Intelligent inquiry recommendation method and system

Abstract

The invention relates to the technical field of intelligent recommendation, in particular to an intelligent inquiry recommendation method and system, and the method comprises the following steps: extracting text organ word frequency features through TF-IDF, generating a matching degree matrix through cosine similarity, screening core features, analyzing positioning parameters through DICOM, constructing a co-occurrence matrix through collaborative filtering, screening associated feature pairs, indexing a recommendation library, and predicting a symptom path through LSTM. And generating a time sequence weight feature set and a PageRank iterative sorting recommendation table. According to the method, a multi-modal association system is constructed by fusing text features and image parameters, semantic tags and space coordinates are combined to filter and screen organ domain features in a collaborative manner, the matching precision of symptoms and resources is improved, a time sequence weight reconstruction model is adopted to capture symptom evolution features, and hidden state transition is adopted to enhance the disease course prediction capability; a three-dimensional decision model is constructed through feature node sorting, time, space and feature dimension unification is achieved, a personalized diagnosis and treatment scheme is optimized by fusing multi-dimensional features and dynamic weights, and the credibility and clinical applicability of a recommendation result are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent recommendation technology, and in particular to an intelligent medical consultation recommendation method and system. Background Art

[0002] The field of intelligent recommendation technology encompasses personalized information delivery systems based on user behavior analysis and data mining. Its core objective is to precisely match user needs with resource availability through algorithmic models. This area involves technical steps such as user profile construction, feature vector extraction, similarity calculation, and recommendation result ranking. It is primarily used in scenarios such as e-commerce, content distribution, and healthcare, utilizing collaborative filtering, deep learning, and knowledge graphs to filter and deliver information.

[0003] The intelligent consultation recommendation method refers to a technical solution that implements diagnosis and treatment plan push through a multi-dimensional matching mechanism between the medical knowledge base and patient symptom characteristics. This method specifically covers three technical aspects: quantization of symptom characteristics, construction of a disease diagnosis rule base, and modeling of patient-doctor bidirectional preferences. It uses a knowledge graph-based reasoning mechanism to map the association between symptoms and diseases, employs a multi-dimensional feature fusion strategy to dynamically match individual patient differences with medical resources, and balances the relationship between symptom severity and department expertise through a dynamic weight allocation mechanism.

[0004] Existing technologies mainly rely on static text processing and single-modal data in the symptom feature analysis stage, and fail to effectively integrate the spatial positioning information of imaging data, resulting in semantic ambiguity and lack of spatial correlation in the feature vector construction process. In the application of disease diagnosis rule bases, traditional collaborative filtering algorithms lack the ability to model the temporal evolution of symptoms, resulting in a decrease in the matching degree between recommendation results and the patient's actual disease course stage. The multi-dimensional feature fusion strategy based on fixed weights is difficult to adapt to the dynamic changes of complex symptom combinations. When modeling patient-doctor two-way preferences, the static knowledge graph reasoning mechanism cannot fully reflect the real-time status changes of the medical resource supply side, which ultimately leads to decision-making deviations in the recommendation system when dealing with multimodal medical data and dynamic disease course scenarios. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose an intelligent medical consultation recommendation method and system.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent medical consultation recommendation method, comprising the following steps: S1: Extract organ-related word frequency features from user self-report texts using the TF-IDF algorithm, perform cosine similarity calculation on feature vectors to generate an organ domain matching matrix, filter and generate an organ domain core feature set based on the organ domain matching matrix, and extract image space coordinate parameters using the DICOM metadata parser to generate an image organ positioning parameter set; S2: Based on the semantic labels of the core feature set of the organ domain and the spatial coordinates of the imaging organ positioning parameter set, a collaborative filtering algorithm is used to construct a feature co-occurrence frequency matrix, threshold comparison is performed on the elements of the feature co-occurrence frequency matrix to generate a set of multimodal associated feature pairs, and the index is used to construct a recommendation candidate feature library; S3: Based on the timestamp sequence of the recommended candidate feature library, the LSTM network forward propagation is used to calculate the hidden state of the symptom duration, a set of symptom evolution path vectors is generated through the hidden state transfer, and a feature set with time series weights is generated based on the similarity; S4: Based on the weight parameters of the feature set with temporal weights and the co-occurrence frequency of the multimodal associated feature pair set, the PageRank algorithm is used to iteratively calculate the PageRank values ​​of the feature nodes, and the three-dimensional recommendation decision table is generated by sorting.

[0007] As a further solution of the present invention, the organ domain core feature set is specifically a word frequency weight and a similarity score. The organ domain core feature set and the image organ positioning parameter set are collaboratively verified through a spatial-semantic mapping relationship to improve the organ positioning accuracy; The LSTM network contains 128 hidden units, and the input gate and forget gate activation functions adopt the sigmoid function.

[0008] As a further solution of the present invention, the image organ positioning parameter set includes spatial coordinates and DICOM tags, the multimodal association feature pair set is specifically a co-occurrence frequency value and a modality type pair, the recommended candidate feature library includes a feature index and an association strength, the symptom evolution path vector set includes a timestamp sequence and a hidden state vector, the feature set with time series weight is specifically a weight value and a time series correlation degree, the three-dimensional recommendation decision table includes a PageRank value, an organ domain label, a time series weight, and a modality correlation degree. As a further solution of the present invention, the step S1 is specifically as follows: S101: Obtain user self-description text, calculate the inverse document frequency weight of organ-related vocabulary using the TF-IDF algorithm, count the frequency distribution of multiple organ terms in the text, multiply the frequency value by the weight value, and generate an organ-related word frequency vector; S102: Based on the organ-related word frequency vectors, a multidimensional feature space coordinate system is constructed. The cosine value of the angle between the vectors is calculated using a cosine similarity function. The calculation results are arranged in a matrix by organ category to generate an organ domain similarity matrix. The interquartile range method is used to set a similarity threshold. Feature dimensions above the threshold in the matrix are screened to form an organ domain core feature set. The multidimensional feature space coordinate system uses the Z-score normalization method to perform dimensionless processing on the organ-related word frequency vector; S103: Call the spatial positioning tag in the DICOM metadata, parse the image layer thickness, pixel spacing, and slice position parameters, convert the spatial coordinates in the Cartesian coordinate system into voxel index values, group and integrate the parameters according to the scan sequence number, and generate an image organ positioning parameter set.

[0009] As a further solution of the present invention, the step S2 is specifically as follows: S201: calling the semantic labels of the core feature set of the organ domain and the spatial coordinates of the imaging organ positioning parameter set, using a collaborative filtering algorithm to count the number of co-occurrences of feature pairs in the spatial dimension, converting the mapping relationship between the semantic labels and the spatial coordinates into a two-dimensional frequency distribution table, and generating a feature co-occurrence frequency matrix; S202: Based on the feature co-occurrence frequency matrix, calculating the Pearson correlation coefficient between the matrix row vector and the column vector, setting a dynamic threshold according to the correlation coefficient distribution, screening elements in the matrix with correlation coefficients higher than the threshold, and generating a set of multimodal correlation feature pairs; S203: According to the feature combination pattern in the multimodal associated feature pair set, an index mapping table of feature codes and spatial coordinates is established, high-frequency associated feature pairs are grouped and stored according to organ categories, and a recommended candidate feature library is generated.

[0010] As a further solution of the present invention, the step of acquiring the feature set with temporal weights is specifically as follows: S301: Calling the timestamp sequence of the recommended candidate feature library, using the LSTM network input gate to calculate the time step weight, controlling the historical state retention ratio through the forget gate, inputting the candidate features into the network in chronological order for forward propagation, and generating the symptom duration hidden state; S302: Based on the hidden state of the symptom duration, a state transition probability matrix is ​​constructed, the transition path between states is calculated using a hidden Markov model, the hidden states of consecutive time steps are connected to form a state transition chain, and a set of symptom evolution path vectors is generated; S303: Calculate the cosine similarity between the path vectors based on the symptom evolution path vector set, assign a temporal weight coefficient according to a time decay function, perform weighted fusion on the similarity value and the weight coefficient, and generate a feature set with temporal weights.

[0011] As a further solution of the present invention, the steps of obtaining the three-dimensional recommendation decision table are specifically as follows: S401: calling the weight parameters of the temporal weight feature set, obtaining the time decay weight value between feature nodes, combining the number of outbound links of the feature nodes, calculating the initial transition probability of each feature node, and generating a feature node weight set; S402: Based on the multimodal associated feature pair set, counting the co-occurrence frequencies between feature nodes, introducing a modality type difference factor, performing modality dimension normalization processing on the co-occurrence frequencies, and establishing a co-occurrence frequency matrix; S403: Based on the transition probability of the feature node weight set and the modal association strength of the co-occurrence frequency matrix, the formula is adopted: ;in, represents the PageRank value of the k+1th iteration of the i-th feature node, is the damping coefficient and its value range is 0.85≤d≤0.95, is the total number of feature nodes, is the temporal weight parameter of feature nodes i to j, is the co-occurrence frequency value of feature nodes i and j, is the length of the eigenvector of the mth modal dimension, is the total number of modal dimensions, Used to strengthen the propagation weight of high-frequency co-occurring feature pairs.

[0012] An intelligent medical consultation recommendation system, which is used to implement the above-mentioned intelligent medical consultation recommendation method, includes: A multimodal feature extraction module is used to extract organ-related word frequency features from user self-report text using the TF-IDF algorithm, perform cosine similarity calculation on feature vectors to generate an organ domain matching matrix, screen and generate an organ domain core feature set based on the matrix element threshold, extract image space coordinate parameters using a DICOM metadata parser to generate an image organ positioning parameter set, and pass the organ domain core feature set and the image organ positioning parameter set to a feature association analysis module; a feature association analysis module, configured to construct a feature co-occurrence frequency matrix using a collaborative filtering algorithm based on the semantic labels of the organ domain core feature set and the spatial coordinates of the imaging organ positioning parameter set, perform threshold comparison on the matrix elements to generate a multimodal associated feature pair set, construct a recommended candidate feature library based on the feature pair index, and pass the recommended candidate feature library and the multimodal associated feature pair set to a temporal evolution modeling module; A temporal evolution modeling module is used to calculate the hidden state of symptom duration using the LSTM network forward propagation based on the timestamp sequence of the recommendation candidate feature library, generate the symptom evolution path vector set through the hidden state transfer, perform cosine similarity calculation on the path vector to generate a feature set with temporal weights, and pass the feature set with temporal weights to the dynamic recommendation decision module; The dynamic recommendation decision module is used to iteratively calculate the PageRank value of the feature node using the PageRank algorithm through the weight parameters of the feature set with temporal weights and the co-occurrence frequency of the multimodal associated feature pair set, and generate a three-dimensional recommendation decision table according to the value range sorting.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are: In the present invention, a multimodal data association system is constructed by integrating text feature extraction and image spatial parameter analysis. The collaborative filtering mechanism of semantic labels and spatial coordinates is combined in the process of organ domain feature screening to effectively improve the matching accuracy of symptom characteristics and medical resources. The temporal weight modeling method is adopted to capture the dynamic change characteristics of the symptom evolution path, and the ability to predict the development law of the disease course is enhanced through the hidden state transfer mechanism. A three-dimensional decision model is established based on the feature node sorting algorithm to achieve the organic unity of the time dimension, space dimension and feature dimension in the process of medical resource recommendation. On the basis of maintaining the core function of the disease diagnosis rule library, the personalized adaptation ability of the diagnosis and treatment plan is optimized through multi-dimensional feature fusion and dynamic weight allocation, which significantly improves the credibility and clinical applicability of the recommendation results in complex symptom scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is the overall flow chart of the intelligent medical consultation recommendation method of the present invention; Figure 2 This is a flow chart of organ feature extraction and positioning of the present invention; Figure 3 A flowchart for multimodal feature association and candidate library construction of the present invention; Figure 4 This is a flow chart for obtaining a feature set with temporal weights according to the present invention; Figure 5 Generate a flow chart for the three-dimensional recommendation decision table of the present invention. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present 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 only used to explain the present invention and are not intended to limit the present invention.

[0016] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined. Example 1

[0017] See also Figure 1 To achieve the above objectives, the present invention adopts the following technical solution: an intelligent medical consultation recommendation method, comprising the following steps: S1: The organ-related word frequency features of the user's self-report text are extracted using the TF-IDF algorithm. The cosine similarity calculation is performed on the feature vector to generate an organ domain matching matrix. The organ domain core feature set is generated based on the organ domain matching matrix. At the same time, the image space coordinate parameters are extracted through the DICOM metadata parser to generate an image organ positioning parameter set. S2: Based on the semantic labels of the core feature set of the organ domain and the spatial coordinates of the imaging organ positioning parameter set, a collaborative filtering algorithm is used to construct a feature co-occurrence frequency matrix. Threshold comparison is performed on the elements of the feature co-occurrence frequency matrix to generate a set of multimodal associated feature pairs, which are then indexed to construct a recommendation candidate feature library. S3: Based on the timestamp sequence of the recommended candidate feature library, the LSTM network forward propagation is used to calculate the hidden state of the symptom duration, and the symptom evolution path vector set is generated through hidden state transfer. A feature set with time series weights is generated based on the similarity. S4: Based on the weight parameters of the feature set with temporal weights and the co-occurrence frequency of the multimodal associated feature pair set, the PageRank algorithm is used to iteratively calculate the PageRank values ​​of the feature nodes, and the ranking is used to generate a three-dimensional recommendation decision table.

[0018] The core feature set of the organ domain is specifically the word frequency weight and similarity score. The core feature set of the organ domain and the image organ positioning parameter set are collaboratively verified through the spatial-semantic mapping relationship to improve the accuracy of organ positioning; The LSTM network contains 128 hidden units, and the input gate and forget gate activation functions use the sigmoid function.

[0019] The image organ positioning parameter set includes spatial coordinates and DICOM tags. The multimodal association feature pair set includes co-occurrence frequency values ​​and modality type pairs. The recommended candidate feature library includes feature index and association strength. The symptom evolution path vector set includes timestamp sequence and hidden state vector. The feature set with time series weight includes weight value and time series correlation. The three-dimensional recommendation decision table includes PageRank value, organ domain label, time series weight and modality correlation. See also Figure 2 , the steps of S1 are as follows: S101: Obtain user self-description text, calculate the inverse document frequency weight of organ-related vocabulary using the TF-IDF algorithm, count the frequency distribution of multiple organ terms in the text, multiply the frequency value by the weight value, and generate an organ-related word frequency vector; The process of obtaining the user's self-reported text is as follows: first, the electronic medical record numbered "HIS20250603001" is retrieved from the medical information system. The medical record records the chief complaint information of the patient "Li Ming" on June 3, 2025. The specific text content is: "I have felt dull pain in the upper right abdomen for the past month, accompanied by fatigue and loss of appetite, and have a history of hepatitis B." This text is used as the user's self-reported text for analysis. Then, based on the preset organ-related vocabulary library, which pre-stores fifty organ terms such as "liver", "gall bladder", "pancreas", "lungs", and "heart", the user's self-reported text is segmented to identify the organ terms that appear in the text. For example, the text recognizes that "upper right abdomen" is usually associated with "liver". Here, the word "liver" is directly identified, and the number of occurrences of each organ term in the current text is counted, which is recorded as the word frequency (TF). For example, in this text, "liver" appears once. Assuming that the total number of words after text segmentation is 50, the word frequency TF value of "liver" is At the same time, based on a benchmark corpus containing 10,000 medical self-report texts, the number of documents in which each organ term appears in different texts is counted, and its inverse document frequency (IDF) weight is calculated. For example, if the word "liver" appears in 200 texts in the benchmark corpus, its IDF value is , where "+1" is used to avoid the denominator being zero. Then, the TF value of each organ term is multiplied by its corresponding IDF weight value to obtain the TF-IDF value of the organ term. For example, the TF-IDF value of "liver" is , the same word frequency statistics and TF-IDF value calculations are performed on all other organ terms in the vocabulary, such as "gall bladder" and "pancreas". If an organ term does not appear in the text, its word frequency is 0 and its TF-IDF value is also 0. For example, "lung" does not appear in this text, its TF value is 0 and its TF-IDF value is 0. Finally, the TF-IDF values ​​of all organ terms are arranged in a preset order to form a numerical vector, namely the organ-related word frequency vector.

[0020] S102: Based on the organ-related word frequency vectors, a multidimensional feature space coordinate system is constructed. The cosine value of the angle between the vectors is calculated using the cosine similarity function. The calculation results are arranged in a matrix by organ category to generate an organ domain similarity matrix. The interquartile range method is used to set the similarity threshold. Feature dimensions above the threshold in the matrix are screened to form the organ domain core feature set. The Z-score normalization method is used in the multidimensional feature space coordinate system to perform dimensionless processing on the organ-related word frequency vectors; The process of obtaining the organ-related word frequency vector is as follows: first, the organ-related word frequency vector generated in step S101 is retrieved. For example, a simplified three-dimensional vector is For this type of vector set containing multiple organ TF-IDF values, for example, simultaneously processing organ-related word frequency vectors from two other user self-report texts and , the values ​​of each dimension of these vectors (corresponding to an organ) are subjected to Z-score normalization in all vectors. Specifically, the mean and standard deviation of each dimension in all vectors are calculated. For example, for the "liver" dimension, its mean is , and its standard deviation is , then the vector The normalized value of the “liver” dimension is , perform this normalization operation on all dimensions in the vector to obtain the dimensionless vector in the multidimensional feature space coordinate system. Then, calculate the cosine similarity between any two normalized organ-related word frequency vectors. The calculation method is to divide the dot product of the two vectors by the product of the two vector moduli. For example, the normalized vector and (The values ​​here are assumed to be standardized results), and their dot product is , The modulus length is , The modulus length is , then the cosine similarity is , the cosine similarity calculation results of all vector pairs are arranged in a matrix according to the organ category (for example, the similarities of multiple subdivided liver regions can be classified into one liver category) to form an organ domain similarity matrix. For example, if there are three main organ categories "liver", "lung", and "stomach", then the elements of the matrix Represents the average similarity between category i and category j. Next, the interquartile range method is applied to all similarity values ​​in the matrix to set the similarity threshold. First, all similarity values ​​in the matrix are sorted, the first quartile (Q1) and the third quartile (Q3) are calculated, and the interquartile range IQR = Q3-Q1 is calculated. The threshold is set to For example, if the similarity value set is sorted, Q1 is 0.2, Q3 is 0.6, then the IQR is 0.4, and the threshold is Since the cosine similarity ranges from -1 to 1, this threshold setting may need to be adjusted, or the Q3 value or a specific percentile can be directly used as the threshold. For example, the 75th percentile value of all similarity values ​​​​is taken as the threshold, and the feature dimensions with values ​​greater than 0.6 in the matrix, that is, organ pairs or organ category pairs, are screened out. These screened feature dimensions together constitute the core feature set of the organ domain.

[0021] S103: Call the spatial positioning tag in the DICOM metadata, parse the image layer thickness, pixel spacing, and slice position parameters, convert the spatial coordinates in the Cartesian coordinate system into voxel index values, group and integrate the parameters according to the scan sequence number, and generate an image organ positioning parameter set.

[0022] The calling process of the spatial positioning tag in the DICOM metadata is as follows: first, access the medical image sequence associated with the patient "Li Ming", such as the abdominal CT scan image data with the serial number "IMG001", and extract the specific spatial positioning tag from the DICOM metadata of each image slice in the sequence, specifically extracting the values ​​of the three parameters "slice thickness" (DICOMTag(0018,0050)), "pixel spacing" (DICOMTag(0028,0030)) and "slice position" (DICOMTag(0020,1041)). For example, for the first slice in the sequence "IMG001", the slice thickness parameter value is 5.0 (unit: mm), and the pixel spacing parameter value is (Indicates that the row spacing and column spacing are both 0.625mm), the slice position parameter value is -120.0 (unit: mm, indicating the position relative to the reference plane), then these spatial coordinate parameters based on the Cartesian coordinate system are converted into voxel index values, the voxel index value They correspond to the positions in the three-dimensional array of image data, where and The index is calculated based on the pixel position and pixel spacing, and The index (layer index) is calculated based on the slice position and layer thickness. For example, if the image origin is at the upper left corner of the image matrix, the value of a pixel in the image is The coordinates are divided by the pixel spacing to get index, and The index can be determined by sorting all slices by their "slice position" value, or by the formula Calculation shows that, assuming the first slice position is -120.0mm and the slice thickness is 5.0mm, the K index of the slice can be regarded as 0. The next slice position is -115.0mm, and its K index is , this parameter extraction and conversion operation is performed on all image slices in the sequence, and these parameters are grouped and integrated according to the scan sequence number, such as "IMG001". Each group corresponds to a complete scan sequence, and finally a set of image organ positioning parameters is generated.

[0023] See also Figure 3 , the steps of S2 are as follows: S201: calling the semantic labels of the core feature set of the organ domain and the spatial coordinates of the imaging organ positioning parameter set, using a collaborative filtering algorithm to count the number of co-occurrences of feature pairs in the spatial dimension, converting the mapping relationship between the semantic labels and the spatial coordinates into a two-dimensional frequency distribution table, and generating a feature co-occurrence frequency matrix; The calling process of the semantic labels of the organ domain core feature set and the spatial coordinates of the image organ positioning parameter set is as follows: first, the organ domain core feature set is obtained from step S102. Each core feature in this set has a semantic label, such as the label "diffuse changes in the liver area" or "nodules in the right lower lobe of the lung". At the same time, the image organ positioning parameter set is obtained from step S103. This set contains the spatial coordinates associated with each image slice, such as the voxel index value Next, for each pair of core feature semantic labels and image space coordinates, the idea of ​​collaborative filtering algorithm is used to count the number of co-occurrences of these feature pairs in the spatial dimension. When executing, all the patient's image data and their corresponding organ domain core features are traversed. If a patient has a specific spatial coordinate (for example, in voxel blocks), As a unit, the voxel block consists of a set of continuous voxel indices) and semantic label A (for example, "liver calcification") and semantic label B (for example, "liver morphological abnormality") appear simultaneously near each other, then A and B are considered to co-occur once in this spatial position, or if the lesion area indicated by the semantic label A (located by image segmentation or other means) overlaps with the spatial range of an organ (for example, "liver") marked in the image organ positioning parameter set in three-dimensional space, then the co-occurrence of the semantic label and the spatial position of the organ is recorded. Here, the collaborative filtering prediction score is not directly used, but the idea of ​​statistical co-occurrence is borrowed to associate each semantic label (for example, "liver lesion feature 1") with each discretized spatial coordinate unit (for example, the entire scan range is divided into The mapping relationship between the spatial cells (each cell has a number) is converted into a two-dimensional frequency distribution table by counting the frequency of their co-occurrence in the entire dataset. The rows of the table represent different semantic labels, and the columns represent different spatial coordinate units (or organ regions). Each cell in the table The value of is the semantic label In space coordinate units For example, if the semantic label "liver inflammation" appears 150 times in the spatial unit "upper segment of right liver lobe", the value of the corresponding cell in the table is 150. By filling in the co-occurrence counts of all semantic labels and all spatial coordinate units, the feature co-occurrence frequency matrix is ​​finally generated.

[0024] S202: Based on the feature co-occurrence frequency matrix, calculate the Pearson correlation coefficient between the matrix row vector and the column vector, set a dynamic threshold according to the correlation coefficient distribution, filter the elements in the matrix with correlation coefficients higher than the threshold, and generate a set of multimodal correlation feature pairs; The process of obtaining the feature co-occurrence frequency matrix is ​​as follows: first, the feature co-occurrence frequency matrix generated in step S201 is retrieved and recorded as matrix , the rows of the matrix represent different semantic labels (e.g., “diffuse changes in the liver area”, “nodules in the right lower lobe of the lung”), the columns represent different spatial coordinate units or organ regions (e.g., “liver area A”, “lung area B”), and the matrix elements Is a semantic tag In the space area Next, we calculate the Pearson correlation coefficient between any two row vectors and any two column vectors in the matrix. The correlation between row vectors indicates the similarity of different semantic labels in spatial distribution, and the correlation between column vectors indicates the similarity of different spatial regions in carrying semantic features. and (corresponding to semantic labels and semantic tags ) as an example, the calculation formula is ,in is the number of columns (number of spatial regions), It's a label In the area frequency, is a row vector The average of the elements in , for example, the row vector and ,but , , the numerator is calculated as , the first term in the square root of the denominator is , the second term is , then the correlation coefficient , according to the overall distribution of all calculated correlation coefficients (e.g., inter-row correlation coefficients and inter-column correlation coefficients), set a dynamic threshold, for example, calculate the mean of all correlation coefficients and standard deviation , set the threshold to ,in is a constant, for example, , if the calculated Pearson correlation coefficient has a mean of 0.5 and a standard deviation of 0.1, then the threshold is , the element pairs with correlation coefficients higher than 0.65 in the matrix (i.e., semantic label pairs or spatial region pairs, or strong correlation between semantic labels and spatial regions) are screened out, and these screened element pairs constitute a set of multimodal correlation feature pairs.

[0025] S203: According to the feature combination pattern in the multimodal associated feature pair set, an index mapping table of feature codes and spatial coordinates is established, high-frequency associated feature pairs are grouped and stored by organ category, and a recommended candidate feature library is generated.

[0026] The process of identifying the feature combination pattern in the multimodal associated feature pair set is as follows: first, the multimodal associated feature pair set generated in step S202 is retrieved, which contains feature pairs with correlation higher than a preset threshold, for example, ("diffuse changes in the liver area", "liver area A"), ("nodules in the right lower lobe of the lung", "apical segment of the lung"), ("diffuse changes in the liver area", "increased jaundice index"), etc., where the features can be semantic labels or spatial area identifiers, and the combination patterns in these feature pairs are analyzed, for example, which semantic labels often appear together and which semantic labels are specifically associated with certain spatial areas. Based on these combination patterns, an index mapping table of feature encoding and spatial coordinates is established, specifically, a unique code (such as "F001") is assigned to each independent semantic feature (such as "diffuse changes in the liver area"), and a unique code (such as "F002") is assigned to each standardized spatial coordinate unit (such as "liver area"). Left lobe cell 035") is also assigned a unique code (such as "S035"). If the feature pair ("diffuse changes in the liver area", "liver area A") is selected, the association between "F001" and "S035" (assuming that "liver area A" corresponds to "S035") is recorded in the index mapping table. The table can be a key-value pair storage, with the key being the feature code and the value being the associated spatial coordinate code list, or a relational table storing feature code pairs and their association strengths. Subsequently, feature pairs with high-frequency associations are grouped and stored according to predefined organ categories, such as "liver", "lung", "heart", etc. For example, all strongly associated feature pairs related to "liver" (such as "cirrhosis" and "portal hypertension", "hepatocellular carcinoma" and "elevated alpha-fetoprotein") are classified under the "liver" category. The set of these grouped and stored feature pairs forms a recommended candidate feature library.

[0027] See also Figure 4 , the steps for obtaining the feature set with temporal weights are as follows: S301: Call the timestamp sequence of the recommended candidate feature library, use the LSTM network input gate to calculate the time step weight, control the historical state retention ratio through the forget gate, input the candidate features into the network in chronological order for forward propagation, and generate the symptom duration hidden state; The calling process of the timestamp sequence of the recommended candidate feature library is as follows: first, the recommended candidate feature library is obtained from step S203. Each feature or feature pair in the library is associated with a timestamp, which indicates the time when the feature was recorded or confirmed. For example, the timestamp of feature "F001" (diffuse changes in the liver area) in the record of patient "Li Ming" is "2025-05-10", and the timestamp of feature "F002" (increased jaundice index) is "2025-05-15". These candidate features are arranged in the order of timestamps to form a time series. For example, for patient "Li Ming", the feature sequence may be , this time series is used as the input of the long short-term memory network (LSTM). At the input gate of the LSTM network, the importance of the input features of each time step is calculated, that is, the time step weight. This calculation process involves the input feature vector of the current time step and the hidden state vector of the previous time step, and operates through a specific weight matrix and activation function (such as the sigmoid function) to generate a weight value between 0 and 1. This weight value determines how much of the current input information is allowed to enter the memory unit. Then, the retention ratio of the historical state (that is, the memory unit state of the previous time step) is controlled by the forget gate. The calculation method of the forget gate is similar to that of the input gate. It also generates a value between 0 and 1 by applying the weight matrix and sigmoid activation function to the current input and the previous hidden state. This value is element-wise multiplied with the previous memory unit state to determine which old information is forgotten. The candidate features are input into the LSTM network one by one in the order of their appearance in the time series for forward propagation. At each time step, the network updates its internal memory unit state and hidden state. Finally, after processing the entire sequence, the LSTM network outputs the hidden state of each time step. These hidden states encode the duration and cumulative impact of the symptoms (features) up to that point in time, forming a symptom duration hidden state sequence.

[0028] S302: Based on the hidden state of symptom duration, a state transition probability matrix is ​​constructed, and the transition path between states is calculated through the hidden Markov model. The hidden states of consecutive time steps are connected to form a state transition chain, and a set of symptom evolution path vectors is generated; The process of acquiring the hidden state of symptom duration is as follows: first, call the hidden state sequence of symptom duration generated in step S301, which is a vector sequence ,in is in the time step The hidden state vectors are constructed based on these hidden states. The specific method can be to count the frequency of transitions from one hidden state (or the state represented by its cluster) to another hidden state (or the state represented by its cluster) in a large amount of patient symptom sequence data, and then normalize the frequency to probability. For example, if hidden state A appears 80 times after 100 times of hidden state B, then the transition probability from A to B is approximately 0.8, forming a Matrix ,in Indicates the slave state Transfer to state The probability of is the total number of discretized hidden states. Subsequently, the idea of ​​the hidden Markov model (HMM) is used, combined with this state transition probability matrix and the emission probability from the hidden state to the observed feature (which can be estimated from the output of the LSTM or the representation of the feature itself), to calculate the transition path between different states. Through decoding methods such as the Viterbi algorithm, the most likely hidden state sequence can be found under a given observation sequence (i.e., the input symptom feature sequence). These most likely hidden states of consecutive time steps are connected to form a state transfer chain. For example, for a patient, the state transfer chain may be "early inflammatory state". Persistent inflammatory state Tissue fibrosis state", this operation is performed on all patients in the dataset, resulting in a set of vectors containing multiple different symptom evolution paths, each path vector is actually a series of ordered hidden states (or the discrete state labels they represent).

[0029] S303: Based on the symptom evolution path vector set, calculate the cosine similarity between the path vectors, assign a temporal weight coefficient according to the time decay function, perform weighted fusion on the similarity value and the weight coefficient, and generate a feature set with temporal weights.

[0030] The process of obtaining the symptom evolution path vector set is as follows: first, the symptom evolution path vector set generated in step S302 is retrieved, where each path vector represents the trajectory of the symptom evolution of a patient or a case over time. For example, path 1 is , path 2 is ,in Represents a symptom state, and calculates the cosine similarity between any two path vectors. If the path vector is a hidden state sequence in a high-dimensional space, the sequence can be directly concatenated or averaged to calculate the cosine similarity between the vectors; if the path is a discrete state sequence, the state sequence can be first converted into a numerical vector (for example, using one-hot encoding or embedding vectors), and then the cosine similarity is calculated. For example, path 1 is vectorized to , path 2 is vectorized to , and its cosine similarity is , get a similarity value, such as 0.75. Next, assign a temporal weight coefficient to the state transition or feature appearance in each path according to the time decay function. The design principle of the time decay function is that recent events (states or features) have higher weights than events that occurred far away. For example, an exponential decay function can be used. ,in is the time difference between the time when the event occurs and the current time. is the decay constant, e.g. , if a feature appeared 3 days ago, its weight coefficient is , if it appeared 10 days ago, the weight coefficient is , the cosine similarity values ​​between the paths (or features) calculated above are weighted fused with the time series weight coefficients obtained according to the time decay function. One fusion method can be to perform weighted fusion on the two features. and , the weighted similarity between them , or when comparing paths, the contribution of each element in the path is adjusted by its temporal weight. Through this weighted fusion, all core features or feature pairs are processed to finally generate a feature set with temporal weights. Each feature or relationship between features in this set is accompanied by a temporal weight value that reflects its current importance.

[0031] See also Figure 5 ,The specific steps for obtaining the three-dimensional recommendation decision table are: S401: Call the weight parameters of the temporal weight feature set to obtain the time decay weight value between feature nodes, calculate the initial transition probability of each feature node based on the number of outbound links of the feature node, and generate a feature node weight set; The calling process of the weight parameters of the temporal weight feature set is as follows: first, a feature set with temporal weights is obtained from step S303. The set includes each feature node (for example, "pain in the liver area", "CT image feature A") and the associations between them, and each feature node or association has a weight value calculated by time decay. For example, the time decay weight value of the feature node "F1" (pain in the liver area) is 0.8, and the time decay weight value of the feature node "F2" (jaundice) is 0.6. The weight values ​​calculated by the time decay function between these feature nodes are obtained. These weight values ​​reflect the changes in the influence of the features over time. Combined with the number of outbound links of each feature node, that is, the number of connections from the feature node to other associated feature nodes, the initial transfer probability of each feature node is calculated. The specific calculation method is to distribute the total temporal weight of a node according to the nodes to which its outbound links point, or to adopt a more complex distribution mechanism. For example, if the feature node The timing weight of , it has a direction and The outgoing link is assigned to The initial transfer impact can be based on and and The specific relationship strength between them, the "initial transfer probability" mentioned here is more inclined to be an initial potential energy reflection of the node's own weight in the network. For example, if the node The timing weight of , its out-degree (number of nodes pointing to other nodes) is , we can simply As its base weight, or normalize it, for example, if the sum of the base weights of all nodes in the network is , then the node The initial influence can be set to ,By performing this operation on all feature nodes, a feature node weight set is generated, in which each feature node is associated with a quantified weight value, representing its initial importance or basic propagation ability in the subsequent PageRank calculation.

[0032] S402: Based on the multimodal correlation feature pair set, the co-occurrence frequency between feature nodes is counted, a modality type difference factor is introduced, the co-occurrence frequency is normalized to the modality dimension, and a co-occurrence frequency matrix is ​​established; The process of obtaining the multimodal associated feature pair set is as follows: first, the multimodal associated feature pair set generated in step S202 is retrieved, which records the association between different features (for example, the symptom "fatigue" in the text description, the imaging feature "uneven liver density"), and the co-occurrence frequency between these feature nodes is counted. For example, in 1,000 medical records and imaging data, the feature "fatigue" and the feature "uneven liver density" appear together 300 times, then their co-occurrence frequency is 300. Next, the modality type difference factor is introduced. This factor is set to adjust the contribution of feature co-occurrences from different modal sources to the calculation of co-occurrence frequencies. For example, the difference factor between the text modality and the imaging modality is set to 0.9, the difference factor of feature co-occurrences within the text modality is 1.0, and the difference factor of feature co-occurrences within the imaging modality is also 1.0. This difference factor It can be set based on experience, or verified through small-scale experimental data. For example, after introducing different difference factor values ​​(such as 0.7, 0.8, 0.9, 1.0), the accuracy of the final recommendation results is compared, and the best performing factor value is selected. Assuming that through experiments, factor 0.9 performs better, the original co-occurrence frequency To normalize the modal dimension, one approach is to: From mode A, features From mode B, the adjusted frequency ,in is the difference factor between mode A and mode B. This step is originally described as "normalizing the co-occurrence frequency to the modal dimension". The more common normalization refers to converting the frequency to a specific range (such as 0 to 1), but combined with the subsequent formula The processing here is more like a weighted adjustment of the original frequency based on the modal type to balance the contribution of different modal information sources. After adjustment, the co-occurrence frequency matrix is ​​established, and the elements of the matrix are Feature Node With feature nodes The adjusted co-occurrence frequency between “fatigue” and “liver density unevenness” .

[0033] S403: Based on the transition probability of the feature node weight set and the modal association strength of the co-occurrence frequency matrix, the formula is used: ; in, represents the PageRank value of the k+1th iteration of the i-th feature node, is the damping coefficient and its value range is 0.85≤d≤0.95, is the total number of feature nodes, is the temporal weight parameter of feature nodes i to j, is the co-occurrence frequency value of feature nodes i and j, is the length of the eigenvector of the mth modal dimension, is the total number of modal dimensions, Used to strengthen the propagation weight of high-frequency co-occurring feature pairs.

[0034] The process of obtaining the transition probability of the feature node weight set and the modal correlation strength of the co-occurrence frequency matrix is ​​as follows: first, the feature node weight set generated in S401 is retrieved, which contains the initial transition probability or weight of each feature node, and the co-occurrence frequency matrix generated in S402 is retrieved, which contains the modal correlation strength between feature nodes, using the formula Calculate the PageRank value of each feature node. The formula parameters are detailed as follows: Representative The feature node is The PageRank value calculated by the iteration is an indicator to measure the importance of the node. is the damping coefficient, and its value range is limited to between 0.85 and 0.95. For example, through experimental comparison, When , the PageRank value of the model basically converges after 100 iterations, and the recommendation accuracy on the test set is 88%; When , 120 iterations are required to converge, and the accuracy is 89%; , 150 iterations are required, and the accuracy is 88.5%. Considering the convergence speed and accuracy, in this embodiment, , is the total number of feature nodes, that is, the number of all independent features (such as symptoms, signs, imaging findings, etc.) involved in the calculation in the network. Assuming that through feature extraction and screening, a total of feature nodes, Feature nodes To feature node The time series weight parameter is obtained from the feature set with time series weight generated in step S303, reflecting the Evolution to characteristics or the temporal strength of the association between the two, for example, if the feature It is a feature If a typical prodromal symptom of Probably higher, assuming (the timing weight from node 2 to node 1), Feature nodes With feature nodes The co-occurrence frequency value of is obtained from the co-occurrence frequency matrix generated in step S402, which represents the adjusted frequency of the two features co-occurring in the multimodal data, for example , It is an exponential operation (1.5 power) on the co-occurrence frequency value. Its purpose is to strengthen the propagation weight of high-frequency co-occurrence feature pairs in PageRank calculation, so that the more co-occurrence feature pairs there are, the greater the influence transmission between them. For the The length of the feature vector of a modality dimension refers to the number of original features that constitute the modality or the dimensionality of its vector representation. For example, the feature vector of the text modality consists of 300 words, then , the image modality obtains a 256-dimensional feature vector through a certain feature extraction method, then , is the total number of modal dimensions, which means that the data processed by the current system contains several different types of information sources. For example, if the system processes two modalities, text and image, then , is the sum of the lengths of the eigenvectors of all modal dimensions. For example, if , (text), (Image), then , The square root of the sum is taken to normalize the overall size of the different modal features so that the PageRank calculation is not overly biased towards systems with a larger total number of feature dimensions. Representative The feature node is The PageRank value calculated at the iteration is In the initial iteration of ,Right now , summation symbol Indicates that for all pointing nodes Node (or all nodes pointed to by node j in standard PageRank, here according to the formula structure, it is more like the contribution from other nodes j to node i) traverse and accumulate, here It should be understood that There is a connection relationship (can be and The size of the node set that has an impact, or the total number of nodes , usually refers to all that can be transferred to of , the operation logic is: the PageRank value of each node consists of two parts, the first part It is a basic value, which indicates the probability of jumping from any node to the current node. The second part is the damping coefficient Multiply all other nodes linked to this node by PageRank value According to the specific weight (i.e. ) is the sum of the values ​​passed through, this particular weight combines the temporal relationship, the co-occurrence strength (and is exponentially enhanced) and the normalization of the multimodal feature scale. The benefit of the formula is that by introducing the temporal weight , so that the PageRank calculation can take into account the temporal dynamics of feature evolution; by co-occurrence frequency conduct The power processing significantly enhances the influence transfer between strongly correlated feature pairs, making it easier for closely related features to obtain high rankings; Normalizing the overall scale of multimodal features avoids the dominance of the sorting results due to the excessive dimension of a certain modal feature, thereby more balanced evaluation of the importance of features from different sources. Now let's demonstrate the example. Suppose that the feature node is calculated In the PageRank value of iteration ,set up , , , assuming node 1 only receives links from nodes 2 and 3, , , for node 2 to node 1: , ,but , the contribution is , for node 3 to node 1: , ,but , the contribution is ,but , the calculation process will continue to iterate until the PageRank values ​​of all nodes converge (that is, the change between two iterations is less than a minimum threshold, e.g. ), the PageRank values ​​of the feature nodes finally obtained will be arranged from high to low according to the values ​​to form an ordered list. The feature nodes with higher PageRank values ​​represent higher comprehensive importance in the current multimodal temporal association network. These feature nodes with high PageRank values ​​and their associations will be used to construct a three-dimensional recommendation decision table. The result shows that feature node 1 obtained a higher PageRank value in this iteration, which means that it has a higher importance in the network. This value will serve as the key basis for measuring the priority of the feature when constructing the three-dimensional recommendation decision table in the future. The value itself directly reflects its importance.

[0035] Table 1 Example table of characteristic node PageRank calculation parameters As shown in Table 1, some parameters and their example values ​​used in calculating the PageRank value of feature node 1 are listed. These parameters together determine the importance ranking of feature nodes in the network.

[0036] The intelligent medical consultation recommendation system is used to implement the above-mentioned intelligent medical consultation recommendation method. The system includes: The multimodal feature extraction module is used to extract organ-related word frequency features from user self-report texts using the TF-IDF algorithm, perform cosine similarity calculations on feature vectors to generate an organ domain matching matrix, filter matrix element thresholds to generate an organ domain core feature set, extract image space coordinate parameters using the DICOM metadata parser to generate an image organ positioning parameter set, and pass the organ domain core feature set and the image organ positioning parameter set to the feature association analysis module; The feature association analysis module is used to construct a feature co-occurrence frequency matrix using the semantic labels of the organ domain core feature set and the spatial coordinates of the imaging organ positioning parameter set using a collaborative filtering algorithm. The threshold comparison is performed on the matrix elements to generate a set of multimodal associated feature pairs. A recommended candidate feature library is constructed based on the feature pair index. The recommended candidate feature library and the multimodal associated feature pair set are then passed to the temporal evolution modeling module. The temporal evolution modeling module is used to calculate the hidden state of symptom duration using the LSTM network forward propagation based on the timestamp sequence of the recommendation candidate feature library, generate a set of symptom evolution path vectors through hidden state transfer, perform cosine similarity calculation on the path vectors to generate a feature set with temporal weights, and pass the feature set with temporal weights to the dynamic recommendation decision module; The dynamic recommendation decision module is used to iteratively calculate the PageRank value of feature nodes using the PageRank algorithm through the weight parameters of the feature set with temporal weights and the co-occurrence frequency of the multimodal associated feature pair set, and generate a three-dimensional recommendation decision table based on the value range sorting.

[0037] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. An intelligent medical consultation recommendation method, characterized in that: The following steps are involved: S1: Extract organ-related word frequency features from user self-report texts using the TF-IDF algorithm, perform cosine similarity calculation on feature vectors to generate an organ domain matching matrix, filter and generate an organ domain core feature set based on the organ domain matching matrix, and extract image space coordinate parameters using the DICOM metadata parser to generate an image organ positioning parameter set; S2: Based on the semantic labels of the core feature set of the organ domain and the spatial coordinates of the imaging organ positioning parameter set, a collaborative filtering algorithm is used to construct a feature co-occurrence frequency matrix, threshold comparison is performed on the elements of the feature co-occurrence frequency matrix to generate a set of multimodal associated feature pairs, and the index is used to construct a recommendation candidate feature library; S3: Based on the timestamp sequence of the recommended candidate feature library, the LSTM network forward propagation is used to calculate the hidden state of the symptom duration, a set of symptom evolution path vectors is generated through the hidden state transfer, and a feature set with time series weights is generated based on the similarity; S4: Based on the weight parameters of the feature set with temporal weights and the co-occurrence frequency of the multimodal associated feature pair set, the PageRank algorithm is used to iteratively calculate the PageRank values ​​of the feature nodes, and the three-dimensional recommendation decision table is generated by sorting.

2. The intelligent medical consultation recommendation method according to claim 1, characterized in that: The organ domain core feature set specifically includes word frequency weights and similarity scores. The organ domain core feature set and the image organ positioning parameter set are collaboratively verified through a spatial-semantic mapping relationship to improve organ positioning accuracy. The LSTM network contains 128 hidden units, and the input gate and forget gate activation functions adopt the sigmoid function.

3. The intelligent medical consultation recommendation method according to claim 2, characterized in that: The image organ positioning parameter set includes spatial coordinates and DICOM tags, the multimodal association feature pair set specifically includes co-occurrence frequency values ​​and modality type pairs, the recommended candidate feature library includes feature indexes and association strengths, the symptom evolution path vector set includes timestamp sequences and hidden state vectors, the feature set with temporal weights specifically includes weight values ​​and temporal correlation, and the three-dimensional recommendation decision table includes PageRank values, organ domain labels, temporal weights, and modality correlations.

4. The intelligent medical consultation recommendation method according to claim 3, characterized in that: The steps of S1 are specifically as follows: S101: Obtain user self-description text, calculate the inverse document frequency weight of organ-related vocabulary using the TF-IDF algorithm, count the frequency distribution of multiple organ terms in the text, multiply the frequency value by the weight value, and generate an organ-related word frequency vector; S102: Based on the organ-related word frequency vectors, a multidimensional feature space coordinate system is constructed. The cosine value of the angle between the vectors is calculated using a cosine similarity function. The calculation results are arranged in a matrix by organ category to generate an organ domain similarity matrix. The interquartile range method is used to set a similarity threshold. Feature dimensions above the threshold in the matrix are screened to form an organ domain core feature set. The multidimensional feature space coordinate system uses the Z-score normalization method to perform dimensionless processing on the organ-related word frequency vector; S103: Call the spatial positioning tag in the DICOM metadata, parse the image layer thickness, pixel spacing, and slice position parameters, convert the spatial coordinates in the Cartesian coordinate system into voxel index values, group and integrate the parameters according to the scan sequence number, and generate an image organ positioning parameter set.

5. The intelligent medical consultation recommendation method according to claim 4, characterized in that: The steps of S2 are specifically as follows: S201: calling the semantic labels of the core feature set of the organ domain and the spatial coordinates of the imaging organ positioning parameter set, using a collaborative filtering algorithm to count the number of co-occurrences of feature pairs in the spatial dimension, converting the mapping relationship between the semantic labels and the spatial coordinates into a two-dimensional frequency distribution table, and generating a feature co-occurrence frequency matrix; S202: Based on the feature co-occurrence frequency matrix, calculating the Pearson correlation coefficient between the matrix row vector and the column vector, setting a dynamic threshold according to the correlation coefficient distribution, screening elements in the matrix with correlation coefficients higher than the threshold, and generating a set of multimodal correlation feature pairs; S203: According to the feature combination pattern in the multimodal associated feature pair set, an index mapping table of feature codes and spatial coordinates is established, high-frequency associated feature pairs are grouped and stored according to organ categories, and a recommended candidate feature library is generated.

6. The intelligent medical consultation recommendation method according to claim 5, characterized in that: The steps for obtaining the feature set with temporal weights are as follows: S301: Calling the timestamp sequence of the recommended candidate feature library, using the LSTM network input gate to calculate the time step weight, controlling the historical state retention ratio through the forget gate, inputting the candidate features into the network in chronological order for forward propagation, and generating the symptom duration hidden state; S302: Based on the hidden state of the symptom duration, a state transition probability matrix is ​​constructed, the transition path between states is calculated using a hidden Markov model, the hidden states of consecutive time steps are connected to form a state transition chain, and a set of symptom evolution path vectors is generated; S303: Calculate the cosine similarity between the path vectors based on the symptom evolution path vector set, assign a temporal weight coefficient according to a time decay function, perform weighted fusion on the similarity value and the weight coefficient, and generate a feature set with temporal weights.

7. The intelligent medical consultation recommendation method according to claim 6, characterized in that: The steps for obtaining the three-dimensional recommendation decision table are specifically as follows: S401: calling the weight parameters of the temporal weight feature set, obtaining the time decay weight value between feature nodes, combining the number of outbound links of the feature nodes, calculating the initial transition probability of each feature node, and generating a feature node weight set; S402: Based on the multimodal associated feature pair set, counting the co-occurrence frequencies between feature nodes, introducing a modality type difference factor, performing modality dimension normalization processing on the co-occurrence frequencies, and establishing a co-occurrence frequency matrix; S403: Based on the transition probability of the feature node weight set and the modal association strength of the co-occurrence frequency matrix, the formula is adopted: ； in, represents the PageRank value of the k+1th iteration of the i-th feature node, is the damping coefficient and its value range is 0.85≤d≤0.95, is the total number of feature nodes, is the temporal weight parameter of feature nodes i to j, is the co-occurrence frequency value of feature nodes i and j, is the length of the eigenvector of the mth modal dimension, is the total number of modal dimensions, Used to strengthen the propagation weight of high-frequency co-occurring feature pairs.

8. An intelligent medical consultation recommendation system, characterized in that: The system is used to implement the intelligent medical consultation recommendation method according to any one of claims 1 to 7, and the system includes: A multimodal feature extraction module is used to extract organ-related word frequency features from user self-report text using the TF-IDF algorithm, perform cosine similarity calculation on feature vectors to generate an organ domain matching matrix, screen and generate an organ domain core feature set based on the matrix element threshold, extract image space coordinate parameters using a DICOM metadata parser to generate an image organ positioning parameter set, and pass the organ domain core feature set and the image organ positioning parameter set to a feature association analysis module; a feature association analysis module, configured to construct a feature co-occurrence frequency matrix using a collaborative filtering algorithm based on the semantic labels of the organ domain core feature set and the spatial coordinates of the imaging organ positioning parameter set, perform threshold comparison on the matrix elements to generate a multimodal associated feature pair set, construct a recommended candidate feature library based on the feature pair index, and pass the recommended candidate feature library and the multimodal associated feature pair set to a temporal evolution modeling module; A temporal evolution modeling module is used to calculate the hidden state of symptom duration using the LSTM network forward propagation based on the timestamp sequence of the recommendation candidate feature library, generate the symptom evolution path vector set through the hidden state transfer, perform cosine similarity calculation on the path vector to generate a feature set with temporal weights, and pass the feature set with temporal weights to the dynamic recommendation decision module; The dynamic recommendation decision module is used to iteratively calculate the PageRank value of the feature node using the PageRank algorithm through the weight parameters of the feature set with temporal weights and the co-occurrence frequency of the multimodal associated feature pair set, and generate a three-dimensional recommendation decision table according to the value range sorting.

Images