0.85) for review. After reviewing the patient's four diagnostic methods (inspection, auscultation and olfaction), the AI's reasoning process, the junior physician's diagnostic opinion, and the basis for correction, the senior physician confirmed the patient's syndrome differentiation as phlegm-dampness accumulation in the lungs. The system used the senior physician's diagnosis as the final diagnosis and calculated the learning weight for this sample: senior physician level weight 2.5, correction strength 0.65, therefore the sample learning weight is 2.5 × 0.65 + 1.0 = 2.625. After learning from this high-value sample, the model's diagnostic accuracy for syndromes easily confused with wind-cold transforming into phlegm-dampness (wind-cold transforming into phlegm-heat) increased from 76.8% to 84.5%.

[0055] Through the application of this invention, primary care physicians in grassroots medical institutions can obtain precise correction guidance from AI and dual support from senior physicians, which significantly improves the accuracy of diagnosis and also promotes the professional growth of primary care physicians.

[0056] To verify the effectiveness of this invention, we conducted a 6-month clinical trial at 3 top-tier tertiary hospitals and 2 primary healthcare institutions. A total of 1200 patients were included, of whom 800 had common diseases and 400 had complex diseases. The results are shown in Table 1. Table 1 Comparison of parameters between existing technology and the method of this invention

[0057] Further analysis shows that the introduction of temporal causal graph convolutional networks improved the diagnostic accuracy of complex diseases by 8.2%, especially for patients with a disease course of more than one month and whose syndrome patterns have evolved, the improvement in diagnostic accuracy was more significant; the two-factor weighted incremental learning of correction intensity and physician level improved the model convergence speed by 45% and effectively avoided the performance degradation of the model on common diseases; the fine-grained correction guidance mechanism based on the orthogonal decomposition of syndrome differentiation elements improved the accuracy of syndrome differentiation correction by 12.3% and shortened the average correction time of physicians by 32.7%, significantly reducing the workload of physicians.

[0058] Clinical validation results show that this invention significantly improves the accuracy, interpretability, and model optimization efficiency of TCM AI diagnosis, shortens the diagnosis and correction time for physicians, and improves patient satisfaction, thus having significant clinical application value.

[0059] Therefore, the present invention adopts the above-mentioned method for reviewing and correcting TCM AI diagnostic results based on human-machine collaborative decision-making, which realizes deep collaboration between AI and TCM doctors in the entire diagnostic process rather than simple sequential review, significantly improving the accuracy and efficiency of syndrome differentiation correction, and significantly improving the clinical credibility and practicality of TCM AI diagnosis.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for reviewing and correcting TCM AI diagnostic results based on human-machine collaborative decision-making, characterized in that: Includes the following steps: S1. Multimodal four diagnostic methods data acquisition and standardized preprocessing; S2. Construct a hierarchical and integrated TCM feature extraction network to perform multi-scale feature extraction and fusion on the standardized four diagnostic methods data to generate a fusion feature vector of the four diagnostic methods. S3. Based on the TCM knowledge graph and the temporal causal reasoning engine, generate preliminary AI diagnosis results, and output the diagnosis reasoning path including the evolution of the syndrome and the confidence level of each diagnosis element. S4. Construct a dynamic weight allocation model with two dimensions: diagnostic confidence and doctor experience matching. The decision weights of AI and doctor are automatically allocated based on the confidence of the AI ​​diagnostic results and the experience level of the doctor currently treating the patient. S5. Human-machine collaborative diagnosis and decision-making: AI shows physicians the diagnosis and reasoning path, key evidence support and similar case references. At the same time, it performs fine-grained conflict detection and correction guidance based on the orthogonal decomposition of diagnosis elements. Physicians adjust the AI ​​results based on clinical experience, and the system calculates the final diagnosis result according to dynamic weights. S6. Define a deep verification mechanism for conflict detection. When there is a significant conflict between the AI ​​and the doctor's diagnosis, the multi-expert consultation process will be automatically triggered and a conflict analysis report will be generated. S7. Incremental knowledge updates and model self-optimization: The correction results of physicians and the conclusions of expert consultations are evaluated and weighted, and used as new training samples to perform weighted incremental learning on the AI ​​model, continuously improving the accuracy of diagnosis.

2. The method for reviewing and correcting TCM AI diagnostic results based on human-machine collaborative decision-making according to claim 1, characterized in that: S1 specifically includes the following steps: S11. Data collection for visual diagnosis includes tongue images, facial images, and body posture images. Standardized lighting conditions and shooting angles are used, and quantitative features of the tongue body, tongue coating, and facial color are extracted through image segmentation algorithms. S12. Auscultation data acquisition includes speech signals and breath sound signals. Speech features are extracted using Mel frequency cepstral coefficients, and breath sound features are extracted using wavelet transform. S13. The data collection of consultations adopts a structured questionnaire combined with natural language processing technology to convert patients' chief complaints and medical history descriptions into standardized TCM symptom terms. S14. Palpation data acquisition uses a digital pulse diagnostic instrument to obtain pulse wave signals and extract the frequency, rhythm, intensity, and morphological quantitative characteristics of the pulse. S15. Map all the data from the four diagnostic methods to a unified TCM symptom ontology database to generate a standardized feature matrix of the four diagnostic methods.

3. The method for reviewing and correcting TCM AI diagnostic results based on human-machine collaborative decision-making as described in claim 1, characterized in that: In S2, the hierarchical fusion TCM feature extraction network specifically includes: The bottom feature extraction layer uses a convolutional neural network to process the raw data of inspection, auscultation and palpation, and a bidirectional long short-term memory network to process the consultation text data. The middle-layer feature fusion layer uses an attention mechanism to weightedly fuse features from different modalities to generate a single-modality feature vector; The high-level semantic fusion layer constructs a semantic fusion network based on the Five Elements theory and the organ differentiation system of traditional Chinese medicine, mapping the single-modal feature vectors to the semantic space of traditional Chinese medicine differentiation, and generating the four diagnostic methods fusion feature vectors.

4. The method for reviewing and correcting TCM AI diagnostic results based on human-machine collaborative decision-making as described in claim 1, characterized in that: In S3, the process of generating preliminary AI diagnostic results based on a traditional Chinese medicine knowledge graph and a temporal causal reasoning engine includes the following steps: S31. Construct a TCM knowledge graph containing at least 2 million combinations of TCM symptoms and tongue and pulse signs, 4,000 syndrome types and corresponding TCM prescriptions, and define the causal and evolutionary transfer relationships between symptoms-syndromes, syndrome types-diseases, syndrome types-prescriptions, and syndrome types. S32. Input the four diagnostic and treatment fusion feature vectors and the patient's historical diagnosis and treatment time series data into the causal reasoning engine. First, use the temporal causal graph convolutional network to extract the dynamic evolution features of symptoms and syndrome types, calculate the evolution and transition probability of each syndrome type, and then combine the Bayesian network to calculate the comprehensive posterior probability of each syndrome type at the current time. S33. Generate a dialectical reasoning path that includes the evolution of syndrome types, and show the complete reasoning process from the appearance of symptoms to the current syndrome type, including the contribution of each symptom to each syndrome type at different time points and the transition probability of syndrome type evolution. S34. Calculate the overall confidence level of the dialectical results and the individual confidence levels of each dialectical element.

5. The method for reviewing and correcting TCM AI diagnostic results based on human-machine collaborative decision-making according to claim 1, characterized in that: In S4, the constructed two-dimensional dynamic weight allocation model of dialectical confidence and physician experience matching specifically includes: The confidence level (C) of the AI-generated identification results is divided into three levels: high confidence, medium confidence, and low confidence. Based on the physician's professional title, years of practice, specialty, and historical diagnostic accuracy, the physician's experience level E is calculated and classified into junior physicians, intermediate physicians, and senior physicians. The decision weight W of the AI ​​is calculated using the following formula. AI The decision weight of the physician W doctor : W AI =α×C+β×(1-E); IN doctor =1-W AI ; Where α and β are adjustment coefficients, α+β=1, and the optimal values ​​are determined through training with clinical data; When the AI's assessment result is low confidence and the physician is a junior physician, the process of review by a senior physician is automatically triggered.

6. The method for reviewing and correcting TCM AI diagnostic results based on human-machine collaborative decision-making according to claim 1, characterized in that: S5 specifically includes the following steps: S51. AI presents physicians with a visualized diagnostic reasoning path diagram, marking the top 5 symptoms with the highest support and the 3 most likely syndrome types, while also displaying a prediction of the patient's syndrome type evolution trend. S52. AI automatically retrieves historical cases with similar symptoms and disease progression to the current patient, and displays the diagnosis results, treatment plans and prognoses of similar cases; S53. AI detects fine-grained conflicts between physician adjustments and AI preliminary results in real time based on orthogonal decomposition of dialectical elements, and pushes targeted evidence and correction guidance for dialectical elements with significant conflicts. S54. Physicians modify the AI's diagnostic results based on the actual diagnosis, including adjusting the syndrome type, adding or deleting diagnostic elements, and correcting the evolution trend of the syndrome type. S55. The AI ​​calculates the final diagnosis result based on dynamic weights. When there is a difference between the doctor's adjustment and the AI's suggestion, the difference is automatically recorded and an adjustment explanation is generated. S56. Generate a complete diagnostic report that includes preliminary AI results, physician adjustments, final diagnosis, and prognostic recommendations.

7. A method for reviewing and correcting TCM AI diagnostic results based on human-machine collaborative decision-making, as described in claim 6, characterized in that: In S53, specifically, the preliminary diagnosis results of AI and the real-time adjustments made by the physician are decomposed into three independent orthogonal diagnostic elements: location of disease, nature of disease, and disease progression. The degree of conflict between AI and physician on each element is calculated. For diagnostic elements with a degree of conflict exceeding the threshold, key supporting evidence, counter-evidence cases, and authoritative expert consensus corresponding to that element are automatically pushed to the physician to guide the physician to make precise element-level corrections rather than overall syndrome adjustment.

8. The method for reviewing and correcting TCM AI diagnostic results based on human-machine collaborative decision-making according to claim 7, characterized in that: The orthogonal decomposition and conflict degree calculation of the diagnostic elements specifically includes: constructing three independent semantic spaces for the location, nature, and progression of the disease; mapping all syndrome names and symptom terms to the three semantic spaces based on the TCM symptom ontology and knowledge graph, obtaining the normalized vector representation of each syndrome and symptom in the three spaces; calculating the cosine distance between the AI ​​diagnostic result vector and the physician adjustment content vector in each semantic space as the conflict degree of the element; when the conflict degree of an element is greater than 0.6, it is determined that the element has a significant conflict, and a targeted correction guidance process is automatically triggered.

9. A method for reviewing and correcting TCM AI diagnostic results based on human-machine collaborative decision-making, as described in claim 1, characterized in that: S6 specifically includes the following steps: S61. Define the conflict threshold: when there are significant differences between the diagnosis results of AI and physician in terms of disease location, disease nature or disease severity, and the confidence level of both is higher than 0.7, it is judged as a significant conflict. S62. Automatically generate conflict analysis reports, providing detailed comparisons of the reasoning paths, evidence, syndrome evolution analysis, and similar case support between AI and physicians; S63. Push conflict cases to at least two senior physicians for consultation. The system automatically summarizes the consultation opinions and calculates the degree of consensus. S64. The consultation conclusion shall be used as the final diagnosis, and the case shall be marked as a difficult case and stored in the knowledge base.

10. A method for reviewing and correcting TCM AI diagnostic results based on human-machine collaborative decision-making, as described in claim 1, characterized in that: S7 specifically includes the following steps: S71. Establish a case quality assessment model to evaluate the quality of physicians' correction results and expert consultation conclusions, and select high-quality training samples. Calculate the learning weight of each training sample based on the two factors of correction intensity and physician level. The correction intensity is quantified according to the degree of difference between the AI ​​and the physician's diagnosis results, and the physician level weight is determined according to the physician's experience level. The correction sample weight of senior physicians is 2-3 times that of junior physicians. S72. The AI ​​model is updated using a weighted incremental learning algorithm. The parameter update magnitude is adjusted according to the sample learning weight. High-value correction samples are learned first, and model parameters are adjusted only using new high-quality samples to avoid catastrophic forgetting. S73. Regularly update the TCM knowledge graph, incorporating newly discovered symptom-syndrome relationships, syndrome evolution patterns, and clinical experience into the knowledge graph; S74. Generate a model optimization report to show the changes in diagnostic accuracy for different syndrome types and the model convergence speed.

Images