A traditional Chinese medicine health assessment method and system for improving stability of diagnosis results

Abstract

The application provides a traditional Chinese medicine health assessment method and system for improving the stability of diagnosis results, which comprises the following steps: collecting traditional Chinese medicine detection physiological characteristic data of a user to obtain a standardized diagnosis input feature vector; inputting the input feature vector into a pre-trained traditional Chinese medicine diagnosis large model to obtain an initial diagnosis result set; judging the initial diagnosis result based on a pre-set constitution transition zone; when the initial diagnosis result is judged to be in an unstable state, starting multi-round reasoning consistency verification based on a pre-set diagnosis constraint rule; comparing the verified diagnosis result with the constitution stable zone to obtain a final diagnosis result; and generating a traditional Chinese medicine health assessment report based on the final diagnosis result. The application can effectively identify the gray area of slight changes or transition periods of constitution, improve the stability of diagnosis results through multi-round independent reasoning and weighted correction, and solve the technical problem of insufficient stability of diagnosis results in the existing AI traditional Chinese medicine health assessment technology.

Description

Technical Field

[0001] This invention relates to the field of digital healthcare, specifically to a method and system for traditional Chinese medicine health assessment that improves the stability of diagnostic results. Background Technology

[0002] With the deep integration of artificial intelligence technology and the digitalization of traditional Chinese medicine, AI-based TCM health assessment and diagnosis technology has been implemented on a large scale. Existing TCM health assessment technologies have constructed TCM syndrome differentiation models based on massive clinical and social samples, realizing the digitalization of TCM constitution identification, organ syndrome differentiation, and qi, blood, and body fluid syndrome differentiation.

[0003] However, the core of current AI-based TCM health assessment technology research and development generally focuses on improving the accuracy of diagnosis results. This is achieved by expanding the training sample size, optimizing model algorithms, and improving the TCM knowledge base, thereby continuously increasing the consistency between single test results and clinical TCM physician diagnoses. However, in actual commercial and clinical applications, a core pain point has been found in the existing technology: when the same user undergoes multiple tests within a short period of time, the diagnosis results are prone to irregular jumps. For example, the first test might indicate phlegm-dampness constitution (heavy dampness), the second test 10 minutes later might indicate qi deficiency constitution, and the third test might revert back to phlegm-dampness constitution. Such jumps lack clinical physiological support and are entirely due to technical instability.

[0004] Analysis revealed that the core reasons for the abrupt change in the diagnostic results included: Input feature critical fluctuation: When the user's detected feature value is in the critical range of the syndrome determination threshold, even a small acquisition error, environmental interference, or physiological fluctuation can cause the feature value to cross the determination threshold, directly triggering a reversal of the syndrome diagnosis result; The probability distributions within the model are close: when the user's physical condition is in the boundary range between the two types of syndromes, the predicted probability values ​​of the two types of syndromes are extremely close. Even a small random perturbation in a single round of reasoning will cause the syndrome label in the final output to switch. Lack of understanding of the patterns of changes in constitution: In traditional Chinese medicine theory, changes in constitution are a long-term and gradual process. Fluctuations in test values ​​within a short period of time (7 days) do not represent actual changes in constitution or syndrome. Current technology does not incorporate this basic theory of traditional Chinese medicine into the model constraints and directly outputs short-term fluctuations as changes in constitution. The output lacks a stable constraint mechanism: the existing large model inference process only performs a single round of inference, without an internal consistency check or the introduction of time-series constraints from user historical data. The error and fluctuation of a single round of inference will be directly reflected in the final output result, without any stability control mechanism.

[0005] In summary, existing AI-based TCM health assessment technologies cannot effectively address the issue of rapidly changing diagnostic results while maintaining diagnostic accuracy. This leads users to question the product's professionalism, reduces user trust, and severely restricts the large-scale commercialization and clinical promotion of TCM digital technologies. Summary of the Invention

[0006] To address the technical problem of insufficient stability of diagnosis results in existing AI-based TCM health assessment technologies and to achieve dynamic correction and long-term stabilization of TCM diagnosis results, this invention provides a TCM health assessment method and system that improves the stability of diagnosis results.

[0007] The technical solution adopted by this invention to solve its technical problem is: A method for TCM health assessment to improve the stability of syndrome differentiation results includes the following steps: S1, collecting users' TCM physiological characteristic data, completing preprocessing and feature extraction to obtain standardized syndrome differentiation input feature vectors; S2, inputting the input feature vectors into a pre-trained TCM syndrome differentiation model to obtain an initial syndrome differentiation result set, calculating the preset confidence interval of each syndrome differentiation result, and judging the initial syndrome differentiation result based on a preset constitution transition zone. When the initial syndrome differentiation result falls into the constitution transition zone, it is judged as an unstable state; S3, when the initial syndrome differentiation result is judged as an unstable state, multi-round inference consistency verification is initiated based on preset syndrome differentiation constraint rules, through multiple independent rounds... The process involves: S1) obtaining multiple sets of diagnostic results and calculating a consistency score; S2) triggering recalculation or weight adjustment when multiple sets of diagnostic results do not meet the diagnostic constraint rules, and performing weighted calibration on the multiple sets of diagnostic results to obtain the verified diagnostic results; S3) acquiring the user's historical diagnostic result sequence, constructing a constitution stability interval model, and comparing the verified diagnostic results with the constitution stability interval under the diagnostic constraint rules. When the constraint conditions are not met or the deviation threshold is exceeded, the diagnostic results are corrected in time series to obtain the final diagnostic result; S4) generating a traditional Chinese medicine health assessment report based on the final diagnostic result and writing the current diagnostic result into the historical sequence to update the constitution stability interval model.

[0008] Preferably, in step S1, the TCM physiological characteristic data includes collected visceral function detection data, tongue image data, user complaint information, and Western medicine detection index data; preprocessing includes outlier removal, K-nearest neighbor algorithm missing value filling, and Z-score standardization; feature extraction includes extracting tongue diagnosis feature vectors and visceral feature vectors, and concatenating them to obtain a standardized input feature vector.

[0009] Furthermore, in step S2, the initial diagnostic result set includes the predicted probability of TCM constitution type and the value of core diagnostic indicators. Based on the preset confidence interval, it is determined whether it falls into the constitution transition zone. If it falls into the constitution transition zone, it is determined to be a high jump risk result and enters the multi-round inference consistency verification. Otherwise, it directly enters the historical constitution smoothing constraint process.

[0010] Optionally, in step S3, each round of independent reasoning is generated by adding a small random perturbation or Gaussian noise to the input feature vector and using different random seeds, and each round of reasoning obtains an independent set of dialectical results and corresponding confidence intervals.

[0011] Preferably, in step S3, the weighted calibration of multiple sets of dialectical results includes: calculating a weighted average value based on the confidence interval width and index correlation of each round of results to form a verified dialectical result for historical interval comparison.

[0012] Furthermore, in step S4, the historical diagnostic result sequence consists of the user's valid diagnostic results within the past 6 months, and the interval between two adjacent tests is not less than 24 hours; the physical stability interval model is formed by calculating the distribution interval of each diagnostic indicator in the historical sequence and setting dynamic boundaries.

[0013] Optionally, in step S4, the time-series correction is achieved through a filtering algorithm, using the historical mean as the prior state and the current verification result as the observed value. The correction value is calculated independently for each dialectical indicator, while maintaining the relative order between indicators and the core syndrome combination relationship unchanged.

[0014] Preferably, in step S5, the TCM health assessment report includes the constitution assessment result, health index, core risk indicators, tongue diagnosis characteristics interpretation, and recommended conditioning plan; the recommended conditioning plan includes a basic conditioning plan and a special conditioning plan.

[0015] This invention also provides a TCM health assessment system to improve the stability of diagnostic results, comprising: a data processing module for collecting users' TCM physiological characteristic data, performing preprocessing and feature extraction to obtain standardized diagnostic input feature vectors; an initial diagnostic module for inputting the standardized diagnostic input feature vectors into a pre-trained TCM diagnostic model to obtain an initial diagnostic result set, and determining whether the initial diagnostic results are in an unstable state based on a preset constitution transition zone; a multi-round verification module for initiating multi-round independent inference consistency verification for unstable diagnostic results falling into the constitution transition zone, generating multiple sets of diagnostic results and recording the confidence intervals of each set of results; a weighted calibration and time-series correction module for performing weighted calibration on the multi-round diagnostic results, comparing the verified diagnostic results with a constitution stability interval model constructed from the user's historical diagnostic result sequence, performing time-series correction on indicators exceeding the range, and obtaining the final diagnostic result; and a report generation and result management module for generating a TCM health assessment report based on the final diagnostic result and writing the current diagnostic result into the historical sequence to update the constitution stability interval model.

[0016] Furthermore, the data processing module further includes a tongue image feature extraction unit and an organ function feature extraction unit, used to extract tongue diagnosis feature vectors and organ function feature vectors respectively, and concatenate them to form a standardized input feature vector; the initial diagnosis module is further configured with a confidence interval calculation unit, used to calculate an independent confidence interval for the predicted probability of each diagnosis indicator and constitution type, and determine the stability of the result based on a preset constitution transition zone; the multi-round verification module further includes a random perturbation generation unit, used to add small random perturbations or Gaussian noise to the input feature vector of each round of independent inference, and record the confidence interval of the diagnosis result of each round; the weighted calibration and time-series correction module further includes a weighted calculation unit and a filtering correction unit, used to perform weighted calibration based on the confidence interval and indicator correlation of the results of each round, and to perform time-series correction on the verified results and the historical constitution stability interval; the report generation and result management module further includes a data storage unit, used to write the final diagnosis result into the historical sequence and update the constitution stability interval model.

[0017] The beneficial effects of the present invention include at least the following: This invention integrates the determination of the constitution transition zone, multi-round inference consistency verification, and a history-driven constitution stability interval model into a single method and system, significantly improving the reliability and practicality of diagnosis. It effectively solves the problem of rapidly changing diagnosis results and avoids misjudgments from single model outputs. Through multi-round inference consistency verification and weighted correction, it significantly improves the accuracy and reliability of diagnosis results. The history-driven constitution stability interval model allows diagnosis accuracy to gradually improve over time, forming a closed-loop optimization mechanism. It also addresses user concerns about the product's professionalism, providing technical support for the large-scale commercialization and clinical promotion of TCM digital technology.

[0018] Compared with existing technologies, this invention can effectively identify gray areas during periods of slight changes or transitions in physical constitution, improve the stability of diagnostic results through multiple rounds of independent reasoning and weighted correction, and utilize historical data to construct a stable interval model of physical constitution to achieve dynamic and adaptive stabilization of diagnostic results, thereby solving the technical problem of insufficient stability of diagnostic results in existing AI TCM health assessment technologies.

[0019] This invention is fully compatible with existing TCM digital hardware and software systems. It can be directly embedded into existing hardware products such as Qiaolangzhong and Daqi AI, store management systems, and CDSS clinical decision support systems without requiring modifications to existing data acquisition hardware or basic models. Upgrade costs are extremely low, enabling rapid large-scale deployment. It is simultaneously adaptable to B2B medical institutions and wellness centers, as well as B2C home smart terminals, addressing the core pain point of insufficient user trust in commercial scenarios. This significantly enhances the product's professional image and provides core technical support for the large-scale promotion of TCM digital technology. Attached Figure Description

[0020] Figure 1 This is a flowchart of a traditional Chinese medicine health assessment method for improving the stability of diagnostic results according to the present invention; Figure 2 This is a flowchart of the data acquisition and feature extraction process in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the comparison between initial dialectical reasoning and confidence intervals in an embodiment of the present invention; Figure 4 This is a flowchart of the multi-round inference consistency verification process in an embodiment of the present invention; Figure 5 This is a flowchart of the historical sequence comparison and timing correction process in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] This invention provides a traditional Chinese medicine health assessment method that improves the stability of diagnostic results, such as... Figure 1 As shown, the steps include: S1. Collect users' TCM physiological characteristic data, complete preprocessing and feature extraction, and obtain standardized syndrome differentiation input feature vectors; S2. Input the input feature vector into the pre-trained TCM syndrome differentiation model to obtain an initial syndrome differentiation result set, calculate the preset confidence interval of each syndrome differentiation result, and judge the initial syndrome differentiation result based on the preset constitution transition zone. When the initial syndrome differentiation result falls into the constitution transition zone, it is judged as an unstable state. The constitution transition zone refers to the numerical range located at the edge of the constitution stable range, used to characterize the stage of slight change or transformation of the user's constitution. S3. When the initial dialectical result is determined to be unstable, a multi-round inference consistency check is initiated based on the preset dialectical constraint rules. Multiple sets of dialectical results are obtained through multiple rounds of independent inference, and a consistency score is calculated. When multiple sets of dialectical results do not meet the dialectical constraint rules, a recalculation or weight adjustment is triggered to perform weighted calibration on the multiple sets of dialectical results to obtain the verified dialectical result. To solve the problem of critical fluctuations in input features, each round of independent inference generates small random perturbations or Gaussian noise by adding them to the input feature vector, so as to avoid the influence of threshold edge fluctuations on a single prediction.

[0023] S4. Obtain the user's historical diagnostic result sequence, construct a physical stability interval model, and compare the verified diagnostic result with the physical stability interval under the diagnostic constraint rules. When the constraint conditions are not met or the deviation threshold is exceeded, perform time-series correction on the diagnostic result to obtain the final diagnostic result. S5. Generate a TCM health assessment report based on the final diagnosis results, and write the diagnosis results into the historical sequence to update the constitution stability interval model.

[0024] The data acquisition and feature extraction in S1, such as Figure 2 As shown, the implementation specifically includes: collecting the user's TCM physiological characteristic data, which includes organ function detection data, tongue image data, pulse characteristic data collected by AIoT smart hardware, user-inputted chief complaint symptom information, and user-provided Western medicine detection indicator data.

[0025] The AIoT smart hardware includes smart health checkup devices, digital diagnostic instruments, tongue image acquisition devices, and electronic blood pressure monitors. The smart health checkup devices and digital diagnostic instruments are equipped with multi-channel physiological signal acquisition units, enabling simultaneous acquisition of the user's organ function detection data and pulse characteristic data. They feature built-in standardized acquisition guidance programs and quality control rules to standardize user operation processes, reduce fluctuations in characteristic data caused by human error, and ensure consistency between single and multiple tests. The tongue image acquisition device is equipped with a high-resolution RGB camera module and a standardized color temperature supplementary lighting unit, enabling the acquisition of tongue surface image data conforming to the standards of traditional Chinese medicine tongue diagnosis. It automatically completes tongue area recognition and removes invalid and environmentally interfering images, ensuring the quality of basic data for tongue feature extraction. Supporting physiological detection devices such as electronic blood pressure monitors can simultaneously acquire the user's heart rate, blood pressure, and other Western medicine detection indicators, achieving simultaneous acquisition of multi-dimensional traditional Chinese medicine physiological characteristics and Western medicine auxiliary indicators, with unified format and timestamp alignment.

[0026] The collected raw data were preprocessed, and outlier screening and removal were performed using the Hat Matrix and Mahalanobis Distance methods to eliminate abnormal data caused by collection errors. The K-nearest neighbor algorithm was used to fill in missing values ​​to ensure data integrity. The Z-score standardization method was used to map all feature data to a standard interval with a mean of 0 and a variance of 1 to eliminate dimensional differences.

[0027] Specifically, for the collected comprehensive TCM physiological characteristic data, including organ function test values, tongue appearance quantitative characteristic values, and Western medicine physiological index data, the data is first matrix-organized. All sample feature data are organized into a unified-dimensional feature matrix, with each row corresponding to the full features of a single sample and each column corresponding to a standardized test indicator. This simultaneously adapts to the correlation between multi-dimensional TCM diagnostic indicators, eliminating judgment bias caused by differences in feature dimensions. Global abnormal sample screening is first performed using Mahalanobis Distance. The Mahalanobis distance of each sample data point relative to the overall feature dataset is calculated. Mahalanobis distance can simultaneously avoid interference from correlations between features. A preset critical threshold for Mahalanobis distance is set, based on the 95th percentile of a chi-square distribution with degrees of freedom equal to the feature dimension. When the Mahalanobis distance of a sample exceeds this critical threshold, it is judged as a globally abnormal sample with no clinical reference significance, and the invalid data of that sample is removed entirely.

[0028] Then, the Hat Matrix is ​​used to accurately screen local outliers. For each effective detection index data in a single sample, a hat matrix based on least squares regression is constructed to calculate the leverage value of each data point. The leverage value is used to quantify the influence of a single index data on the overall feature fitting result. A threshold for the judgment of the leverage value is preset. The threshold is set to 3 times the ratio of the feature dimension to the sample size. When the leverage value of a single index data point exceeds the threshold, it is judged as a local outlier caused by the collection error. The outlier index data is removed, while the other normal index data in the same sample are retained.

[0029] After completing the entire process of outlier screening and removal, the missing data bits generated after removal are marked simultaneously. This provides a clear basis for the subsequent use of the K-nearest neighbor algorithm to fill in the missing values, ensuring the integrity of the preprocessed data and the effectiveness of the subsequent feature extraction process.

[0030] Feature extraction is performed on the preprocessed data. For tongue images, a pre-trained convolutional neural network (CNN) is used to extract core tongue diagnosis features such as tongue color, tongue coating, tongue shape, teeth marks, and enlargement, outputting a 128-dimensional tongue diagnosis feature vector. For visceral function detection data, competitive adaptive reweighted sampling (CARS) and partial least squares linear regression (PLS) are used for feature selection to extract core features strongly correlated with TCM syndrome differentiation, outputting a 64-dimensional visceral feature vector. The tongue diagnosis feature vector and the visceral feature vector are concatenated to obtain a 192-dimensional standardized syndrome differentiation input feature vector X.

[0031] The pre-trained convolutional neural network adopts a residual network architecture optimized to meet the needs of TCM tongue image feature extraction. It is pre-trained based on tongue image samples labeled by clinical doctors in hospitals and tongue image samples from the general population. During the pre-training process, supervised training is completed simultaneously with tongue image feature labels labeled by TCM doctors to ensure that the extracted features are strongly correlated with the needs of TCM diagnosis. The network is composed of five cascaded components: input layer, multi-scale residual convolutional feature extraction module, pooling layer, fully connected layer and feature output layer.

[0032] The input layer receives the preprocessed standardized tongue surface RGB image, scales the image resolution to 224×224, completes pixel value normalization, maps pixel values ​​to the range of 0 to 1, and performs color space standardization correction to eliminate interference caused by color shift and lighting differences from different acquisition devices, ensuring the consistency of input data.

[0033] The multi-scale residual convolution feature extraction module consists of four cascaded residual convolutional units. Each residual convolutional unit contains two sets of convolutional kernels, a batch normalization layer, an activation function layer, and a shortcut direct connection branch. The first set of convolutional kernels is 1×1 in size and is used to adjust the dimensionality of the feature channels. The second set of convolutional kernels is 3×3 in size and is used to extract local detail features of the tongue image. The batch normalization layer is used to alleviate the gradient vanishing problem during network training. The ReLU activation function is used to improve the nonlinear fitting ability of the network. The shortcut direct connection branch is used to ensure the complete transmission of feature information and avoid the feature degradation problem of deep networks. The four residual convolutional units output multi-scale feature maps of 64, 128, 256, and 512 dimensions respectively, corresponding to the global contour features of the tongue image, the details of the tongue shape and teeth marks, the color features of the tongue and the color of the tongue coating, and the thickness and texture of the tongue coating, fully covering the core feature dimensions required for TCM tongue diagnosis.

[0034] At the end of each residual convolutional unit, a global average pooling layer with a stride of 2 is set to downsample the feature map of the corresponding dimension, compress the spatial size of the feature map, retain the core effective feature information, and reduce the computational cost and parameter count of the network, ensuring the real-time performance of feature extraction and adapting to the low-computing deployment requirements of home smart terminals and store inspection equipment.

[0035] The fully connected layer and feature output layer take the 512-dimensional global features output from the last pooling layer and input them into two cascaded fully connected layers. The first fully connected layer maps the feature dimension to 256 dimensions, and the second fully connected layer further maps the feature dimension to 128 dimensions. At the same time, a dropout layer is set between the two fully connected layers with a dropout ratio of 0.3 to alleviate the overfitting problem of the network and improve the generalization ability of feature extraction. Finally, a standardized 128-dimensional tongue diagnosis feature vector is output, which can be directly concatenated with the visceral feature vector to construct the standardized input feature vector required for diagnosis.

[0036] For the organ function test data after outlier removal, missing value imputation and Z-score standardization, the gold standard results of TCM syndrome differentiation labeled by clinical TCM physicians were used as the response variable. The gold standard results include the results of 9 TCM constitution types and 20 core organ and qi, blood and body fluid differentiation indicators. First, a regression model between feature variables and syndrome differentiation response variables was constructed using PLS to quantify the correlation between each collected feature and TCM syndrome differentiation results. At the same time, the inherent multicollinearity interference among TCM multidimensional test indicators was eliminated, feature components with significant explanatory power for syndrome differentiation results were extracted, and noisy features and redundant data that are not clinically related to TCM syndrome differentiation were filtered out.

[0037] Based on this, CARS is used to carry out iterative feature optimization. 50 Monte Carlo sampling cycles are set. In each cycle, 80% of the samples are randomly selected as the training set and the remaining 20% ​​of the samples are selected as the validation set. PLS dialectical regression model is built on the training set. The absolute value of the regression coefficient of each feature variable is calculated. This is used as the core evaluation basis for the contribution of the feature to the dialectical result. The feature variables with high contribution are gradually screened and retained through the exponential decay function. At the same time, low-value and easily disturbed features with absolute regression coefficients close to 0 are removed. The dialectical prediction error of the corresponding feature subset is calculated through 10-fold cross-validation in each sampling cycle. Finally, the optimal feature subset with the smallest cross-validation error is selected.

[0038] The optimal feature subset obtained by screening is normalized by a pre-trained PLS mapping model, which maps the high-dimensional feature data to a 64-dimensional standardized vector space and outputs a fixed-dimensional visceral feature vector. This vector can be directly concatenated with the 128-dimensional tongue diagnosis feature vector to construct the standardized input feature vector required for diagnosis. At the same time, redundant variables that are prone to cause fluctuations in results are removed by feature screening, which reduces the interference of irrelevant data disturbances on the stability of diagnosis results from the feature input level and reduces the problem of jumps in diagnosis results caused by critical fluctuations in features.

[0039] like Figure 3 As shown, S2 specifically includes the following steps in implementation: inputting the standardized input feature vector X into the pre-trained TCM syndrome differentiation model. The pre-trained TCM syndrome differentiation model is continuously iteratively trained based on 157,000 hospital clinical samples and 8 million social samples. It adopts a hybrid modeling method combining partial least squares linear regression (PLS) and support vector machine (SVM) to output the initial syndrome differentiation result set Y.

[0040] The initial diagnostic result set Y includes numerical results for 9 TCM constitution types and 20 core diagnostic indicators. The predicted probability values ​​P=[p1,p2,...,p9] for the 9 TCM constitution types correspond to the balanced constitution, yang deficiency constitution, yin deficiency constitution, qi deficiency constitution, phlegm-dampness constitution, damp-heat constitution, blood stasis constitution, qi stagnation constitution, and special constitution, respectively. The numerical results S=[s1,s2,...,s9] for the 20 core diagnostic indicators are... 20 These correspond to spleen deficiency, lung deficiency, kidney deficiency, liver deficiency, heart deficiency, stomach deficiency, large intestine deficiency, small intestine deficiency, bladder deficiency, gallbladder deficiency, qi deficiency, qi stagnation, blood deficiency, blood stasis, body fluid deficiency, body fluid stagnation, cold, heat, yin deficiency, and yang deficiency, respectively. Each indicator is matched with a clinically validated standard reference range.

[0041] Specifically, the hybrid architecture of the TCM syndrome differentiation model adopts a two-level cascade structure of pre-feature mapping and post-syndrome differentiation reasoning. First, the standardized input feature vector X is input into the first-level PLS feature mapping module. This module uses the clinical gold standard syndrome differentiation result as the response variable to construct a multiple linear regression mapping relationship between the input features and the syndrome differentiation target. Simultaneously, it eliminates the inherent multicollinearity interference among TCM multidimensional detection indicators, filters out noisy features and redundant variables that are not clinically related to the syndrome differentiation target, extracts the core feature components that have significant explanatory power for the syndrome differentiation result, and outputs a low-dimensional, high-discrimination syndrome differentiation feature matrix with regular dimensions. This reduces the impact of irrelevant perturbations on the stability of subsequent reasoning results from the feature input level.

[0042] The second level is a parallel dual-branch SVM diagnostic reasoning module. The low-dimensional diagnostic feature matrix output from the first level is synchronously input into two parallel branches: classification and regression. The classification branch uses a radial basis function optimized to fit the characteristics of TCM constitution classification as the kernel function to complete the probability prediction of 9 TCM constitution types. It outputs the standardized predicted probability value for each constitution type: balanced, yang deficiency, yin deficiency, qi deficiency, phlegm-dampness, damp-heat, blood stasis, qi stagnation, and special constitution. The regression branch uses a linear kernel function optimized to fit the numerical values ​​of the diagnostic indicators to complete the numerical fitting of 20 core diagnostic indicators of heart and visceral deficiency and qi, blood, and body fluid status. It outputs the standardized quantitative result for each indicator. The outputs of the two branches together form a complete initial diagnostic result set Y.

[0043] During model pre-training and iterative optimization, ten-fold cross-validation was used to coordinate the optimization of the two-level modules. Based on the feature selection results of the PLS module, the kernel function parameters, penalty coefficients and slack variables of the two branches of SVM were optimized simultaneously to minimize the random fluctuation of the model inference results while ensuring the consistency between the diagnostic results and the clinical gold standard.

[0044] Furthermore, confidence intervals are calculated for the initial diagnostic result set Y. Specifically, in the pre-trained TCM diagnostic model, a corresponding confidence interval calculation mechanism is constructed for the prediction results of each diagnostic indicator and constitution type to characterize the uncertainty range of the model output results.

[0045] For the numerical results of the 20 core dialectical indicators S=[s1,s2,...,s20], calculate the confidence interval [CI] for each indicator. lowi CI highi For the predicted probabilities P = [p1, p2, ..., p9] of the nine TCM constitution types, calculate the confidence interval [PI] for each constitution type. lowi ,PI highi The confidence interval is used to reflect the stability and fluctuation range of the current diagnostic results.

[0046] After obtaining the confidence intervals for each diagnostic result, the initial diagnostic result is determined based on a preset constitution transition zone. The constitution transition zone is a numerical range located at the edge of the stable constitution range, used to characterize a user's constitution in a stage of slight change or transformation. It is preset according to the standard reference range and historical statistical characteristics of each diagnostic indicator.

[0047] Specifically, the values ​​of each diagnostic indicator and the predicted probability of constitution type in the initial diagnostic result set Y are compared with the corresponding constitution transition zone. When the confidence interval of any diagnostic indicator or the confidence interval of the predicted probability of constitution type falls into the constitution transition zone, the initial diagnostic result is determined to be in an unstable state; when all diagnostic results do not fall into the constitution transition zone, it is determined to be in a stable state.

[0048] When the state is determined to be unstable, the process proceeds to the multi-round inference consistency verification process in step S3; when the state is determined to be stable, the process proceeds directly to the physical stability interval model constraint process in step S4.

[0049] like Figure 4 As shown, step S3 is used to perform stability verification and optimization on the preliminary identification result Y obtained in step S2, so as to reduce the impact of high jump risk and random disturbance on the final identification result.

[0050] Assume that step S2 has generated an initial diagnostic result set Y for a patient, containing indicator matrices for constitution differentiation, visceral differentiation, qi, blood, and body fluid differentiation, eight-principle differentiation, and corresponding symptom-symptom dimensions. In this example, the preliminary results show that the probability of the main constitution syndrome type falls within the preset critical range, and the matching degree between visceral syndrome and the eight-principle differentiation dimensions is insufficient. Step S3 is executed according to the following process: S31. Preparation for multi-round independent reasoning: Take the initial dialectical result Y output from S2 as input, and generate 5 sets of perturbation input feature vectors X1, X2, ..., X5 based on it. Each set uses an independent random seed (e.g., seed1=123, seed2=456, ...). Add Gaussian perturbation with an amplitude of 1% of the feature standard deviation to the core features to reduce the impact of the randomness of single-round reasoning on the results.

[0051] S32. Independent Inference for Differentiation Dimensions: Each set of perturbation input vectors is input into the TCM differentiation model, resulting in 5 sets of independent differentiation results Y1, Y2, ..., Y5. Each result contains a complete indicator matrix for the five differentiation dimensions. In this example, the probability of the main syndrome in the constitution differentiation dimension is between 0.45 and 0.48, two core indicators in the organ differentiation dimension fall into the critical range, the difference between cold and heat in the eight principles differentiation dimension is only 0.08, and the symptom-syndrome matching degree is less than 70%. These situations trigger the dimension enhancement mechanism, increasing the number of independent inference rounds for the relevant differentiation dimensions and adjusting the feature weights for weighted calibration.

[0052] S33. Consistency Score Calculation: A stratified weighted rank consistency score is calculated for the results of the five rounds of diagnosis. Rank consistency is calculated for core critical features (probability of main constitution syndrome, key organ indicators); numerical consistency is calculated by globally normalizing ordinary features. The overall consistency score SC is obtained through the deviation between the two layers of features. In this example, the preliminary calculation result SC=0.82, which is lower than the preset threshold of 0.95, indicating that there are significant differences in the results across multiple rounds, requiring calibration.

[0053] S34. Consistency Calibration: Based on the confidence interval of each round's results, assign weights to the 5 groups of dialectical results (higher confidence levels result in greater weights), perform a weighted average, and generate the final consistency verification dialectical result Y. cal In this example, after calibration, the probability of the main constitution syndrome type stabilized at 0.46, the matching degree of the viscera and the eight principles indicators increased to over 80%, and the consistency between the syndrome dimension and other dimensions was significantly enhanced.

[0054] like Figure 5 As shown, S4 is used in implementation to further improve the temporal stability and reliability of the diagnostic results by comparing and correcting them with historical diagnostic sequences and stable constitution intervals. Kalman filtering or weighted moving average algorithms are used to perform temporal correction on each diagnostic indicator, while maintaining the core syndrome combination relationship and the relative order between indicators. The specific implementation is as follows: S41. Acquisition and Construction of Historical Dialectical Sequences The system retrieves the historical diagnostic result sequence {Y} of the same patient from the patient health management database. t 2,Y t 1,Y t (This may include the results of the most recent 3 or more diagnosis sessions), where each result contains a constitution probability matrix of five diagnosis dimensions and the values ​​of each diagnosis indicator. Based on these historical results, a patient constitution stability interval model B is constructed, defining the high and low fluctuation range of the probability of each constitution type and the allowable deviation threshold of the core diagnosis indicators.

[0055] S42. Preliminary comparison and deviation identification The Ycal output from step S3 is compared dimension by dimension with the body stability interval B: For the physical constitution dimension: determine whether the probability of Ycal's main certificate type falls within the stable range of this dimension; For the dimensions of internal organs, qi, blood and body fluids, eight principles, and symptoms: compare whether the core indicator values ​​are within the allowable fluctuation range of the historical interval.

[0056] If the result of any dimension exceeds the preset deviation threshold, it is marked as a "deviation event" and the correction process is triggered.

[0057] S43. Timing Correction and Weighting Adjustment For dimensions of labeling bias, the system performs the following correction strategy: Weighted time series fusion: combining historical dialectical results with the current Y cal The results are weighted averaged, and the weights can be set according to the confidence level of historical results and the time decay function. For example, the most recent result has a higher weight, and the more recent result has a lower weight.

[0058] Core indicator protection: For core indicators with high risk of jumps, priority is given to maintaining the core evidence type direction after calibration in S3 reasoning, and only non-core indicators are numerically adjusted to ensure that the main evidence type does not change in the opposite direction.

[0059] Overall consistency assessment: Calculate the cosine similarity or rank consistency score between the corrected result and the historical sequence to ensure that the corrected result still meets the consistency principle of multi-round inference.

[0060] S44. Output the final dialectical result: After correction, the system outputs the final dialectical result Y. final .

[0061] Specifically, S5 includes the following in its implementation: based on the final dialectical result Y final The system completes the final TCM syndrome differentiation and constitution classification, generating a standardized TCM health assessment report. The report includes: constitution assessment results, health index, core risk indicators, tongue diagnosis characteristics interpretation, and recommended treatment plans. These recommended plans include basic treatment plans (dietary recommendations, acupoint therapy, health exercises) and specialized treatment plans (medicinal foods, health-preserving teas, moxibustion, massage, and prepared Chinese medicines), fully adaptable to clinical application scenarios. The report generation is based on the final syndrome differentiation results after multiple rounds of verification and historically driven constitution stability range correction, thus ensuring the scientific validity and stability of the output report.

[0062] At the same time, the final diagnosis results will be stored in the user's historical database, and the user's physical stability interval model will be updated to provide a basis for time-series smoothing constraints for subsequent tests, forming a closed-loop continuous optimization mechanism.

[0063] In one specific embodiment, this embodiment is a complete method implementation for a home smart terminal scenario. The main implementing entity is the Qiaolangzhong Home Traditional Chinese Medicine Physical Examination Instrument and its supporting cloud service system. The specific implementation steps are as follows: S1, Data Acquisition and Feature Extraction: In this example, the user completes organ function testing using the Qiaolangzhong Home-use Traditional Chinese Medicine Physical Examination Instrument, captures tongue images using the tongue image acquisition module of the accompanying APP, and simultaneously fills in the user's chief complaint information within the APP. The raw data collected by the system includes: Internal organ function test data: raw test values ​​of 20 core diagnostic indicators, including spleen deficiency, lung deficiency, kidney deficiency, liver deficiency, heart deficiency, and stomach deficiency; Tongue image: A 1080×1440 resolution RGB image of the tongue surface, including complete information on the tongue body and tongue coating; Chief complaint information: The user filled in three core symptoms: "chills, fatigue, and loss of appetite".

[0064] Preprocessing of the raw data includes outlier filtering, missing value imputation, and standardization: The Hat Matrix and Mahalanobis Distance methods were used to screen out outliers for 20 organ detection indicators, and data that exceeded the clinically reasonable range were removed. No outliers were found in this embodiment. The K-nearest neighbor algorithm is used to fill in the missing items in the chief complaint information. In this embodiment, there are no missing values. The Z-score standardization method was used to map 20 organ detection indicators to a standard interval with a mean of 0 and a variance of 1, thus eliminating dimensional differences.

[0065] Feature extraction is performed on the preprocessed data: The preprocessed tongue image is input into a pre-trained ResNet50 convolutional neural network to extract core features such as tongue color, tongue coating thickness, tongue coating color, tongue teeth marks, and tongue enlargement, outputting a 128-dimensional tongue diagnosis feature vector. The CARS and PLS methods are used to filter features from 20 standardized organ detection indicators, extracting 16 core features that are strongly correlated with syndrome differentiation, and outputting a 64-dimensional organ feature vector after dimensionality reduction. The 128-dimensional tongue diagnosis feature vector is concatenated with the 64-dimensional organ feature vector, and the unique heat encoding features of the chief complaint symptoms are also incorporated to finally obtain a 192-dimensional standardized syndrome differentiation input feature vector X.

[0066] S2, input the input feature vector X into the pre-trained TCM syndrome differentiation model, and the model outputs the initial syndrome differentiation result set Y.

[0067] The core content of the initial diagnostic result set Y includes the predicted probability of constitution type: Yang deficiency constitution 0.52, Qi deficiency constitution 0.48, balanced constitution 0.03, and the probability of other constitution types is less than 0.01; Core diagnostic indicators: Spleen deficiency 34.12 (reference range 0-33.85), Qi deficiency 35.62 (reference range 0-38.26), Cold 37.82 (reference range 0-38.26), and all other indicators are within the reference range.

[0068] Confidence interval calculation: The Bootstrap method was used to perform 1000 resampling fits on the initial diagnostic result set. The 95% confidence intervals for the predicted probability of Yang deficiency were [0.46, 0.58]; the 95% confidence intervals for the spleen deficiency index were [33.76, 34.48]; and the 95% confidence intervals for the predicted probability of Qi deficiency were [0.42, 0.54].

[0069] Determination of constitution transition zone: The predicted probability of Yang deficiency constitution is 0.52, falling into the constitution transition zone [0.45, 0.55]; the spleen deficiency index is 34.12, falling into the constitution transition zone [33.95, 34.20]; In summary, the initial diagnostic result falls into the constitution transition zone and is judged to be in an unstable state, thus initiating a multi-round consistency check of reasoning.

[0070] S3, Multi-round inference consistency check. In this example, five rounds of independent inference consistency checks are initiated, and the specific execution is as follows: Five independent random seeds were set up: 123, 456, 789, 101112, and 131415. A Gaussian perturbation with an amplitude of 1% of the feature standard deviation was added to the input feature vector X, resulting in five independent perturbation input feature vectors X1-X5. X1-X5 were then input into a large-scale TCM syndrome differentiation model, yielding five independent sets of syndrome differentiation results. The core spleen deficiency index results were 33.98, 34.12, 34.25, 34.09, and 34.17, respectively, and the predicted probabilities of Yang deficiency were 0.51, 0.52, 0.53, 0.52, and 0.52, respectively. The cosine similarity between each pair of the five results was calculated, with an average value of 0.992, i.e., a consistency score SC = 0.992. A preset consistency score threshold of 0.95 was set. An SC = 0.992 greater than or equal to 0.95 indicates good consistency across multiple rounds of results. The median of the five results was taken as the validated syndrome differentiation result Y. cal The spleen deficiency index was 34.12, and the predictive probability of yang deficiency was 0.52.

[0071] S4. Retrieve the user's valid diagnostic result sequence from the past 6 months from the user's historical database. A total of 4 sets of valid detection data are obtained, with time intervals of 7 days or more. The core data is shown in Table 1 below:

[0072] Table 1. Valid detection data from the past 6 months in the user history database. The historical mean of the spleen deficiency index was μ=33.65, the standard deviation was σ=0.082, and the 95% stability interval was [33.486, 33.814]; the historical mean of the probability of yang deficiency was μ=0.505, the standard deviation was σ=0.013, and the 95% stability interval was [0.479, 0.531].

[0073] Deviation comparison and correction: The spleen deficiency index after this verification is 34.12, with an absolute deviation of Δ=0.47 from the historical average. The allowable deviation threshold 2σ=0.164, Δ>2σ, which exceeds the allowable deviation range; the probability of Yang deficiency in this case is 0.52, which falls within the stable constitution range [0.479,0.531], and the deviation is within the allowable range.

[0074] Time-series correction was performed: Using the Kalman filter algorithm, with the historical mean of the spleen deficiency index (33.65) as the prior value and the current verification result (34.12) as the observed value, the Kalman gain K was calculated as K = 0.082² / (0.082² + 0.18²) = 0.172. The final corrected spleen deficiency index is as follows: Y final =33.65+0.172×(34.12-33.65)=33.73; The corrected probability of Yang deficiency constitution remained unchanged at 0.52, and the core syndrome direction was verified to have not changed. The initial syndrome was Yang deficiency constitution, and it remained Yang deficiency constitution after correction, which met the boundary condition requirements.

[0075] S5, In this embodiment, based on the final corrected diagnostic results, a Traditional Chinese Medicine health assessment report is generated, including the following: Constitution Assessment: Yang Deficiency Constitution (Unbalanced Constitution), Health Index 79; Core Risk Indicators: Spleen Deficiency 33.73, Qi Deficiency 35.62, Cold 37.82, Body Fluid Deficiency 5.9; Tongue Diagnosis Interpretation: White tongue coating, pale red tongue, teeth marks on the tongue, enlarged tongue, indicating the presence of cold and dampness in the body, and weakened Qi and blood function; Recommended Treatment Plan: Basic treatment includes dietary suggestions for warming Yang and strengthening the spleen, Guanyuan / Zusanli acupoint therapy, and Baduanjin exercise; Specialized treatment includes health-preserving teas for warming Yang and replenishing Qi, moxibustion therapy, and massage therapy.

[0076] At the same time, the final diagnosis results will be stored in the user's historical database, the user's stable physical condition range model will be updated, and the entire process of this test will be completed.

[0077] This application also discloses a traditional Chinese medicine health assessment system to improve the stability of diagnostic results, including: The data processing module is used to collect users' TCM physiological characteristic data, complete preprocessing and feature extraction, and obtain standardized syndrome differentiation input feature vectors. In this embodiment, the module includes a tongue image feature extraction unit and an organ function feature extraction unit, which are used to extract tongue diagnosis feature vectors and organ function feature vectors respectively, and splice them to form standardized input feature vectors.

[0078] The initial diagnosis module is used to input the standardized diagnosis input feature vector into the pre-trained TCM diagnosis model to obtain an initial diagnosis result set, and to determine whether the initial diagnosis result is in an unstable state based on a preset constitution transition zone. In this embodiment, the module includes a confidence interval calculation unit, which is used to calculate an independent confidence interval for the predicted probability of each diagnosis indicator and constitution type, and to determine the stability of the result based on the preset constitution transition zone.

[0079] The multi-round verification module is used to initiate multi-round independent reasoning consistency verification for unstable dialectical results that fall into the transition zone of constitution, generate multiple sets of dialectical results and record the confidence interval of each set of results; in this embodiment, the module includes a random perturbation generation unit, which is used to add small random perturbations or Gaussian noise to the input feature vector of each round of independent reasoning and record the confidence interval of each round of dialectical results.

[0080] The weighted calibration and time-series correction module is used to perform weighted calibration on the results of multiple rounds of diagnosis, compare the verified diagnosis results with the physical stability interval model constructed from the user's historical diagnosis result sequence, and perform time-series correction on indicators that exceed the range to obtain the final diagnosis result. In this embodiment, the module includes a weighted calculation unit and a filtering correction unit, which are used to perform weighted calibration based on the confidence interval and indicator correlation of the results of each round, and perform time-series correction on the verified results and the historical physical stability interval.

[0081] The report generation and result management module is used to generate a TCM health assessment report based on the final diagnosis result and write the diagnosis result into the historical sequence to update the constitution stability interval model. In this embodiment, the module includes a data storage unit for writing the final diagnosis result into the historical sequence and updating the constitution stability interval model.

[0082] This application also provides an embodiment of an electronic device. The electronic device is manifested in the form of a general-purpose computing device. The components of the electronic device may include, but are not limited to: one or more processors or processing units, memory, and buses connecting different components (including memory and processing units).

[0083] A bus refers to one or more of several bus architectures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of the various bus architectures. For example, these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MCA) bus, the Enhanced ISA bus, the Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect (PCI) bus.

[0084] Electronic devices typically include a variety of computer-readable media. These media can be any available media that can be accessed by the electronic device, including volatile and non-volatile media, and removable and non-removable media.

[0085] The memory may include computer-readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. Electronic devices may further include other removable / non-removable, volatile / non-volatile computer device storage media. By way of example only, the storage system may be used to read and write non-removable, non-volatile magnetic media.

[0086] The electronic device can also communicate with one or more external devices (e.g., keyboard, pointing device, camera, etc.), may include a display, and may communicate with one or more devices that enable a user to interact with the electronic device, and / or with any device that enables the electronic device to communicate with one or more other computing devices (e.g., network card, modem, etc.). This communication can be performed via an input / output (I / O) interface. Furthermore, the electronic device can communicate with one or more networks (e.g., Local Area Network (LAN), Wide Area Network (WAN)) and / or public networks, such as the Internet, via a network adapter. The network adapter communicates with other modules of the electronic device via a bus. The processor executes various functional applications and data processing by running programs stored in memory, such as implementing the TCM health assessment method for improving the stability of diagnostic results provided in the above embodiments of the present invention.

Claims

1. A traditional Chinese medicine health assessment method to improve the stability of diagnostic results, characterized in that, Includes the following steps: S1. Collect users' TCM physiological characteristic data, complete preprocessing and feature extraction, and obtain standardized syndrome differentiation input feature vectors; S2. Input the input feature vector into the pre-trained TCM syndrome differentiation model to obtain an initial syndrome differentiation result set, calculate the preset confidence interval of each syndrome differentiation result, and judge the initial syndrome differentiation result based on the preset constitution transition zone. When the initial syndrome differentiation result falls into the constitution transition zone, it is judged as an unstable state. The constitution transition zone refers to the numerical range located at the edge of the constitution stable range, used to characterize the stage of slight change or transformation of the user's constitution. S3. When the initial dialectical result is determined to be unstable, a multi-round inference consistency check is initiated based on the preset dialectical constraint rules. Multiple sets of dialectical results are obtained through multiple rounds of independent inference, and a consistency score is calculated. When multiple sets of dialectical results do not meet the dialectical constraint rules, a recalculation or weight adjustment is triggered. The weighted average is calculated based on the confidence interval width and index correlation of the results of each round to obtain the verified dialectical result. In this process, each round of independent inference is generated by adding a small random perturbation or Gaussian noise to the input feature vector and using different random seeds. Each round of inference yields an independent set of dialectical results and a corresponding confidence interval. S4. Obtain the user's historical diagnostic result sequence, construct a constitution stability interval model, and compare the verified diagnostic result with the constitution stability interval under the diagnostic constraint rules. When the constraint conditions are not met or the deviation threshold is exceeded, perform time-series correction on the diagnostic result to obtain the final diagnostic result. The time-series correction is implemented through a filtering algorithm, using the historical mean as the prior state and the current verification result as the observation value. The correction value is calculated independently for each diagnostic indicator, while maintaining the relative order between indicators and the core syndrome combination relationship. S5. Generate a TCM health assessment report based on the final diagnosis results, and write the diagnosis results into the historical sequence to update the constitution stability interval model.

2. The method for TCM health assessment to improve the stability of diagnostic results according to claim 1, characterized in that, In step S1, the TCM physiological characteristic data includes collected visceral function detection data, tongue image data, user complaint information, and Western medicine detection index data; preprocessing includes outlier removal, K-nearest neighbor algorithm missing value filling, and Z-score standardization; feature extraction includes extracting tongue diagnosis feature vectors and visceral feature vectors, and concatenating them to obtain a standardized input feature vector.

3. The TCM health assessment method for improving the stability of diagnostic results according to claim 1, characterized in that, In step S2, the initial diagnostic result set includes the predicted probability of TCM constitution type and the value of core diagnostic indicators. Based on the preset confidence interval, it is determined whether it falls into the constitution transition zone. If it falls into the constitution transition zone, it is determined as a high jump risk result and enters the multi-round inference consistency verification. Otherwise, it directly enters the historical constitution smoothing constraint process.

4. The TCM health assessment method for improving the stability of diagnostic results according to claim 1, characterized in that, In step S4, the historical diagnosis result sequence is the user's valid diagnosis results within the past 6 months, and the interval between two adjacent tests is not less than 24 hours; the physical stability interval model is formed by calculating the distribution interval of each diagnosis indicator in the historical sequence and setting dynamic boundaries.

5. A method for TCM health assessment to improve the stability of diagnostic results according to claim 1, characterized in that, In step S5, the TCM health assessment report includes the constitution assessment results, health index, core risk indicators, interpretation of tongue diagnosis characteristics, and recommended conditioning plan; the recommended conditioning plan includes a basic conditioning plan and a special conditioning plan.

6. A TCM health assessment system for improving the stability of diagnostic results, employing the method described in any one of claims 1-5, characterized in that, include: The data processing module is used to collect users' TCM physiological characteristic data, complete preprocessing and feature extraction, and obtain standardized syndrome differentiation input feature vectors. The initial diagnosis module is used to input the standardized diagnosis input feature vector into the pre-trained TCM diagnosis model to obtain the initial diagnosis result set, and to determine whether the initial diagnosis result is in an unstable state based on the preset constitution transition zone. The multi-round verification module is used to initiate multiple rounds of independent reasoning consistency verification for unstable dialectical results that fall into the transition zone of constitution, generate multiple sets of dialectical results and record the confidence interval of each set of results; The weighted calibration and time-series correction module is used to perform weighted calibration on the results of multiple rounds of diagnosis, compare the verified diagnosis results with the physical stability interval model constructed by the user's historical diagnosis result sequence, perform time-series correction on indicators that exceed the range, and obtain the final diagnosis result. The report generation and results management module is used to generate a TCM health assessment report based on the final diagnosis results and write the diagnosis results into the historical sequence to update the constitution stability interval model.

7. The TCM health assessment system for improving the stability of diagnostic results according to claim 6, characterized in that: The data processing module further includes a tongue image feature extraction unit and an organ function feature extraction unit, which are used to extract tongue diagnosis feature vectors and organ function feature vectors respectively, and splice them together to form a standardized input feature vector. The initial diagnostic module is further configured with a confidence interval calculation unit, which is used to calculate an independent confidence interval for the predicted probability of each diagnostic indicator and constitution type, and to determine the stability of the result based on the preset constitution transition zone. The multi-round verification module further includes a random perturbation generation unit, which is used to add small random perturbations or Gaussian noise to the input feature vector of each round of independent reasoning, and record the confidence interval of the verification result of each round. The weighted calibration and time-series correction module further includes a weighted calculation unit and a filtering correction unit, which are used to perform weighted calibration based on the confidence interval and index correlation of the results of each round, and to perform time-series correction on the verified results and the historical physical stability interval. The report generation and result management module further includes a data storage unit, used to write the final diagnostic results into the historical sequence and update the physical stability interval model.

Images