Large-model-driven traditional Chinese medicine curative effect multi-dimensional intelligent evaluation system and device

Abstract

The invention provides a large-model-driven traditional Chinese medicine curative effect multi-dimensional intelligent evaluation system and device, and relates to the technical field of traditional Chinese medicine curative effect intelligent evaluation. The system mainly comprises a data layer, a knowledge layer, a model layer and an application layer which are arranged in sequence; the data layer is used for collecting tongue images, pulse condition data, cardiovascular data, structured data, unstructured data and the like, and carrying out data treatment operations such as structured transformation, block chain evidence storage and abnormal value correction; the knowledge layer comprises a traditional Chinese medicine diagnosis and treatment knowledge map, a traditional Chinese medicine prescription compatibility relationship library, a curative effect evaluation standard library, a component-target network, a semantic mapping engine and a syndrome-treatment association map; the model layer comprises a multi-modal feature extraction module and a curative effect prediction module; the application layer comprises a doctor end and a patient end. According to the scheme, quantitative analysis of traditional Chinese medicine syndrome differentiation rules is realized through a large model, and the syndrome classification consistency of patients is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent evaluation technology of TCM efficacy, and in particular to a large-model-driven multi-dimensional intelligent evaluation system and device for TCM efficacy. Background Technology

[0002] Existing TCM efficacy evaluation systems often have the following problems:

[0003] Problem 1: Strong reliance on subjective experience: Physicians mainly rely on individual experience to make diagnostic decisions, and lack quantitative standards for symptom descriptions (such as "wiry and slippery" pulse, "pale purple" tongue, etc.). Research data shows that the consistency of diagnosis among different physicians for the same patient is only 62% (sample size n=500, significance level α=0.05), which leads to significant subjective bias in efficacy evaluation.

[0004] Problem 2: Insufficient data standardization: Regarding equipment heterogeneity, for example, existing tongue imaging instruments have a resolution of ≤12 million pixels under the ISO 19264 standard, resulting in a 30% decrease in the accuracy of microvascular texture recognition, which may not be able to fully capture microvascular texture features; Regarding parameter non-standardization, for example, the sampling rate of pulse sensors is ≤200 Hz, which may miss the harmonic components of pulse waves; Regarding data silos, for example, the proportion of unstructured text in electronic medical records is >65%, the mapping error rate of Chinese and Western medicine terminology is >22%, and there is a semantic gap and lack of standards between the Chinese and Western medicine evaluation systems;

[0005] Problem 3: The evaluation dimensions are limited: they focus on the symptom relief rate (such as the total effective rate evaluation adopted in the "Guiding Principles for Clinical Research of New Traditional Chinese Medicine"), but ignore the pattern of syndrome evolution (such as the dynamic evolution path of "phlegm turbidity → blood stasis") and quality of life assessment.

[0006] Question 4: Traditional machine learning models have limitations in modeling complex systems in traditional Chinese medicine: time series models (such as LSTM models) are limited by fixed-length inputs, and the error rate of predicting therapeutic efficacy trajectories exceeding 7 days increases significantly; spatiotemporal alignment methods (such as the DTW algorithm) cause the time axis offset of tongue and pulse data to be greater than 15 milliseconds;

[0007] Question 5: Deficiencies exist in the clinical validation system: Existing systems only support single-time-point efficacy evaluation and cannot dynamically track syndrome evolution (e.g., monitoring of the "phlegm turbidity → blood stasis" evolution window is missing); the existing efficacy evidence meets the CONSORT statement requirements at a rate of <53% (The Lancet 2024 Special Issue on Traditional Chinese Medicine); traditional models (e.g., SVM, LSTM models) are difficult to analyze the nonlinear relationships of TCM syndrome differentiation and have weak dynamic tracking capabilities (unable to predict the 14-day efficacy trajectory); spatiotemporal alignment of multimodal data (e.g., tongue appearance, pulse appearance, prescriptions, etc.) is difficult, resulting in low efficiency in generating efficacy evidence (single evaluation time >48 hours); there is a lack of a dynamic evaluation framework that conforms to the holistic view of TCM, which cannot support individualized treatment feedback and compound prescription optimization. Summary of the Invention

[0008] The purpose of this invention is to provide a large-model-driven multi-dimensional intelligent evaluation system and device for TCM efficacy, so as to solve at least one of the above-mentioned technical problems existing in the prior art.

[0009] In the first aspect, to solve the above-mentioned technical problems, the present invention provides a large model-driven multi-dimensional intelligent evaluation system for TCM efficacy, which includes a data layer, a knowledge layer, a model layer and an application layer arranged in sequence.

[0010] The data layer is used to collect tongue images, pulse data, cardiovascular data, structured data and unstructured data, and to perform data governance operations such as structure transformation, blockchain notarization and outlier correction.

[0011] The knowledge layer includes a TCM diagnosis and treatment knowledge graph, a TCM prescription compatibility database, a efficacy evaluation standard database, a component-target network, a semantic mapping engine, and a syndrome-treatment association graph.

[0012] The model layer includes a multimodal feature extraction module and a therapeutic effect prediction module;

[0013] The application layer includes a doctor's side and a patient's side; the doctor's side includes a diagnostic consistency evaluation module, a treatment principle compliance analysis module, and an interpretability feedback module; the patient's side includes a efficacy indicator fusion module, an intelligent grading and intervention module, and a patient interaction optimization module.

[0014] In one feasible implementation, the tongue image is specifically acquired using a tongue imager.

[0015] In one feasible implementation, the pulse data is collected using a pulse sensor.

[0016] In one feasible implementation, the cardiovascular data is collected via a wearable health monitoring device.

[0017] In one feasible implementation, the structured data includes electronic medical records, clinical trial data, etc.; the unstructured data includes ancient books and medical notes, etc.

[0018] In one feasible implementation, the structuring transformation refers to parsing unstructured text using NLP (Natural Language Processing) methods and mapping it to ICD-11-TCM encoding.

[0019] In one feasible implementation, the blockchain-based evidence storage is implemented by constructing a medical data evidence storage system based on the FHIR R5 standard and using blockchain technology. This enables full lifecycle traceability of data, ensuring the authenticity and immutability of the data, while also meeting the compliance requirements of GDPR, the General Data Protection Regulation.

[0020] In one feasible implementation, the outlier correction is achieved through intelligent repair using a Generative Adversarial Network (GAN), a method certified by the ISO / TC215 committee, which can generate reasonable imputation values ​​based on data distribution characteristics, thereby improving data quality.

[0021] In one feasible implementation, the multimodal feature extraction module includes a tongue image feature extraction unit and a pulse image temporal unit.

[0022] In one feasible implementation, the tongue image feature extraction unit specifically includes a DenseNet-201 model, and optimizes the convolutional layers of the DenseNet-201 model by combining grouped convolution and dilated convolution:

[0023] Grouped convolution divides the input channels into 16 groups, and each group performs convolution operations independently; this reduces the computational cost by 35% and the number of parameters by 40% while maintaining feature extraction capabilities.

[0024] Dilated convolution uses a kernel with a dilation rate of 2; this expands the receptive field and enhances the ability to capture subtle texture features in tongue images without increasing the number of parameters.

[0025] During the training of the DenseNet-201 model, the Adam optimizer was used; the learning rate was set to 0.001 and the batch size was set to 32; after 200 epochs of training, the measurement error of tongue coating thickness was reduced to <0.05mm.

[0026] In one feasible implementation, the pulse timing unit specifically includes a WaveNet-LSTM hybrid model, which combines the advantages of two network architectures to achieve in-depth analysis of the waveform features of pulse data. Specifically:

[0027] The WaveNet model constructs a multi-scale receptive field through dilated convolutional layers. This allows it to capture short-scale features such as the main pulse wave and dicrotic wave morphology, as well as long-interval rhythm features, without increasing computational cost. Compared to traditional CNN models, it significantly improves the extraction efficiency of long sequence features and reduces the number of parameters by 40%. By using the Morlet wavelet transform method to generate a 128-dimensional feature vector, and combining it with the "position, number, shape, and momentum" theory of traditional Chinese medicine pulse diagnosis for dimensional mapping, it accurately quantifies the time-frequency features of non-stationary pulse signals. Compared to Fourier transform, its time-frequency localization capability is improved by 35%.

[0028] The LSTM (Long Short-Term Memory) model, combined with a bidirectional gated recurrent unit (BiGRU), learns the pulse wave propagation pattern and rhythmic periodic changes through forward and reverse paths, respectively. For example, the forward path captures the rising edge features of the waveform, while the reverse path analyzes the trend of continuous pulse intervals. Compared with unidirectional LSTM, the F1-score in the late pulse recognition task is improved from 0.72 to 0.91, significantly improving accuracy. By using residual connections and layer normalization methods, the training difficulties of deep networks are solved, and an adaptive forgetting gating method is introduced to optimize the memory weights of abnormal rhythms, thereby avoiding long-interval anomaly misses.

[0029] In one feasible implementation, the WaveNet-LSTM hybrid model also includes a probability value-pulse rate conversion formula, which can transform the machine learning output into diagnostic indicators interpretable by Traditional Chinese Medicine (TCM), avoiding the "black box" problem of the model. Simultaneously, thanks to its lightweight dilated convolution design, the model can be deployed on portable pulse diagnosis devices while maintaining high accuracy, balancing hardware efficiency and diagnostic accuracy. The specific formula includes:

[0030] The formula for absolute difference is as follows:

[0031] △T i =|T i+1 -T i |

[0032] Where, ΔT i T represents the absolute difference in the i-th pulse cycle; i+1 T represents the (i+1)th pulse cycle; i This represents the i-th pulse cycle;

[0033] The confidence interval formula, used to calculate the upper limit of the 95% confidence interval when assuming the absolute difference follows a normal distribution, is as follows:

[0034]

[0035] Among them, CI 95％ This represents the upper limit of the 95% confidence interval; σ represents the mean of the absolute differences. ΔT The standard error represents the absolute difference.

[0036] The unit conversion formula is used to convert pulse rate standards that conform to the "Traditional Chinese Medicine Diagnostics" into time intervals. Specifically:

[0037]

[0038] Where t represents the time interval; f represents the pulse rate, in beats per minute;

[0039] The initial threshold formula is as follows:

[0040] Thres = max(CI) 95％ ,t 标准 ±Δt 误差 )

[0041] Where Thres represents the initial threshold, t 标准 Indicates the standard time interval; Δt 误差 Indicates the allowable error for the time interval;

[0042] The loss function of the bidirectional gated recurrent unit is:

[0043]

[0044] Where MAE represents the mean absolute error; m represents the total number of samples; y j This represents the true value of the j-th sample; This represents the predicted value of the j-th sample;

[0045] In this way, the model parameters can be adjusted with the goal of minimizing the mean absolute error until the mean absolute error stabilizes at a level that meets the needs of clinical application, thereby achieving accurate temporal prediction of pulse characteristics.

[0046] In one feasible implementation, the multimodal feature extraction module specifically achieves accurate analysis and matching between tongue image and pulse data through a dynamic time warping (DTW) algorithm and a hardware / software collaborative mechanism, specifically including:

[0047] In the time dimension, the image features of the tongue image and the waveform features of the pulse data are first normalized; then, by fusing the composite metric function of image feature distance and Euclidean distance, combined with the dynamic time warping algorithm, the time synchronization error can be controlled within 10 milliseconds.

[0048] In the spatial dimension, hardware calibration is completed using a high-resolution camera, infrared markers, and a multi-channel pressure sensor; through a deep learning model, tongue feature points are extracted and 3D reconstruction is used to correct deformation, while pulse data is converted into a spatial matrix. After iterative optimization, the spatial positioning error can be reduced to less than 0.1 square millimeters.

[0049] By employing the aforementioned "normalization-composite metric-DTW" strategy in the time dimension and the "hardware calibration-deep learning correction" strategy in the spatial dimension, a composite error control model is constructed. This model centers on a "spatiotemporal feedback closed loop": the time submodule utilizes a composite metric function to optimize dynamic time warping, achieving 10-millisecond-level time synchronization between tongue and pulse images; the spatial submodule relies on hardware calibration and deep learning to achieve sub-millimeter-level spatial positioning; and the two submodules, through a collaborative mechanism of "time reference constraining spatial acquisition and spatial feature feedback for time correction," uniformly control the spatiotemporal errors of cross-modal data.

[0050] In this way, by constructing a spatiotemporal joint alignment framework for cross-modal tongue vein data, the application limitations of the traditional single-modal DTW algorithm are overcome; through a composite error control model, millisecond-level time synchronization and sub-millimeter-level spatial positioning are achieved; and through the collaborative mechanism of hardware clock synchronization and deep learning feature extraction, a standardized foundation for multimodal data fusion is laid, which has significant advantages over existing technologies in terms of time accuracy, spatial positioning and cross-modal adaptability.

[0051] In one feasible implementation, the efficacy prediction module includes a large efficacy prediction model, which is used to predict the clinical efficacy of traditional Chinese medicine through architecture design, time-aware coding and parameter optimization; specifically, it includes an input layer, a time-aware coding layer, a Transformer-XL layer and an output layer arranged in sequence.

[0052] The input layer is used to input multi-source clinical data and perform data standardization, including medication records, acupuncture time, disease data, and treatment timestamps; the data standardization refers to standardizing numerical data using the Z-score method and encoding discrete events using the one-hot encoding method.

[0053] The time-aware coding layer is used to perform feature fusion after absolute time coding and event time-series coding;

[0054] The absolute time encoding uses sine and cosine functions to encode the timestamp, and the specific formula is as follows:

[0055] PE(t,2i)=sin(t / 10000 2i / 512 ),PE(t,2i+1)=cos(t / 10000 2i / 512 )

[0056] Where PE(t, 2i) represents the time encoding vector with even index in time interval t; PE(t, 2i+1) represents the time encoding vector with odd index in time interval t;

[0057] The event timing coding refers to binding treatment events with time intervals to generate event vectors with time tags, and using L2 regularization to control coding errors, thereby ensuring that the coding error is ≤0.5%.

[0058] The feature fusion refers to concatenating the event vector and the time-encoded vector, and then compressing them to 512 dimensions through a linear transformation;

[0059] The architectural parameters of the Transformer-XL layer include:

[0060] Hidden layer dimension: 512 dimensions, used to support complex feature representation;

[0061] Attention heads: 8 heads, used for parallel processing of multi-scale features, i.e., 64 dimensions per head;

[0062] Relative position encoding formula: This scheme improves upon the traditional absolute position encoding formula of Transformer. The specific formula is as follows:

[0063]

[0064] Among them, a i，j This represents the attention score between the i-th word and the j-th word; This represents the content encoding of the j-th word; W represents the relative distance encoding between the i-th word and the j-th word. 0 Represents the projection matrix; q i T The query vector representing the i-th word;

[0065] Layer structure: Each layer contains a self-attention mechanism and a feedforward neural network (FFN), supplemented by residual connections and layer normalization, as shown in the following expression:

[0066] FFN(x′)=max(0,x′W1+b1)W2+b2

[0067] Where FFN(x′) represents the output of the feedforward neural network; b1 represents the first layer bias vector; b2 represents the second layer bias vector; W1 represents the first layer weight matrix, used to project the input into a high-dimensional space; W2 represents the second layer weight matrix, used to compress the high-dimensional features back to the original dimension; and x′ represents the input vector.

[0068] The output layer maps the output of the Transformer-XL layer to a preset dimension through a fully connected layer; the preset dimension is specifically t' days x efficacy index; the specific expression is:

[0069]

[0070] Among them, h t’ b represents the hidden state vector; out W represents the output bias vector; Softmax represents the normalization function; out Indicates the output projection matrix; This represents the predicted probability of the efficacy indicator; t' can take the value of an integer between 1 and 14.

[0071] Thus, in the large-scale efficacy prediction model, the time-aware encoding layer transforms events such as "medication time-dosage" and "acupuncture time-acupoint" into time-labeled vectors, performs feature fusion, and obtains a fused vector. After the fused vector enters the Transformer-XL layer, it learns the temporal correlation of "treatment time → efficacy change" through an attention mechanism, thereby realizing the correlation fusion of treatment and time. The Transformer-XL layer captures both short-term and long-term features through relative position encoding. The short-term features can be the impact of acupuncture intervention on current symptoms within a preset period. The long-term features can be the lag relationship between the cumulative dosage of medication and efficacy within a preset period. The Transformer-XL layer extracts features from different time scales through 8-head attention and achieves multi-dimensional fusion through multi-head output splicing.

[0072] In one feasible implementation, the specific evaluation method of the diagnostic consistency evaluation module includes:

[0073] Step a1: Using the generator G and discriminator D of the adversarial generative network, simulate the diagnostic logic of different doctors, and optimize the minimum-maximum game loss function, specifically expressed as:

[0074] min G max D E x～pdata (x)[logD(x)]+E z～pz(z) [log(1-D(G(z)))]

[0075] Where pdata(x) represents the distribution of the true dialectical results; pz(z) represents the noise distribution; min G max D E represents a minimax game; x～pdata (x) represents the expected distribution of the true dialectical results; D(x) represents the probability that the discriminator judges the input sample to be from the true dialectical results; E z～pz(z)G(z) represents the expected value of the noise distribution; G(z) represents the generator mapping noise to samples similar to the real dialectical results; D(G(z)) represents the probability that the discriminator misclassifies the generated sample as the real dialectical result.

[0076] After training with a large amount of clinical diagnostic data, the generator learns the distribution characteristics of physicians' diagnostic thinking and outputs the distribution of diagnostic results PG(x), thus approximating the probability distribution of different physicians' diagnostic results. The specific expression is as follows:

[0077]

[0078] Where δ represents the Dirac function; N represents the number of generated samples;

[0079] Step a2: Map the generated dialectical result X to a two-dimensional space using a dimensionality reduction algorithm, and estimate the probability density of each coordinate point using a kernel density function. The specific expression is as follows:

[0080]

[0081] in, Kσ represents the probability density estimate at position (u,v); Kσ represents the Gaussian kernel function, specifically expressed as:

[0082]

[0083] Where d represents the Euclidean distance between the input point and the center point; σ represents the smoothing parameter of the kernel function;

[0084] The formula for generating a heatmap and calculating its colors is as follows:

[0085]

[0086] Wherein, Color(u,v) represents the heatmap color at position (u,v). The warmer the color, the higher the probability of a diagnostic result appearing at that position. This represents the minimum probability density estimate; This represents the largest probability density estimate;

[0087] Step a3: Calculate the distribution P of the actual diagnostic conclusions. real (x) and the standard distribution of the heatmap P std (x) is compared; the deviation is quantified using KL divergence, the specific expression of which is:

[0088]

[0089] Here, Deviation represents the degree of deviation.

[0090] In one feasible implementation, the diagnostic consistency evaluation module further includes a dynamic correction method for identifying syndrome contradictions, specifically including:

[0091] Step b1: Convert the RGB chromaticity values ​​of the tongue image to the HSV color space. The specific expression is as follows:

[0092]

[0093] Where H represents chromaticity; S represents saturation; and V represents brightness.

[0094] Step b2: Calculate the average chromaticity of the tongue region in the tongue image. With average saturation The specific formulas include:

[0095]

[0096] Where n represents the total number of pixels in the tongue image region; H i S represents the chromaticity of the i-th pixel; i This represents the saturation of the i-th pixel;

[0097] Step b3: Calculate the amplitude ratio of the main wave and the depth of the diphthoplasty notch of the pulse wave characteristics;

[0098] The specific formula for the amplitude ratio of the main wave is:

[0099]

[0100] Where A represents the amplitude ratio of the main wave; H1 represents the height of the main wave; and H0 represents the baseline height.

[0101] The specific formula for calculating the depth of the diphtheria wave notch is as follows:

[0102]

[0103] Where A' represents the depth of the diphtheria wave notch; H2 represents the height of the diphtheria wave trough.

[0104] Step b4: Calculate the string slip S using the string slip exponential model. string The specific formula is as follows:

[0105] S string =w1×A+w2×D′+w3×f peak

[0106] Where w1, w2, and w3 are all weighting coefficients; D' represents the differential characteristic of the pulse waveform; f peak Indicates the peak frequency of the pulse wave;

[0107] Step b5: Measure the consistency of symptoms using a composite metric function. Specific formulas include:

[0108]

[0109] Where d(T, P) represents the Euclidean distance between the tongue image feature vector and the pulse image feature vector; T represents the tongue image feature vector, specifically: P represents the pulse feature vector, specifically: P = [S string f peak ];μ H Indicates the mean chromaticity under normal symptoms; μ S μ represents the mean saturation value under normal symptoms. P This represents the mean value of the string slip under normal symptoms; the three means can be obtained based on relevant historical data and statistical analysis.

[0110] In one feasible implementation, the rule compliance analysis module includes an association rule base and a time-series monitoring model;

[0111] The association rule base is used to integrate the "principal, assistant, adjuvant, and guide" rules of classic prescriptions in ancient Chinese medicine books to construct a syndrome-treatment association graph.

[0112] The time-series monitoring model uses an LSTM network (Long Short-Term Memory network) to perform deep learning on prescription adjustments, tongue and pulse changes, and prescription target data in continuous time series. Based on the correlation coefficient between tongue and pulse changes and prescription target data, the model determines the correlation between prescription adjustments and syndrome evolution. If the correlation coefficient is less than a preset threshold, it indicates insufficient correlation, and a treatment deviation report is generated.

[0113] In one feasible implementation, the interpretability feedback module includes a logic breakpoint detection unit and a correction suggestion generation unit;

[0114] The logical breakpoint detection unit, based on the interpretability feedback mechanism of the knowledge graph comparison method, maps the diagnosis results and physical sign data (specifically including tongue images, pulse data and cardiovascular data collected from the data layer) to the TCM diagnosis and treatment knowledge graph respectively. By analyzing the relevant nodes and relationships, it calculates the probability of logical conflict between the diagnosis results and physical sign data and determines the contradiction point: if the probability of logical conflict is greater than the preset threshold, it is determined to be a contradiction point.

[0115] The correction suggestion generation unit retrieves relevant original texts from ancient Chinese medicine books based on the syndrome differentiation results, and screens RCT research evidence with a Jadad score ≥4; it integrates the relevant original texts, RCT (randomized controlled trial) research evidence with the current syndrome differentiation results, and outputs an optimized treatment plan through a recommendation formula, thereby obtaining scientific and reasonable correction suggestions.

[0116] In one feasible implementation, the diagnostic consistency evaluation module further includes a standard mapping library based on national standard documents, which is constructed as a two-dimensional space to map the knowledge layer; this can cover 820 diseases and syndromes with a coding mapping error rate of less than 6%.

[0117] In one feasible implementation, the logic breakpoint detection unit calculates the logic conflict probability using a logistic regression model, with the specific formula including:

[0118]

[0119] Among them, P conflict Let S represent the probability of logical conflict; D represent the mapping vector of the diagnostic result; θ represent the correlation benchmark threshold; e represent the natural constant; λ represent the correlation influence coefficient; and Associate(S, D) represents the correlation strength between the diagnostic result and the vital signs data, specifically expressed as:

[0120]

[0121] Where W represents the weight matrix, which is used to perform a linear transformation on the original vector space, assigning different importance to different dimensions (features) or adjusting the strength of their correlation.

[0122] In one feasible implementation, the recommendation formula in the correction suggestion generation unit specifically includes:

[0123]

[0124] Wherein, Tx recommendation represents the recommended modified treatment plan (such as recommended Chinese herbal medicine prescriptions, treatment methods, etc., which is the output of the optimal choice after formula calculation); Indicates t pairs The degree of support; t represents the candidate treatment plan; This represents the query conditions formed by combining patient's treatment needs and diagnostic results; β represents the conflict penalty coefficient; T represents the set of candidate treatment options; ω total The integrated weights are represented by the following expression:

[0125] ω total =α·ω ancient +(1-α)·ω RCT

[0126] Where α represents the adjustment coefficient; ω ancient ω represents the weight of the relevant original text (based on traditional knowledge such as ancient Chinese medicine books and classic treatises, used to measure the degree to which t conforms to the recommendations of ancient books); RCTThis indicates the weight of evidence from RCT studies (based on evidence from modern medical research such as randomized controlled trials, used to measure the degree to which t is supported by modern research).

[0127] In one feasible implementation, the efficacy index fusion module includes an acute disease model and a chronic disease model;

[0128] The acute disease model is used to collect pain index in real time through wearable health monitoring devices and calculate the improvement rate of the area under the characteristic curve (AUC) before and after treatment. The specific formula is defined as follows:

[0129]

[0130] Among them, AUC 治疗前 The area under the characteristic curve of the pain index sequence before treatment intervention is used to reflect the overall pain load; AUC 治疗中 The area under the characteristic curve of the pain index sequence during the treatment phase; the improvement rate, used to quantify the degree of relief of acute pain symptoms;

[0131] The chronic disease model integrates tongue and pulse data, laboratory indicators, and the SF-36 scale (Short Scale for Health Survey) using a composite efficacy index formula to calculate the composite efficacy index. The specific formula includes:

[0132]

[0133] Wherein, CEI represents the composite efficacy index; ΔSSIM represents the change in tongue veins, which can be obtained by comparing the tongue structure similarity index (SSIM) before and after treatment, that is:

[0134] ΔSSIM=SSIM 治疗后 -SSIM 治疗前

[0135] SSIM focuses on analyzing the structural features of the tongue image, indirectly reflecting the overall physiological state adjustment of the tongue and veins through changes in the tongue image; the baseline value represents the similarity index of the tongue image structure before treatment; the ratio of ΔSSIM to the baseline value is used to reflect the relative changes in tongue and vein features; the RF decrease rate represents the proportion of decrease in laboratory indicators related to the pathogenesis; the SF-36 total score represents the patient's total score on the SF-36 scale; 100 represents the maximum total score on the SF-36 scale; the ratio of the SF-36 total score to 100 is used to reflect the contribution of the patient's quality of life to the composite efficacy index; 0.4 and 0.3 represent the corresponding weights, respectively.

[0136] In one feasible implementation, the intelligent grading and intervention module includes response criteria and intervention mechanisms;

[0137] The response criteria are based on the K-means clustering method and are divided into four levels of response.

[0138] The intervention mechanism includes:

[0139] For patients in grade III: Based on the past syndrome data of patients in grade III, the time series characteristics and patterns of the syndrome data are learned through the LSTM model to predict the direction of syndrome evolution of patients over several days.

[0140] For Level IV patients: Associate with a pre-set emergency medical knowledge base.

[0141] In one feasible implementation, the patient interaction optimization module includes an adaptive scale and a compliance check.

[0142] The adaptive scale assesses patients' cognitive level, transforms professional medical statements into corresponding quantitative scores, and continuously iterates and adjusts the scale logic to enhance the reliability of subjective assessment data, focusing on personalized information collection in doctor-patient collaboration scenarios.

[0143] The compliance check establishes a time-series correlation analysis mechanism between medication behavior and tongue appearance evolution: when the correlation between the medication record sequence and the tongue appearance feature change sequence is below a threshold, a medication check dialogue flow is automatically triggered. Relying on time-series data mining technology, intelligent monitoring and intervention of the closed-loop diagnosis and treatment behavior are achieved, ensuring the quality of treatment plan implementation from the patient interaction side.

[0144] Secondly, based on the same inventive concept, this application also provides a large-model-driven multi-dimensional intelligent evaluation device for TCM efficacy, including a processor, a memory, and a bus. The memory stores instructions and data read by the processor, and the processor is used to call the instructions and data in the memory to realize the large-model-driven multi-dimensional intelligent evaluation system for TCM efficacy as described above. The bus connects the various functional components for transmitting information.

[0145] In one feasible implementation, the device further includes a tongue imager, a pulse sensor, and a wearable health monitoring device.

[0146] By adopting the above technical solution, the present invention has the following beneficial effects:

[0147] This invention provides a large-scale model-driven multi-dimensional intelligent evaluation system and device for TCM efficacy. Through the large-scale model, it realizes the quantitative analysis of TCM syndrome differentiation rules and improves the consistency of syndrome classification for patients. Through dynamic tracking capabilities, it can predict the evolution path of syndromes in real time and improve the sensitivity to resist treatment risks. The efficacy evidence generated by this solution is compatible with the ICD-11-TCM standard. This solution supports multiple daily optimizations of treatment plans, improving patient medication adherence. Attached Figure Description

[0148] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0149] Figure 1 A diagram illustrating a large-model-driven multi-dimensional intelligent evaluation system for TCM efficacy provided in an embodiment of the present invention;

[0150] Figure 2 This is a schematic diagram of the structure of the large-scale efficacy prediction model provided in an embodiment of the present invention;

[0151] Figure 3 A flowchart of the evaluation method provided in the embodiments of the present invention;

[0152] Figure 4 A flowchart of the dynamic correction method provided in an embodiment of the present invention. Detailed Implementation

[0153] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0154] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0155] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0156] The present invention will be further explained below with reference to specific embodiments.

[0157] Example 1:

[0158] like Figure 1 As shown in the figure, this embodiment provides a large model-driven multi-dimensional intelligent evaluation system for TCM efficacy, which includes a data layer, a knowledge layer, a model layer and an application layer arranged sequentially.

[0159] The data layer is used to collect tongue images, pulse data, cardiovascular data, structured data and unstructured data, and to perform data governance operations such as structure transformation, blockchain notarization and outlier correction.

[0160] The knowledge layer includes a TCM diagnosis and treatment knowledge graph, a TCM prescription compatibility database, a efficacy evaluation standard database, a component-target network, a semantic mapping engine, and a syndrome-treatment association graph.

[0161] The model layer includes a multimodal feature extraction module and a therapeutic effect prediction module;

[0162] The application layer includes a doctor's side and a patient's side; the doctor's side includes a diagnostic consistency evaluation module, a treatment principle compliance analysis module, and an interpretability feedback module; the patient's side includes a efficacy indicator fusion module, an intelligent grading and intervention module, and a patient interaction optimization module.

[0163] Furthermore, the tongue image is specifically acquired using a tongue imaging device;

[0164] The tongue imaging system employs a 24-megapixel microscopic imaging system with an RGB colorimetric error of ≤0.8% to ensure the authenticity of tongue color information; it also boasts a spatial resolution of 0.1mm. 2 The ability to analyze microvascular textures is used to capture the microscopic structural features of the tongue and provide data support for the objectification of tongue diagnosis.

[0165] Furthermore, the pulse data is collected using a pulse sensor;

[0166] The pulse sensor is specifically a piezoelectric array sensor with a sampling frequency of 1kHz and a harmonic resolution of 0.5Hz, and a pulse wave error of ≤3%, so as to achieve high-frequency and accurate acquisition of pulse signals. In the signal processing stage, the original signal is denoised by an adaptive filtering algorithm, which improves the signal-to-noise ratio (SNR) to 42dB. Feature extraction is performed based on Morlet wavelet transform, focusing on the 5-20Hz frequency band, which has an energy concentration of ≥82%, thereby effectively separating pulse features.

[0167] Furthermore, the cardiovascular data is collected through wearable health monitoring devices;

[0168] The wearable health monitoring device integrates a PPG+ECG dual-modal monitoring module, which can collect cardiovascular data in real time; and through 5G MEC edge computing technology, it can achieve millisecond-level synchronous transmission of data to ensure data timeliness.

[0169] Furthermore, the structured data includes electronic medical records, clinical trial data, etc.; the unstructured data includes ancient books and medical notes, etc.

[0170] Furthermore, the structuring transformation refers to parsing unstructured text (such as "liver stagnation and spleen deficiency") using NLP (Natural Language Processing) methods and mapping it to ICD-11-TCM encoding. Specifically, this includes: first, extracting TCM terminology features from the unstructured text using techniques such as word vectors to construct a TCM terminology knowledge base; then, matching the TCM terminology knowledge base with the ICD-11-TCM encoding mapping table using a trained sequence labeling model; and finally, verifying and correcting the mapping manually or through algorithms to obtain an accurate mapping between the unstructured text and the ICD-11-TCM encoding.

[0171] Furthermore, the blockchain-based evidence storage is based on the FHIR R5 standard and is implemented through blockchain technology to build a medical data evidence storage system; the timestamp accuracy is ±1 millisecond; in this way, the entire lifecycle of data can be traced, ensuring the authenticity and immutability of the data, while meeting the compliance requirements of GDPR, the General Data Protection Regulation.

[0172] Furthermore, the outlier correction is achieved through intelligent repair using a generative adversarial network (GAN). This method has been certified by ISO / TC215 and can generate reasonable imputation values ​​based on data distribution characteristics, thereby improving data quality.

[0173] Furthermore, the TCM diagnosis and treatment knowledge graph is based on the BERT-ECM model. It is obtained by deep learning the compatibility rules of classic prescriptions in ancient Chinese medicine books (such as "Treatise on Febrile Diseases") and analyzing the rules of monarch, minister, assistant and guide. This is to explore the application theory of prescriptions, deconstruct the compatibility logic of monarch, minister, assistant and guide from the basic level, and build the underlying support for the TCM prescription knowledge system.

[0174] Furthermore, the aforementioned database of traditional Chinese medicine prescription compatibility relationships is used to provide basic compatibility data support for the model layer, assist doctors in diagnosis and treatment decisions, and enhance the synergy of the knowledge layer. Specifically, it integrates data on the principal, assistant, adjuvant, and guide rules, the proportion of medicinal materials, contraindications (such as "Eighteen Incompatibilities" and "Nineteen Antagonisms"), and synergistic relationships of efficacy of classic prescriptions in ancient Chinese medicine books (such as prescriptions recorded in ancient books such as "Treatise on Febrile Diseases" and "Compendium of Materia Medica") and modern clinically effective prescriptions (covering more than 3,000 classic prescriptions and more than 5,000 clinically verified compatibility combinations). It is digitally stored through structured coding (mapped to the ICD-11-TCM standard).

[0175] Furthermore, the component-target network integrates component-target-pathway data (KEGG pathway enrichment p < 0.01) of 599 Chinese herbal medicines. Through KEGG pathway enrichment analysis (p < 0.01), the molecular mechanism of action of Chinese herbal medicines is obtained, thus providing a data foundation for network pharmacology research.

[0176] Furthermore, the syndrome-treatment correlation map is obtained by integrating the rules of principal, assistant, adjuvant, and guide formulas of classic prescriptions (more than 3,000 in ancient Chinese medicine books such as "Treatise on Cold Damage") with clinical syndrome-treatment scenarios, thereby realizing the implementation of prescription theory into clinical application. It forms a complementary hierarchy of "theoretical analysis - clinical application" with the TCM diagnosis and treatment knowledge map. The TCM diagnosis and treatment knowledge map delves into the principles of rules, while the syndrome-treatment correlation map focuses on the practice of rules in syndrome-treatment matching. Together, they serve the TCM diagnosis and treatment knowledge system, with no content duplication and synergistic effects.

[0177] Furthermore, the semantic mapping engine is based on the ICD-11-TCM standard and constructs a two-way mapping between traditional Chinese medicine terminology and Western medicine disease classification system.

[0178] Furthermore, the efficacy evaluation standard library is based on the national standard "Terminology for Clinical Diagnosis and Treatment in Traditional Chinese Medicine" (GB / T 16751). It uses NLP methods to parse unstructured text and establishes a "symptom-syndrome-treatment principle" coding mapping relationship, which can cover 820 kinds of TCM diseases and syndromes, with a coding mapping error rate of <6%.

[0179] Furthermore, the multimodal feature extraction module includes a tongue image feature extraction unit and a pulse image temporal unit.

[0180] Furthermore, the tongue image feature extraction unit specifically includes a DenseNet-201 model, and the convolutional layers of the DenseNet-201 model are optimized by grouped convolution (16 groups) and dilated convolution (dilation rate 2):

[0181] Grouped convolution divides the input channels into 16 groups, and each group performs convolution operations independently; this reduces the computational cost by 35% and the number of parameters by 40% while maintaining feature extraction capabilities.

[0182] Dilated convolution uses a kernel with a dilation rate of 2; this expands the receptive field and enhances the ability to capture subtle texture features in tongue images without increasing the number of parameters.

[0183] During the training of the DenseNet-201 model, the Adam optimizer was used; the learning rate was set to 0.001 and the batch size was set to 32; after 200 epochs of training, the measurement error of tongue coating thickness was reduced to <0.05mm.

[0184] Furthermore, the pulse timing unit specifically includes a WaveNet-LSTM hybrid model, which combines the advantages of two network architectures to achieve in-depth analysis of the waveform features of pulse data. Specifically:

[0185] The WaveNet model constructs multi-scale receptive fields through dilated convolutional layers (e.g., dilation rates of 1, 2, 4, etc.). This allows it to simultaneously capture short-scale features such as the main wave and diatonic wave morphology of the pulse wave (millisecond level) and long-interval rhythm features (pulse intervals changing on the second level) without increasing computational cost. Compared to traditional CNN models, it significantly improves the extraction efficiency of long sequence features and reduces the number of parameters by 40%. Through the Morlet wavelet transform method, it generates 128-dimensional feature vectors and performs dimensional mapping based on the "position, number, shape, and momentum" theory of traditional Chinese medicine pulse diagnosis, accurately quantifying the time-frequency features of non-stationary pulse signals. Compared to Fourier transform, its time-frequency localization capability is improved by 35%.

[0186] The LSTM (Long Short-Term Memory) model, combined with a bidirectional gated recurrent unit (BiGRU), learns the pulse wave propagation pattern and rhythmic periodic changes through forward and reverse paths, respectively. For example, the forward path captures the rising edge features of the waveform, while the reverse path analyzes the trend of continuous pulse intervals. Compared with unidirectional LSTM, the F1-score in the late pulse recognition task is improved from 0.72 to 0.91, significantly improving accuracy. By using residual connections and layer normalization methods, the training difficulties of deep networks are solved, and an adaptive forgetting gating method is introduced to optimize the memory weights of abnormal rhythms, thereby avoiding long-interval anomaly misses.

[0187] Furthermore, the WaveNet-LSTM hybrid model also includes a probability value-pulse rate conversion formula, which can transform the machine learning output into diagnostic indicators that can be interpreted by Traditional Chinese Medicine (TCM), avoiding the "black box" problem of the model. Simultaneously, thanks to its lightweight dilated convolution design, the model can be deployed on portable pulse diagnosis devices while maintaining high accuracy, balancing hardware efficiency and diagnostic accuracy. Specifically, this includes:

[0188] absolute difference formula ΔT i =|T i+1 -T i |; Used to calculate the fluctuation amplitude of continuous pulse cycles, providing a basis for identifying abnormal rhythms (such as rapid pulse, knotted pulse);

[0189] Confidence interval formula Based on the pulse rate distribution characteristics of the normal population, a threshold for judging rhythm abnormalities is set.

[0190] Unit conversion formula By linking the pulse rate (beats / minute) in Traditional Chinese Medicine with the time interval (seconds), the compatibility of indicators between Traditional Chinese Medicine and Western medicine is achieved.

[0191] loss function By minimizing the deviation between the predicted value and the true pulse rate, the model achieves an error of ≤3% in recognizing typical pulse patterns such as slow pulse and rapid pulse.

[0192] In this way, the model parameters can be adjusted with the goal of minimizing the mean absolute error until the mean absolute error stabilizes at a level that meets the needs of clinical application, thereby achieving accurate temporal prediction of pulse characteristics.

[0193] Furthermore, the multimodal feature extraction module specifically achieves accurate analysis and matching between tongue image and pulse data through the Dynamic Time Warping (DTW) algorithm and a hardware / software collaborative mechanism, specifically including:

[0194] In the time dimension, the image features of the tongue image (such as texture, RGB color, etc.) and the waveform features of the pulse data are first normalized; a composite metric function of image feature distance and Euclidean distance is fused to enable the dynamic time warping algorithm to more accurately align the time series of the two and solve the problem of time synchronization of multimodal data.

[0195] In the spatial dimension, hardware calibration is completed using a high-resolution camera, infrared markers, and a multi-channel pressure sensor; through a deep learning model, tongue feature points are extracted and 3D reconstruction is used to correct deformation, while pulse data is converted into a spatial matrix. After iterative optimization, the spatial positioning error can be reduced to less than 0.1 square millimeters.

[0196] In this way, by constructing a spatiotemporal joint alignment framework for cross-modal tongue vein data, the application limitations of the traditional single-modal DTW algorithm are overcome; through a composite error control model, millisecond-level time synchronization and sub-millimeter-level spatial positioning are achieved; and through the collaborative mechanism of hardware clock synchronization and deep learning feature extraction, a standardized foundation for multimodal data fusion is laid, which has significant advantages over existing technologies in terms of time accuracy, spatial positioning and cross-modal adaptability.

[0197] Furthermore, the efficacy prediction module includes a large-scale efficacy prediction model, used to predict the clinical efficacy of traditional Chinese medicine through architecture design, time-aware coding, and parameter optimization; such as Figure 2 As shown, it specifically includes an input layer, a time-aware coding layer, a Transformer-XL layer, and an output layer arranged sequentially.

[0198] The input layer is used to input multi-source clinical data and perform data standardization, including medication records (e.g., drug name, dosage, time, etc.), acupuncture time (e.g., acupoints, frequency, acupuncture time, etc.), disease data (e.g., laboratory indicators, symptom scores, etc.), and treatment timestamps (e.g., consultation time, treatment execution time, etc.). The data standardization refers to standardizing numerical data using the Z-score method and encoding discrete events using the one-hot encoding method.

[0199] The time-aware coding layer is used to perform feature fusion after absolute time coding and event time-series coding;

[0200] The absolute time encoding uses sine and cosine functions to encode the treatment timestamps (e.g., 24 hours after medication, 72 hours after acupuncture) so that the model can capture the temporal correlation between "short-term intervention (within 3 days)" and "long-term cumulative effect (within 14 days)".

[0201] The event timing coding refers to binding treatment events (such as medication) with time intervals to generate event vectors with time tags, and using L2 regularization to control coding errors, thereby ensuring that the coding error is ≤0.5%.

[0202] The feature fusion refers to concatenating the event vector and the time-encoded vector, and then compressing them to 512 dimensions through a linear transformation;

[0203] The architectural parameters of the Transformer-XL layer include:

[0204] Hidden layer dimension: 512 dimensions, used to support complex feature representation;

[0205] Attention heads: 8 heads, used for parallel processing of multi-scale features, i.e., 64 dimensions per head;

[0206] Relative position encoding formula: This scheme improves upon the traditional absolute position encoding formula of Transformer. The specific formula is as follows:

[0207]

[0208] Among them, a i，j This represents the attention score between the i-th word and the j-th word; This represents the content encoding of the j-th word; W represents the relative distance encoding between the i-th word and the j-th word. 0 Represents the projection matrix; q i T The query vector representing the i-th word;

[0209] Layer structure: Each layer contains a self-attention mechanism and a feedforward neural network (FFN), supplemented by residual connections and layer normalization, as shown in the following expression:

[0210] FFN(x′)=max(0,x′W1+b1)W2+b2

[0211] Where FFN(x′) represents the output of the feedforward neural network; b1 represents the first layer bias vector; b2 represents the second layer bias vector; W1 represents the first layer weight matrix, used to project the input into a high-dimensional space; W2 represents the second layer weight matrix, used to compress the high-dimensional features back to the original dimension; and x′ represents the input vector.

[0212] The output layer uses a normalization function. The output of the Transformer-XL layer is mapped to a preset dimension of 14 days × efficacy indicators; the normalization function ensures that the sum of daily efficacy probabilities is 1, facilitating intuitive clinical judgment of trends in indicators such as "pain relief" and "improvement in tongue appearance"; W out Indicates the output projection matrix; This represents the predicted probability of the efficacy indicator; t' can take the value of an integer between 1 and 14.

[0213] Thus, in the large-scale efficacy prediction model, the time-aware encoding layer transforms events such as "medication time-dosage" and "acupuncture time-acupoint" into time-labeled vectors, performs feature fusion, and obtains a fusion vector. After the fusion vector enters the Transformer-XL layer, it learns the temporal correlation of "treatment time → efficacy change" (such as the symptom improvement rate 48 hours after medication) through an attention mechanism, thereby realizing the correlation fusion of treatment and time. The Transformer-XL layer captures both short-term and long-term features through relative position encoding. The short-term features can be the impact of acupuncture intervention on current symptoms within a preset period (e.g., the past 3 days). The long-term features can be the lag relationship between the cumulative dosage of medication and efficacy within a preset period (e.g., the past 7 days). The Transformer-XL layer extracts features from different time scales through 8-head attention and achieves multi-dimensional fusion through multi-head output splicing.

[0214] Furthermore, such as Figure 3 As shown, the specific evaluation method of the diagnostic consistency evaluation module includes:

[0215] Step a1: Using the generator G and discriminator D of the adversarial generative network, simulate the diagnostic logic of different doctors, and optimize the minimum-maximum game loss function, specifically expressed as:

[0216] min G max D E x～pdata(x) [logD(x)]+E z～pz(z) [log(1-D(G(z)))]

[0217] Where pdata(x) represents the distribution of the true dialectical results; pz(z) represents the noise distribution; min G max D E represents a minimax game; x～pdata (x) represents the expected distribution of the true dialectical results; D(x) represents the probability that the discriminator judges the input sample to be from the true dialectical results; E z～pz(z) G(z) represents the expected value of the noise distribution; G(z) represents the generator mapping noise to samples similar to the real dialectical results; D(G(z)) represents the probability that the discriminator misclassifies the generated sample as the real dialectical result.

[0218] After training with a large amount of clinical diagnostic data, the generator learns the distribution characteristics of physicians' diagnostic thinking and outputs the distribution of diagnostic results PG(x), thus approximating the probability distribution of different physicians' diagnostic results. The specific expression is as follows:

[0219]

[0220] Where δ represents the Dirac function; N represents the number of generated samples;

[0221] Step a2: Map the generated dialectical result X (e.g., the dialectical vector) to a two-dimensional space using a dimensionality reduction algorithm (e.g., t-SNE), and estimate the probability density of each coordinate point using the kernel density function. The specific expression is as follows:

[0222]

[0223] in, Kσ represents the probability density estimate at position (u,v); Kσ represents the Gaussian kernel function, specifically expressed as:

[0224]

[0225] Where d represents the Euclidean distance between the input point and the center point; σ represents the smoothing parameter of the kernel function;

[0226] The formula for generating a heatmap and calculating its colors is as follows:

[0227]

[0228] Wherein, Color(u,v) represents the heatmap color at position (u,v). The warmer the color (e.g., red), the higher the probability of a diagnostic result appearing at that position. This represents the minimum probability density estimate; This represents the largest probability density estimate;

[0229] Step a3: Calculate the distribution P of the actual diagnostic conclusions. real (x) (e.g., the probability of a one-hot encoding) and the standard distribution P of the heatmap std (x) is compared; the deviation is quantified using KL divergence, the specific expression of which is:

[0230]

[0231] Where Deviation represents the degree of deviation;

[0232] For example, if the probability distribution of the actual diagnosis of "wind-cold exterior syndrome" is P real = [1, 0, 0], while the probability distribution of "wind-cold exterior syndrome" in the heat map is P. std =0.6; then the deviation is The larger the deviation value, the more significant the difference between the actual diagnostic conclusion and the physician's diagnostic reasoning.

[0233] Furthermore, such as Figure 4As shown, the diagnostic consistency evaluation module also includes a dynamic correction method for identifying syndrome contradictions, specifically including:

[0234] Step b1: Convert the RGB chromaticity values ​​of the tongue image to the HSV color space. The specific expression is as follows:

[0235]

[0236] Where H represents the chromaticity value (such as the quantitative index of pale red tongue, purple tongue, etc.); S represents saturation; and V represents brightness.

[0237] Step b2: Calculate the average chromaticity of the tongue region in the tongue image. With average saturation The specific formulas include:

[0238]

[0239] Where n represents the total number of pixels in the tongue image region; H i S represents the chromaticity of the i-th pixel; i This represents the saturation of the i-th pixel;

[0240] Step b3: Calculate the amplitude ratio of the main wave and the depth of the diphthoplasty notch of the pulse wave characteristics;

[0241] The specific formula for the amplitude ratio of the main wave is:

[0242]

[0243] Where A represents the amplitude ratio of the main wave; H1 represents the height of the main wave; and H0 represents the baseline height.

[0244] The specific formula for calculating the depth of the diphtheria wave notch is as follows:

[0245]

[0246] Where A' represents the depth of the diphtheria wave notch; H2 represents the height of the diphtheria wave trough.

[0247] Step b4: Calculate the string slip S using the string slip exponential model. string The specific formula is as follows:

[0248] S string =w1×A+w2×D'+w3×f peak

[0249] Where w1, w2, and w3 are 0.4, 0.3, and 0.3 respectively, all weighting coefficients determined through clinical training; D' represents the differential characteristics of the pulse waveform, used to reflect the rising and falling speed of the pulse wave, such as the rapid rising and falling characteristics of a slippery pulse; f peakIt represents the peak frequency of the pulse wave, which is used in conjunction with the traditional Chinese medicine theory of "smooth pulse like a pearl" to reflect the smoothness of the pulse wave (the higher the value, the smoother the pulse shape).

[0250] Step b5: Measure the consistency of symptoms using a composite metric function, specifically applied to tongue image feature vectors. With pulse feature vector P = [S string f peak The distance between the values ​​is used to reflect the consistency of the "tongue and pulse combined assessment". When the distance is greater than 0.6 (an empirical threshold based on 500 clinical cases), the system uses a composite metric function to determine if there is a contradiction in the syndrome (such as inconsistencies in tongue and pulse characteristics, such as pale tongue with white coating but no pulse rate), and then initiates a correction process.

[0251] Furthermore, the correction process includes:

[0252] Logical breakpoint location: Based on the TCM diagnosis and treatment knowledge graph, the probability of logical conflict between the diagnosis results and the physical signs data is calculated through a logistic regression model to accurately locate the contradiction points (such as the mismatch between tongue appearance features and pulse appearance features).

[0253] Evidence retrieval and integration: The correction suggestion generation unit retrieves original texts related to the current syndrome from ancient Chinese medicine books and screens RCT research evidence with Jadad scores ≥4 as the basis for correction.

[0254] Generate optimized solutions: The optimized treatment plan is output through the recommendation formula (which combines the weights of relevant original texts and the evidence weights of RCT studies). It includes specific suggestions such as prescription adjustments and corrections of diagnosis and treatment methods to help doctors calibrate their diagnosis or treatment plan.

[0255] Furthermore, the governance principle compliance analysis module includes an association rule base and a time-series monitoring model;

[0256] The association rule base is used to integrate the principal, assistant, adjuvant, and guide rules of classic prescriptions in ancient Chinese medicine books (such as "Treatise on Febrile Diseases") (more than 3,000) to construct a syndrome-treatment association graph;

[0257] The time-series monitoring model learns the temporal correlation between prescription adjustments, changes in tongue and pulse, and prescription target data through an LSTM network. Based on the correlation coefficient judgment logic (threshold set at 0.6), it quantifies the correlation between prescription adjustments and syndrome evolution. For example, when the correlation coefficient between the tongue and pulse changes (increased redness of the tongue) and the prescription target data (IL-6 inhibition rate) of "increased dosage of Bupleurum" is <0.6, it is determined that the treatment principle has deviated, and adjustment suggestions are generated.

[0258] Furthermore, the interpretability feedback module includes a logic breakpoint detection unit and a correction suggestion generation unit;

[0259] The logical breakpoint detection unit, based on the interpretability feedback mechanism of the knowledge graph comparison method, maps the diagnosis results and physical sign data to the TCM diagnosis and treatment knowledge graph respectively. By analyzing the relevant nodes and relationships, it calculates the probability of logical conflict between the diagnosis results and physical sign data (e.g., red tongue with little coating, but no diagnosis of Yin deficiency) and determines the contradiction point: if the probability of logical conflict is greater than a preset threshold (e.g., 75%), it is determined to be a contradiction point.

[0260] The correction suggestion generation unit retrieves relevant original texts from ancient Chinese medicine books (such as "Wenbing Tiaobian") with a citation accuracy rate of 95%, and screens RCT research evidence with a Jadad score ≥4. It then integrates the relevant original texts, RCT (randomized controlled trial) research evidence, and current diagnosis results, and outputs an optimized treatment plan through a recommendation formula, thereby obtaining scientific and reasonable correction suggestions.

[0261] Furthermore, the diagnostic consistency evaluation module also includes a standard mapping library for constructing a mapping knowledge layer based on national standard documents (such as GB / T 16751 "Clinical Terminology of Traditional Chinese Medicine"), which serves as a two-dimensional space; this can cover 820 diseases and syndromes with a coding mapping error rate of less than 6%.

[0262] Furthermore, the logic breakpoint detection unit calculates the probability of logical conflict using a logistic regression model; for example, when the association strength Associate(S,D) between "red tongue with little coating" and "Yang deficiency syndrome" is less than θ (e.g., θ equals 0.3), the probability of logical conflict P is... conflict If the percentage is greater than 75%, it is identified as a contradiction point and the correction suggestion generation unit is activated.

[0263] Furthermore, in the revised suggestion generation unit, the recommendation formula specifically includes:

[0264]

[0265] Where Tx recommendation represents the recommended modified treatment plan; Indicates t pairs The degree of support; t represents the candidate treatment plan; This represents the query conditions formed by combining patient's treatment needs and diagnostic results; β represents the conflict penalty coefficient; T represents the set of candidate treatment options; ω total The integrated weights are represented by the following expression:

[0266] ω total =α·ω ancient +(1-α)·ω RCT

[0267] Where α represents the adjustment coefficient; ω ancient Indicates the weight of the relevant original text; ωRCT This indicates the weight of evidence from RCT studies.

[0268] Furthermore, the efficacy indicator fusion module includes an acute disease model and a chronic disease model;

[0269] The acute disease model is used to collect pain index (digitized VAS scale, i.e., digitized visual analog scale) in real time through wearable health monitoring devices (sampling rate of 100 Hz) and calculate the improvement rate of the area under the characteristic curve (AUC) before and after treatment.

[0270] The chronic disease model calculates a composite efficacy index based on tongue and pulse data, laboratory indicators, and the SF-36 questionnaire (Short Scale for Health Survey). The specific formula includes:

[0271]

[0272] in, The weight of the metric used to reflect changes in tongue structure before and after treatment (such as the improvement rate of tongue coating thickness) is 0.4, in order to reflect the TCM theory of "tongue diagnosis is essential". Used to quantify the degree to which pathological indicators such as rheumatoid factor approach the normal range (weighted at 0.3); Used to reflect the patient's subjective quality of life (weighted at 0.3); the three work together to achieve a multi-dimensional evaluation of "physiological-pathological-subjective".

[0273] Furthermore, the intelligent grading and intervention module includes response criteria and intervention mechanisms;

[0274] The response criteria are based on the K-means clustering method (profile coefficient ≥ 0.65), and are divided into four levels of response, as detailed in Table 1;

[0275] Table 1

[0276]

[0277] The intervention mechanism includes:

[0278] For patients in grade III: Based on the past syndrome data of patients in grade III, the time series characteristics and patterns of the syndrome data are learned through the LSTM model to predict the direction of syndrome evolution within a few days (e.g., 3 days).

[0279] For Level IV patients: Associate with a pre-set emergency medical knowledge base.

[0280] Furthermore, the patient interaction optimization module includes an adaptive scale and a compliance check;

[0281] The adaptive scale assesses the patient's cognitive level and converts professional medical terms (such as "fatigue and weakness") into corresponding quantitative scores (such as a six-level activity difficulty score) to achieve the dynamic adjustment function of the adaptive scale and improve its reliability.

[0282] The adherence check focuses on the dynamic correlation between patient medication behavior and changes in vital signs. When the temporal correlation between the time series of medication records and the evolution of tongue image features is lower than a preset threshold (e.g., 0.4), the medication check dialogue procedure is initiated to ensure the implementation of the treatment plan from the treatment execution stage. The analysis of temporal correlation in this process focuses more on the matching monitoring at the clinical behavior level, the molecular mechanism research of the action targets of the prescription components, or the statistical analysis of correlation coefficients that simply measure the correlation between variables, each with its own emphasis. This system has been shown in multi-center clinical trials (≥5 tertiary hospitals):

[0283] The consistency of dialectical relationships can reach 89%;

[0284] The sensitivity of early warning for syndrome evolution can reach 92.1%;

[0285] The generated efficacy evidence is 94% compatible with the ICD-11-TCM standard;

[0286] The concordance rate between the efficacy prediction and the gold standard reached 85.6%;

[0287] System response latency can be less than 50 milliseconds;

[0288] Patient medication adherence can be improved by 37%;

[0289] Metabolomics methods were used to verify that the concordance rate of Guizhi Decoction in inhibiting the IL-6 inflammatory pathway reached 81.3%.

[0290] Example 2:

[0291] This embodiment provides a large-model-driven multi-dimensional intelligent evaluation device for TCM efficacy, including a processor, a memory, and a bus. The memory stores instructions and data read by the processor, and the processor is used to call the instructions and data in the memory to realize the large-model-driven multi-dimensional intelligent evaluation system for TCM efficacy as described above. The bus connects the various functional components for information transmission.

[0292] Furthermore, the device also includes a tongue imager, a pulse sensor, and a wearable health monitoring device.

[0293] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A large-model-driven multi-dimensional intelligent evaluation system for the efficacy of traditional Chinese medicine, characterized in that, This includes the data layer, knowledge layer, model layer, and application layer, which are set up sequentially. The data layer is used to collect tongue images, pulse data, cardiovascular data, structured data and unstructured data, and to perform structure transformation, blockchain notarization and outlier correction. The knowledge layer includes a TCM diagnosis and treatment knowledge graph, a TCM prescription compatibility database, a efficacy evaluation standard database, a component-target network, a semantic mapping engine, and a syndrome-treatment association graph. The model layer includes a multimodal feature extraction module and a therapeutic effect prediction module; The application layer includes a doctor's side and a patient's side; the doctor's side includes a diagnostic consistency evaluation module, a treatment principle compliance analysis module, and an interpretability feedback module; the patient's side includes a efficacy indicator fusion module, an intelligent grading and intervention module, and a patient interaction optimization module.

2. The system according to claim 1, characterized in that, The multimodal feature extraction module includes a tongue image feature extraction unit and a pulse image temporal unit; The tongue image feature extraction unit specifically includes a DenseNet-201 model, and optimizes the convolutional layers of the DenseNet-201 model by combining grouped convolution and dilated convolution: Grouped convolution divides the input channels into 16 groups, and each group is convolutionally processed independently. Dilated convolution uses a convolution kernel with a dilation rate of 2; The pulse temporal unit specifically includes a WaveNet-LSTM hybrid model: The WaveNet model constructs a multi-scale receptive field through dilated convolutional layers; it generates a 128-dimensional feature vector through the Morlet wavelet transform method, and performs dimensional mapping in combination with the theory of pulse diagnosis in traditional Chinese medicine to quantify the time-frequency characteristics of non-stationary pulse signals. The LSTM model, combined with bidirectional gated recurrent units, learns the pulse wave propagation pattern and rhythm periodic changes through forward and reverse paths, respectively; it is trained using residual connections and layer normalization methods; and an adaptive forgetting gating method is introduced to optimize the memory weights of abnormal rhythms.

3. The system according to claim 2, characterized in that, The WaveNet-LSTM hybrid model also includes a probability value-pulse rate conversion formula, specifically: The formula for absolute difference is as follows: ΔT i =|T i+1 -T i |; Where, ΔT i T represents the absolute difference in the i-th pulse cycle; i+1 Ti represents the (i+1)th pulse cycle; Ti represents the ith pulse cycle. The confidence interval formula is as follows: Among them, CI 95％ This represents the upper limit of the 95% confidence interval; σ represents the mean of the absolute differences. ΔT The standard error represents the absolute difference. The unit conversion formula is as follows: Where t represents the time interval; f represents the pulse rate, in beats per minute; The initial threshold formula is as follows: Thres=max(CI 95％ ,t 标准 ±Δt 误差 ) Where Thres represents the initial threshold, t 标准 Δt represents the standard time interval. 误差 Indicates the allowable error for the time interval; The loss function of the bidirectional gated recurrent unit is: Where MAE represents the mean absolute error; m represents the total number of samples; y j This represents the true value of the j-th sample; This represents the predicted value of the j-th sample.

4. The system according to claim 1, characterized in that, The efficacy prediction module includes a large efficacy prediction model, which specifically includes an input layer, a time-aware coding layer, a Transformer-XL layer, and an output layer arranged sequentially. The input layer is used to input multi-source clinical data and standardize the data, including medication records, acupuncture time, disease data and treatment timestamps; The time-aware coding layer is used to perform feature fusion after absolute time coding and event time-series coding; The absolute time encoding is specifically formulated as follows: PE(t,2i)=sin(t / 10000 2i / 512 ),PE(t,2i+1)=cos(t / 10000 2i / 512 ) Where PE(t,2i) represents the time encoding vector with even index in time interval t; PE(t,2i+1) represents the time encoding vector with odd index in time interval t. The event timing coding refers to binding treatment events with time intervals to generate event vectors with time labels, and controlling coding errors through L2 regularization. The feature fusion refers to concatenating the event vector and the time-encoded vector, and then compressing them to 512 dimensions through a linear transformation; The architectural parameters of the Transformer-XL layer include: Hidden layer dimensions: 512; Attention count: 8; The relative position encoding formula is as follows: Among them, a i，j This represents the attention score between the i-th word and the j-th word; This represents the content encoding of the j-th word; W represents the relative distance encoding between the i-th word and the j-th word. 0 Represents the projection matrix; q i T The query vector representing the i-th word; The specific expression for the layer structure is: FFN(x′)=max(0,x′W1+b1)W2+b2 Where FFN(x′) represents the output of the feedforward neural network; b1 represents the first layer bias vector; b2 represents the second layer bias vector; W1 represents the first layer weight matrix, used to project the input into a high-dimensional space; W2 represents the second layer weight matrix, used to compress the high-dimensional features back to the original dimension; and x′ represents the input vector. The output layer maps the output of the Transformer-XL layer to a preset dimension through a fully connected layer; the preset dimension is specifically t' days x efficacy index; the specific expression is: Among them, h t’ b represents the hidden state vector; out W represents the output bias vector; Softmax represents the normalization function; out Indicates the output projection matrix; t' represents the predicted probability of the efficacy indicator; t' takes the value of an integer between 1 and 14.

5. The system according to claim 1, characterized in that, The specific evaluation methods for the diagnostic consistency evaluation module include: Step a1: Using the generator G and discriminator D of the adversarial generative network, simulate the diagnostic logic of different doctors, and optimize the minimum-maximum game loss function, specifically expressed as: min G max D E x～pdata(x) [logD(x)]+E z～pz(z) [log(1-D(G(z)))] Where pdata(x) represents the distribution of the true dialectical results; pz(z) represents the noise distribution; min G max D E represents a minimax game; x～pdata(x) Let represent the expected distribution of the true dialectical results; D(x) represents the probability that the discriminator judges the input sample to be from the true dialectical results; E z～pz(z) G(z) represents the expected value of the noise distribution; G(z) represents the generator mapping noise to samples similar to the real dialectical results; D(G(z)) represents the probability that the discriminator misclassifies the generated sample as the real dialectical result. After training with clinical diagnostic data, the generator learns the distribution characteristics of physicians' diagnostic reasoning and outputs the distribution of diagnostic results PG(x), with the specific expression as follows: Where δ represents the Dirac function; N represents the number of generated samples; Step a2: Map the generated dialectical result X to a two-dimensional space using a dimensionality reduction algorithm, and estimate the probability density of each coordinate point using a kernel density function. The specific expression is as follows: in, Kσ represents the probability density estimate at position (u, v); Kσ represents the Gaussian kernel function, specifically expressed as: Where d represents the Euclidean distance between the input point and the center point; σ represents the smoothing parameter of the kernel function; The formula for generating a heatmap and calculating its colors is as follows: Wherein, Color(u,v) represents the heatmap color at position (u,v). The warmer the color, the higher the probability of a diagnostic result appearing at that position. This represents the minimum probability density estimate; This represents the largest probability density estimate; Step a3: Calculate the distribution P of the actual diagnostic conclusions. real (x) and the standard distribution of the heatmap P std (x) is compared; the deviation is quantified using KL divergence, the specific expression of which is: Here, Deviation represents the degree of deviation.

6. The system according to claim 1, characterized in that, The diagnostic consistency evaluation module also includes a dynamic correction method, specifically including: Step b1: Convert the RGB chromaticity values ​​of the tongue image to HSV space. The specific expression is as follows: Where H represents chromaticity; S represents saturation; and V represents brightness. Step b2: Calculate the average chromaticity of the tongue region in the tongue image. With average saturation The specific formulas include: Where n represents the total number of pixels in the tongue image region; H i S represents the chromaticity of the i-th pixel; i This represents the saturation of the i-th pixel; Step b3: Calculate the amplitude ratio of the main wave and the depth of the diphthoplasty notch; The specific formula for the amplitude ratio of the main wave is: Where A represents the amplitude ratio of the main wave; H1 represents the height of the main wave; and H0 represents the baseline height. The specific formula for calculating the depth of the diphtheria wave notch is as follows: Where A' represents the depth of the diphtheria wave notch; H2 represents the height of the diphtheria wave trough. Step b4: Calculate the chord slip S using the chord slip exponential model. string The specific formula is as follows: S string =w1×A+w2×D’+w3×f peak ; Where w1, w2, and w3 are all weighting coefficients; D' represents the differential characteristic of the pulse waveform; f peak Indicates the peak frequency of the pulse wave; Step b5: Measure the consistency of symptoms using a composite metric function. Specific formulas include: Where d(T, P) represents the Euclidean distance between the tongue image feature vector and the pulse image feature vector; T represents the tongue image feature vector, specifically: P represents the pulse feature vector, specifically: P = [S string f peak ];μ H Indicates the mean chromaticity under normal symptoms; μ S μ represents the mean saturation value under normal symptoms. P This represents the mean value of string smoothness under normal symptoms.

7. The system according to claim 1, characterized in that, The governance principle compliance analysis module includes an association rule base and a time-series monitoring model; The association rule base is used to integrate the rules of monarch, minister, assistant and guide in classic prescriptions in ancient Chinese medicine books with clinical syndrome-treatment scenarios to construct a syndrome-treatment association map. The time-series monitoring model uses an LSTM network to perform deep learning on prescription adjustments, tongue and pulse changes, and prescription target data in continuous time series. Based on the correlation coefficient between changes in tongue and pulse and the target data of the prescription, the correlation between prescription adjustment and syndrome evolution is determined: if the correlation coefficient is less than the preset threshold, it indicates that the correlation is insufficient and a treatment deviation report is generated.

8. The system according to claim 1, characterized in that, The interpretability feedback module includes a logic breakpoint detection unit and a correction suggestion generation unit; The logical breakpoint detection unit, based on the interpretability feedback mechanism of the knowledge graph comparison method, maps the diagnosis results and physical signs data to the TCM diagnosis and treatment knowledge graph respectively. By analyzing the relevant nodes and relationships, it calculates the probability of logical conflict between the diagnosis results and physical signs data and determines the contradiction point: if the probability of logical conflict is greater than the preset threshold, it is determined to be a contradiction point. The correction suggestion generation unit retrieves relevant original texts from ancient Chinese medicine books based on the syndrome differentiation results, and screens RCT research evidence with Jadad scores ≥4; it then integrates the relevant original texts, RCT research evidence, and current syndrome differentiation results, and outputs an optimized treatment plan through a recommendation formula.

9. The system according to claim 8, characterized in that, In the logic breakpoint detection unit, the probability of logical conflict is calculated using a logistic regression model, and the specific formula includes: Among them, P conflict Let S represent the probability of logical conflict; D represent the mapping vector of the diagnostic results; θ represent the correlation benchmark threshold; e represent the natural constant; λ represent the correlation influence coefficient; Associate(S,D) represents the correlation strength between the diagnostic results and the physical characteristics data, and its specific expression is as follows: Where W represents the weight matrix.

10. A large-scale model-driven multi-dimensional intelligent evaluation device for TCM efficacy, characterized in that, The system includes a processor, a memory, and a bus. The memory stores instructions and data read by the processor, and the processor is used to call the instructions and data in the memory to implement the system as described in any one of claims 1-9. The bus connects the functional components for transmitting information.

Images