A traditional chinese medicine misdiagnosis risk multidimensional early warning method and system

Abstract

This invention relates to the fields of medical information technology and artificial intelligence-assisted decision-making, and particularly to a multi-dimensional early warning method and system for misdiagnosis risk in Traditional Chinese Medicine (TCM). It constructs a complete early warning closed loop through four interrelated steps: patient state modeling, syndrome differentiation logic verification, prescription risk quantification, and system self-healing diagnosis. First, a dynamic multi-dimensional patient state profile is constructed; second, the integrity verification and logical consistency analysis of the syndrome differentiation path are performed; third, the risk quantification assessment of individualized prescription compatibility is performed; and finally, the dynamic diagnosis of the early warning system's own health status is performed. In summary, this method achieves full-chain, multi-dimensional early warning of misdiagnosis risk from "patient state → syndrome differentiation process → prescription compatibility → system itself," significantly improving the safety and intelligence level of TCM diagnosis and treatment.

Description

Technical Field

[0001] This invention relates to the fields of medical information technology and artificial intelligence-assisted decision-making, and in particular to a multi-dimensional early warning method and system for the risk of misdiagnosis in traditional Chinese medicine. Background Technology

[0002] Traditional Chinese medicine (TCM) possesses unique advantages in the management of chronic diseases, the treatment of complex illnesses, and health maintenance. The core of TCM diagnosis and treatment lies in "syndrome differentiation and treatment," which involves collecting patient information through the four diagnostic methods of observation, auscultation and olfaction, inquiry, and palpation, comprehensively analyzing this information to arrive at a syndrome differentiation conclusion, and then prescribing medication accordingly. However, the TCM diagnostic and treatment process highly relies on the doctor's personal experience and accumulated knowledge, and presents the following significant technical challenges: First, existing TCM-assisted decision-making systems mostly analyze information based on a single dimension. For example, some systems recommend prescriptions solely based on the patient's chief complaint, while others conduct health assessments based only on physiological indicators (such as blood pressure and heart rate), neglecting the influence of environmental factors (seasons, geography, and climate) on the condition in the TCM concept of "harmony between man and nature." They also fail to effectively integrate multi-source heterogeneous information such as the patient's past medical history and living environment, resulting in an incomplete portrayal of the patient's condition and a lack of individualized adaptation in the early warning results.

[0003] Second, existing systems have limitations in the scope of risk warning coverage. Most systems only focus on the final stage of "whether the prescription compatibility violates the Eighteen Incompatibilities and Nineteen Antagonisms," failing to monitor the risks at the front end of TCM diagnosis and treatment—the syndrome differentiation process. Syndrome differentiation is the soul of TCM diagnosis and treatment; if the syndrome differentiation is incorrect (such as misdiagnosing a cold syndrome as a heat syndrome), the subsequent prescription will inevitably be incorrect. However, current technology lacks automated verification methods for the completeness of doctors' consultation information and the consistency of syndrome differentiation logic, making it impossible to provide early warnings of misdiagnosis risks at the syndrome differentiation stage.

[0004] Third, existing risk warning systems use static, fixed warning thresholds, which cannot dynamically adjust the warning sensitivity according to the severity of the patient's condition. For critically ill patients, fixed thresholds may lead to a large number of invalid alerts (false positives), interfering with doctors' judgment; for patients with mild symptoms, they may miss potential risks (false negatives). This "one-size-fits-all" warning model is ineffective in clinical practice.

[0005] Fourth, existing systems lack the ability to perceive and maintain their own health status. Early warning models rely on the quality of historical training data, but in clinical practice, training data can be contaminated due to varying levels of expertise among experts, shifting diagnostic criteria, and data entry errors. Furthermore, the predictive performance of the model gradually declines after deployment as the clinical environment changes. Current technologies cannot automatically monitor the model's reliability, nor can they identify and clean contaminated data, making it difficult to guarantee the long-term stability and reliability of the system.

[0006] In conclusion, there is an urgent need for a TCM misdiagnosis risk warning method that can cover the entire diagnosis and treatment process (from syndrome differentiation to prescription), integrate multi-dimensional information, and possess dynamic adaptive and self-repair capabilities, in order to systematically reduce the TCM misdiagnosis rate and improve the safety of clinical medication. Summary of the Invention

[0007] In view of this, the purpose of this invention is to construct a multi-dimensional early warning method that can cover the entire chain of TCM diagnosis and treatment (from patient status modeling, syndrome differentiation process verification to prescription compatibility risk quantification), and at the same time enable the early warning system to have dynamic diagnosis and self-repair capabilities for its own health status (prediction credibility, training data source quality), thereby systematically reducing the risk of misdiagnosis in TCM.

[0008] To achieve the above objectives, the first technical solution adopted by the present invention is as follows: A multi-dimensional early warning method for the risk of misdiagnosis in Traditional Chinese Medicine includes the following steps: Step 1: Construct a dynamic multidimensional state profile of the patient, specifically: collect the four diagnostic methods information, vital sign parameters, living environment data and past medical history information of the target object; extract features and numerically encode the collected data, and fuse the encoded feature vectors to generate a multidimensional state profile vector that represents the current comprehensive physiological-pathological-environmental state of the target object. Step Two: Perform a diagnostic path integrity check and logical consistency analysis. Specifically, the four diagnostic methods information is input into a pre-trained diagnostic reasoning network. The diagnostic reasoning network simulates the derivation process from symptoms to syndrome types based on a graph structure reasoning mechanism. During the derivation process, the diagnostic reasoning network performs the following two checks: If it detects that the currently input four diagnostic methods information is missing key symptom nodes relative to the preset standard diagnostic knowledge graph, and the absence of such nodes will cause the confidence of syndrome type inference to decrease by more than a preset threshold, then an incomplete diagnostic information warning is output; If the system syndrome type derived by the diagnostic reasoning network itself is inconsistent with the syndrome type diagnosed by the externally input doctor, then a diagnostic logic deviation warning is output. Step 3: Perform individualized prescription compatibility risk quantification assessment, specifically: obtain the traditional Chinese medicine prescription issued for the target object, map each herb in the prescription into a herb feature vector integrating its properties, flavors, meridian tropism, chemical components, and target pathways; based on the multidimensional state profile vector and the herb feature vectors, use a bilinear interaction function to calculate the compatibility risk score between any two herbs in the prescription; when the compatibility risk score exceeds a preset confidence threshold, mark the pair of herbs as a suspicious compatibility combination and output a prescription compatibility risk warning; Step 4: Perform dynamic health diagnosis of the early warning system itself, specifically: periodically collect real clinical outcome data, compare the various early warning results output in Step 2 and / or Step 3 with the real clinical outcome data, calculate the overall false alarm rate and false negative rate of the system, and generate a prediction reliability index; at the same time, perform periodic verification based on temporal consistency on the original training sample library used to train the dialectical deduction network and the bilinear interaction function, and evaluate the stability and reliability of each data source; when the prediction reliability index is lower than the first quality threshold, or when the temporal consistency verification of a data source fails, output a system self-healing trigger signal and start part or all of the retraining process of the corresponding model.

[0009] Furthermore, the specific methods for generating the multi-dimensional state profile vector in step one include: The collected information from the four diagnostic methods is converted into syndrome feature sub-vectors through a preset syndrome coding table; The collected vital signs parameters, including pulse, blood pressure, body temperature, and respiratory rate, are input into a pre-trained physiological state assessment network. The physiological state assessment network adopts a multi-layer fully connected structure to output a physiological state feature value between 0 and 1, which constitutes a physiological feature sub-vector. The training objective of the physiological state assessment network is to minimize the mean square error between the predicted physiological state and the healthy baseline. The collected living environment data is encoded into environmental feature sub-vectors; The syndrome feature sub-vector, physiological feature sub-vector, and environmental feature sub-vector are vertically concatenated to generate the multidimensional state profile vector.

[0010] Furthermore, the dialectical reasoning network in step two is pre-trained in the following way: Construct a TCM syndrome knowledge graph, in which nodes include three types of medical entities: symptoms, syndrome elements, and syndrome types, and edges represent the derivation or membership relationships between entities; The knowledge graph is used to learn representations to obtain the embedding vector of each node; By performing a depth-first search-based random walk on the knowledge graph, multiple derivation path samples from symptom nodes through symptom element nodes to symptom type nodes are obtained. Using the derivation path samples as training data, the graph attention network is fine-tuned using a graph neural network sequence prediction task, so that the graph attention network has the ability to infer and output corresponding syndrome nodes from the input symptom node set along the graph path. After training convergence, the dialectical deduction network is obtained.

[0011] Furthermore, the triggering condition for the incomplete dialectical information warning in step two is as follows: During its internal reasoning process, the dialectical deduction network identifies a pre-defined strong correlation edge in the knowledge graph between the currently input symptom node set and a certain uninputted symptom node. Through simulated reasoning, it determines that if the uninputted symptom node is used as an additional input, the probability of the syndrome type prediction output by the dialectical deduction network will increase beyond a preset gain threshold. Therefore, it is determined that there is a lack of key information and generates a warning of incomplete dialectical information.

[0012] Furthermore, the preset trust threshold in step three adopts a dynamic adjustment strategy: Obtain the historical multidimensional state profile vector and the current multidimensional state profile vector of the target object, and calculate the cosine similarity or Euclidean distance between the two as the state change amplitude. If the magnitude of the state change indicates that the current state has significantly deteriorated relative to the historical baseline, the value of the confidence threshold is automatically lowered to enhance the sensitivity of prescription risk warning. If the magnitude of the state change indicates that the current state is stable or improving, then the confidence threshold is maintained or increased to reduce false alarms.

[0013] Furthermore, the method for calculating the compatibility risk score using a bilinear interaction function in step three includes: For any two medicinal materials in the prescription and Obtain the feature vectors of the medicinal materials respectively. and and the multidimensional state profile vector ; Will and Perform a fusion operation to obtain the fusion vector of the medicinal materials. ; The original risk estimate is calculated using the following bilinear function: in, Given a trainable bilinear parameter matrix, For bias terms; Original risk estimate The compatibility risk score is obtained by mapping to the (0,1) interval using the Sigmoid function.

[0014] Furthermore, the specific method for evaluating the stability and reliability of the data source based on time-series consistency verification in step four is as follows: For each data contributor in the original training sample library, the samples they provided are divided into multiple time segments according to the collection time, and each time segment contains several diagnostic samples. For each time segment, calculate the statistical distribution of the diagnostic conclusions within the corresponding time segment; Using a sliding window approach, the Jensen-Shannon divergence of the statistical distribution between adjacent time segments is calculated to obtain the temporal inconsistency curve; Calculate the cumulative change of the time-series inconsistency curve; when the cumulative change exceeds a preset time-series stability threshold and cannot be attributed to external medical knowledge updates, determine that the data source credibility of the data contributor is unqualified.

[0015] Furthermore, the specific strategy for initiating part or all of the retraining process for the corresponding model in step four is as follows: When the triggering reason is that the prediction confidence index is lower than the first quality threshold, the collection time window of real clinical outcome data is automatically expanded, and the attention model associated with the bilinear interaction function is incrementally fine-tuned based on the expanded positive and negative samples. When the triggering reason is that the temporal consistency check of the data source fails, all samples contributed by the unqualified data source are automatically removed from the original training sample library, and the dialectical deduction network is fully retrained based on the cleaned sample library.

[0016] Furthermore, it also includes step five: multi-dimensional risk linkage analysis and auxiliary decision-making output, specifically: When steps two and three simultaneously output diagnostic logic deviation warnings and prescription compatibility risk warnings in the same diagnosis and treatment event, it is determined to be a high-risk misdiagnosis scenario, generating a first-level linkage alarm and blocking prescription output or forcing a second review. When only the system self-healing trigger signal is output in step four, it is determined to be a system maintenance scenario, the new early warning service is suspended and a model maintenance prompt is generated; When only a prescription compatibility risk warning is output, a traditional Chinese medicine knowledge graph retrieval engine is activated to find alternative medicinal materials with similar efficacy to the target medicinal material in the suspected compatibility combination and without known compatibility contraindications. At least one medicinal material replacement suggestion is output in combination with the multi-dimensional state profile vector, and the basis for the replacement is noted.

[0017] The second technical solution adopted in this invention is as follows: A multi-dimensional early warning system for misdiagnosis risk in traditional Chinese medicine (TCM) is used to implement the aforementioned multi-dimensional early warning method for misdiagnosis risk in TCM, comprising: a state profile construction unit, a syndrome differentiation logic verification unit, a compatibility risk quantification unit, a system self-healing diagnosis unit, and a risk linkage management unit. The state profile construction unit is used to perform step one; The dialectical logic verification unit integrates the dialectical deduction network based on graph attention network, which is used to execute step two. The matching risk quantification unit integrates the bilinear interaction function and its associated trainable parameters, and is used to perform step three. The system self-healing diagnostic unit is used to perform the prediction confidence calculation and data source time series consistency verification in step four. The risk linkage management unit is used to execute step five and display the warning level, blocking decision and auxiliary optimization information to the user.

[0018] The beneficial effects of this invention are: This invention provides a multi-dimensional early warning method for misdiagnosis risk in Traditional Chinese Medicine (TCM). It constructs a complete early warning closed loop through four interrelated steps: patient state modeling, syndrome differentiation logic verification, prescription risk quantification, and system self-healing diagnosis. First, a dynamic multi-dimensional patient state profile is constructed, fusing heterogeneous data from multiple sources, including the four diagnostic methods (inspection, auscultation and olfaction, palpation, and olfaction), vital signs, living environment, and past medical history, into a unified state vector. This overcomes the limitation of traditional TCM diagnosis relying solely on single-dimensional information, providing a comprehensive and quantitative foundation of patient characteristics for subsequent risk analysis. Second, the method performs integrity verification and logical consistency analysis of the syndrome differentiation path. Utilizing a graph-structured reasoning network to simulate the TCM derivation process from symptoms to syndrome types, it proactively identifies key symptom information omitted by doctors during consultation (incomplete syndrome differentiation information warning) and promptly detects deviations between the doctor's diagnostic conclusions and the system's reasoning results (syndrome differentiation logic deviation warning), thereby reducing the risk of misdiagnosis due to diagnostic errors at the source. Third, the method implements individualized quantitative assessment of prescription compatibility risks. By combining patient profiles with medicinal herb characteristics through a bilinear interaction function, it achieves individualized and quantitative assessment of the compatibility risks of pairwise herbs in traditional Chinese medicine compound prescriptions, overcoming the shortcomings of traditional compatibility contraindications which rely solely on empirical qualitative assessment. Finally, the method performs dynamic health diagnosis of the early warning system itself. By retrospectively analyzing real clinical outcomes to calculate false positive and false negative rates and performing temporal consistency checks on training data sources, the system acquires self-awareness and self-repair capabilities. It can promptly detect model performance degradation or training data contamination and trigger retraining, thus maintaining the accuracy and reliability of early warnings over the long term. In summary, this method achieves full-chain, multi-dimensional early warning of misdiagnosis risks from "patient status → diagnosis process → prescription compatibility → system itself," significantly improving the safety and intelligence level of TCM diagnosis and treatment.

[0019] This invention also provides a multi-dimensional early warning system for misdiagnosis risks in Traditional Chinese Medicine (TCM). Corresponding to the aforementioned early warning methods, the system architecture consists of five collaborative modules: a state profile construction unit, a syndrome differentiation logic verification unit, a compatibility risk quantification unit, a system self-healing diagnosis unit, and a risk linkage management unit. This architecture offers the following technical advantages: First, the modular design allows each unit to be developed, tested, and upgraded independently. For example, the syndrome differentiation logic verification unit can be replaced with a more advanced inference network without affecting other units, reducing system maintenance costs. Second, the data flow between units is clear: the multi-dimensional vectors output by the state profile construction unit simultaneously serve both the syndrome differentiation logic verification unit and the compatibility risk quantification unit, achieving feature reuse and avoiding redundant calculations. Third, the system self-healing diagnosis unit simultaneously monitors the training data of the other two core models (syndrome differentiation inference network and bilinear interaction function) and can trigger targeted retraining, forming a complete closed loop of data-model-feedback-optimization. Fourth, the risk linkage management unit, as a unified exit point, can integrate early warning signals from different dimensions and execute linked decisions, providing users with concise and clear treatment suggestions. This architecture is highly scalable and can be easily integrated with more multi-dimensional risk analysis units (such as drug dosage risk, risk of combined use of Chinese and Western medicines) in the future, providing a solid foundation for building a comprehensive and intelligent safety protection network for TCM diagnosis and treatment. Attached Figure Description

[0020] Figure 1 The diagram shows a flowchart of the steps of a multi-dimensional early warning method for the risk of misdiagnosis in traditional Chinese medicine according to the present invention. Figure 2 The diagram shown is a structural block diagram of a multi-dimensional early warning system for the risk of misdiagnosis in traditional Chinese medicine according to the present invention. Explanation of icon numbers: 1-Status profile construction unit; 2-Dialectical logic verification unit; 3-Compatibility risk quantification unit; 4-System self-healing diagnosis unit; 5-Risk linkage management unit. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and specific embodiments: like Figure 1 As shown, the present invention provides a multi-dimensional early warning method for the risk of misdiagnosis in traditional Chinese medicine, comprising the following steps: Step 1: Construct a dynamic multidimensional state profile of the patient, specifically: collect the four diagnostic methods information, vital sign parameters, living environment data and past medical history information of the target object; extract features and numerically encode the collected data, and fuse the encoded feature vectors to generate a multidimensional state profile vector that represents the current comprehensive physiological-pathological-environmental state of the target object. Step Two: Perform a diagnostic path integrity check and logical consistency analysis. Specifically, the four diagnostic methods information is input into a pre-trained diagnostic reasoning network. The diagnostic reasoning network simulates the derivation process from symptoms to syndrome types based on a graph structure reasoning mechanism. During the derivation process, the diagnostic reasoning network performs the following two checks: If it detects that the currently input four diagnostic methods information is missing key symptom nodes relative to the preset standard diagnostic knowledge graph, and the absence of such nodes will cause the confidence of syndrome type inference to decrease by more than a preset threshold, then an incomplete diagnostic information warning is output; If the system syndrome type derived by the diagnostic reasoning network itself is inconsistent with the syndrome type diagnosed by the externally input doctor, then a diagnostic logic deviation warning is output. Step 3: Perform individualized prescription compatibility risk quantification assessment, specifically: obtain the traditional Chinese medicine prescription issued for the target object, map each herb in the prescription into a herb feature vector integrating its properties, flavors, meridian tropism, chemical components, and target pathways; based on the multidimensional state profile vector and the herb feature vectors, use a bilinear interaction function to calculate the compatibility risk score between any two herbs in the prescription; when the compatibility risk score exceeds a preset confidence threshold, mark the pair of herbs as a suspicious compatibility combination and output a prescription compatibility risk warning; Step 4: Perform dynamic health diagnosis of the early warning system itself, specifically: periodically collect real clinical outcome data, compare the various early warning results output in Step 2 and / or Step 3 with the real clinical outcome data, calculate the overall false alarm rate and false negative rate of the system, and generate a prediction reliability index; at the same time, perform periodic verification based on temporal consistency on the original training sample library used to train the dialectical deduction network and the bilinear interaction function, and evaluate the stability and reliability of each data source; when the prediction reliability index is lower than the first quality threshold, or when the temporal consistency verification of a data source fails, output a system self-healing trigger signal and start part or all of the retraining process of the corresponding model.

[0022] The beneficial effects of this invention are: This invention provides a multi-dimensional early warning method for misdiagnosis risk in Traditional Chinese Medicine (TCM). It constructs a complete early warning closed loop through four interrelated steps: patient state modeling, syndrome differentiation logic verification, prescription risk quantification, and system self-healing diagnosis. First, a dynamic multi-dimensional patient state profile is constructed, fusing heterogeneous data from multiple sources, including the four diagnostic methods (inspection, auscultation and olfaction, palpation, and olfaction), vital signs, living environment, and past medical history, into a unified state vector. This overcomes the limitation of traditional TCM diagnosis relying solely on single-dimensional information, providing a comprehensive and quantitative foundation of patient characteristics for subsequent risk analysis. Second, the method performs integrity verification and logical consistency analysis of the syndrome differentiation path. Utilizing a graph-structured reasoning network to simulate the TCM derivation process from symptoms to syndrome types, it proactively identifies key symptom information omitted by doctors during consultation (incomplete syndrome differentiation information warning) and promptly detects deviations between the doctor's diagnostic conclusions and the system's reasoning results (syndrome differentiation logic deviation warning), thereby reducing the risk of misdiagnosis due to diagnostic errors at the source. Third, the method implements individualized quantitative assessment of prescription compatibility risks. By combining patient profiles with medicinal herb characteristics through a bilinear interaction function, it achieves individualized and quantitative assessment of the compatibility risks of pairwise herbs in traditional Chinese medicine compound prescriptions, overcoming the shortcomings of traditional compatibility contraindications which rely solely on empirical qualitative assessment. Finally, the method performs dynamic health diagnosis of the early warning system itself. By retrospectively analyzing real clinical outcomes to calculate false positive and false negative rates and performing temporal consistency checks on training data sources, the system acquires self-awareness and self-repair capabilities. It can promptly detect model performance degradation or training data contamination and trigger retraining, thus maintaining the accuracy and reliability of early warnings over the long term. In summary, this method achieves full-chain, multi-dimensional early warning of misdiagnosis risks from "patient status → diagnosis process → prescription compatibility → system itself," significantly improving the safety and intelligence level of TCM diagnosis and treatment.

[0023] Furthermore, the specific methods for generating the multi-dimensional state profile vector in step one include: The collected information from the four diagnostic methods is converted into syndrome feature sub-vectors through a preset syndrome coding table; The collected vital signs parameters, including pulse, blood pressure, body temperature, and respiratory rate, are input into a pre-trained physiological state assessment network. The physiological state assessment network adopts a multi-layer fully connected structure to output a physiological state feature value between 0 and 1, which constitutes a physiological feature sub-vector. The training objective of the physiological state assessment network is to minimize the mean square error between the predicted physiological state and the healthy baseline. The collected living environment data is encoded into environmental feature sub-vectors; The syndrome feature sub-vector, physiological feature sub-vector, and environmental feature sub-vector are vertically concatenated to generate the multidimensional state profile vector.

[0024] As described above, the specific method for generating the multidimensional state profile vector in step one is further defined. By converting the information from the four diagnostic methods into syndrome feature sub-vectors, a numerical expression of the qualitative descriptions in traditional Chinese medicine (such as tongue coating and pulse) is realized, facilitating subsequent calculations. A pre-trained physiological state assessment network is introduced, using objective vital signs parameters such as pulse, blood pressure, body temperature, and respiratory rate as input. Employing a multi-layer fully connected structure, it outputs physiological state feature values ​​between 0 and 1. This network, trained on a healthy baseline, can quantitatively characterize the patient's current physiological deviation, offering greater objectivity and accuracy compared to traditional empirical judgments. Environmental data is encoded as environmental feature sub-vectors, reflecting the impact of environmental factors on health in the traditional Chinese medicine concept of "harmony between man and nature." Finally, the three sub-vectors are vertically concatenated to form a unified patient state profile. This fusion strategy preserves the independence of each modality's information while providing structured input for the subsequent bilinear interaction function. This allows the compatibility risk assessment to simultaneously consider the patient's syndrome attributes, physiological function, and external environmental factors, greatly enhancing the individualized adaptation capability of risk warning.

[0025] Furthermore, the dialectical reasoning network in step two is pre-trained in the following way: Construct a TCM syndrome knowledge graph, in which nodes include three types of medical entities: symptoms, syndrome elements, and syndrome types, and edges represent the derivation or membership relationships between entities; The knowledge graph is used to learn representations to obtain the embedding vector of each node; By performing a depth-first search-based random walk on the knowledge graph, multiple derivation path samples from symptom nodes through symptom element nodes to symptom type nodes are obtained. Using the derivation path samples as training data, the graph attention network is fine-tuned using a graph neural network sequence prediction task, so that the graph attention network has the ability to infer and output corresponding syndrome nodes from the input symptom node set along the graph path. After training convergence, the dialectical deduction network is obtained.

[0026] As described above, the pre-training method for the syndrome differentiation and deduction network is limited. By constructing a TCM syndrome knowledge graph containing three types of nodes—symptoms, syndrome elements, and syndrome types—along with deductive relationship edges, discrete TCM theoretical knowledge is transformed into structured graph data. Using a graph attention network for representation learning, the importance weights of different neighboring nodes to the current node can be automatically learned, which better reflects the characteristic of different symptoms contributing differently to syndrome types in TCM syndrome differentiation compared to traditional graph convolutional networks. By performing a depth-first search-based random walk on the knowledge graph, multiple deduction path samples from symptoms to syndrome types are obtained, solving the problem of scarce training data for TCM syndrome differentiation sequence annotation. Finally, a graph neural network sequence prediction task is used for fine-tuning, enabling the network to learn the ability to reason along graph paths. The syndrome differentiation and deduction network trained using this method can not only accurately infer syndrome types based on input symptoms but also output intermediate results (syndrome elements) during the reasoning process, exhibiting good interpretability and providing a reliable reasoning engine for subsequent warnings of incomplete syndrome information and logical deviations.

[0027] Furthermore, the triggering condition for the incomplete dialectical information warning in step two is as follows: During its internal reasoning process, the dialectical deduction network identifies a pre-defined strong correlation edge in the knowledge graph between the currently input symptom node set and a certain uninputted symptom node. Through simulated reasoning, it determines that if the uninputted symptom node is used as an additional input, the probability of the syndrome type prediction output by the dialectical deduction network will increase beyond a preset gain threshold. Therefore, it is determined that there is a lack of key information and generates a warning of incomplete dialectical information.

[0028] As described above, the triggering conditions for the incomplete diagnostic information warning are specifically defined. This method utilizes the knowledge graph structure within the diagnostic reasoning network to automatically identify strong correlations between currently input symptom nodes and nodes without input symptoms. It then uses simulation reasoning to determine the extent to which the missing symptom increases the probability of predicting the syndrome type. When the gain exceeds a preset threshold, the system determines that key information is missing and issues a warning. This mechanism has the following significant effects: First, it achieves proactive and intelligent consultation assistance, automatically prompting for potentially overlooked important consultation points after the doctor completes the chief complaint collection, reducing misdiagnosis due to incomplete information collection; second, the gain threshold can be flexibly adjusted according to clinical scenarios (e.g., a lower threshold can be set in emergency situations to improve sensitivity), balancing the sensitivity and specificity of the warning; third, the warning not only outputs "what information is missing" but also implicitly provides a quantitative prompt on "how much benefit supplementing this information will bring," helping doctors quickly assess the clinical value of the information and optimize consultation efficiency.

[0029] Furthermore, the preset trust threshold in step three adopts a dynamic adjustment strategy: Obtain the historical multidimensional state profile vector and the current multidimensional state profile vector of the target object, and calculate the cosine similarity or Euclidean distance between the two as the state change amplitude. If the magnitude of the state change indicates that the current state has significantly deteriorated relative to the historical baseline, the value of the confidence threshold is automatically lowered to enhance the sensitivity of prescription risk warning. If the magnitude of the state change indicates that the current state is stable or improving, then the confidence threshold is maintained or increased to reduce false alarms.

[0030] As described above, a dynamic adjustment strategy for the reliability threshold in prescription incompatibility risk warning is defined. Traditional warning systems use fixed thresholds, which can easily lead to too many invalid alerts (high false positive rate) in critically ill patients or to missed potential risks in mildly ill patients (high false negative rate). This solution quantitatively assesses the degree of disease deterioration by calculating the cosine similarity or Euclidean distance between the patient's current multidimensional state profile vector and the historical baseline. When the condition deteriorates significantly, the reliability threshold is automatically lowered, making the system more sensitive to prescription incompatibility risks and ensuring that high-risk patients receive adequate warning protection. When the condition stabilizes or improves, the threshold is maintained or raised to reduce unnecessary interfering alerts and avoid the "crying wolf" effect for doctors. This dynamic adaptive mechanism allows the same warning system to be applied to patient groups with different severity levels, significantly improving clinical usability and user experience.

[0031] Furthermore, the method for calculating the compatibility risk score using a bilinear interaction function in step three includes: For any two medicinal materials in the prescription and Obtain the feature vectors of the medicinal materials respectively. and and the multidimensional state profile vector ; Will and Perform a fusion operation to obtain the fusion vector of the medicinal materials. ; The original risk estimate is calculated using the following bilinear function: in, Given a trainable bilinear parameter matrix, For bias terms; Original risk estimate The compatibility risk score is obtained by mapping to the (0,1) interval using the Sigmoid function.

[0032] As described above, the specific mathematical model for calculating the compatibility risk score using a bilinear interaction function is defined. This scheme employs a bilinear form. , where the patient state vector Fusion vector with medicinal materials Through trainable matrices Compared to traditional dot-product attention or fully connected concatenation methods, the bilinear form can explicitly model the second-order interaction between patient state and drug compatibility, which better aligns with the complex mapping principle of "treatment tailored to the individual" in traditional Chinese medicine. (Medicinal herb pair fusion vector) It can be designed in various forms such as concatenation and element-wise multiplication, offering high flexibility. Finally, a risk score between 0 and 1 is output through the Sigmoid function, facilitating threshold setting and comparison. The advantages of this calculation method are: first, the number of parameters is controllable, reducing the risk of overfitting; second, it uses a bilinear matrix. It can be optimized end-to-end through clinical feedback data, enabling the model to continuously learn new compatibility knowledge; third, it has high computational efficiency and is suitable for real-time prescription risk screening scenarios.

[0033] Furthermore, the specific method for evaluating the stability and reliability of the data source based on time-series consistency verification in step four is as follows: For each data contributor in the original training sample library, the samples they provided are divided into multiple time segments according to the collection time, and each time segment contains several diagnostic samples. For each time segment, calculate the statistical distribution of the diagnostic conclusions within the corresponding time segment; Using a sliding window approach, the Jensen-Shannon divergence of the statistical distribution between adjacent time segments is calculated to obtain the temporal inconsistency curve; Calculate the cumulative change of the time-series inconsistency curve; when the cumulative change exceeds a preset time-series stability threshold and cannot be attributed to external medical knowledge updates, determine that the data source credibility of the data contributor is unqualified.

[0034] As described above, the method for assessing data source credibility based on temporal consistency verification is limited. Traditional data quality control often relies on cross-validation or expert scoring, which struggles to detect diagnostic style drift or quality degradation that can occur among data contributors over time. This solution innovatively introduces temporal analysis technology: it segments the samples of each data contributor according to the collection time, calculates the statistical distribution of diagnostic conclusions (such as syndrome frequency and medication preference) within each segment, and then uses the Jensen-Shannon divergence to measure the degree of change in distribution between adjacent time segments. The cumulative change is used to determine whether abnormal drift has occurred. This method can effectively identify data quality degradation caused by fatigue, outdated knowledge, or deliberate falsification. Furthermore, when external medical knowledge updates (such as the release of new guidelines) lead to changes in global diagnostic standards, the system can avoid misjudgments by tagging external events. The high-quality, highly consistent training data selected through this method can significantly improve the generalization ability and long-term stability of the syndrome differentiation network and the compatibility risk model.

[0035] Furthermore, the specific strategy for initiating part or all of the retraining process for the corresponding model in step four is as follows: When the triggering reason is that the prediction confidence index is lower than the first quality threshold, the collection time window of real clinical outcome data is automatically expanded, and the attention model associated with the bilinear interaction function is incrementally fine-tuned based on the expanded positive and negative samples. When the triggering reason is that the temporal consistency check of the data source fails, all samples contributed by the unqualified data source are automatically removed from the original training sample library, and the dialectical deduction network is fully retrained based on the cleaned sample library.

[0036] As described above, the specific strategy for initiating retraining after triggering system self-healing is defined, and differentiated repair solutions are adopted for two different causes of failure. When the prediction reliability indicators (false positive rate and false negative rate) decrease, it indicates that the model's adaptability to the current clinical environment has deteriorated. At this time, it is not necessary to discard historical knowledge. Instead, the collection window of real clinical outcome data is expanded to obtain more recent positive and negative samples, and the bilinear interaction function model is incrementally fine-tuned. This incremental learning method retains the general matching rules learned by the model while quickly adapting to new data distributions, resulting in low training costs and fast convergence. When the temporal consistency check of a certain data source is found to be unsatisfactory, it indicates that there is a systematic bias in the data source. At this time, a more aggressive repair measure is taken: all samples from the original training sample library of the data source are removed, and then the dialectical inference network is fully retrained. This hierarchical repair strategy not only ensures the system's rapid response to common performance degradation but also ensures that the impact of data contamination can be completely eradicated, reflecting the health management concept of system self-healing.

[0037] Furthermore, it also includes step five: multi-dimensional risk linkage analysis and auxiliary decision-making output, specifically: When steps two and three simultaneously output diagnostic logic deviation warnings and prescription compatibility risk warnings in the same diagnosis and treatment event, it is determined to be a high-risk misdiagnosis scenario, generating a first-level linkage alarm and blocking prescription output or forcing a second review. When only the system self-healing trigger signal is output in step four, it is determined to be a system maintenance scenario, the new early warning service is suspended and a model maintenance prompt is generated; When only a prescription compatibility risk warning is output, a traditional Chinese medicine knowledge graph retrieval engine is activated to find alternative medicinal materials with similar efficacy to the target medicinal material in the suspected compatibility combination and without known compatibility contraindications. At least one medicinal material replacement suggestion is output in combination with the multi-dimensional state profile vector, and the basis for the replacement is noted.

[0038] As described above, the introduction of multi-dimensional risk linkage analysis and auxiliary decision-making output is one of the core innovations of this invention. This step breaks through the limitation of independent risk analysis in traditional early warning systems. By judging the combined state of different early warning signals, it triggers differentiated response measures, achieving an upgrade from "information prompts" to "decision support." Specific effects include: First, when both the dialectical logic deviation warning and the prescription compatibility risk warning occur simultaneously, the system determines it as a high-risk misdiagnosis scenario. At this point, it not only prompts the risk but also proactively blocks prescription output or mandates secondary review, preventing high-risk prescriptions from reaching pharmacies or patients, significantly improving medication safety. Second, when only the system self-healing trigger signal appears, the system enters maintenance mode, suspending new early warning services to avoid misleading results in an unstable model state, and prompting the administrator to perform maintenance, demonstrating the system's responsible attitude towards its own health. Third, when only prescription compatibility risks exist, the system activates the knowledge graph retrieval engine to intelligently recommend alternative medicines with similar efficacy and no contraindications, and adds the reasoning basis, realizing a one-stop auxiliary decision-making process from "risk discovery" to "risk resolution", which greatly improves the efficiency of doctors' work and the accuracy of prescription adjustments.

[0039] The second technical solution adopted in this invention is as follows: A multi-dimensional early warning system for misdiagnosis risk in traditional Chinese medicine (TCM) is used to implement the aforementioned multi-dimensional early warning method for misdiagnosis risk in TCM, comprising: a state profile construction unit, a syndrome differentiation logic verification unit, a compatibility risk quantification unit, a system self-healing diagnosis unit, and a risk linkage management unit. The state profile construction unit is used to perform step one; The dialectical logic verification unit integrates the dialectical deduction network based on graph attention network, which is used to execute step two. The matching risk quantification unit integrates the bilinear interaction function and its associated trainable parameters, and is used to perform step three. The system self-healing diagnostic unit is used to perform the prediction confidence calculation and data source time series consistency verification in step four. The risk linkage management unit is used to execute step five and display the warning level, blocking decision and auxiliary optimization information to the user.

[0040] As can be seen from the above description, the present invention has the following beneficial effects: The system architecture for implementing the aforementioned early warning method is defined, consisting of five collaborative modules: a state profile construction unit, a dialectical logic verification unit, a compatibility risk quantification unit, a system self-healing diagnosis unit, and a risk linkage management unit. This architecture offers the following technical advantages: First, the modular design allows each unit to be developed, tested, and upgraded independently. For example, the dialectical logic verification unit can be replaced with a more advanced inference network without affecting other units, reducing system maintenance costs. Second, the data flow between units is clear: the multi-dimensional vectors output by the state profile construction unit simultaneously serve both the dialectical logic verification unit and the compatibility risk quantification unit, achieving feature reuse and avoiding redundant computation. Third, the system self-healing diagnosis unit simultaneously monitors the training data of the other two core models (dialectical inference network and bilinear interaction function) and can trigger targeted retraining, forming a complete closed loop of data-model-feedback-optimization. Fourth, the risk linkage management unit, as a unified exit point, can integrate early warning signals from different dimensions and execute linked decisions, providing users with concise and clear handling suggestions. This architecture is highly scalable and can be easily integrated with more multi-dimensional risk analysis units (such as drug dosage risk, risk of combined use of Chinese and Western medicines) in the future, providing a solid foundation for building a comprehensive and intelligent safety protection network for TCM diagnosis and treatment.

[0041] The following are several preferred embodiments or application embodiments to help those skilled in the art better understand the technical content of the present invention and the technical contributions made by the present invention compared with the prior art: Preferred embodiment 1: This embodiment provides a multi-dimensional early warning method for the risk of misdiagnosis in Traditional Chinese Medicine. For example... Figure 1 As shown, the method includes steps one through five.

[0042] Step 1: Constructing a dynamic, multidimensional patient profile In this embodiment, the target subject is a 45-year-old male patient who sought medical attention for long-term stomach discomfort.

[0043] The collected multi-source heterogeneous information includes: Information from the four diagnostic methods: collected through a human-computer interaction interface (patient consultation) and a tongue and pulse diagnostic instrument. Patient's chief complaint: "Dull pain in the epigastrium, relieved by warmth and pressure, worsening on an empty stomach and relieved after eating; fatigue and weakness; loose stools." Tongue examination results: Pale and swollen tongue with teeth marks on the edges; white and slippery tongue coating. Pulse examination results: Deep, thready, and weak pulse.

[0044] Vital signs parameters: The patient's pulse (78 beats / min), blood pressure (118 / 76 mmHg), body temperature (36.7℃), and respiratory rate (16 breaths / min) were collected via a wearable wristband.

[0045] Living environment data: obtained through location and network access: the patient's location is a coastal city in southern China, the current solar term is the Beginning of Winter, the weather is cloudy and rainy, and the humidity is 85%. At the same time, the patient's occupation was extracted from the electronic medical record as an office clerk, who usually sits for long periods of time and prefers cold drinks.

[0046] Past medical history information: obtained from the hospital information system: The patient has a 3-year history of chronic superficial gastritis and reported that the previous TCM diagnosis was "spleen and stomach deficiency and cold syndrome".

[0047] Feature extraction and encoding: The information from the four diagnostic methods (inspection, auscultation and olfaction, palpation, and olfaction) is converted into syndrome feature sub-vectors according to a pre-defined syndrome coding table (based on the national standard "Terminology for Clinical Diagnosis and Treatment in Traditional Chinese Medicine"). For example, "dull stomach pain" is encoded as symptom ID 0421, "preferring warmth and pressure" as symptom ID 0573, "pale and swollen tongue" as tongue appearance ID 1102, and "deep and thready pulse" as pulse appearance ID 3021, etc. Each symptom corresponds to a unique coding dimension, ultimately forming a 200-dimensional syndrome feature sub-vector. .

[0048] Vital signs parameters are input into a pre-trained physiological state assessment network. This network employs a three-layer fully connected structure: the input layer has a dimension of 4 (pulse, systolic blood pressure, body temperature, and respiratory rate), the hidden layers have 64 and 32 nodes respectively, and the output layer has 1 node, outputting a physiological state feature value between 0 and 1 through sigmoid activation. The network training objective is to minimize the mean squared error between the predicted physiological state and the healthy baseline (the target value of the average physiological parameters of a healthy population after network output). Upon input from the current patient, the network outputs a physiological state feature value of 0.65 (the healthy baseline is calibrated as 0.5, where 0.65 indicates a slight deviation), forming a 1-dimensional physiological feature sub-vector. .

[0049] The living environment data is encoded into environmental feature sub-vectors: geographical location (southern coastal area) is encoded as a 2-bit binary number, the solar term (Beginning of Winter) is encoded as a 24-choose-1 unique heat vector, and humidity (85%) is normalized to 0.85, ultimately forming a 26-dimensional environmental feature sub-vector. .

[0050] Will (200 dimensions) (1-dimensional) (26 dimensions) are vertically stitched together to obtain a 227-dimensional multidimensional state profile vector. This vector serves as the basis input for all subsequent steps in this embodiment.

[0051] Step Two: Perform dialectical path integrity verification and logical consistency analysis The dialectical reasoning network was pre-trained using the following methods: First, a knowledge graph of TCM syndromes was constructed. Three categories of medical entities were extracted from textbooks on TCM diagnostics and TCM clinical terminology: symptoms (such as "stomachache," "preferring warmth and pressure," "loose stools"), syndrome elements (such as "cold," "deficiency," "spleen," "stomach," etc.), and syndrome types (such as "spleen and stomach deficiency-cold syndrome," "spleen yang deficiency syndrome," etc.). Derivation edges were established using expert rules: symptom → syndrome element (e.g., "preferring warmth and pressure" → "cold"), syndrome element → syndrome type (e.g., "spleen" + "cold" → "spleen and stomach deficiency-cold syndrome"). The graph contains 526 symptom nodes, 78 syndrome element nodes, 312 syndrome type nodes, and 4089 directed edges.

[0052] Secondly, graph attention networks are used for representation learning of the knowledge graph. In each layer of the graph attention network, nodes... The update formula is: ; in The attention coefficient is calculated by the dot product of the query vector and the key vector. After passing through two graph attention layers (each layer outputting an embedding dimension of 128), a 128-dimensional embedding vector is obtained for each node.

[0053] Then, a depth-first search-based random walk is performed on the knowledge graph. Starting from any symptom node, a path leading to the symptom element node and then to the symptom type node is randomly selected (the path depth does not exceed 5 steps). 1000 paths are collected for each path, resulting in approximately 30,000 derivation path samples. Each sample is in the form of... .

[0054] Finally, a graph neural network sequence prediction task was used for fine-tuning. The derived path samples were used as training data, with the first few nodes of the path as input and the symptom node at the end of the path as the prediction target. The graph attention network was trained using the cross-entropy loss function. After training convergence, the network was able to infer and output the corresponding symptom node along the graph path, starting from the input set of symptom nodes.

[0055] Input the four diagnostic information (symptom set) collected in step one into the pre-trained syndrome differentiation and deduction network.

[0056] Completeness check: During internal reasoning, the network identified "stomach pain," "preferring warmth and pressure," and "loose stools" among the current input symptom nodes. However, the symptom "relief from eating" (with a strong correlation edge to "stomach pain" in the knowledge graph, with a co-occurrence strength > 0.8), was not input. The network performed simulated reasoning: If the current symptom set was used as input, the probability of outputting the syndrome type "spleen and stomach deficiency-cold syndrome" was 0.72. If "relief from eating" was added as an additional input, the output probability increased to 0.91, an increase of 0.19 exceeding the preset gain threshold (0.15). Therefore, the network determined that key information was missing, generating an incomplete diagnostic information warning, prompting the doctor: "The patient has the typical symptom 'relief from eating' which has not been collected; supplementing this information can increase the confidence of spleen and stomach deficiency-cold syndrome by 19%."

[0057] Logical consistency analysis: The network automatically infers the system syndrome type as "Spleen and Stomach Deficiency-Cold Syndrome" (probability 0.72) based on the input symptom set. At this time, the doctor's externally input diagnostic syndrome type is "Spleen and Stomach Damp-Heat Syndrome". The system compares the system syndrome type output by the network with the syndrome type input by the doctor. If an inconsistency is found, a diagnostic logic deviation warning is output: "The system infers Spleen and Stomach Deficiency-Cold Syndrome, but you are diagnosed with Spleen and Stomach Damp-Heat Syndrome. Please review the four diagnostic methods and diagnostic basis." Step 3: Perform individualized prescription compatibility risk quantitative assessment Suppose the doctor's final prescription for traditional Chinese medicine is: Astragalus membranaceus 15g, Atractylodes macrocephala 12g, Poria cocos 12g, Zingiber officinale 6g, and Coptis chinensis 3g.

[0058] The feature vectors of medicinal materials are generated as follows: For each medicinal herb in the prescription, a herb characteristic database is pre-constructed. This database integrates the herb's properties and meridian tropism (from *Zhonghua Bencao*), chemical components (from the TCMSP database), and target pathways (extracted from PubMed literature). For example: Astragalus characteristics: warm in nature, sweet in taste, and enters the spleen and lung meridians. It contains astragaloside IV, and its targets include STAT3 and NF-κB. It is mapped to a 128-dimensional vector through an embedding layer. .

[0059] Characteristics of dried ginger: It is hot in nature, pungent in taste, and enters the spleen and stomach meridians. It contains 6-gingerol and targets TRPV1, etc.

[0060] Coptis chinensis characteristics: cold in nature, bitter in taste, enters the heart, liver and stomach meridians, contains berberine, and targets AMPK, etc.

[0061] All medicinal materials are mapped to 128-dimensional feature vectors. .

[0062] The bilinear interaction function is used to calculate the compatibility risk score, as detailed below: For any two herbs in the prescription, such as Astragalus (i) and Coptis (j): Obtain the feature vector of medicinal materials , and the multi-dimensional state profile vector generated in step one. (227 dimensions).

[0063] Will and Perform a fusion operation (in this embodiment, the Hadamard product is used element-wise) to obtain the fusion vector of the medicinal material pairs. , dimension 128.

[0064] Bilinear parameter matrix The dimension is 227×128, where 227 is the number of elements. Dimensions, 128 The matrix is ​​dimensional. It was pre-trained using a large amount of historical prescription and adverse reaction data.

[0065] Calculate the original risk estimate: In this example, the calculation is as follows: .

[0066] Risk score mapped to the (0,1) interval using the Sigmoid function. .

[0067] For all possible pairs of herbs in the prescription (a total of C(5,2)=10 pairs), a risk score was calculated for each pair. The initial confidence threshold was set at 0.75. Since Astragalus (warm in nature) and Coptis (cold in nature) have opposite medicinal properties, and the patient's condition profile shows "spleen and stomach deficiency and cold" (the deficiency and cold-related features in the syndrome subvector are activated), the bilinear interaction function gives a high-risk score of 0.91, which exceeds the threshold. Therefore, this pair of herbs is marked as a suspicious combination, and the system outputs a prescription combination risk warning: "The combination of Astragalus and Coptis involves both cold and hot, tonifying and purging, which may aggravate spleen and stomach deficiency and cold. It is recommended to use it with caution or adjust the dosage."

[0068] Step 4: Perform dynamic health assessment of the early warning system itself. The prediction reliability index is calculated as follows: The system is deployed in a top-tier tertiary hospital of traditional Chinese medicine and collects real clinical outcome data monthly. A total of 1200 cases of diagnosis and treatment using this early warning system within the past 30 days were collected. For each case, the improvement of symptoms and the presence of adverse reactions were tracked for 7 days after medication (using electronic medical records and follow-up records). The early warning results output from steps two and three (such as the aforementioned early warnings for incomplete syndrome differentiation information, syndrome differentiation logic deviation, and prescription compatibility risks) were compared with the actual outcome of the case.

[0069] False alarm rate = Number of cases where a warning indicates a risk but no adverse reaction actually occurs / Total number of warnings = 42 / 380 ≈ 11.1%.

[0070] The underreporting rate = number of cases where adverse reactions actually occurred but were not indicated by the early warning / total number of adverse reaction cases = 5 / 65 ≈ 7.7%.

[0071] The first quality threshold is set as a false alarm rate ≤ 15% and a false negative rate ≤ 10%. The current false alarm rate is 11.1% ≤ 15% and the false negative rate is 7.7% ≤ 10%. The prediction reliability index is qualified, and the system self-healing will not be triggered for the time being.

[0072] Data source timing consistency verification, as detailed below: The system periodically (quarterly) performs time-series consistency checks on each data contributor in the original training sample library (e.g., 50 TCM experts who provided historical diagnostic data to the system). Taking one expert surnamed Li as an example: The 3,000 diagnostic samples provided by the expert over the past three years were divided into 12 time segments based on the collection time, with each segment containing 250 samples.

[0073] For each time segment, calculate the statistical distribution of the diagnostic conclusions (syndrome type categories) within that segment (a total of 20 common syndrome types, and the frequency of occurrence of each syndrome type).

[0074] A sliding window (window size 2) is used to calculate the Jensen-Shannon divergence of the statistical distribution between two adjacent time segments. The Jensen-Shannon divergence is defined as follows: in Let KL divergence be denoted as KL divergence.

[0075] Eleven JSD values ​​were obtained, and a timing inconsistency curve was plotted.

[0076] Calculate the cumulative change (area under the curve). When the cumulative change exceeds the preset time stability threshold (0.8) (the cumulative change is calculated to be 1.2 in this example), and it is confirmed that no new version of the TCM diagnosis and treatment guidelines has been published during this period (which cannot be attributed to external knowledge updates), then the credibility of the expert's data source is deemed unqualified.

[0077] In this embodiment, the determination that the expert surnamed Li is unqualified triggers the system's self-healing signal, and step four outputs "System self-healing triggered: Data source [Expert Li XX] credibility is unqualified". The retraining process will then be initiated.

[0078] Step 5: Multi-dimensional risk linkage analysis and auxiliary decision-making output This embodiment combines the results of the above steps to demonstrate three typical scenarios: Scenario A (High-Risk Misdiagnosis): A patient simultaneously triggers both the diagnostic logic deviation warning in step two (the system infers "liver stagnation and spleen deficiency," while the doctor diagnoses "liver-stomach disharmony") and the prescription compatibility risk warning in step three (the risk score for Bupleurum and White Peony is 0.88). The risk linkage management unit detects the simultaneous presence of both, classifies it as a high-risk misdiagnosis scenario, generates a level-one linkage alarm, and forcibly blocks the issuance of the prescription in the electronic medical record system, requiring the doctor to complete a second review (e.g., fill out a review explanation or invite a senior doctor for consultation). The prescription function can only be unlocked after the second review is passed.

[0079] Scenario B (System Maintenance): When step four outputs a system self-healing trigger signal (data source unqualified), the risk linkage management unit determines it as a system maintenance scenario, suspends new early warning services (displaying "Self-healing in progress, real-time early warnings not provided temporarily" in the system backend), and simultaneously displays a maintenance prompt to the front-end administrator: "Data source expert Li XX's credibility has decreased, and full retraining of the dialectical deduction network has been initiated." Services will automatically resume after maintenance is completed.

[0080] Scenario C (Prescription Compatibility Risk Only): Regarding the Astragalus-Coptis compatibility risk warning output in step three of this embodiment, since no simultaneous diagnostic deviation warning appears, the system does not forcibly block it. Instead, it activates the traditional Chinese medicine knowledge graph retrieval engine. This knowledge graph stores the main therapeutic effects, properties, meridian tropism, and compatibility contraindications of medicinal materials. The retrieval engine searches for medicinal materials with similar effects to Coptis (clearing heat and drying dampness) and no compatibility contraindications with Astragalus, resulting in the candidate material "Scutellaria" (cold in nature but leaning towards clearing heat in the upper burner, with relatively high safety when used with Astragalus). The system outputs a replacement suggestion: "It is recommended to replace Coptis with Scutellaria. Reason: The patient has spleen and stomach deficiency and cold. Coptis is bitter and cold, which can damage the stomach. Scutellaria has a slightly milder heat-clearing effect, and there is literature support for the Astragalus-Scutellaria drug pair when used with Astragalus (from "Lan Shi Mi Cang"). The dosage reference range is 3-9g." Doctors can adopt the suggestion and directly modify the prescription.

[0081] Preferred embodiment two: This embodiment is basically the same as the first embodiment above, except that the specific implementation of the dynamic adjustment strategy for the preset confidence threshold in step three and the triggering details of the retraining process in step four are different.

[0082] In this embodiment, the dynamic threshold adjustment is as follows: For an elderly patient admitted to the hospital due to acute heart failure, their historical multidimensional state profile vector (Collected one month ago) and the current multidimensional state profile vector A Euclidean distance of 2.3 (the normal variation range after vector normalization is generally within 0.5) indicates a significant deterioration in the condition. The system automatically lowers the preset confidence threshold in step three from 0.75 to 0.55. At this point, the herbal medicine pair that originally scored 0.60 (which did not trigger an alert at the previous threshold of 0.75) is marked as suspicious due to a score of 0.60 > 0.55, triggering an alert. This ensures that critically ill patients receive more sensitive protection.

[0083] When the same patient's condition improves and When the distance is reduced to 0.3, the system restores the threshold to 0.75, reducing unnecessary alarms.

[0084] In this embodiment, the retraining process is as follows: When the prediction confidence index is detected to be lower than the first quality threshold (false positive rate > 15% or false negative rate > 10%) in step four, the system automatically expands the collection time window of real clinical outcome data (from 30 days to 90 days) and performs incremental fine-tuning training on the bilinear interaction function model based on the expanded positive and negative samples. That is, the original parameters are retained as initialization, and the new data is used to continue training for 10 epochs with a small learning rate (0.0001) to quickly adapt to the current clinical environment.

[0085] When a data source fails the time-series consistency check (such as the expert surnamed Li in Example 1), the system automatically removes all samples contributed by that expert from the original training sample library. Then, based on the cleaned sample library, the dialectical reasoning network is fully retrained (training the graph attention network from scratch until the loss converges). Full retraining takes longer, but it can completely eliminate the impact of contaminated data.

[0086] Preferred embodiment three: like Figure 2 As shown, the early warning system corresponding to the multi-dimensional early warning method for misdiagnosis risk in traditional Chinese medicine in the above embodiment includes: a status profile construction unit 1, a syndrome differentiation logic verification unit 2, a compatibility risk quantification unit 3, a system self-healing diagnosis unit 4, and a risk linkage management unit 5. The state profile construction unit 1 is used to perform step one; The dialectical logic verification unit 2 integrates the dialectical deduction network based on graph attention network, which is used to execute step two. The matching risk quantification unit 3 integrates the bilinear interaction function and its associated trainable parameters, and is used to execute step three. The system self-healing diagnostic unit 4 is used to perform the prediction confidence calculation and data source time series consistency verification in step four. The risk linkage management unit 5 is used to execute step five and display the warning level, blocking decision and auxiliary optimization information to the user.

[0087] The present invention has been described with reference to the foregoing embodiments and accompanying drawings; however, the foregoing embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, modifications and equivalents included within the spirit and scope of the claims are all included within the scope of the present invention.

Claims

1. A traditional Chinese medicine misdiagnosis risk multidimensional early warning method, characterized in that, Includes the following steps: Step 1: Construct a dynamic multidimensional state profile of the patient, specifically: collect the four diagnostic methods information, vital sign parameters, living environment data and past medical history information of the target object; extract features and numerically encode the collected data, and fuse the encoded feature vectors to generate a multidimensional state profile vector that represents the current comprehensive physiological-pathological-environmental state of the target object. Step Two: Perform a diagnostic path integrity check and logical consistency analysis. Specifically, the four diagnostic methods information is input into a pre-trained diagnostic reasoning network. The diagnostic reasoning network simulates the derivation process from symptoms to syndrome types based on a graph structure reasoning mechanism. During the derivation process, the diagnostic reasoning network performs the following two checks: If it detects that the currently input four diagnostic methods information is missing key symptom nodes relative to the preset standard diagnostic knowledge graph, and the absence of such nodes will cause the confidence of syndrome type inference to decrease by more than a preset threshold, then an incomplete diagnostic information warning is output; If the system syndrome type derived by the diagnostic reasoning network itself is inconsistent with the syndrome type diagnosed by the externally input doctor, then a diagnostic logic deviation warning is output. Step 3: Perform individualized prescription compatibility risk quantification assessment, specifically: obtain the traditional Chinese medicine prescription issued for the target object, map each herb in the prescription into a herb feature vector that integrates its properties, flavors, meridian tropism, chemical components, and target pathways; based on the multidimensional state profile vector and the herb feature vectors, use a bilinear interaction function to calculate the compatibility risk score between any two herbs in the prescription; When the compatibility risk score exceeds a preset confidence threshold, the pair of medicinal materials is marked as a suspicious compatibility combination, and a prescription compatibility risk warning is output. Step 4: Perform dynamic health diagnosis of the early warning system itself, specifically: periodically collect real clinical outcome data, compare the various early warning results output in Step 2 and / or Step 3 with the real clinical outcome data, calculate the overall false alarm rate and false negative rate of the system, and generate a prediction reliability index; at the same time, perform periodic verification based on temporal consistency on the original training sample library used to train the dialectical deduction network and the bilinear interaction function, and evaluate the stability and reliability of each data source; when the prediction reliability index is lower than the first quality threshold, or when the temporal consistency verification of a data source fails, output a system self-healing trigger signal and start part or all of the retraining process of the corresponding model.

2. The risk early warning method for TCM misdiagnosis according to claim 1, characterized in that, The specific methods for generating the multidimensional state profile vector in step one include: The collected information from the four diagnostic methods is converted into syndrome feature sub-vectors through a preset syndrome coding table; The collected vital signs parameters, including pulse, blood pressure, body temperature, and respiratory rate, are input into a pre-trained physiological state assessment network. The physiological state assessment network adopts a multi-layer fully connected structure to output a physiological state feature value between 0 and 1, which constitutes a physiological feature sub-vector. The training objective of the physiological state assessment network is to minimize the mean square error between the predicted physiological state and the healthy baseline. The collected living environment data is encoded into environmental feature sub-vectors; The syndrome feature sub-vector, physiological feature sub-vector, and environmental feature sub-vector are vertically concatenated to generate the multidimensional state profile vector.

3. The risk early warning method for TCM misdiagnosis according to claim 1, characterized in that, The dialectical reasoning network in step two is pre-trained in the following way: Construct a TCM syndrome knowledge graph, in which nodes include three types of medical entities: symptoms, syndrome elements, and syndrome types, and edges represent the derivation or membership relationships between entities; The knowledge graph is used to learn representations to obtain the embedding vector of each node; By performing a depth-first search-based random walk on the knowledge graph, multiple derivation path samples from symptom nodes through symptom element nodes to symptom type nodes are obtained. Using the derivation path samples as training data, the graph attention network is fine-tuned using a graph neural network sequence prediction task, so that the graph attention network has the ability to infer and output corresponding syndrome nodes from the input symptom node set along the graph path. After training convergence, the dialectical deduction network is obtained.

4. The risk of TCM misdiagnosis multi-dimensional early warning method according to claim 1, characterized in that, The triggering condition for the incomplete dialectical information warning in step two is: During its internal reasoning process, the dialectical deduction network identifies a pre-defined strong correlation edge in the knowledge graph between the currently input symptom node set and a certain uninputted symptom node. Through simulated reasoning, it determines that if the uninputted symptom node is used as an additional input, the probability of the syndrome type prediction output by the dialectical deduction network will increase beyond a preset gain threshold. Therefore, it is determined that there is a lack of key information and generates a warning of incomplete dialectical information.

5. The risk of TCM misdiagnosis multi-dimensional early warning method according to claim 1, characterized in that, The preset trust threshold in step three adopts a dynamic adjustment strategy: Obtain the historical multidimensional state profile vector and the current multidimensional state profile vector of the target object, and calculate the cosine similarity or Euclidean distance between the two as the state change amplitude. If the magnitude of the state change indicates that the current state has significantly deteriorated relative to the historical baseline, the value of the confidence threshold is automatically lowered to enhance the sensitivity of prescription risk warning. If the magnitude of the state change indicates that the current state is stable or improving, then the confidence threshold is maintained or increased to reduce false alarms.

6. The risk of TCM misdiagnosis multi-dimensional early warning method according to claim 1, characterized in that, Step three involves using a bilinear interaction function to calculate the compatibility risk score, including: For any two medicinal materials in the prescription and , respectively, obtain their medicinal material feature vectors and , and the multidimensional state portrait vector ; Will and Perform a fusion operation to obtain the fusion vector of the medicinal materials. ; The original risk estimate is calculated using the following bilinear function: in, Given a trainable bilinear parameter matrix, For bias terms; Original risk estimate The compatibility risk score is obtained by mapping to the (0,1) interval using the Sigmoid function.

7. The multi-dimensional early warning method for the risk of misdiagnosis in traditional Chinese medicine according to claim 1, characterized in that, The specific method for evaluating the stability and reliability of the data source based on time-series consistency verification in step four is as follows: For each data contributor in the original training sample library, the samples they provided are divided into multiple time segments according to the collection time, and each time segment contains several diagnostic samples. For each time segment, calculate the statistical distribution of the diagnostic conclusions within the corresponding time segment; Using a sliding window approach, the Jensen-Shannon divergence of the statistical distribution between adjacent time segments is calculated to obtain the temporal inconsistency curve; Calculate the cumulative change of the time-series inconsistency curve; when the cumulative change exceeds a preset time-series stability threshold and cannot be attributed to external medical knowledge updates, determine that the data source credibility of the data contributor is unqualified.

8. The multi-dimensional early warning method for TCM misdiagnosis risk according to claim 1, characterized in that, The specific strategy for initiating partial or complete retraining of the corresponding model in step four is as follows: When the triggering reason is that the prediction confidence index is lower than the first quality threshold, the collection time window of real clinical outcome data is automatically expanded, and the attention model associated with the bilinear interaction function is incrementally fine-tuned based on the expanded positive and negative samples. When the triggering reason is that the temporal consistency check of the data source fails, all samples contributed by the unqualified data source are automatically removed from the original training sample library, and the dialectical deduction network is fully retrained based on the cleaned sample library.

9. The multi-dimensional early warning method for the risk of misdiagnosis in traditional Chinese medicine according to claim 1, characterized in that, It also includes step five: multi-dimensional risk linkage analysis and auxiliary decision-making output, specifically: When steps two and three simultaneously output diagnostic logic deviation warnings and prescription compatibility risk warnings in the same diagnosis and treatment event, it is determined to be a high-risk misdiagnosis scenario, generating a first-level linkage alarm and blocking prescription output or forcing a second review. When only the system self-healing trigger signal is output in step four, it is determined to be a system maintenance scenario, the new early warning service is suspended and a model maintenance prompt is generated; When only a prescription compatibility risk warning is output, a traditional Chinese medicine knowledge graph retrieval engine is activated to find alternative medicinal materials with similar efficacy to the target medicinal material in the suspected compatibility combination and without known compatibility contraindications. At least one medicinal material replacement suggestion is output in combination with the multi-dimensional state profile vector, and the basis for the replacement is noted.

10. A multi-dimensional early warning system for the risk of misdiagnosis in Traditional Chinese Medicine, characterized in that, The method for implementing the method as described in any one of claims 1 to 9 includes: a state profile construction unit, a dialectical logic verification unit, a compatibility risk quantification unit, a system self-healing diagnosis unit, and a risk linkage management unit; The state profile construction unit is used to perform step one; The dialectical logic verification unit integrates the dialectical deduction network based on graph attention network, which is used to execute step two. The matching risk quantification unit integrates the bilinear interaction function and its associated trainable parameters, and is used to perform step three. The system self-healing diagnostic unit is used to perform the prediction confidence calculation and data source time series consistency verification in step four. The risk linkage management unit is used to execute step five and display the warning level, blocking decision and auxiliary optimization information to the user.

Images