Traditional Chinese medicine diagnosis and treatment aid decision-making method and system

Abstract

The invention discloses a traditional Chinese medicine diagnosis and treatment aided decision-making method and system, and relates to the technical field of traditional Chinese medicine intelligent diagnosis and treatment aided decision-making, and the method comprises the steps: carrying out the wavelet transformation analysis of pulse condition waveform time series data, extracting a pulse condition dynamic change feature sequence, inputting a tongue image color change rate and a pulse condition feature change rate into a space-time convolution network, and obtaining a time-space convolution network; generating a tongue pulse dynamic association index and outputting a preliminary syndrome classification; fourier transform analysis is carried out on the spleen channel acupoint impedance spectrum, false-true state judgment is carried out in combination with tongue pulse dynamic correlation indexes, preliminary syndrome classification is corrected, and a false-true state quantification result and corrected syndrome classification are output; and matching the corrected syndrome classification with the traditional Chinese medicine knowledge map, adjusting the dosage according to a virtual-real state quantification result, reversely verifying the drug-syndrome specificity through a curative effect database, screening an optimal treatment scheme, and outputting a personalized diagnosis and treatment scheme. According to the method, the syndrome classification precision is improved, and an objective basis is provided for personalized dose adjustment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent diagnosis and treatment decision-making in traditional Chinese medicine, and in particular to a traditional Chinese medicine diagnosis and treatment decision-making method and system. BACKGROUND

[0002] In the field of traditional Chinese medicine diagnosis and treatment, the objectification and quantification of traditional four examinations (inspection, auscultation and olfaction, interrogation, and palpation) have become a key direction for the development of modern traditional Chinese medicine. In recent years, with the progress of computer vision, signal processing and artificial intelligence technology, auxiliary diagnostic tools such as tongue image analysis and pulse diagnosis instrument have been gradually applied in clinical practice. In the prior art, methods such as HSV color space analysis based on static tongue image and wavelet feature extraction of pulse waveform have been proposed, and preliminary classification of syndromes is realized through single modal data processing. Similarly, in the field of in vitro diagnosis, such as the development of sexually transmitted disease detection kits, methods such as colloidal gold labeling technology and nitrocellulose membrane coating optimization are also committed to improving the sensitivity and specificity of detection, which reflects the common needs of cross-field technology for precision medicine.

[0003] The existing auxiliary diagnosis technology of traditional Chinese medicine still has limitations. On the one hand, static tongue analysis is difficult to capture the relevance of dynamic changes in tongue color and evolution of syndromes, and isolated analysis of pulse characteristics ignores the dynamic physiological significance of tongue and pulse coordination; on the other hand, the existing methods rely on a single data source (such as only tongue or pulse), lack of spatiotemporal correlation modeling of multi-modal data, resulting in insufficient accuracy of virtual and real state determination. If only the static features of tongue or pulse are relied on in traditional Chinese medicine diagnosis and treatment without integrating multi-dimensional physiological parameters such as spleen meridian acupoint impedance, the overall state of the patient may not be fully reflected, thereby affecting the accuracy of syndrome differentiation and the individualization degree of treatment plan. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a traditional Chinese medicine diagnosis and treatment auxiliary decision-making method to solve the problem of inaccurate virtual and real state determination due to single data source dependence and insufficient dynamic feature modeling.

[0006] To solve the above technical problems, the present application provides the following technical solutions: In a first aspect, the present application provides a traditional Chinese medicine diagnosis and treatment auxiliary decision-making method, which comprises: acquiring a tongue video sequence, and collecting pulse waveform time series data and spleen meridian acupoint impedance spectrum; performing image processing on the tongue video sequence, extracting the HSV color space features of each frame of tongue surface, and calculating the tongue color change rate between adjacent frames; performing wavelet transform analysis on the pulse waveform time series data, extracting the pulse dynamic change feature sequence, and inputting the tongue color change rate and pulse feature change rate into a spatiotemporal convolution network to generate a tongue-pulse dynamic correlation index and output a preliminary syndrome classification; Fourier transform analysis is performed on the spleen meridian acupoint impedance spectrum, the dynamic correlation index of the tongue pulse is combined to determine the deficiency and excess state, the preliminary syndrome classification is corrected, and the quantitative results of the deficiency and excess state and the corrected syndrome classification are output; The corrected syndrome classification is matched with the traditional Chinese medicine knowledge graph, the dosage is adjusted according to the quantitative results of the deficiency and excess state, the drug-syndrome specificity is verified reversely through the efficacy database, and the optimal treatment scheme is screened to output the individualized diagnosis and treatment scheme.

[0007] As a preferred scheme of the traditional Chinese medicine diagnosis and treatment auxiliary decision method, the tongue video sequence includes tongue dynamic morphology, tongue color change, tongue fur thickness and dampness and dryness degree. The pulse waveform time sequence data includes pulse position floating and sinking, pulse rate delay and pulse force strength. The spleen meridian acupoint impedance spectrum includes the resistance impedance amplitude and phase angle of the spleen meridian acupoint under different frequencies.

[0008] As a preferred scheme of the traditional Chinese medicine diagnosis and treatment auxiliary decision method, the tongue color change rate between adjacent frames is calculated, and the specific steps are as follows, The tongue video sequence is subjected to bilateral filtering denoising processing to obtain a tongue surface image with complete edge information, and is input into a pre-trained deep segmentation model for region extraction to output a tongue surface binary mask. The tongue surface image is cropped using the tongue surface binary mask to extract a pure tongue surface region, color calibration is performed in combination with a standard color card to obtain a white balance tongue surface image, and the image is converted to an HSV color space to extract the HSV average features of each frame. The tongue color change rate is calculated according to the difference between the HSV average features of adjacent two frames.

[0009] As a preferred scheme of the traditional Chinese medicine diagnosis and treatment auxiliary decision method, the pulse waveform time sequence data is subjected to wavelet transform analysis to extract the pulse dynamic change feature sequence, and the specific steps are as follows, Baseline drift correction and band-pass filtering processing are performed on the pulse waveform time sequence data to obtain high signal-to-noise ratio pulse preprocessing waveform time sequence data, and resampling and amplitude normalization are performed to generate uniform amplitude time sequence data, and the discrete wavelet decomposition is performed to separate the pulse detail feature components; The pulse detail feature components are subjected to envelope extraction to obtain an envelope curve reflecting the pulse force change, peak value detection is performed in combination with a physiological rhythm to set a pulse peak value detection threshold, the peak value position and amplitude information in each pulse fluctuation period are extracted to form a pulse position dynamic time sequence feature and a peak value amplitude sequence, respectively; The pulse dynamic change feature sequence is obtained through maximum minimum normalization standardization processing.

[0010] As a preferred scheme of the traditional Chinese medicine diagnosis and treatment auxiliary decision method, the specific steps of generating the tongue and pulse dynamic correlation index and outputting the preliminary syndrome classification are as follows, The feature value difference of adjacent time points is calculated according to the sampling time stamp sequence of the pulse condition dynamic change characteristic sequence, and the pulse condition feature change rate is calculated in combination with the time interval of the feature value difference of adjacent time points, and the time alignment and standardization processing are performed on the pulse condition feature change rate and the tongue condition color change rate, to generate a synchronous input sequence; The synchronous input sequence is input into the parallel branch of the double-flow space-time convolution network, local time sequence features are extracted by using time convolution, key structure information is encoded by using space convolution, and residual connection and normalization are superimposed in multiple convolution layers to extract preliminary space-time features of the tongue condition and the pulse condition; The preliminary space-time features of the tongue condition and the pulse condition are respectively converted into query, key and value vectors, the attention relationship of the tongue condition to the pulse condition is established to extract the correlation features of the tongue to the pulse, and the attention relationship of the pulse condition to the tongue condition is established to extract the response features of the pulse to the tongue, the two are cross-modal fused to obtain tongue-pulse interaction features; The tongue-pulse interaction features are iteratively fused through multiple layers of cross-attention and residual connection to generate refined space-time correlation features, and global aggregation is performed to output the tongue-pulse dynamic correlation index, and the preliminary syndrome classification is output through the classifier.

[0011] As a preferred scheme of the traditional Chinese medicine diagnosis and treatment auxiliary decision method, the specific steps of outputting the virtual and real state quantitative results and the corrected syndrome classification are as follows, The impedance spectrum of the spleen meridian acupoint is analyzed by fast Fourier transform to extract the impedance phase angle sequence of each frequency point, a time sequence phase vector is generated, and the feature dimension is spliced with the tongue and pulse dynamic correlation index to form a multi-modal node feature; The similarity between the multi-modal node features is calculated by cosine similarity, and a dynamic space-time graph is constructed, and the potential similar syndrome groups are identified through the high similarity of adjacent nodes in the dynamic space-time graph; The dynamic space-time graph is updated by message passing through the graph attention layer, the cross interference between similar syndromes is corrected by adaptive weight, the graph attention update is repeated, and the graph features of different levels are fused through the jump connection to generate refined node representations; The refined node representations are graphically aggregated to generate a virtual and real state quantitative vector, and the corrected syndrome classification is output through the classifier.

[0012] As a preferred scheme of the traditional Chinese medicine diagnosis and treatment auxiliary decision method, the specific steps of outputting the individualized diagnosis and treatment scheme are as follows, The corrected syndrome classification is taken as a query keyword, the corresponding syndrome entity is matched in the traditional Chinese medicine knowledge graph, and the associated prescriptions and recommended dosages are obtained along the treatment relationship to generate a prescription initial dosage; According to the fuzzy inference rule, the initial dose of the prescription is adjusted and mapped according to the virtual and real state quantization result, and a personalized dose is generated; The individualized dose is queried in the therapeutic effect database to obtain a historical therapeutic effect record, a drug-syndrome specificity score is calculated, a high specificity combination is screened through a specificity score threshold, and the optimal prescription combination is screened as a candidate scheme in descending order of the score and in combination with a patient file constraint, so that an individualized diagnosis and treatment scheme is output.

[0013] In the second aspect, the present application provides a traditional Chinese medicine diagnosis and treatment auxiliary decision system, comprising: a collection module, which is used for acquiring a tongue image video sequence and collecting pulse waveform time series data and spleen meridian acupoint impedance spectrum; An extraction module is used for image processing on the tongue image video sequence, extracting HSV color space features of each frame of tongue surface, and calculating tongue color change rate between adjacent frames; An analysis module is used for wavelet transform analysis on the pulse waveform time series data, extracting a pulse dynamic change feature sequence, inputting the tongue color change rate and the pulse feature change rate into a space-time convolution network, generating a tongue-pulse dynamic correlation index, and outputting a preliminary syndrome classification; A correction module is used for Fourier transform analysis on the spleen meridian acupoint impedance spectrum, combining the tongue-pulse dynamic correlation index to determine a virtual and real state, correcting the preliminary syndrome classification, and outputting a virtual and real state quantization result and a corrected syndrome classification; An output module is used for matching the corrected syndrome classification with a traditional Chinese medicine knowledge graph, adjusting a drug amount according to the virtual and real state quantization result, verifying drug-syndrome specificity in reverse through a therapeutic effect database, screening an optimal treatment scheme, and outputting an individualized diagnosis and treatment scheme.

[0014] In the third aspect, the present application provides a computer device, comprising a memory and a processor, and the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the traditional Chinese medicine diagnosis and treatment auxiliary decision method according to the first aspect of the present application is realized.

[0015] In the fourth aspect, the present application provides a computer readable storage medium, which stores a computer program, wherein: when the computer program is executed by the processor, any step of the traditional Chinese medicine diagnosis and treatment auxiliary decision method according to the first aspect of the present application is realized.

[0016] The present application has the following beneficial effects: by dynamically collecting tongue image, pulse image and spleen meridian acupoint impedance data, using a space-time convolution network and a cross-modal attention mechanism to establish a tongue-pulse dynamic correlation index, realizing digital expression of traditional Chinese medicine "tongue-pulse combined consultation", and combining frequency domain features of the spleen meridian impedance spectrum to construct a dynamic space-time graph, quantizing virtual and real syndrome differentiation results, forming a "macroscopic dynamic-microscopic energy" dual verification system, not only improving the accuracy of syndrome classification, but also providing an objective basis for individualized drug amount adjustment. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only constitute some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without any creative effort.

[0018] Fig. 1 The flow chart of the TCM diagnosis and treatment auxiliary decision method.

[0019] Fig. 2 The schematic diagram of the TCM diagnosis and treatment auxiliary decision system.

[0020] Fig. 3 The flow chart of the wavelet transform analysis of the pulse waveform time series data.

[0021] Fig. 4 The flow chart of the spatiotemporal convolution network generation of the tongue and pulse dynamic correlation index. DETAILED DESCRIPTION

[0022] In order to make the above objectives, features and advantages of the present application more apparent and comprehensible, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0023] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the connotation of the present application, so the present application is not limited to the specific embodiments disclosed below.

[0024] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or selective embodiment that excludes other embodiments.

[0025] Reference Figs. 1-4 For one embodiment of the present application, the embodiment provides a TCM diagnosis and treatment auxiliary decision method, comprising the following steps: S1, acquiring a tongue video sequence, and collecting pulse waveform time series data and spleen meridian acupoint impedance spectrum.

[0026] The tongue video sequence includes tongue body dynamic morphology, tongue color change, and tongue fur thickness and dampness.

[0027] The pulse waveform time series data includes pulse position floating and sinking, pulse rate delay number, and pulse force strength.

[0028] The impedance spectrum of the acupoint of the spleen meridian includes the resistance impedance amplitude and phase angle of the acupoint of the spleen meridian at different frequencies.

[0029] S2, image processing is performed on the tongue video sequence, HSV color space features of each frame of tongue surface are extracted, and tongue color change rates between adjacent frames are calculated.

[0030] The tongue video sequence is subjected to bilateral filtering denoising processing to obtain a tongue surface image with complete edge information, and is input into a pre-trained deep segmentation model for region extraction, and a tongue surface binary mask is output.

[0031] It should be noted that based on the pixel spatial information and color similarity in each frame of tongue image, the tongue image is subjected to bilateral filtering operation, which effectively removes image noise while keeping the image edge clear, so that the tongue surface edge profile in the tongue image remains complete. The tongue image after bilateral filtering is input frame by frame into the pre-trained deep segmentation model, and the tongue surface region in each frame of tongue image is classified pixel by pixel by the deep segmentation model, and the segmentation result corresponding to the tongue surface region is extracted. The tongue surface region in the segmentation result is further subjected to binary processing, the tongue surface pixels in the tongue image are set as foreground using a fixed threshold, and the rest is set as background, to obtain a tongue surface binary mask containing only the tongue surface region. It should be noted that the deep segmentation model is trained in a supervised learning manner, the cross-entropy loss function is used to measure the difference between the predicted mask and the true label, and the deep segmentation model parameters are optimized through the back propagation algorithm. In the training process, the tongue image and its corresponding tongue surface label are used as input and supervision information, the segmentation accuracy of the deep segmentation model for the tongue surface region is improved through repeated iteration, and finally the deep segmentation model for tongue surface region extraction is obtained. The fixed threshold is a gray or probability threshold used for binary processing of the tongue surface segmentation result output by the deep segmentation model, which is commonly used to determine whether each pixel belongs to the tongue surface region. The specific form of this fixed threshold is usually a real value between 0 and 1.

[0032] The tongue surface binary mask is used to crop the tongue surface image to extract a pure tongue surface region, color calibration is performed in combination with a standard color card to obtain a white balanced tongue surface image, and the HSV average features of each frame are extracted by converting to the HSV color space.

[0033] It should be noted that after obtaining the tongue surface binary mask, the tongue surface image is subjected to pixel-level masking processing using the tongue surface binary mask, the corresponding pixel values of the tongue surface region are retained and the non-tongue surface region is shielded, the cropping operation of the complete tongue surface image is realized, and the image content containing only the tongue surface region is obtained as a pure tongue surface region image. In the pure tongue surface area image, combined with the standard color card image placed in the shooting area beforehand, the image pixel value of each standard color area on the standard color card is identified and compared with the known standard color value. The gain adjustment coefficient of the color channel is calculated, and the gain calibration is performed on the RGB channels respectively to realize the color calibration processing of the pure tongue surface area image and obtain the tongue surface image after white balance correction. After white balance correction, the tongue image is converted from RGB color space to HSV color space according to the channel conversion formula. The three channels of hue, saturation and value are separated. The average value of the three channels is calculated on the tongue area pixels of each frame of the tongue image, which is used as the HSV average feature of the tongue image of that frame.

[0034] The tongue color change rate is calculated based on the average HSV feature difference between two adjacent frames.

[0035] It should be noted that after color calibration and conversion to the HSV color space of the tongue image sequence, an HSV image is obtained. In the HSV image, each pixel is represented as a three-dimensional vector, where the first dimension is the hue value, the second dimension is the saturation value, and the third dimension is the brightness value. By separating the pixel matrix of the HSV image by channel, the first channel of all pixels is extracted to form the hue channel matrix, the second channel to form the saturation channel matrix, and the third channel to form the brightness channel matrix. By summing the values ​​of all pixels in each channel matrix, the average hue value representing the hue feature of the entire tongue image, the average saturation value representing the saturation feature, and the average brightness value representing the brightness feature are obtained. The color change information is then fused using Euclidean distance calculation to calculate the tongue image color change rate, expressed as: ; In the formula, It is the first Frame relative to the first The rate of tongue color change in the frame. It is the first Average hue value within the ROI on the frame tongue surface. It is the first Average hue value within the ROI on the frame tongue surface. It is the first Average saturation value within the ROI on the frame's tongue surface. It is the first Average saturation value within the ROI on the frame's tongue surface. It is the first The average brightness (luminance) value within the ROI on the frame's tongue surface. It is the first The average brightness (luminance) value within the ROI on the frame's tongue surface. is the first frame and the second time interval between the frames, is the index of the frame in the video sequence.

[0036] S3, wavelet transform analysis is performed on the pulse wave form time series data, pulse dynamic change characteristic sequences are extracted, tongue color change rate and pulse characteristic change rate are input into a space-time convolution network, tongue and pulse dynamic correlation indexes are generated, and preliminary syndrome classification is output.

[0037] Baseline drift correction and band-pass filtering processing are performed on the pulse wave form time series data to obtain high signal-to-noise ratio pulse preprocessing wave form time series data.

[0038] It should be noted that the pulse wave form time series data is first subjected to baseline drift correction, and the local trend component of the pulse wave form time series data is obtained by using the moving average method or the polynomial fitting method, and the local trend component is subtracted from the original pulse wave form time series data point by point to eliminate non-physiological low-frequency drift caused by external factors. After completing the baseline drift correction, the pulse wave form time series data is subjected to band-pass filtering processing, and existing filtering methods such as Butterworth filter or wavelet transform are selected, and the frequency passband range is set, for example, the example value is 0.5 Hz to 8 Hz, only the effective frequency component in the pulse wave form time series data within the frequency band is retained, thereby suppressing noise interference and false signal, improving the signal-to-noise ratio of the pulse wave form time series data, and finally obtaining high signal-to-noise ratio pulse preprocessing wave form time series data.

[0039] The pulse preprocessing wave form time series data is resampled and amplitude normalized to generate uniform amplitude time series data, and discrete wavelet decomposition is performed to separate the pulse detail feature components.

[0040] It should be noted that the pulse preprocessing wave form time series data is subjected to interpolation algorithm to unify the number of sampling points of the pulse preprocessing wave form time series data to the same number, so that all pulse preprocessing wave form time series data have comparability in time dimension. After resampling, the resampled pulse preprocessing wave form time series data is subjected to amplitude normalization processing, and the amplitude values of all sampling points are uniformly mapped to a specified standard amplitude interval, for example, the example value is [0, 1], to eliminate the influence caused by the amplitude scale difference between different samples, and obtain uniform amplitude time series data. After obtaining the uniform amplitude time series data, the uniform amplitude time series data is subjected to discrete wavelet decomposition by using the discrete wavelet transform method, and the uniform amplitude time series data is decomposed into approximation components and detail components at different scales, and the detail component part is retained to separate the pulse detail feature components.

[0041] The envelope curve reflecting the pulse change is obtained by envelope extraction of the pulse detail feature component, and the peak value detection threshold is set in combination with the physiological rhythm to perform peak value detection, the peak position and amplitude information in each pulse fluctuation period are extracted to form the pulse position dynamic time sequence feature and the peak amplitude sequence respectively.

[0042] It should be noted that the Hilbert transform operation is input according to the original time sequence order for the pulse detail feature component, the complex signal composed of the real part and the imaginary part is obtained by calculating the analytic signal form of the time sequence signal, and the amplitude of the complex signal is calculated on this basis, that is, the complex signal at each time is taken as a modulus to form a continuous waveform sequence of all positive numbers, and the obtained waveform is the envelope curve reflecting the trend of pulse change; The expected pulse interval range is calculated by combining the physiological rhythm and the pulse waveform time sequence data sampling frequency, the pulse peak value detection threshold is set by using the local peak and valley mean value, and each peak value higher than the pulse peak value detection threshold and meeting the minimum time interval constraint is identified by scanning the envelope curve, the time index is recorded at each identified peak position and arranged in time sequence to obtain the pulse position dynamic time sequence feature, and the envelope curve amplitude corresponding to the same position is extracted and combined in time sequence to form the peak amplitude sequence; It should also be noted that the physiological rhythm refers to the periodic change law of the pulse fluctuation in the natural state, which is usually characterized by the rhythm of the repeated appearance of the pulse wave in a certain time range with a relatively stable frequency; by analyzing the time interval and amplitude distribution trend of the fluctuation peak in each cycle of the envelope curve reflecting the pulse detail feature component, and combining the average cycle length and amplitude change range of multiple consecutive pulse fluctuation cycles, the pulse fluctuation rhythm characteristics of the current individual in the resting state can be derived.

[0043] The pulse dynamic change feature sequence is obtained by maximum and minimum normalization standardization processing.

[0044] It should be noted that after obtaining the pulse position dynamic time sequence feature and the peak amplitude sequence, maximum and minimum normalization standardization processing is performed on each pulse peak value feature value in the pulse position dynamic time sequence feature and the peak amplitude sequence respectively, the original pulse peak value feature value is transformed to a unified pulse peak value feature value interval according to the elements, the maximum value corresponds to the normalized maximum value, the minimum value corresponds to the normalized minimum value, and other pulse peak value feature values are mapped to the corresponding normalized values according to their positions in the original range, so as to eliminate the interference caused by individual feature differences or different original amplitudes of the signal, so that the pulse position dynamic time sequence feature and the peak amplitude sequence have a unified feature expression scale, thereby forming a pulse dynamic change feature sequence that can stably reflect the trend of pulse change; The pulse peak value feature value is detected by detecting the peak position and corresponding amplitude in each pulse fluctuation period in the envelope curve, and each extracted amplitude is the pulse peak value feature value.

[0045] The characteristic value difference between adjacent moments is calculated according to the sampling timestamp sequence of the pulse dynamic change characteristic sequence, and the pulse characteristic change rate is calculated by combining the time interval of the characteristic value difference between adjacent moments.

[0046] It should be noted that the characteristic value difference between adjacent moments in the pulse dynamic change characteristic sequence is calculated according to the sampling timestamp order, and the expression is as follows: ; In the formula, It is the first The feature value difference between adjacent sampling points of each sampling point It is the first The feature values ​​of each sampling point It is the first The feature values ​​of each sampling point It is the index of the sampling point; Based on the corresponding sampling timestamps, the time interval between two sampling points is calculated using the following expression: ; In the formula, It is the first The time interval between sampling points It is the first The timestamp of each sampling point It is the first Timestamp of each sampling point; Normalize the change over a time scale to obtain the pulse feature change rate, which describes the speed of change of the features. The expression is: ; In the formula, It is the first The rate of change of pulse characteristics at each sampling point.

[0047] The tongue color change rate and pulse feature change rate are time-aligned and standardized to generate a synchronous input sequence.

[0048] It should be noted that, based on the frame timestamps of the tongue image video sequence and the sampling timestamps of the pulse waveform time series data, the tongue image color change rate and the pulse feature change rate are aligned along the time axis to ensure that the two have a corresponding relationship at the same time point. During the alignment process, if the timestamps do not match perfectly, the feature values ​​of the missing time points are supplemented by linear interpolation. The tongue color change rate and pulse feature change rate of the aligned tongue image are standardized by Z-score. The standardized tongue color change rate and pulse feature change rate are concatenated in chronological order to generate a synchronous input sequence with a unified time scale and dimension.

[0049] The synchronous input sequence is input into the parallel branches of the dual-flow spatio-temporal convolution network, local temporal features are extracted by time convolution, key structural information is encoded by spatial convolution, and residual connections and normalization are added in multiple convolution layers to extract preliminary spatio-temporal features of tongue and pulse images.

[0050] It should be noted that the synchronous input sequence is input into the two parallel branches of the dual-flow spatio-temporal convolution network for processing: in the tongue branch, the time convolution kernel slides along the time dimension to capture the local temporal pattern of continuous multiple frames in the tongue color change rate; in the pulse branch, the time convolution kernel extracts the short-time fluctuation features of the pulse feature change rate in the same way; The spatial convolution layers of the two branches respectively encode the spatial distribution of the tongue mask area and the morphological structure of the pulse waveform, and perform batch normalization operation after each convolution layer to stabilize the data distribution and pass the shallow layer features to the deep layer through the residual connection; The feature maps output by the dual-flow branches are concatenated in the channel dimension to form a preliminary spatio-temporal feature representation containing the correlation between tongue dynamic changes and pulse fluctuations; It should also be noted that the dual-flow spatio-temporal convolution network uses a cross-entropy loss function to supervise training, uses labeled data of tongue video frames and pulse waveform pairs, optimizes the parameters of the dual-flow spatio-temporal convolution network through back propagation, and obtains the trained dual-flow spatio-temporal convolution network.

[0051] The preliminary spatio-temporal features of tongue and pulse are respectively converted into query, key and value vectors to establish the attention relationship between tongue and pulse, extract the correlation features of tongue and pulse, and establish the attention relationship between pulse and tongue to extract the response features of pulse and tongue. Cross-modal fusion is performed on the two to obtain tongue-pulse interaction features.

[0052] It should be noted that the preliminary spatio-temporal features of tongue are transformed into query vectors through a trainable weight matrix, and the preliminary spatio-temporal features of pulse are transformed into key vectors and value vectors respectively. The dot product of the tongue query vector and the pulse key vector is calculated to obtain an attention score matrix, which is scaled and normalized by softmax to obtain an attention weight matrix, which is then weighted and summed with the pulse value vector to generate the correlation features of tongue and pulse. At the same time, the preliminary spatio-temporal features of pulse are transformed into query vectors, and the preliminary spatio-temporal features of tongue are transformed into key vectors and value vectors. The dot product of the pulse query vector and the tongue key vector is calculated to obtain an attention score matrix, which is processed in the same way and then weighted and summed with the tongue value vector to generate the response features of pulse and tongue. The preliminary fusion features are obtained by element-by-element addition of the tongue-pulse correlation features and the pulse-tongue response features, followed by linear transformation through a fully connected layer (the dimension of the weight matrix is exemplified as 512x1024), application of a GeLU activation function for nonlinear mapping, and finally reduction of the feature dimension through an output layer (the dimension of the weight matrix is exemplified as 1024x512) to generate tongue-pulse interaction features containing cross-modal interaction information.

[0053] The tongue-pulse interaction features are iteratively fused through multiple layers of cross-attention and connected in residual to generate refined spatiotemporal correlation features, which are globally aggregated to output tongue-pulse dynamic correlation indicators and preliminary syndrome classification through a classifier.

[0054] It should be noted that the tongue-pulse interaction features are input into a multi-layer cross-attention structure. Each layer first splits the tongue-pulse interaction features into query vectors, key vectors, and value vectors. The attention scores are calculated through matrix multiplication of the query vectors and the key vectors, normalized by softmax, and then weighted summed with the value vectors to obtain updated tongue-pulse interaction features. The updated tongue-pulse interaction features are connected in residual with the tongue-pulse interaction features, normalized by layer, and after multiple iterations (exemplified as 6 layers), the refined spatiotemporal correlation features are finally output. The final output of the refined spatiotemporal correlation features is input into a global average pooling layer for feature compression to obtain tongue-pulse dynamic correlation indicators of fixed dimension. The tongue-pulse dynamic correlation indicators are input into a fully connected layer classifier (exemplified as output dimension of 8 classes) to calculate the probability of each syndrome class through a softmax activation function and output the preliminary syndrome classification.

[0055] S4, Fourier transform analysis is performed on the spleen meridian acupoint impedance spectrum, and the virtual-actual state is determined in combination with the tongue-pulse dynamic correlation indicators to correct the preliminary syndrome classification and output the virtual-actual state quantification results and the corrected syndrome classification.

[0056] The spleen meridian acupoint impedance spectrum is analyzed by fast Fourier transform, the impedance phase angle sequence of each frequency point is extracted, a time series phase vector is generated, and the feature dimension is spliced with the tongue-pulse dynamic correlation indicators to form a multi-modal node feature.

[0057] It should be noted that the spleen meridian acupoint impedance spectrum is decomposed into a frequency domain representation by fast Fourier transform, and the impedance phase angle of each frequency point at each sampling time is extracted within the frequency range (exemplified as 1Hz-100kHz) to form a time series phase vector in chronological order. The tongue-pulse dynamic correlation indicators are time-aligned time series features, and the time series phase vector is spliced in the feature dimension to form a multi-modal node feature containing both the spleen meridian acupoint impedance spectrum phase information and the tongue-pulse dynamic correlation indicators. Before splicing, both are normalized by maximum and minimum to ensure consistent feature scales.

[0058] By calculating the similarity between features of multimodal nodes using cosine similarity, a dynamic spatiotemporal graph is constructed, and potential similar syndrome groups are identified from the high similarity between adjacent nodes in the dynamic spatiotemporal graph.

[0059] It should be noted that the similarity between features of multimodal nodes is calculated using cosine similarity, and the expression is as follows: ; In the formula, It is the first The features of the multimodal node and the first Similarity between features of multimodal nodes It is the first Multimodal node features, It is the first Multimodal node features, It is the first The Euclidean norm of multimodal node features, It is the first The Euclidean norm of a multimodal node feature; Based on the similarity between multimodal node features, the k-nearest neighbor graph construction method (with an example value of 8) is used to select the neighboring nodes with the highest similarity for each multimodal node feature to establish connection edges, forming a dynamic spatiotemporal graph; In a dynamic spatiotemporal graph, a depth-first search algorithm is applied to traverse connected subgraphs. Connected subgraphs with node degrees greater than the population size threshold (example value: 5) are marked, and the multimodal node feature set corresponding to all nodes in the connected subgraph is extracted. Consistency within the population is verified by calculating the ratio of the average similarity between nodes in the connected subgraph to the median of the global similarity (example value: 1.2 times). Finally, the set of nodes that meet the consistency condition is output as a potential similarity syndrome group.

[0060] The dynamic spatiotemporal graph is updated by message passing using a graph attention layer. Cross-interference between similar symptoms is corrected by adaptive weights. The graph attention update is repeated and combined with skip connections to fuse graph features at different levels to generate a refined node representation.

[0061] It should be noted that the dynamic spatiotemporal graph is input into the multi-head graph attention layer (example value: 8 heads). Each attention head generates query vector, key vector and value vector for the multimodal node features of each node in the dynamic spatiotemporal graph through a trainable weight matrix. The attention coefficients of the node and its neighboring nodes are calculated and normalized. The current node features are updated by weighted aggregation of the value vectors of the neighboring nodes. After multi-layer graph attention update (example value: 3 layers), the output of each layer is added to the initial multi-modal node feature through a jump connection, and then processed by layer normalization. The node features output by each layer are concatenated in the channel dimension (example value: dimension 1536 after concatenation), and then compressed by a fully connected layer (example value: output dimension 512) to generate refined node representations.

[0062] The refined node representations are aggregated at the graph level to generate virtual-real state quantization vectors, and a corrected syndrome classification is output by a classifier.

[0063] It should be noted that the feature vectors of all nodes are aggregated by graph-level average pooling operation based on the refined node representations, and the graph-level feature representation with unchanged dimension is generated. The graph-level feature representation with unchanged dimension is input into a virtual-real state quantization module composed of a fully connected layer (example value: hidden units 256) and a ReLU activation function, and an 8-dimensional virtual-real state quantization vector is output. At the same time, the graph-level feature representation with unchanged dimension is input into another fully connected layer classifier (example value: output dimension corresponding to the number of syndrome categories), and the probability distribution of each syndrome category is calculated by a softmax function, and the corrected syndrome classification result is output.

[0064] S5, match the corrected syndrome classification with the traditional Chinese medicine knowledge graph, and adjust the dosage according to the virtual-real state quantization result. The drug-syndrome specificity is verified reversely through the efficacy database, and the optimal treatment plan is selected to output the individualized diagnosis and treatment plan.

[0065] The corrected syndrome classification is used as a query keyword to match the corresponding syndrome entity in the traditional Chinese medicine knowledge graph and obtain associated prescriptions and recommended dosages along the treatment relationship to generate the initial dosage of the prescription.

[0066] It should be noted that the corrected syndrome classification is input into the traditional Chinese medicine knowledge graph query interface as a query keyword, and an accurate match is performed in the syndrome entity node set of the traditional Chinese medicine knowledge graph to find the corresponding syndrome entity node. All prescription entity nodes directly connected to the syndrome entity node are obtained by traversing the "treatment" relationship edge defined in the traditional Chinese medicine knowledge graph, and the medicinal material entities connected by the "constituent medicinal material" relationship edge and their "recommended dosage" attribute values (example value: Huangqi 10g, Danggui 6g) are extracted from the prescription entity nodes. The recommended dosages of each medicinal material are summarized according to the type of medicinal material to generate the initial dosage of the prescription containing the name and standard dosage of the medicinal material. It should also be noted that the traditional Chinese medicine knowledge graph is a structured knowledge base constructed based on authoritative data sources such as classical works of traditional Chinese medicine, historical medical case literature, pharmacopoeia standards, and clinical diagnosis and treatment guidelines.

[0067] According to the fuzzy inference rule, the initial dosage of the prescription and the virtual-real state quantization result are adjusted by rule mapping to generate an individualized dosage.

[0068] It should be noted that the standard doses of each herb in the initial prescription are input into the fuzzy inference engine and matched with the eight dimensions of the virtual-real state quantification results. The Mamdani-type fuzzy inference method is used to fuzzify the virtual-real state quantification results through triangular membership functions, and then match them with the rule antecedents for similarity matching to activate the corresponding rule consequents. The activated dose adjustment suggestions are defuzzified to obtain the precise adjustment coefficients of each herb. The adjustment coefficients are multiplied by the initial prescription dose to generate a personalized dose that takes into account the individual virtual-real state of the patient.

[0069] For personalized dosages, historical efficacy records are queried in the efficacy database, and drug-syndrome specificity scores are calculated. High-specificity combinations are then selected by using specificity score thresholds.

[0070] It should be noted that the revised syndrome classification and personalized dosage combination are used as search criteria to match historical records in the efficacy database. The efficacy database stores efficacy scores related to the same syndrome classification and drug combination, including quantitative indicators such as the degree of symptom relief and recurrence rate after treatment. By extracting efficacy scores from the matching records, the average efficacy score of the drug combination under the target syndrome classification is calculated. Combined with the dosage ratio of each drug in the drug combination, the drug-syndrome specificity score is calculated, expressed as: ; In the formula, It is the drug-symptom specificity score. It is the first The average efficacy score of the drug under the target syndrome classification (calculated by matching historical efficacy scores recorded in the efficacy database). It is the first The dosage ratio of a drug in a personalized dosage combination (i.e., the drug) (the proportion of the total drug combination dose) It is a normalization factor. It is the first Weighting factors for each drug It is the total number of drugs. It is an index variable for drugs; A specificity score threshold is set, and drug combinations with drug-syndrome specificity scores higher than the specificity score threshold are selected to form a high-specificity combination candidate set. It should also be noted that the specificity score threshold is determined by analyzing the drug-syndrome specificity score distribution of the drug combination to distinguish between drug combinations with high specificity and those with low specificity. This is done by calculating the specificity scores corresponding to the historical efficacy records of all relevant drug combinations, and then selecting an appropriate specificity score threshold based on the score distribution, such as 0.6 in the example.

[0071] The high specificity combination is ranked in descending order according to the drug-syndrome specificity score, and the optimal prescription combination is screened as a candidate scheme in combination with the patient file constraint, and an individualized diagnosis and treatment scheme is output.

[0072] It should be noted that the high specificity combination is ranked in descending order according to the drug-syndrome specificity score from high to low to form a candidate drug combination list; In combination with the patient file constraint, including the patient's age, gender, body type, underlying disease, allergy history and current disease development stage and other information (all the above are agreed by the user and used for legal purposes), each combination in the candidate drug combination list is adaptively evaluated; for example, for patients with specific underlying diseases, drug combinations with potential adverse reaction risks are excluded; for patients with weak constitution, the proportion of harsh drugs is appropriately reduced or replaced with mild drugs; under the premise of meeting the individualized restriction conditions of the patient, the drug-syndrome specificity score of the highest drug combination is selected as the optimal prescription combination as a candidate scheme, and an individualized diagnosis and treatment scheme containing specific drug composition, individualized dosage and decoction method is output.

[0073] The embodiment also provides a traditional Chinese medicine diagnosis and treatment auxiliary decision-making system, comprising: an acquisition module, configured to acquire a tongue image video sequence, and acquire pulse waveform time series data and spleen meridian acupoint impedance spectrum; An extraction module is configured to perform image processing on the tongue image video sequence, extract HSV color space features of each frame of tongue surface, and calculate tongue image color change rate between adjacent frames; An analysis module is configured to perform wavelet transform analysis on the pulse waveform time series data, extract a pulse dynamic change feature sequence, input the tongue image color change rate and the pulse feature change rate into a space-time convolution network, generate a tongue-pulse dynamic correlation index, and output a preliminary syndrome classification; A correction module is configured to perform Fourier transform analysis on the spleen meridian acupoint impedance spectrum, determine virtual and real states in combination with the tongue-pulse dynamic correlation index, correct the preliminary syndrome classification, and output a virtual and real state quantization result and a corrected syndrome classification; An output module is configured to match the corrected syndrome classification with a traditional Chinese medicine knowledge graph, adjust the drug dosage according to the virtual and real state quantization result, verify the drug-syndrome specificity through a curative effect database, screen an optimal treatment scheme, and output an individualized diagnosis and treatment scheme.

[0074] The embodiment also provides a computer device suitable for the traditional Chinese medicine diagnosis and treatment auxiliary decision-making method, comprising a memory and a processor; the memory is configured to store computer executable instructions, and the processor is configured to execute the computer executable instructions to implement the traditional Chinese medicine diagnosis and treatment auxiliary decision-making method proposed in the above embodiment.

[0075] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected by a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is configured to perform wired or wireless communication with an external terminal. The wireless communication can be achieved by WIFI, an operator network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0076] The embodiment also provides a storage medium having a computer program stored thereon, the program being executed by a processor to implement the method for assisting decision-making in traditional Chinese diagnosis and treatment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk or an optical disk.

[0077] In summary, the present application achieves the digital expression of the traditional Chinese "tongue and pulse combined reference" by dynamically collecting tongue image, pulse image and impedance data of the spleen meridian acupoints, establishing a dynamic correlation index of tongue and pulse by using a space-time convolution network and a cross-modal attention mechanism, and quantifying the virtual and real syndrome differentiation results by constructing a dynamic space-time graph based on the frequency domain characteristics of the impedance spectrum of the spleen meridian, thereby forming a "macro-dynamic-micro-energy" dual verification system, which not only improves the accuracy of syndrome classification, but also provides an objective basis for individualized dosage adjustment.

[0078] It should be noted that the above examples are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced, without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A method for assisting decision-making in traditional Chinese medicine diagnosis and treatment, characterized in that: include, Acquire tongue image video sequences, and collect pulse waveform time-series data and spleen meridian acupoint impedance spectra; Image processing was performed on the tongue image video sequence to extract the HSV color space features of the tongue surface in each frame and to calculate the tongue image color change rate between adjacent frames. Wavelet transform analysis was performed on the pulse waveform time series data to extract the pulse dynamic change feature sequence. The tongue color change rate and the pulse feature change rate were input into the spatiotemporal convolutional network to generate tongue and pulse dynamic correlation index and output preliminary syndrome classification. Fourier transform analysis was performed on the impedance spectrum of spleen meridian acupoints, and the deficiency and excess states were determined by combining the dynamic correlation index of tongue and pulse. The preliminary syndrome classification was corrected, and the quantitative results of deficiency and excess states and the corrected syndrome classification were output. The revised syndrome classification is matched with the TCM knowledge graph, and the dosage is adjusted according to the quantitative results of deficiency and excess states. The drug-syndrome specificity is verified by reverse verification through the efficacy database, and the optimal treatment plan is selected and a personalized treatment plan is output.

2. The TCM diagnostic and treatment auxiliary decision-making method as described in claim 1, characterized in that: The tongue image video sequence includes the dynamic morphology of the tongue, changes in the color of the tongue, and the thickness and moisture of the tongue coating. The pulse waveform time sequence data includes pulse position, pulse rate, and pulse strength. The impedance spectrum of the spleen meridian acupoints includes the impedance amplitude and phase angle of the spleen meridian acupoints at different frequencies.

3. The TCM diagnostic and treatment auxiliary decision-making method as described in claim 1, characterized in that: The specific steps for calculating the tongue image color change rate between adjacent frames are as follows. Bilateral filtering is applied to the tongue image video sequence to denoise, and a tongue surface image with complete edge information is obtained. The image is then input into a pre-trained depth segmentation model for region extraction, and a binarized mask of the tongue surface is output. The tongue image is cropped using a tongue binarization mask to extract the clean tongue area. Color calibration is performed using a standard color chart to obtain a white-balanced tongue image. The image is then converted to the HSV color space to extract the average HSV features of each frame. The tongue color change rate is calculated based on the average HSV feature difference between two adjacent frames.

4. The TCM diagnostic and treatment auxiliary decision-making method as described in claim 2, characterized in that: The specific steps for performing wavelet transform analysis on the pulse waveform time-series data to extract the dynamic change feature sequence of the pulse are as follows. Baseline drift correction and bandpass filtering are performed on the pulse waveform time series data to obtain pulse preprocessed waveform time series data with high signal-to-noise ratio. Resampling and amplitude normalization are then performed to generate uniform amplitude time series data. Discrete wavelet decomposition is used to separate the components that reflect the detailed features of the pulse. Envelope extraction is performed on the components reflecting the detailed features of the pulse to obtain the envelope curve reflecting the changes in pulse force. The peak detection threshold is set in combination with the physiological rhythm to detect the peak. The peak position and amplitude information in each pulse fluctuation cycle are extracted to form the dynamic temporal features of pulse position and the peak amplitude sequence, respectively. The pulse dynamic change feature sequence is obtained through the normalization process of maximal and minimum normalization.

5. The TCM diagnostic and treatment auxiliary decision-making method as described in claim 1, characterized in that: The specific steps for generating dynamic correlation indicators of tongue and pulse and outputting preliminary syndrome classification are as follows: The characteristic value difference between adjacent moments is calculated according to the sampling timestamp sequence of the pulse dynamic change feature sequence. The pulse feature change rate is calculated by combining the time interval of the characteristic value difference between adjacent moments. The pulse feature change rate is then time-aligned and standardized with the tongue color change rate to generate a synchronous input sequence. The synchronous input sequence is input into the parallel branch of the dual-stream spatiotemporal convolutional network. Temporal convolution is used to extract local temporal features, spatial convolution is used to encode key structural information, and residual connections and normalization are superimposed in multiple convolutional layers to extract preliminary spatiotemporal features of tongue and pulse. The preliminary spatiotemporal features of tongue and pulse are converted into query, key and value vectors respectively. The attention relationship between tongue and pulse is established to extract the association features between tongue and pulse. The attention relationship between pulse and tongue is established to extract the response features between pulse and tongue. The two are fused across modalities to obtain the tongue and pulse interaction features. The tongue and pulse interaction features are iteratively fused through multi-layer cross-attention and residual connection to generate refined spatiotemporal correlation features. These features are then globally aggregated to output dynamic correlation indicators of the tongue and pulse, and a preliminary syndrome classification is output through a classifier.

6. The TCM diagnostic and treatment auxiliary decision-making method as described in claim 1, characterized in that: The specific steps for outputting the virtual and real state quantification results and the corrected syndrome classification are as follows. Fast Fourier transform analysis was performed on the impedance spectrum of spleen meridian acupoints to extract the impedance phase angle sequence of each frequency point, generate a time-series phase vector, and splice it with the tongue pulse dynamic correlation index to form multimodal node features. By calculating the similarity between features of multimodal nodes using cosine similarity, a dynamic spatiotemporal graph is constructed, and potential similar syndrome groups are identified by the high similarity between adjacent nodes in the dynamic spatiotemporal graph. The graph attention layer is used to update the dynamic spatiotemporal graph through message passing. The cross interference between similar symptoms is corrected by adaptive weights. The graph attention update is repeated and combined with skip connections to fuse graph features at different levels to generate a refined node representation. The refined node representations are aggregated at the graph level to generate virtual and real state quantization vectors, and the corrected syndrome classification is output through a classifier.

7. The TCM diagnostic and treatment auxiliary decision-making method as described in claim 1, characterized in that: The specific steps for generating personalized treatment plans are as follows. The revised syndrome classification is used as the query keyword. The corresponding syndrome entity is matched in the TCM knowledge graph and related prescriptions and recommended dosages are obtained along the treatment relationship to generate the initial dosage of the prescription. Based on the fuzzy inference rules, the initial dose of the prescription is mapped and adjusted to the quantification results of the virtual and real states to generate a personalized dose. For personalized dosage, historical efficacy records are queried in the efficacy database, and drug-syndrome specificity scores are calculated. High-specificity combinations are selected by using specificity score thresholds, and the optimal prescription combination is selected as a candidate scheme by sorting the scores in descending order and combining patient file constraints. Personalized treatment plans are then output.

8. A TCM diagnosis and treatment auxiliary decision-making system, based on the TCM diagnosis and treatment auxiliary decision-making method according to any one of claims 1 to 7, characterized in that: include, The acquisition module is used to acquire tongue image video sequences, pulse waveform time series data, and spleen meridian acupoint impedance spectrum. The extraction module is used to perform image processing on the tongue image video sequence, extract the HSV color space features of the tongue surface in each frame, and calculate the tongue image color change rate between adjacent frames. The analysis module is used to perform wavelet transform analysis on pulse waveform time series data, extract the pulse dynamic change feature sequence, and input the tongue color change rate and pulse feature change rate into the spatiotemporal convolutional network to generate tongue and pulse dynamic correlation index and output preliminary syndrome classification. The correction module is used to perform Fourier transform analysis on the impedance spectrum of spleen meridian acupoints, combine tongue and pulse dynamic correlation indicators to determine the deficiency and excess state, correct the preliminary syndrome classification, and output the quantitative results of deficiency and excess state and the corrected syndrome classification. The output module is used to match the corrected syndrome classification with the TCM knowledge graph, adjust the dosage according to the quantitative results of deficiency and excess states, verify the drug-syndrome specificity through the efficacy database, and output personalized treatment plans after screening the optimal treatment plan.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the TCM diagnosis and treatment auxiliary decision-making method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the TCM diagnosis and treatment auxiliary decision-making method according to any one of claims 1 to 7.

Images