AI-based traditional Chinese medicine nursing recommendation system

Abstract

The invention relates to the technical field of intelligent traditional Chinese medicine nursing, in particular to an AI-based traditional Chinese medicine nursing recommendation system. The AI-based traditional Chinese medicine nursing recommendation system comprises a tongue image feature extraction module, a tongue partition anomaly recognition module, a meridian mapping positioning module, a drug property tropism matching module and a nursing intervention recommendation module. According to the method, depth feature extraction and partition quantization are carried out on the tongue body image, visual information such as contours, textures and colors is converted into accurate numerical parameters, and an objective judgment reference is set according to statistical distribution of large-scale health samples, so that uncertainty and difference caused by traditional evaluation depending on subjective experience are overcome, and the accuracy of evaluation is improved. Standardized recognition of tongue picture abnormity is achieved, a calculation mode of spatial position contribution degree is introduced, the method can be automatically focused on the tongue body partition with the most remarkable pathological significance, then accurate mapping from the tongue surface abnormal partition to the body meridian path is constructed, local observation and an overall meridian system are closely associated, and the accuracy of the tongue picture abnormity recognition is improved. And the problem of generalization fuzziness of previous recommendation is solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent traditional Chinese medicine nursing technology, and in particular to an AI-based traditional Chinese medicine nursing recommendation system. Background Technology

[0002] The field of intelligent TCM nursing technology involves the integration of TCM nursing theory with modern information technology. Core aspects include the formulation and adjustment of TCM nursing plans, the selection of compatibility between TCM prescriptions and nursing measures, the application of syndrome differentiation in the nursing process, and the collection and utilization of individual health data.

[0003] Among them, the TCM nursing recommendation system refers to a tool that selects appropriate TCM nursing plans based on the patient's basic health information and the key points of TCM syndrome differentiation. It usually addresses technical matters such as the patient's symptoms, tongue and pulse characteristics, and constitution type.

[0004] Current technologies for processing patient symptoms and tongue and pulse characteristics often rely on qualitative descriptions based on manual observation and experience. This approach is easily influenced by the operator's subjectivity, leading to inconsistent and unrepeatable recommendations. Furthermore, the analysis of tongue appearance is usually a holistic, macroscopic assessment, failing to precisely quantify and locate subtle abnormalities in specific areas of the tongue. For example, the same yellow tongue coating can have drastically different pathological meanings depending on whether it occurs at the tip or root of the tongue. Current technologies struggle to distinguish these spatial differences, thus failing to establish a precise correspondence between abnormal areas and specific meridians, resulting in weakly targeted nursing plans. Moreover, the process from identifying diagnostic points to selecting nursing plans is often a simple, pre-defined macroscopic match, lacking a hierarchical and refined matching logic that considers pathological characteristics, medicinal properties and meridian tropism, and the intervention area of ​​the nursing method. This results in insufficient correlation between recommended nursing content and the specific affected body parts and treatment needs. For instance, recommending nursing content with heat-clearing and dampness-removing effects based solely on damp-heat symptoms fails to specify that the nursing method should primarily target the body areas related to the diseased meridians, making truly personalized and precise intervention difficult. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an AI-based traditional Chinese medicine nursing recommendation system.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an AI-based traditional Chinese medicine nursing recommendation system, the system comprising:

[0007] The tongue image feature extraction module acquires tongue image data and extracts the image features of the outline edges, texture structure and color distribution of the tongue tip, tongue sides, tongue middle and tongue root regions to obtain tongue region feature maps;

[0008] The tongue body partition anomaly recognition module identifies the anomaly types of each tongue body partition in the tongue body partition feature map, such as broken contour edges, disordered texture structure, and color distribution shift, and generates tongue image anomaly partition annotation data.

[0009] The meridian mapping and positioning module extracts the names of abnormal parts in the tongue body partitions from the tongue abnormal partition annotation data, matches the abnormal part names with the corresponding body meridian paths according to the TCM meridian attribution relationship, and outputs meridian influence path data.

[0010] The medicinal property meridian matching module, referring to the meridian influence path data, filters the information of Chinese medicinal materials that belong to the corresponding meridian and whose range of action covers the influence path, and obtains medicinal property association data;

[0011] The nursing intervention recommendation module compares the drug property association data with a preset nursing intervention knowledge base, filters nursing content that matches the drug property characteristics, and outputs traditional Chinese medicine nursing recommendation results.

[0012] The present invention is improved in that the tongue body partition feature map specifically includes tongue body partition identification information, image feature vector expression, and feature space coordinate set; the tongue image abnormal partition annotation data includes abnormal partition location, abnormal type label, and feature classification mark; the meridian influence path data specifically includes meridian path name, tongue body part correspondence, and affiliated organ number; the drug property association data includes drug name, meridian number, and action level information; and the traditional Chinese medicine nursing recommendation result specifically refers to nursing action name and intervention site location.

[0013] The present invention is improved in that the tongue image feature extraction module includes:

[0014] The image feature extraction submodule acquires tongue image data and inputs it into a residual convolutional neural network with a cross-layer connection structure. It extracts features of contour edges, texture structure and color distribution of the image in parallel, and performs numerical dimensional calculation on the extracted multi-channel features to obtain a multi-dimensional feature vector.

[0015] The partition feature calibration submodule divides the multidimensional feature vector into corresponding partitions based on the preset coordinate masks of the tongue tip, tongue side, tongue middle and tongue root partitions, and calculates the curvature value of the contour edge, the complexity of the texture structure and the mean value of the color distribution in each partition to generate a tongue domain feature parameter set.

[0016] The feature map fusion construction submodule integrates the tongue domain feature parameter set through multi-channel results and maps and arranges them according to the spatial position relationship of the partitions to establish a tongue body partition feature map.

[0017] The present invention is improved in that the tongue body partition abnormality recognition module includes:

[0018] The spatial feature calculation submodule calls the tongue body partition feature map, calculates the spatial position features of each partition of the tongue tip, tongue side, tongue middle and tongue root through the position attention mechanism function, and quantifies the contribution of the contour edge, texture structure and color distribution feature points in the partition as the partition spatial attention weight.

[0019] The partition anomaly judgment submodule, based on the partition spatial attention weight, compares the features of contour edge breakage, texture structure disorder and color distribution shift of each partition in the tongue partition feature map with the preset anomaly judgment benchmark value one by one, performs partition anomaly type judgment, and obtains partition anomaly category set.

[0020] The anomaly partition annotation submodule marks the identified partition anomaly category set and the corresponding contour edge breakage, texture structure disorder and color distribution shift features, and associates and classifies them with the category information of the tongue tip, tongue side, tongue middle and tongue root partitions to generate tongue image anomaly partition annotation data.

[0021] The present invention is improved in that the meridian mapping and positioning module includes:

[0022] The abnormal part extraction submodule identifies and extracts the names of abnormal parts marked in the tongue image abnormal partition annotation data, and performs a deduplication operation on the repeated names to obtain a list of abnormal part names;

[0023] The meridian relationship matching submodule, based on the list of abnormal part names, calls the TCM meridian attribution relationship, and performs meridian mapping operation on each part name in the list according to the rule that the tip of the tongue maps to the heart meridian, the sides of the tongue maps to the liver and gallbladder meridians, the middle of the tongue maps to the spleen and stomach meridians, and the root of the tongue maps to the kidney and bladder meridians, to establish a set of meridian mapping relationships;

[0024] The influence path construction submodule associates the name of each abnormal part in the meridian mapping relationship set with the matching body meridian path, constructs the correspondence between tongue appearance and meridians, and obtains meridian influence path data.

[0025] The present invention is improved in that the TCM meridian attribution relationship is a preset structured lookup table. The structured lookup table includes two data fields: tongue body partition and meridian path, and pre-stores key-value pair records of tongue tip and heart meridian, tongue side and liver and gallbladder meridian, tongue middle and spleen and stomach meridian, tongue root and kidney and bladder meridian.

[0026] The present invention is improved in that the drug property meridian matching module includes:

[0027] The meridian path analysis submodule performs structured analysis on the path information in the meridian influence path data, extracts the meridian name and corresponding organ type, and summarizes the identified meridian name and organ type to obtain a meridian organ list.

[0028] The medicinal material information filtering submodule compares the meridian names in the meridian organ list with the meridian tropism attributes of all medicinal materials in the medicinal property meridian tropism information database, filters the Chinese medicinal material entries corresponding to the meridians, and establishes a candidate medicinal material information set.

[0029] The drug property feature matching submodule, based on the candidate drug information set and the meridian influence path data, determines whether the meridian tropism and action level of each drug are consistent with the requirements in the influence path, extracts drug information whose action range can cover the influence path, determines the drug property features related to the current tongue appearance, and generates drug property association data.

[0030] An improvement of this invention is that the meridian tropism attribute is a text set consisting of multiple standardized meridian names, and the text set is compared with the meridian names in the meridian organ list to determine the meridian tropism matching.

[0031] The present invention is improved in that the nursing intervention recommendation module includes:

[0032] The drug property characteristic analysis submodule performs structured analysis on the drug property characteristics and range of action in the drug property association data, identifies the required conditioning direction and body intervention site, and generates a conditioning requirement list;

[0033] The nursing content filtering submodule compares the applicable treatment direction and intervention area of ​​each nursing content in the preset nursing intervention knowledge base with the needs in the treatment needs list, filters out nursing content that matches the drug properties, and establishes a set of matching nursing items.

[0034] The recommendation result generation submodule, based on the matching nursing item set, sorts the nursing action content in the items according to the correlation between the range of drug conditioning and the intervention area of ​​nursing action, and outputs the traditional Chinese medicine nursing recommendation result.

[0035] The present invention is improved in that the applicable treatment direction is a text set consisting of multiple treatment direction keywords, and the text set is compared with the treatment demand list as the input item for matching judgment; the intervention area is a text set consisting of multiple body part names, and the text set is matched with the body intervention parts in the treatment demand list during the comparison operation.

[0036] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0037] In this invention, deep feature extraction and zoning quantification of tongue images are performed to transform visual information such as contour, texture, and color into precise numerical parameters. An objective judgment benchmark is established based on the statistical distribution of a large-scale healthy sample, overcoming the uncertainty and variability inherent in traditional subjective experience-based assessments. This achieves standardized identification of tongue abnormalities. The introduction of a spatial location contribution calculation method automatically focuses on the tongue regions with the most significant pathological significance, making diagnostic analysis more targeted. Furthermore, by constructing a precise mapping from abnormal tongue regions to the body's meridian pathways, local observation is closely linked to the overall meridian system, solving the problem of generalized ambiguity in previous recommendations. Based on this, a multi-level screening and matching process is created, from meridian influence pathways to drug tropism and drug characteristics, ensuring that the scope of action of selected Chinese medicinal materials is highly consistent with the pathological needs reflected in the tongue image. Finally, the analyzed treatment direction and the body intervention site are associated, matched, and ranked with specific non-drug interventions, forming personalized nursing recommendations that link medicinal materials, efficacy, and operation methods, significantly improving the precision and clinical practical value of the recommendation results. Attached Figure Description

[0038] Figure 1 This is a system module diagram of the present invention;

[0039] Figure 2 This is a system framework diagram of the present invention;

[0040] Figure 3 This is a schematic diagram of the tongue image feature extraction module of the present invention;

[0041] Figure 4 This is a schematic diagram of the tongue body partition anomaly recognition module of the present invention;

[0042] Figure 5 This is a schematic diagram of the meridian mapping and positioning module of the present invention;

[0043] Figure 6 This is a schematic diagram of the drug property meridian matching module of the present invention;

[0044] Figure 7 This is a schematic diagram of the nursing intervention recommendation module of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0046] Please see Figure 1 This invention provides a technical solution: an AI-based traditional Chinese medicine nursing recommendation system, the system comprising:

[0047] The tongue image feature extraction module acquires tongue image data and extracts the image features of the outline edges, texture structure and color distribution of the tongue tip, tongue sides, tongue middle and tongue root regions to obtain tongue region feature maps;

[0048] The tongue region anomaly recognition module identifies the anomaly types of each tongue region in the tongue region feature map, such as broken contour edges, disordered texture structure, and color distribution shift, and generates tongue image anomaly region annotation data.

[0049] The meridian mapping and positioning module extracts the names of abnormal parts in the tongue body partitions from the abnormal part labeling data, matches the abnormal part names with the corresponding body meridian paths according to the meridian attribution relationship of traditional Chinese medicine, and outputs meridian influence path data.

[0050] The drug property and meridian matching module refers to the meridian influence path data, filters the information of Chinese medicinal materials that belong to the corresponding meridian and whose range of action covers the influence path, and obtains drug property association data.

[0051] The nursing intervention recommendation module compares the drug property-related data with the preset nursing intervention knowledge base, filters out nursing content that matches the drug property characteristics, and outputs traditional Chinese medicine nursing recommendation results.

[0052] The tongue body zoning feature map specifically includes tongue body zoning identification information, image feature vector expression, and feature space coordinate set. The tongue image abnormal zoning annotation data includes abnormal zoning location, abnormality type label, and feature classification mark. The meridian influence path data specifically includes meridian path name, tongue body part correspondence, and affiliated organ number. The drug property association data includes drug name, meridian number, and action level information. The traditional Chinese medicine nursing recommendation results specifically refer to the nursing action name and intervention site location.

[0053] Please see Figure 2 and Figure 3 The tongue image feature extraction module includes:

[0054] The image feature extraction submodule acquires tongue image data and inputs it into a residual convolutional neural network with a cross-layer connection structure. It extracts features of contour edges, texture structure and color distribution of the image in parallel, and performs numerical dimensional calculation on the extracted multi-channel features to obtain a multi-dimensional feature vector.

[0055] The tongue image acquisition data consisted of a high-resolution digital color image of the front of the tongue acquired using standardized acquisition equipment under uniform lighting and distance conditions. The image was input into a pre-trained deep residual convolutional neural network with fifty convolutional layers and a cross-layer shortcut connection structure. The network structure included sequentially connected convolutional layers, batch normalization layers, ReLU activation function layers, and pooling layers. The cross-layer shortcut connections allowed the input signal to skip multiple layers and directly pass to subsequent layers, thus maintaining the intensity of the gradient signal in the deep network and performing parallel feature extraction. First, the color image data was converted from RGB space to a single-channel grayscale image space. A 5x5 Sobel operator was applied to each pixel in the grayscale image. The gradient magnitude was obtained by calculating approximate values ​​of the horizontal and vertical first derivatives of the pixel intensity in the neighborhood of the pixel. The gradient magnitudes of all pixels together constituted a contour edge. The first step involves convolving the grayscale image with a pre-defined Gabor filter bank containing four directions (0°, 45°, 90°, and 135°) and five frequency scales to generate a set of twenty-channel texture structure feature maps. The response intensity of each channel characterizes the texture details of the tongue surface in a specific direction and scale. Next, the original RGB color image is converted to the HSV color space model, and the H channel (hue) and S channel (saturation) are separated to serve as two independent color distribution feature maps. Finally, the generated single-channel contour edge feature map, twenty-channel texture structure feature map, and H and S dual-channel color distribution feature map are expanded at the pixel level, converting their respective two-dimensional matrix forms into one-dimensional vectors. All one-dimensional vectors are then concatenated in a predetermined order to obtain a multi-dimensional feature vector containing contour, texture, and color information.

[0056] The partition feature calibration submodule divides the multidimensional feature vector into corresponding partitions based on the preset coordinate masks of the tongue tip, tongue side, tongue middle and tongue root partitions, and calculates the curvature value of the contour edge, the complexity of the texture structure and the mean of the color distribution in each partition to generate a tongue domain feature parameter set.

[0057] Based on a pre-defined set of coordinate masks for the tongue tip, sides, middle, and root regions, the multidimensional feature vectors generated in the preceding steps are regionalized and calibrated. These coordinate masks are binary image templates representing the standard spatial positions of each region of the tongue in an average adult, extracted from a database containing tens of thousands of standardized tongue images through image registration and statistical overlay analysis. Pixels with a value of 1 represent the interior of a region, and pixels with a value of 0 represent the exterior. In practice, the multidimensional feature vectors are first remapped back to their original multiple two-dimensional feature maps. Then, each feature map is subjected to a pixel-level bitwise AND operation with the coordinate masks for the tongue tip, sides, middle, and root. This operation clears the feature map data at pixel positions with a mask value of 0, thus accurately isolating the effective feature data within each region. Subsequently, parameter calculations are performed within the data of each region. Taking the tongue root region as an example, the parameters are calculated at its contour edges. On the feature map, along the extracted edge lines, the second derivative of the local curve is calculated point by point to obtain the curvature. The arithmetic mean of the curvature values ​​of all edge points within the partition is then calculated to obtain a curvature value that quantifies the smoothness of the partition edge. Next, on the twenty texture structure feature map channels of the tongue root partition, the standard deviation of all non-zero pixel values ​​in each channel is calculated. These twenty standard deviations are then weighted and averaged, with the weights set according to the frequency of the corresponding Gabor filter, with higher frequencies receiving higher weights. This yields a complexity value that characterizes the texture complexity of the partition. Finally, on the H and S channel color distribution maps of the tongue root partition, the arithmetic mean of all non-zero pixel values ​​is calculated to obtain the hue mean and saturation mean that characterize the dominant hue and color vibrancy of the partition. These parameters calculated for the tongue root partition, along with the parameters obtained by performing the same calculation process on the other three partitions, are collectively compiled into a tongue domain feature parameter set.

[0058] The feature map fusion construction submodule integrates the tongue domain feature parameter set through multi-channel results and maps and arranges them according to the spatial position relationship of the partitions to establish tongue body partition feature maps;

[0059] The tongue region feature parameter set, which includes the curvature values, texture complexity, hue mean, and saturation mean of each of the four regions of the tongue (tip, side, middle, and root), is integrated through multi-channel results. This process creates a 4x4 matrix structure as the data container, with the number of rows equal to the number of tongue regions (four) and the number of columns equal to the number of feature parameter types (four). Then, based on the standard spatial relationship of tongue anatomy—tip at the top, root at the bottom, middle in the center, and side on the sides—the row indices of the matrix are semantically mapped. The first row of the matrix is ​​defined as the corresponding tongue tip region, and the second row... The third row corresponds to the tongue side, the fourth row to the tongue center, and the fifth row to the tongue root. Similarly, the column indices of the matrix are defined as corresponding to the four feature parameters: curvature value, texture complexity, hue mean, and saturation mean. Finally, the values ​​of each parameter in the tongue region feature parameter set are precisely filled into the unique position in the matrix determined by its partition index and parameter index. For example, the curvature value of the tongue root partition is placed in the first column of the fourth row. In this way, a series of discrete numerical parameters are transformed into a structured data matrix that can simultaneously reflect spatial relationships and feature types, thus establishing a tongue body partition feature map.

[0060] Please see Figure 2 and Figure 4 The tongue body partition anomaly recognition module includes:

[0061] The spatial feature calculation submodule calls the tongue body partition feature map and calculates the spatial position features of each partition of the tongue tip, tongue side, tongue middle and tongue root through the position attention mechanism function. It also quantifies the contribution of the contour edge, texture structure and color distribution feature points in the partition as the partition spatial attention weight.

[0062] The tongue body feature map is invoked. The feature map is a matrix with tongue body regions as rows and feature parameters as columns. A positional attention mechanism function is used to calculate the spatial positional features of each region of the tongue tip, sides, middle, and root. This quantifies the contribution of each region's features to the overall judgment in the current diagnostic context. The calculation process is defined by the following formula: In this formula, the meanings of each letter and symbol are explained in the context of the current situation as follows: Represents the calculated first The spatial attention weight for each partition is a scalar value between 0 and 1; a higher value indicates a higher level of attention. The more relevant the features presented by each partition are to the current pathological pattern being queried, the greater their importance in diagnosis. It is an index used to identify the partition currently being computed, and its value ranges from 1 to 1. ,in The total number of partitions, in this embodiment , Corresponding to the tip of the tongue, Corresponding to the tongue side, Corresponding to the middle of the tongue, Corresponds to the root of the tongue. This refers to the first section of the tongue body regional feature diagram. The eigenvectors contained in the row, the vectors are formed by the first row. The four feature parameters of each partition (curvature value, texture complexity, hue mean, and saturation mean) are arranged in order. It is a preset query weight vector, whose dimensions are the same as those of the feature vector. Similarly, this vector represents an idealized feature pattern for a specific TCM syndrome (such as "damp-heat syndrome"). Its values ​​are not hypothetical but derived from analyzing a database of thousands of cases diagnosed as "damp-heat syndrome" by authoritative TCM physicians. Specifically, a multivariate logistic regression model is established using four tongue appearance features as independent variables and whether it is "damp-heat syndrome" (yes = 1, no = 0) as the dependent variable. The regression coefficients of each independent variable after model training reflect the contribution of each feature to the diagnosis of "damp-heat syndrome." After normalizing these coefficients, the query weight vector is formed. . This represents a scoring function, which specifically executes the feature vector. With query weight vector The dot product operation between them can be geometrically understood as... exist The projected length in the direction, therefore the higher the score, the better the number of projected lengths. The more similar the actual characteristics of each partition are to the queried "humid and hot" pattern, the better. It is the natural exponential function Its function is to convert all scores into positive numbers and to amplify the differences between scores, so that the high-scoring partitions have a significantly larger weight in the subsequent normalization. It is the summation symbol, indicating that all partitions (from...) are summed. arrive The scores obtained by the scoring function are summed after being calculated using an exponential function, and this sum is used as the normalized denominator. The entire formula functionally implements the Softmax function, which transforms a set of arbitrary real score values ​​into a probability distribution with a sum of 1, ensuring that the sum of the attention weights for all partitions equals 1.

[0063] Suppose that the feature vectors of the four partitions calculated through the aforementioned steps are as follows: (Tip of tongue) = [0.12, 25.0, 20.0, 30.0], (Tongue edge) = [0.15, 30.0, 22.0, 35.0], (Middle of tongue) = [0.10, 28.0, 21.0, 33.0], (Tongue root) = [0.18, 43.15, 25.5, 40.2], and simultaneously, assume the query weight vector for "damp-heat syndrome". Based on the aforementioned experiments and regression analysis, it was determined that: =[0.05, 0.5, 0.15, 0.3], the weights indicate that texture complexity (0.5) and saturation (0.3) are key indicators for judging damp-heat syndrome.

[0064] The calculation process is as follows: First, calculate the score for each partition. : , , , .

[0065] The second step is to perform an exponential calculation on each score value: , , , .

[0066] The third step is to calculate the sum of all exponent values ​​(to avoid floating-point overflow that may occur when directly calculating large values, stabilization techniques such as subtracting the maximum value are used in actual calculations; this is just a conceptual explanation of the principle): .

[0067] The fourth step is to calculate the final attention weights for each partition. : , , , The final set of weight values ​​[0.00002, 0.0017, 0.0003, 0.99798] is the partition space attention weight.

[0068] The partition anomaly judgment submodule, based on the partition spatial attention weight, compares the features of contour edge breakage, texture structure disorder and color distribution shift of each partition in the tongue partition feature map with the preset anomaly judgment benchmark value one by one, performs partition anomaly type judgment, and obtains the partition anomaly category set.

[0069] The spatial attention weights of the tongue body regions and the feature values ​​in the feature maps of the tongue body regions will be compared item by item with preset abnormality judgment benchmarks to determine the type of abnormality in each region. These abnormality judgment benchmarks are set based on statistical analysis of a large-scale, rigorously screened database of tongue images from healthy individuals. Specifically, all sample values ​​for each tongue body region and each feature parameter (e.g., the texture complexity of the tongue root) in the database are statistically analyzed, their probability density distribution is calculated, and the 95th percentile of the distribution is selected as the upper limit of the normal range, i.e., the benchmark value for abnormality judgment. For example, if statistical analysis shows that 95% of healthy individuals have a tongue root texture complexity value below 35.0, then 35.0 is set as the abnormality judgment benchmark value for this feature. In addition, a spatial attention weight threshold of 0.1 is set, based on historical case data. By plotting Receiver Operating Characteristic (ROC) curves, we sought weight values ​​that could distinguish between core and non-core clinical lesion areas with the highest true positive rate and the lowest false positive rate. In the judgment process, we first traversed the spatial attention weights of each partition. If the weight of the tongue tip (0.00002) was less than 0.1, the judgment of the tongue tip was skipped. If the weight of the tongue root (0.99798) was greater than 0.1, a detailed abnormality judgment of the tongue root partition was initiated. Subsequently, the feature values ​​of the tongue root partition, such as the texture complexity value corresponding to texture structure disorder, were compared with its corresponding benchmark value of 35.0. If the calculated value was greater than the benchmark value, "texture structure disorder" was determined to exist. If the calculated value was not greater than the benchmark value, it was determined to be normal. This comparison operation was performed on all features with weights exceeding the threshold within the partition. All categories judged as abnormal were aggregated to obtain the partition abnormality category set.

[0070] The abnormal partition labeling submodule marks the identified partition abnormality category set and the corresponding contour edge breakage, texture structure disorder and color distribution shift features, and associates and classifies them with the category information of the tongue tip, tongue side, tongue middle and tongue root partitions to generate tongue image abnormal partition labeling data.

[0071] The system meticulously labels and associates the anomaly category set and the original features that led to the judgment. Specifically, for each feature judged as an anomaly, a structured data record is created. This record explicitly includes four key information fields: First, the "Partition Name" field records the specific partition name where the anomaly occurred, such as "tongue root"; second, the "Anomaly Category" field records the name of the identified anomaly category, such as "texture structure disorder"; third, the "Feature Item" field records the quantified feature item that triggered this judgment, such as "texture complexity"; and fourth, the "Quantified Value" field records the actual calculated value of the feature item exceeding the baseline value. Subsequently, these generated, fully-informed label records are finally associated and classified with the original partition category information—tongue tip, tongue side, tongue middle, and tongue root—and output in a hierarchical manner with a complete traceability chain, generating tongue image anomaly partition annotation data.

[0072] Please see Figure 2 and Figure 5 The meridian mapping and positioning module includes:

[0073] The abnormal part extraction submodule identifies and extracts the names of abnormal parts marked in the tongue image abnormal partition annotation data, and performs a deduplication operation on the duplicate names to obtain a list of abnormal part names;

[0074] The process involves identifying and extracting the names of abnormal parts from the tongue image abnormality partition annotation data. First, it iterates through the entire list of structured records, locating and extracting the value of the "partition name" field from each record element. All extracted partition names, such as "tongue root" or "tongue root," are then compiled into a temporary initial list. Since the same partition may be identified as having multiple different types of abnormalities during the anomaly detection phase, resulting in duplicate partition names in the initial list, a deduplication operation is performed on the initial list. This is achieved by constructing a set data structure that does not allow duplicate elements. List elements are added to the set one by one, and the set is finally converted back into a list, ensuring that each partition name in the list is unique, thus obtaining the list of abnormal part names.

[0075] The meridian relationship matching submodule, based on the list of abnormal part names, calls the TCM meridian attribution relationship. According to the rule that the tip of the tongue maps to the Heart Meridian, the sides of the tongue maps to the Liver and Gallbladder Meridians, the middle of the tongue maps to the Spleen and Stomach Meridians, and the root of the tongue maps to the Kidney and Bladder Meridians, it performs meridian mapping operation on each part name in the list to establish a set of meridian mapping relationships. The TCM meridian attribution relationship is a preset structured query table. The structured query table contains two data fields: tongue body partition and meridian path. It also pre-stores key-value pair records of the tip of the tongue and the Heart Meridian, the sides of the tongue and the Liver and Gallbladder Meridians, the middle of the tongue and the Spleen and Stomach Meridians, and the root of the tongue and the Kidney and Bladder Meridians.

[0076] Based on the list of abnormal body part names, a pre-defined structured lookup table for TCM meridian attribution is invoked. This lookup table is a knowledge base reviewed and solidified by a team of TCM experts, based on the theories of "the tongue as the sprout of the heart" and the attribution of organs on the tongue surface in core TCM classics such as the *Huangdi Neijing*. It pre-stores deterministic mapping rules between tongue body regions and meridian paths in the form of key-value pairs. For example, the key "tongue root" corresponds to the value "[Kidney Meridian, Bladder Meridian]". During execution, each body part name in the list of abnormal body part names (such as "tongue root") is used as the query key. A precise match is performed in the lookup table to retrieve the list of meridian path values ​​corresponding to the key (i.e., "[Kidney Meridian, Bladder Meridian]"). Then, the retrieved body part names are paired and associated with their corresponding meridian path lists to establish a set of meridian mapping relationships.

[0077] The influence path construction submodule associates the name of each abnormal part with the matching body meridian path in the meridian mapping relationship set, constructs the correspondence between tongue appearance and meridians, and obtains meridian influence path data;

[0078] By centralizing the meridian mapping relationships, each abnormal location name is explicitly associated with its corresponding meridian path, establishing a correspondence between abnormal tongue manifestations and the body's internal meridian system. Specifically, the mapping relationship set is traversed. If an abnormal location (such as "tongue root") corresponds to a list containing multiple meridian paths (such as "[Kidney Meridian, Bladder Meridian]"), then an influence path record is constructed separately for each meridian path in the list. This generates two independent path records: "{Abnormal Location: Tongue Root, Affected Meridian: Kidney Meridian}" and "{Abnormal Location: Tongue Root, Affected Meridian: Bladder Meridian}". These records together constitute a complete data list pointing from the specific tongue manifestation to all potentially affected meridian channels, thus obtaining the meridian influence path data.

[0079] Please see Figure 2 and Figure 6 The drug property and meridian tropism matching module includes:

[0080] The meridian path analysis submodule performs structured analysis on the path information in the meridian influence path data, extracts the meridian name and corresponding organ type, and summarizes the identified meridian name and organ type to obtain a meridian organ list.

[0081] Structured parsing is performed on each path information in the meridian influence path data. This process first traverses the list of path records. For each record in the list, the value of the "influence meridian" field is extracted. Then, based on the theory of viscera and the theory of the connection relationship of the twelve meridians in traditional Chinese medicine, a pre-set meridian-viscera association knowledge base is invoked, with the meridian name as the key and the main viscera it connects to as the value, to identify the core organ type corresponding to each meridian. For example, when parsing "Kidney Meridian", the knowledge base is queried to identify its corresponding organ type as "Kidney". When parsing "Bladder Meridian", its corresponding organ type is identified as "Bladder". All identified meridian names and their corresponding organ types are summarized and organized to obtain the meridian organ list.

[0082] The medicinal material information filtering submodule compares the meridian names in the meridian organ list with the meridian tropism attributes of all medicinal materials in the medicinal property meridian tropism information database, filters the Chinese medicinal material entries corresponding to the meridians, and establishes a candidate medicinal material information set. Among them, the meridian tropism attribute is a text set composed of multiple standardized meridian names. The text set is compared with the meridian names in the meridian organ list to determine the matching of meridian tropism.

[0083] The meridian names in the list of meridian organs are compared and screened with a digitized database of medicinal properties and meridian tropism information, sourced from the Pharmacopoeia of the People's Republic of China and authoritative monographs on traditional Chinese medicine. This database labels each entry of thousands of Chinese medicinal materials with its "meridian tropism" attribute, which is a text set consisting of multiple standardized meridian names. The specific screening process involves using the meridian names in the list of meridian organs (such as "Kidney Meridian" and "Bladder Meridian") as query conditions, traversing each Chinese medicinal material entry in the database, and checking whether the "meridian tropism" attribute text set of the medicinal material contains any of the meridian names in the query conditions. This check is performed by calculating whether the intersection of the two sets is empty. If the intersection is not empty, it means that the medicinal material belongs to at least one target meridian, and the medicinal material entry is considered a candidate and screened out. In this way, a candidate medicinal material information set containing all potentially applicable medicinal materials is established.

[0084] The drug property feature matching submodule, based on the candidate drug information set and meridian influence path data, determines whether the meridian tropism and action level of each drug are consistent with the requirements in the influence path, extracts drug information whose action range can cover the influence path, determines the drug property features related to the current tongue appearance, and generates drug property association data.

[0085] Based on the candidate medicinal material information set and meridian influence path data, a deep matching judgment is performed on the candidate medicinal materials. This process not only determines whether the direction of the medicinal material's meridian tropism covers the meridians in the influence path, but also determines whether its medicinal properties are consistent with the pathological state reflected by the tongue appearance. This matching process is divided into two steps. The first step is the scope of action matching. If the influence path requires coverage of the "Kidney Meridian" and the "Bladder Meridian", then a medicinal material that includes both of these meridians in its meridian tropism attribute (such as Phellodendron bark) will have a higher matching score than a medicinal material that only includes one of these meridians (such as Anemarrhena rhizome). The second step is the medicinal property matching. This step compares the medicinal material's properties and effects (such as "clearing heat and drying dampness" of Phellodendron bark) with the pathological requirements initially analyzed from the tongue appearance features (such as "damp-heat accumulation" represented by a thick, greasy yellow coating at the root of the tongue). Only when the main effects of the medicinal material are highly matched with the pathological requirements are they finally selected. The medicinal material information that can accurately match the influence path requirements in terms of scope of action and medicinal property features is extracted, the medicinal property features most relevant to the current tongue appearance are determined, and medicinal property association data is generated.

[0086] Please see Figure 2 and Figure 7 The nursing intervention recommendation module includes:

[0087] The drug property characteristics analysis submodule performs structured analysis on the drug property characteristics and scope of action in the drug property association data, identifies the required treatment direction and body intervention site, and generates a list of treatment needs;

[0088] The data on the correlation of medicinal properties is structured and analyzed to identify the underlying treatment directions and body intervention sites. This analysis process first processes the "medicinal property characteristics" field in the data. Through a pre-defined medicinal property terminology-treatment direction mapping table, professional pharmaceutical terms are converted into easily understandable treatment needs. For example, "clearing heat and drying dampness" is mapped to "the need to clear damp heat from the body." Second, the "scope of action" field in the data is analyzed, i.e., the affected meridians (such as "Kidney Meridian" and "Bladder Meridian"). Based on the meridian circulation theory knowledge base, the surface circulation paths of these meridians are mapped to specific anatomical regions of the human body to identify the body parts that need intervention. For example, the Foot Shaoyin Kidney Meridian and the Foot Taiyang Bladder Meridian both run in the lower part of the body, so their main intervention sites are identified as "lower jiao" and further specified as "waist and abdomen" and "foot". These two analyzed treatment directions and body intervention sites are integrated to generate a list of treatment needs.

[0089] The nursing content filtering submodule compares the applicable treatment direction and intervention area of ​​each nursing content in the preset nursing intervention knowledge base with the needs in the treatment needs list, filters the nursing content that matches the drug properties, and establishes a set of matching nursing items. Among them, the applicable treatment direction is a text set consisting of multiple treatment direction keywords. The text set is compared with the treatment needs list as the input for matching judgment. The intervention area is a text set consisting of multiple body part names. The text set is matched with the body intervention parts in the treatment needs list during the comparison operation.

[0090] The needs in the treatment needs list are compared and filtered against a pre-built nursing intervention knowledge base, jointly constructed by senior TCM doctors and rehabilitation therapists, which includes various non-drug intervention measures. The knowledge base labels each measure (such as acupressure, moxibustion, dietary therapy, and guided exercises) with a set of keywords for its "applicable treatment direction" and a set of names for its "effective body intervention area". During the filtering, a double-condition query is performed: First, the set of keywords for the "applicable treatment direction" of the nursing measure must have semantic overlap or subordinate relationship with the "treatment direction" of the need (such as "clearing damp heat"). Second, the set of "effective body intervention areas" of the nursing measure must intersect with the "body intervention area" of the need (such as "waist and abdomen" or "foot"). Only nursing content that meets both of these conditions is selected. For example, when the need is "clearing damp heat" and "foot", the "traditional Chinese medicine foot bath" item in the knowledge base is selected because its applicable direction includes "clearing heat and promoting diuresis" and its intervention area is "foot". All the filtered items are gathered to establish a matching nursing item set.

[0091] The recommendation result generation submodule, based on the matching nursing item set, sorts the nursing action content in the items according to the correlation between the range of drug conditioning and the intervention area of ​​nursing action, and outputs the traditional Chinese medicine nursing recommendation results;

[0092] Based on the matching nursing item set and the correlation between the scope of medicinal conditioning and the intervention area of ​​nursing actions, the items are sorted and refined to generate the final recommendation results. Items with high relevance, such as foot baths, which directly affect the kidney and bladder meridians running through the feet, will be given higher recommendation priority. Subsequently, the specific operating procedures of high-priority items are extracted from the knowledge base and dynamically integrated with the specific medicinal material information obtained in this analysis to generate personalized recommendation content. For example, if the matching nursing item is "traditional Chinese medicine foot bath" and the associated medicinal material is "Phellodendron bark", then the recommendation will be more personalized. If the result is "...", the generated recommendation will be a specific instruction: "It is recommended to use the traditional Chinese medicine foot bath method. Take 30 grams of Phellodendron bark, add water and decoct for 15 minutes, remove the dregs and take the juice. Add the juice to warm water at 40 degrees Celsius and soak both feet for 20 minutes to help clear damp heat in the lower abdomen." If the matching is acupoint massage, the recommendation will be: "It is recommended to massage the Yongquan acupoint (Kidney Meridian) on the sole of the foot and the Kunlun acupoint (Bladder Meridian) behind the outer ankle. Press each acupoint for 5 minutes every day to draw damp heat downward." If the matching is dietary therapy, the recommendation will be: "You can add an appropriate amount of Job's tears and red beans to your daily diet to cook porridge to strengthen the spleen and remove dampness." The output will be the traditional Chinese medicine care recommendation result.

[0093] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An AI-based traditional Chinese medicine nursing recommendation system, characterized in that: The system includes: The tongue image feature extraction module acquires tongue image data and extracts the image features of the outline edges, texture structure and color distribution of the tongue tip, tongue sides, tongue middle and tongue root regions to obtain tongue region feature maps; The tongue body partition anomaly recognition module identifies the anomaly types of each tongue body partition in the tongue body partition feature map, such as broken contour edges, disordered texture structure, and color distribution shift, and generates tongue image anomaly partition annotation data. The meridian mapping and positioning module extracts the names of abnormal parts in the tongue body partitions from the tongue abnormal partition annotation data, matches the abnormal part names with the corresponding body meridian paths according to the TCM meridian attribution relationship, and outputs meridian influence path data. The medicinal property meridian matching module, referring to the meridian influence path data, filters the information of Chinese medicinal materials that belong to the corresponding meridian and whose range of action covers the influence path, and obtains medicinal property association data; The nursing intervention recommendation module compares the drug property association data with a preset nursing intervention knowledge base, filters nursing content that matches the drug property characteristics, and outputs traditional Chinese medicine nursing recommendation results.

2. The AI-based traditional Chinese medicine nursing recommendation system according to claim 1, characterized in that: The tongue body partition feature map specifically includes tongue body partition identification information, image feature vector expression, and feature space coordinate set. The tongue image abnormal partition annotation data includes abnormal partition location, abnormal type label, and feature classification mark. The meridian influence path data specifically includes meridian path name, tongue body part correspondence, and affiliated organ number. The drug property association data includes drug name, meridian number, and action level information. The traditional Chinese medicine nursing recommendation result specifically refers to nursing action name and intervention site location.

3. The AI-based traditional Chinese medicine nursing recommendation system according to claim 1, characterized in that: The tongue image feature extraction module includes: The image feature extraction submodule acquires tongue image data and inputs it into a residual convolutional neural network with a cross-layer connection structure. It extracts features of contour edges, texture structure and color distribution of the image in parallel, and performs numerical dimensional calculation on the extracted multi-channel features to obtain a multi-dimensional feature vector. The partition feature calibration submodule divides the multidimensional feature vector into corresponding partitions based on the preset coordinate masks of the tongue tip, tongue side, tongue middle and tongue root partitions, and calculates the curvature value of the contour edge, the complexity of the texture structure and the mean value of the color distribution in each partition to generate a tongue domain feature parameter set. The feature map fusion construction submodule integrates the tongue domain feature parameter set through multi-channel results and maps and arranges them according to the spatial position relationship of the partitions to establish a tongue body partition feature map.

4. The AI-based traditional Chinese medicine nursing recommendation system according to claim 1, characterized in that: The tongue body partition anomaly recognition module includes: The spatial feature calculation submodule calls the tongue body partition feature map, calculates the spatial position features of each partition of the tongue tip, tongue side, tongue middle and tongue root through the position attention mechanism function, and quantifies the contribution of the contour edge, texture structure and color distribution feature points in the partition as the partition spatial attention weight. The partition anomaly judgment submodule, based on the partition spatial attention weight, compares the features of contour edge breakage, texture structure disorder and color distribution shift of each partition in the tongue partition feature map with the preset anomaly judgment benchmark value one by one, performs partition anomaly type judgment, and obtains partition anomaly category set. The anomaly partition annotation submodule marks the identified partition anomaly category set and the corresponding contour edge breakage, texture structure disorder and color distribution shift features, and associates and classifies them with the category information of the tongue tip, tongue side, tongue middle and tongue root partitions to generate tongue image anomaly partition annotation data.

5. The AI-based traditional Chinese medicine nursing recommendation system according to claim 1, characterized in that: The meridian mapping and positioning module includes: The abnormal part extraction submodule identifies and extracts the names of abnormal parts marked in the tongue image abnormal partition annotation data, and performs a deduplication operation on the repeated names to obtain a list of abnormal part names; The meridian relationship matching submodule, based on the list of abnormal part names, calls the TCM meridian attribution relationship, and performs meridian mapping operation on each part name in the list according to the rule that the tip of the tongue maps to the heart meridian, the sides of the tongue maps to the liver and gallbladder meridians, the middle of the tongue maps to the spleen and stomach meridians, and the root of the tongue maps to the kidney and bladder meridians, to establish a set of meridian mapping relationships; The influence path construction submodule associates the name of each abnormal part in the meridian mapping relationship set with the matching body meridian path, constructs the correspondence between tongue appearance and meridians, and obtains meridian influence path data.

6. The AI-based traditional Chinese medicine nursing recommendation system according to claim 5, characterized in that: The TCM meridian attribution relationship is a pre-defined structured lookup table. The structured lookup table contains two data fields: tongue body partition and meridian path. It also pre-stores key-value pair records of the tongue tip and the heart meridian, the tongue side and the liver and gallbladder meridians, the tongue middle and the spleen and stomach meridians, and the tongue root and the kidney and bladder meridians.

7. The AI-based traditional Chinese medicine nursing recommendation system according to claim 1, characterized in that: The drug property and meridian tropism matching module includes: The meridian path analysis submodule performs structured analysis on the path information in the meridian influence path data, extracts the meridian name and corresponding organ type, and summarizes the identified meridian name and organ type to obtain a meridian organ list. The medicinal material information filtering submodule compares the meridian names in the meridian organ list with the meridian tropism attributes of all medicinal materials in the medicinal property meridian tropism information database, filters the Chinese medicinal material entries corresponding to the meridians, and establishes a candidate medicinal material information set. The drug property feature matching submodule, based on the candidate drug information set and the meridian influence path data, determines whether the meridian tropism and action level of each drug are consistent with the requirements in the influence path, extracts drug information whose action range can cover the influence path, determines the drug property features related to the current tongue appearance, and generates drug property association data.

8. The AI-based traditional Chinese medicine nursing recommendation system according to claim 7, characterized in that: The meridian tropism attribute is a text set consisting of multiple standardized meridian names. The text set is compared with the meridian names in the meridian organ list to determine the meridian tropism matching.

9. The AI-based traditional Chinese medicine nursing recommendation system according to claim 1, characterized in that: The nursing intervention recommendation module includes: The drug property characteristic analysis submodule performs structured analysis on the drug property characteristics and range of action in the drug property association data, identifies the required conditioning direction and body intervention site, and generates a conditioning requirement list; The nursing content filtering submodule compares the applicable treatment direction and intervention area of ​​each nursing content in the preset nursing intervention knowledge base with the needs in the treatment needs list, filters out nursing content that matches the drug properties, and establishes a set of matching nursing items. The recommendation result generation submodule, based on the matching nursing item set, sorts the nursing action content in the items according to the correlation between the range of drug conditioning and the intervention area of ​​nursing action, and outputs the traditional Chinese medicine nursing recommendation result.

10. The AI-based traditional Chinese medicine nursing recommendation system according to claim 9, characterized in that: The applicable treatment direction is a text set consisting of multiple treatment direction keywords. The text set is compared with the treatment requirement list as the input for matching judgment. The intervention area is a text set consisting of multiple body part names. The text set is matched with the body intervention parts in the treatment requirement list during the comparison operation.

Images