Patient desensitization image-driven brain treatment resource recommendation method and system

Abstract

The application discloses a patient desensitization portrait-driven brain treatment resource recommendation method and system, adopts a two-level desensitization mechanism and a privacy enhancement method to perform desensitization processing on data, and completes patient privacy protection; through a medical Transformer model and a bidirectional LSTM network, decoupling of patient conditions and preference characteristics is realized, a cross-modal fusion method is used for semantic alignment to generate a patient personalized portrait representation vector; a brain diagnosis and treatment heterogeneous knowledge graph is constructed, an RGCN and a KGAT two-level graph neural network are used to model all medical institution diagnosis and treatment resource representation vectors; a double-tower deep matching model is adopted to calculate diagnosis and treatment adaptation degrees of patients to diagnosis and treatment resources and diagnosis and treatment resource load adaptation degrees of medical institutions to patients based on the patient personalized portrait representation vector and the diagnosis and treatment resource representation vector, and then an adaptive inertia weight discrete particle swarm optimization algorithm is used to generate a brain diagnosis and treatment resource recommendation list based on a Pareto thought.

Description

Technical Field

[0001] This invention relates to the field of medical recommendation, specifically to a method and system for recommending brain therapy resources driven by patient desensitization profiles. Background Technology

[0002] With the incidence of brain diseases rising year by year, the mismatch between supply and demand and uneven regional distribution of brain diagnosis and treatment resources are becoming increasingly prominent. Existing brain diagnosis and treatment resource matching schemes often focus on a single optimization goal of treatment suitability, making it difficult to balance the conflicts between multiple objectives such as individualized patient needs and balanced medical resource load. This easily leads to local optima, resulting in insufficient matching accuracy and unbalanced resource allocation. Furthermore, cross-institutional collaboration of diagnosis and treatment data requires the aggregation of patients' original sensitive information, posing a serious risk of privacy leaks.

[0003] Therefore, it is essential to design a brain treatment resource recommendation method and system driven by patient desensitization profiles to improve the accuracy and scientific nature of brain diagnosis and treatment resource recommendations while protecting patient privacy, optimize patients' medical experience for brain diseases, and achieve efficient allocation of high-quality brain medical resources. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by proposing a method and system for recommending brain treatment resources driven by patient desensitization profiles. It employs a two-stage desensitization mechanism and privacy enhancement methods to desensitize data, ensuring patient privacy. Simultaneously, it models both patients and treatment resources and uses a dual-tower deep matching model for bidirectional adaptation verification. Finally, it utilizes an adaptive inertial weighted discrete particle swarm optimization algorithm based on Pareto principles to generate a recommended list of brain treatment resources, thereby improving the accuracy and scientific rigor of the recommendations.

[0005] The technical solution to achieve the objective of this invention is as follows: A patient desensitization profiling-driven approach to recommending brain therapy resources includes the following steps: For each medical institution, medical text data and patient attribute data are collected, and statistical desensitization methods are used to desensitize the patient attribute data to obtain desensitized patient attribute data. The BERT model is used to extract the relationships between medical entities from the medical text data and output medical semantic embedding vectors. The desensitized patient attribute data is cleaned by standardization algorithms and anomaly handling. Subsequently, privacy enhancement is performed to obtain enhanced medical semantic embedding vectors and enhanced patient attribute data. A multimodal fusion-based personalized patient profile representation vector is constructed based on enhanced medical semantic embedding vectors and enhanced patient attribute data. The enhanced medical semantic embedding vectors are then fed into a medical Transformer model to capture long-distance dependencies between different medical entities, resulting in a patient condition representation vector. A dual-branch structure is adopted, using an embedding layer and a bidirectional LSTM layer to extract a personalized patient preference representation vector from the enhanced patient attribute data and diagnostic text data. The patient condition representation vector and the personalized patient preference representation vector are then input into a cross-modal fusion Transformer, and a cross-attention mechanism is used to obtain the personalized patient profile representation vector. Based on the medical entities and relationships, a heterogeneous knowledge graph for brain diagnosis and treatment is constructed. RGCN is used to process the various types of nodes and relationships in the heterogeneous knowledge graph for brain diagnosis and treatment, generating an initial aggregated feature vector for the nodes. Subsequently, KGAT is used to introduce an attention mechanism, and based on the initial aggregated feature vector, the importance of different neighbor nodes in the heterogeneous knowledge graph for the target node is adaptively learned to obtain a representation vector of diagnosis and treatment resources. A dual-tower deep matching model is constructed, in which the patient's personalized profile representation vector is input into the patient tower sub-network, and the treatment resource representation vector is input into the resource tower sub-network. The treatment fit and treatment resource load fit are calculated. Then, an adaptive inertial weighted discrete particle swarm optimization algorithm is used, with the treatment fit objective function and the treatment resource load fit objective function as the fitness function, to generate a brain treatment resource recommendation list based on the Pareto principle.

[0006] Furthermore, for each medical institution, medical records and patient attribute data are collected with the patient's informed consent. The medical records include the patient's electronic medical records, outpatient records, brain imaging reports, inpatient records, consultation records, and examination records. The patient attribute data includes the patient's identification, age, gender, geographical location, and year of treatment. A two-stage desensitization mechanism is used to process patient attribute data. For patient identifiers that can directly identify patient information, an irreversible desensitization process is performed using a hash algorithm, adding a random salt value to the patient identifier before hashing. For basic patient information data such as age, gender, geographical location, and year of treatment that can be indirectly combined to identify patient information, generalization processing is performed so that the equivalence class after generalization contains no fewer than 10 different patient records, that is, any combination of basic patient information data corresponds to at least 10 different patients. After the above two-stage desensitization mechanism, desensitized patient attribute data is obtained. Data preprocessing was performed on the medical text data and the anonymized patient attribute data. For the medical text data, a pre-trained BERT model was used. First, the medical text data was segmented into sentences and words to generate a token sequence. The token sequence was then input into the BERT model to encode the text semantics through a bidirectional Transformer layer, outputting a medical semantic embedding vector for the entire token sequence. This medical semantic embedding vector was then input into a NER decoder to output the medical entity label corresponding to each token. Subsequently, the medical semantic embedding vector and the entity labels output by the NER decoder were input into a RE decoder to extract the relationships between medical entities. For the anonymized patient attribute data, a... The guidelines perform outlier detection and removal on the desensitized patient attribute data, fill out the outlier data with the median, and perform Z-score standardization on the data after outlier removal. For the preprocessed desensitized patient attribute data, controllable noise following a Laplace distribution is added to perform differential privacy enhancement to obtain enhanced patient attribute data. For the medical semantic embedding vector, small noise following a normal distribution is added, and privacy enhancement is performed through small perturbations in the feature space to obtain enhanced medical semantic embedding vector.

[0007] Furthermore, constructing a multimodal fusion-based personalized patient profile representation vector includes the following steps: A three-layer MLP is used to perform nonlinear mapping on the enhanced patient attribute data to obtain auxiliary attribute features. The enhanced medical semantic embedding vector and auxiliary attribute features are input into the medical Transformer model. A bidirectional Transformer structure is adopted, and an attention bias matrix is ​​introduced to output the patient condition representation vector. A dual-branch structure is adopted, and patient static preference feature vector and patient temporal feature vector are extracted from enhanced patient attribute data and diagnosis and treatment text data respectively through Embedding layer and bidirectional LSTM layer. The two feature vectors are concatenated and fused through fully connected layer to obtain patient personalized preference representation vector. The patient's condition representation vector and the patient's personalized preference representation vector are input into the cross-modal fusion Transformer, and the patient's personalized profile representation vector is obtained through intra-modal self-attention encoding and cross-modal cross-attention fusion. Furthermore, a 3-layer MLP is used to perform nonlinear mapping on the enhanced patient attribute data to obtain auxiliary attribute features. Then, a sine and cosine position coding method is used to apply a sine function to even positions in the token sequence and a cosine function to odd positions to obtain the position code of the token sequence corresponding to the enhanced medical semantic embedding vector. The enhanced medical semantic embedding vector, position code, and auxiliary attribute features are concatenated and fused to obtain the encoder input. The encoder layer consists of eight bidirectional Transformer layers, each including a multi-head self-attention layer and an FFN layer. An attention bias matrix based on the relationships between medical entities is introduced into the multi-head self-attention layer to obtain its output. The output of the self-attention layer is then input into the FFN layer, passing through two fully connected layers and a GELU activation function to obtain its output. These single-layer Transformer layers are stacked eight times, with the input of each layer being the output of the previous layer, resulting in the global output of the encoder layer. The global output of the encoder layer is then input into the prediction layer for patient condition prediction, serving as the patient condition representation vector. Furthermore, enhanced patient attribute data is input into the Embedding layer to output a static patient preference feature vector; the patient's historical medical records are compiled from the medical text data and input into a bidirectional LSTM layer to obtain a patient temporal feature vector; then the patient's static preference feature vector and the patient's temporal feature vector are fused to obtain a patient's personalized preference representation vector. Furthermore, a cross-modal fusion Transformer framework is constructed. The patient's condition representation vector and the patient's personalized preference representation vector are respectively mapped through linear layers to obtain condition feature maps and preference feature maps. For intramodal self-attention, the condition feature maps and preference feature maps are concatenated and then self-attention encoded to obtain the intramodal encoded output. For cross-modal cross-attention, it includes two sub-modules: condition cross-attention and preference cross-attention, which respectively obtain enhanced condition features and enhanced preference features. Then, the enhanced condition features and enhanced preference features are concatenated and then residually connected with the intramodal encoded output to obtain the cross-modal fusion output. The intramodal encoded output and the cross-modal fusion output are concatenated, and nonlinear transformation and normalization are performed to obtain the final output patient personalized profile representation vector.

[0008] Furthermore, obtaining the representation vector of the medical resources corresponding to the medical institution includes the following steps: Based on the relationships between the medical entities, a heterogeneous knowledge graph for brain diagnosis and treatment is constructed, with medical entities as nodes and the relationships between medical entities as edges. The node features of each node are the semantic vectors of the corresponding medical entity token. The relational graph convolutional network RGCN is adopted to complete the aggregation and updating of node features in the heterogeneous knowledge graph of brain diagnosis and treatment through message passing mechanism. After multiple layers of message passing, the initial aggregated feature vector of each node in the heterogeneous knowledge graph of brain diagnosis and treatment is output. The KGAT knowledge graph attention network is adopted, and an attention mechanism is introduced. Based on the initial aggregated feature vector of the node, the importance of different neighbor nodes in the heterogeneous knowledge graph of brain diagnosis and treatment is adaptively learned to obtain the representation vector of diagnosis and treatment resources. Furthermore, for the initial aggregated feature vector of nodes obtained through RGCN, the Knowledge Graph Attention Network (KGAT) is adopted to introduce an attention mechanism and adaptively learn the normalized attention weights of different neighbor nodes on the target node. Based on the normalized attention weights, the feature of the target node is updated. After updating the feature of all nodes in the heterogeneous knowledge graph of brain diagnosis and treatment, the node features are processed by dimensional mapping and normalization through a linear layer to obtain the final diagnostic and treatment resource representation vector.

[0009] Furthermore, a brain diagnostic and treatment resource recommendation list is generated based on the patient's personalized profile representation vector and the diagnostic and treatment resource representation vector, including the following steps: The aforementioned method is used to obtain the treatment resource representation vectors for each medical institution and the personalized patient profile representation vectors for all patients. A dual-tower deep matching model is then constructed. The personalized patient profile representation vector is input into the patient tower sub-network, outputting the patient profile encoded feature vector. The treatment resource representation vector is input into the resource tower sub-network, outputting the treatment resource encoded feature vector. The cosine similarity between the patient profile encoded feature vector and the treatment resource encoded feature vector is calculated as the treatment fit between the medical institution's treatment resources and the patient's treatment needs. Simultaneously, the treatment resource load fit after the medical institution receives the patient is calculated. After calculating the matching degree between patients' treatment needs and the treatment resources of various medical institutions through the dual-tower deep matching model, considering the problem of treatment resource load in each medical institution, in order to achieve the global optimal balance between patients' personalized treatment needs and the treatment resource load of medical institutions, an adaptive inertial weighted discrete particle swarm optimization algorithm was designed. The fitness function is based on the treatment matching degree objective function and the treatment resource load matching degree objective function. Based on the Pareto idea, a brain treatment resource recommendation list is generated. Furthermore, assuming there are N medical institutions in total, in the adaptive inertia weighted discrete particle swarm optimization algorithm, each particle corresponds to a sorting scheme, i.e., an N-dimensional integer permutation. The fitness functions are the objective functions for diagnosis and treatment fit and the objective function for diagnosis and treatment resource load fit. The optimization steps of the adaptive inertia weighted discrete particle swarm optimization algorithm are as follows: First, initial particles are generated by randomly generating P initial particles that conform to the aforementioned encoding rules. Each particle corresponds to a medical institution ranking scheme. For the i-th particle... There are particles, and their corresponding initial particle position vectors are: At the same time The initial particle velocity vector corresponding to each particle is empty. Subsequently, regarding the first Each particle has its initial position set to its initial optimal position. Then calculate the first one respectively. The objective function value of the diagnostic and therapeutic fit for each particle Objective function value of adaptation to medical resource load All non-dominated solutions are selected from all initial particles and added to the Pareto optimal solution external archive. ; After completing the particle swarm initialization, generate two random numbers between 0 and 1 that satisfy a uniform distribution. Then update the particle's velocity vector. ; after obtaining the updated velocity vector Based on this, add the particle's current position vector Obtain the updated particle position vector For the updated particle position vector Then calculate the corresponding two fitness function values ​​and set the particle's current position. Best position of historical individuals According to the Pareto dominance rule, if Pareto Domination Then update ,like Pareto Domination Then keep After updating the individual optimal positions of all particles according to the above steps, the external archive of the current position vectors of all particles and the Pareto optimal solutions from the previous round is used. The external archive set of all non-dominated connections is selected as the Pareto optimal solution for this round of optimization. Subsequently, a roulette wheel method was used to extract from the external archive of Pareto optimal solutions. Choose one solution to update the global optimal position. The maximum number of iterations has been reached in the current iteration round. Or Pareto optimal solution external archive set Compared to the previous round When the solution is completely consistent and remains unchanged for 10 consecutive iterations, the iteration ends, and the final Pareto optimal solution is output as an external archive set.

[0010] Based on the external archive set of the final Pareto optimal solution For each non-dominated solution, a decision matrix is ​​constructed based on the fitness function value, and the final output list of recommended brain diagnostic and treatment resources is determined based on the decision matrix.

[0011] Furthermore, this invention proposes a brain therapy resource recommendation system driven by patient desensitization profiles, including a data desensitization module, a profile construction module, a resource representation module, and a recommendation decision module; The data anonymization module collects medical text data and patient attribute data for each medical institution, and uses statistical anonymization methods to anonymize the patient attribute data to obtain anonymized patient attribute data. It then uses the BERT model to extract the relationships between medical entities from the medical text data and outputs medical semantic embedding vectors. The anonymized patient attribute data is cleaned using standardization algorithms and anomaly handling. Finally, privacy enhancement is performed to obtain enhanced medical semantic embedding vectors and enhanced patient attribute data. The patient profile construction module constructs a multimodal fusion personalized patient profile representation vector based on enhanced medical semantic embedding vectors and enhanced patient attribute data. It then inputs the enhanced medical semantic embedding vectors into a medical Transformer model to capture long-distance dependencies between different medical entities, obtaining a patient condition representation vector. A dual-branch structure is employed, using an embedding layer and a bidirectional LSTM layer to extract patient personalized preference representation vectors from enhanced patient attribute data and diagnostic text data. Finally, the patient condition representation vector and the patient personalized preference representation vector are input into a cross-modal fusion Transformer, utilizing a cross-attention mechanism to obtain the patient personalized profile representation vector. The resource representation module constructs a heterogeneous knowledge graph for brain diagnosis and treatment based on the medical entities and relationships. It uses RGCN to process the various types of nodes and relationships in the heterogeneous knowledge graph for brain diagnosis and treatment, and generates an initial aggregated feature vector for the nodes. Then, it uses KGAT to introduce an attention mechanism, and adaptively learns the importance of different neighbor nodes in the heterogeneous knowledge graph for the target node based on the initial aggregated feature vector, to obtain the diagnosis and treatment resource representation vector. The recommendation decision module constructs a dual-tower deep matching model, inputting the patient's personalized profile representation vector into the patient tower sub-network and the treatment resource representation vector into the resource tower sub-network to calculate the treatment fit and treatment resource load fit. Subsequently, an adaptive inertial weighted discrete particle swarm optimization algorithm is used, with the treatment fit objective function and the treatment resource load fit objective function as the fitness function, to generate a brain treatment resource recommendation list based on the Pareto principle.

[0012] Compared with the prior art, the beneficial effects of this invention are as follows: A two-stage desensitization mechanism and privacy enhancement methods are employed to desensitize data, ensuring patient privacy. A medical Transformer model and a bidirectional LSTM network are used to decouple patient condition and preference features. A cross-modal fusion method is then used for semantic alignment to generate personalized patient profile representation vectors. Subsequently, a heterogeneous knowledge graph for brain diagnosis and treatment is constructed. A two-level graph neural network (RGCN and KGAT) is used to accurately model the representation vectors of diagnostic and treatment resources for all medical institutions. A dual-tower deep matching model is employed to calculate the patient's suitability for diagnostic and treatment resources and the medical institution's suitability for the patient's resource load, based on the patient's personalized profile representation vector and the diagnostic and treatment resource representation vector. Finally, an adaptive inertial weighted discrete particle swarm optimization algorithm, based on the Pareto principle, is used to generate a recommended list of brain diagnosis and treatment resources. Attached Figure Description

[0013] Figure 1 Flowchart of a brain therapy resource recommendation method driven by patient desensitization profiling; Figure 2 Flowchart for obtaining personalized patient profile representation vectors; Figure 3 Flowchart for obtaining the representation vector of diagnostic and treatment resources; Figure 4 Flowchart for generating a recommended list of brain diagnosis and treatment resources; Figure 5 A module diagram of a brain therapy resource recommendation system driven by patient desensitization profiling. Detailed Implementation

[0014] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0015] like Figure 1 As shown, a specific embodiment of the present invention discloses a method for recommending brain therapy resources driven by patient desensitization profiles, comprising the following steps: For each medical institution, medical text data and patient attribute data are collected, and statistical desensitization methods are used to desensitize the patient attribute data to obtain desensitized patient attribute data. The BERT model is used to extract the relationships between medical entities from the medical text data and output medical semantic embedding vectors. The desensitized patient attribute data is cleaned by standardization algorithms and anomaly handling. Subsequently, privacy enhancement is performed to obtain enhanced medical semantic embedding vectors and enhanced patient attribute data. A multimodal fusion-based personalized patient profile representation vector is constructed based on enhanced medical semantic embedding vectors and enhanced patient attribute data. The enhanced medical semantic embedding vectors are then fed into a medical Transformer model to capture long-distance dependencies between different medical entities, resulting in a patient condition representation vector. A dual-branch structure is adopted, using an embedding layer and a bidirectional LSTM layer to extract a personalized patient preference representation vector from the enhanced patient attribute data and diagnostic text data. The patient condition representation vector and the personalized patient preference representation vector are then input into a cross-modal fusion Transformer, and a cross-attention mechanism is used to obtain the personalized patient profile representation vector. Based on the medical entities and relationships, a heterogeneous knowledge graph for brain diagnosis and treatment is constructed. RGCN is used to process the various types of nodes and relationships in the heterogeneous knowledge graph for brain diagnosis and treatment, generating an initial aggregated feature vector for the nodes. Subsequently, KGAT is used to introduce an attention mechanism, and based on the initial aggregated feature vector, the importance of different neighbor nodes in the heterogeneous knowledge graph for the target node is adaptively learned to obtain a representation vector of diagnosis and treatment resources. A dual-tower deep matching model is constructed, in which the patient's personalized profile representation vector is input into the patient tower sub-network and the treatment resource representation vector is input into the resource tower sub-network to calculate the treatment fit degree and the treatment resource load fit degree. Subsequently, an adaptive inertial weighted discrete particle swarm optimization algorithm is adopted, with the treatment fit degree objective function and the treatment resource load fit degree objective function as the fitness function, and a brain treatment resource recommendation list is generated based on the Pareto principle.

[0016] Furthermore, for each medical institution, medical records and patient attribute data are collected with the patient's informed consent. The medical records include the patient's electronic medical records, outpatient records, brain imaging reports, inpatient records, consultation records, and examination records. The patient attribute data includes the patient's identification, age, gender, geographical location, and year of treatment. To further protect patient privacy, a two-stage desensitization mechanism is employed to process patient attribute data. For patient identity identifiers that directly identify the patient, an irreversible desensitization process is performed using a hash algorithm, adding a random salt value before hashing. For basic patient information such as age, gender, geographical location, and year of treatment that can be indirectly combined to identify the patient, a statistical desensitization method is used. Specifically, the aforementioned basic patient information data is generalized, with age and year of treatment using interval generalization and geographical location using region generalization, ensuring that the equivalence classes after generalization contain no fewer than 10 different patient records. That is, any combination of basic patient information data corresponds to at least 10 different patients. After this two-stage desensitization mechanism, desensitized patient attribute data is obtained. Data preprocessing was performed on medical text data and de-identified patient attribute data. For the medical text data, a pre-trained BERT model was used. Specifically, the medical text data was first segmented into sentences and words to generate a token sequence. This token sequence was then input into the BERT model, where a bidirectional Transformer layer encoded the text semantics, obtaining a semantic vector for each token. This resulted in the output of a medical semantic embedding vector for the entire token sequence. This medical semantic embedding vector was then input into a NER decoder, where a linear layer and a conditional random field structure were used to classify the token sequence, outputting a medical entity label for each token. Subsequently, the medical semantic embedding vector and the entity labels output from the NER decoder were input into a RE decoder, where a fully connected layer extracted the relationships between medical entities. For the de-identified patient attribute data, data cleaning was performed using standardization algorithms and anomaly handling methods. Specifically, a... The guidelines perform outlier detection and removal on the desensitized patient attribute data, and use the median to fill in the outliers to avoid the influence of outliers on the data distribution. Subsequently, the data after outlier removal is standardized by Z-score to map the data to a standard normal distribution with a mean of 0 and a standard deviation of 1 to eliminate the dimensional differences between different data. After data preprocessing, privacy enhancement methods are used to further de-identify the data to improve patient privacy and security. For the preprocessed de-identified patient attribute data, controllable noise following a Laplace distribution is added to perform differential privacy enhancement to obtain enhanced patient attribute data. For the medical semantic embedding vector, small noise following a normal distribution is added, and privacy enhancement is performed through small perturbations in the feature space to obtain enhanced medical semantic embedding vector.

[0017] Furthermore, such as Figure 2 As shown, constructing a multimodal fusion-based personalized patient profile representation vector includes the following steps: A three-layer MLP is used to perform nonlinear mapping on the enhanced patient attribute data to obtain auxiliary attribute features. The enhanced medical semantic embedding vector and auxiliary attribute features are input into the medical Transformer model. A bidirectional Transformer structure is adopted, and long-distance dependencies in the diagnosis and treatment text data are captured through self-attention. Attention bias matrix is ​​introduced at the same time to output the patient condition representation vector. A dual-branch structure is adopted, and patient static preference feature vector and patient temporal feature vector are extracted from enhanced patient attribute data and diagnosis and treatment text data respectively through Embedding layer and bidirectional LSTM layer. The two feature vectors are concatenated and fused through fully connected layer to obtain patient personalized preference representation vector. The patient's condition representation vector and the patient's personalized preference representation vector are input into the cross-modal fusion Transformer, and the patient's personalized profile representation vector is obtained through intra-modal self-attention encoding and cross-modal cross-attention fusion. Furthermore, to compensate for the shortcomings of the medical Transformer model in representing patient attribute data features and to achieve complementarity between text data and attribute data, a 3-layer MLP is used to perform nonlinear mapping on the enhanced patient attribute data to obtain auxiliary attribute features. Then, a sine and cosine positional encoding method is used, applying a sine function to even positions and a cosine function to odd positions in the token sequence to obtain the positional encoding of the token sequence corresponding to the enhanced medical semantic embedding vector. The enhanced medical semantic embedding vector, positional encoding, and auxiliary attribute features are concatenated and fused to obtain the encoder input. The encoder layer consists of eight bidirectional Transformer layers. Each bidirectional Transformer layer includes a multi-head self-attention layer and an FFN layer. The multi-head self-attention layer introduces an attention bias matrix to assign higher initial attention weights to tokens with strong correlations, suppressing interference from irrelevant noise. For example, three independent linear layers generate a first query vector, a first key vector, and a first value vector based on the encoder input. These three vectors are then split into multiple parallel attention heads to obtain the components of the three vectors. Simultaneously, an attention bias matrix is ​​generated based on the relationships between medical entities. The specific values ​​of the matrix elements are as follows: when there is a strong correlation between any two tokens and their corresponding medical entities, the corresponding matrix element is set to 0.5; when there is a weak correlation between any two tokens and their corresponding medical entities, the corresponding matrix element is set to 0.2; when there is a weak correlation between any two tokens and their corresponding medical entities, the corresponding matrix element is set to 0.5. When there is no correlation between them, the corresponding matrix elements are set to 0. Then, for any attention head, the first query vector component is multiplied by the first key vector component and divided by the square root of the attention head dimension. After adding the attention bias matrix, it is normalized using softmax and then multiplied by the first value vector component to obtain the attention output of the attention head. Finally, the attention outputs of each attention head are concatenated and projected through a linear layer to obtain the output of the self-attention layer. The output of the self-attention layer is input into the FFN layer and passed through two fully connected layers and the GELU activation function to obtain the output. The above single-layer Transformer layers are stacked 8 times, with the input of each layer being the output of the previous layer, to obtain the global output of the encoder layer. The global output of the encoder layer is input into the prediction layer to predict the patient's condition. Cross-entropy loss is used to train and optimize the medical Transformer model, and the global output of the encoder layer is used as the patient's condition representation vector. Furthermore, enhanced patient attribute data is input into the Embedding layer, mapping the high-dimensional sparse enhanced patient attribute data into a low-dimensional dense vector. A three-layer MLP network then performs nonlinear transformation and deep feature interaction, outputting a static patient preference feature vector. Based on historical data such as the patient's electronic medical records, outpatient records, and inpatient records from the diagnostic and treatment text data, the patient's historical medical treatment sequence is compiled and arranged chronologically to generate a time-series sequence. This sequence is input into a bidirectional LSTM layer to capture the temporal dependencies and continuous treatment needs of the patient's medical behavior. The outputs of the forward and backward LSTMs are concatenated to obtain the patient's time-series feature vector, reflecting the patterns of the patient's medical behavior. Subsequently, the static patient preference feature vector and the time-series feature vector are fused through a fully connected layer to obtain a personalized patient preference representation vector. The Embedding layer, bidirectional LSTM layer, and fully connected layer feature fusion are all mature deep learning techniques and are not the focus of this application's technical solution, so they will not be elaborated upon further. Furthermore, a cross-modal fusion Transformer framework is constructed to achieve semantic alignment and fusion of multimodal features through a cross-attention mechanism, addressing the semantic gap between different modal features and obtaining personalized patient profile representation vectors. Specifically, the patient's condition representation vector and personalized patient preference representation vector are mapped to a unified 128-dimensional feature space through linear layers to obtain condition feature maps and preference feature maps. For intramodal self-attention, the condition feature maps and preference feature maps are concatenated and then self-attention encoded using a two-layer Transformer to obtain the intramodal encoded output. For cross-modal cross-attention, it includes two sub-modules: condition cross-attention and preference cross-attention, where condition... Cross-attention uses disease feature mapping as the query and preference feature mapping as the key and value. Based on the self-attention mechanism, it calculates the attention weight of disease features on preference features, performs weighted aggregation to obtain enhanced disease features, and uses a similar method to obtain enhanced preference features. Then, it concatenates the enhanced disease features and enhanced preference features and performs residual connection with the intramodal encoding output to obtain cross-modal fusion output. The intramodal encoding output and cross-modal fusion output are concatenated and processed through a linear layer with GELU as the activation function for nonlinear transformation and normalization to obtain the final output patient personalized profile representation vector. Among these, the linear layer mapping, self-attention mechanism, and residual connection are all mature model mechanisms and will not be elaborated on in detail.

[0018] Furthermore, such as Figure 3 As shown, obtaining the representation vector of medical resources corresponding to a medical institution includes the following steps: Based on the relationships between the medical entities, a heterogeneous knowledge graph for brain diagnosis and treatment is constructed, with medical entities as nodes and the relationships between medical entities as edges. The node features of each node are the semantic vectors of the corresponding medical entity token. The relational graph convolutional network RGCN is adopted to complete the aggregation and updating of node features in the heterogeneous knowledge graph of brain diagnosis and treatment through message passing mechanism. After multiple layers of message passing, the initial aggregated feature vector of each node in the heterogeneous knowledge graph of brain diagnosis and treatment is output. The KGAT knowledge graph attention network is adopted, and an attention mechanism is introduced. Based on the initial aggregated feature vector of the node, the importance of different neighbor nodes in the heterogeneous knowledge graph of brain diagnosis and treatment is adaptively learned to obtain the representation vector of diagnosis and treatment resources. Furthermore, the heterogeneous knowledge graph of brain diagnosis and treatment is input into a relational graph convolutional neural network (RGCN). Based on the relationships between edges between nodes, a message passing mechanism is used to aggregate and update node features. For example, for a target node, its set of neighboring nodes is first determined based on the relationships between edges between nodes in the heterogeneous knowledge graph of brain diagnosis and treatment. Then, the message passing mechanism is used to aggregate the node features of each neighboring node in the set of neighboring nodes to obtain the updated target node features. After message passing and feature aggregation through multiple layers of RGCN, an initial aggregated feature vector containing the node's own attribute features and the neighboring node association features is output. The message passing mechanism is a mature technology in graph neural networks and will not be elaborated here. Furthermore, to further improve the accuracy of medical institution resource representation by using the initial aggregated feature vector obtained through RGCN, a knowledge graph attention network (KGAT) is adopted. This network introduces an attention mechanism to adaptively learn the importance of different neighboring nodes to the target node. First, the attention score between nodes is calculated based on the initial aggregated feature vectors of each node in the heterogeneous knowledge graph for brain diagnosis and treatment. Specifically, for the target node and one of its neighboring nodes, the initial aggregated feature vectors of the two nodes are multiplied by a mapping matrix, concatenated, and then multiplied with a randomly initialized vector. The LeakyReLU activation function is then used to obtain the attention score between the target node and its neighboring node. Subsequently, the attention scores of all neighboring nodes of the target node are normalized using the softmax function to obtain normalized attention weights. Based on these normalized attention weights, the features of the target node's neighboring nodes are weighted and aggregated, and then the target node's own features are added to complete the update of the target node's features. After updating the features of all nodes in the heterogeneous knowledge graph for brain diagnosis and treatment, the node features are dimensionally mapped and normalized through a linear layer to ensure they are in the same dimensional space as the patient's personalized profile representation vector, resulting in the final medical resource representation vector.

[0019] Furthermore, such as Figure 4As shown, the process of generating a brain diagnosis and treatment resource recommendation list based on patient personalized profile representation vectors and diagnosis and treatment resource representation vectors includes the following steps: The aforementioned method is used to obtain the treatment resource representation vectors for each medical institution and the personalized patient profile representation vectors for all patients. A dual-tower deep matching model is then constructed. The personalized patient profile representation vectors are input into the patient tower sub-network, undergoing nonlinear transformation and dimensional mapping through three fully connected layers to output the patient profile encoding feature vector. The treatment resource representation vectors are input into the resource tower sub-network, undergoing nonlinear transformation and dimensional mapping through two fully connected layers to output the treatment resource encoding feature vector. The cosine similarity between the patient profile encoding feature vector and the treatment resource encoding feature vector is calculated as the treatment fit between the medical institution's treatment resources and the patient's treatment needs. Simultaneously, the absolute value of the medical institution's load rate after receiving the patient minus its optimal load rate is calculated, and this absolute value is subtracted from 1 to obtain the treatment resource load fit between the medical institution's treatment resources and the patient's treatment needs. The medical institution's load rate is calculated by dividing the current total number of patients by its maximum number of patients it can receive. After calculating the matching degree between patients' treatment needs and the treatment resources of various medical institutions through the dual-tower deep matching model, considering the problem of treatment resource load in each medical institution, in order to achieve the global optimal balance between patients' personalized treatment needs and the treatment resource load of medical institutions, an adaptive inertial weighted discrete particle swarm optimization algorithm was designed. The fitness function is based on the treatment matching degree objective function and the treatment resource load matching degree objective function. Based on the Pareto idea, a brain treatment resource recommendation list is generated. Furthermore, assuming there are N medical institutions in total, in the adaptive inertia weighted discrete particle swarm optimization algorithm, each particle corresponds to a sorting scheme, i.e., an N-dimensional integer permutation. The order of the elements in the arrangement represents the recommendation priority of the medical institutions corresponding to each element. The medical institution ranked first has the highest priority. The fitness functions are the treatment fit objective function and the treatment resource load fit objective function. The treatment fit objective function is calculated using exponentially decaying weights to ensure that the fit of higher-ranked medical resources has a greater impact. The calculation formula is as follows: ,in , representing the cosine similarity between the feature vector of the medical resources encoding of the nth medical institution in the ranking scheme and the feature vector of the patient profile encoding, i.e., the degree of medical resource matching between medical institutions and patient medical needs. The exponential decay coefficient; the objective function for the adaptation of medical resource load. The calculation formula is similar to the calculation formula for the objective function of diagnosis and treatment fit, except that... Replace with a representation of the suitability of medical resource load.

[0020] The optimization steps of the adaptive inertia weighted discrete particle swarm optimization algorithm are as follows: First, initial particles are generated by randomly generating P initial particles that conform to the aforementioned encoding rules. Each particle corresponds to a medical institution ranking scheme. For the i-th particle... There are particles, and their corresponding initial particle position vectors are: At the same time The initial particle velocity vector corresponding to each particle is empty. Subsequently, regarding the first Each particle has its initial position set to its initial optimal position. Then calculate the first one respectively. The objective function value of the diagnostic and therapeutic fit for each particle Objective function value of adaptation to medical resource load All non-dominated solutions are selected from all initial particles and added to the Pareto optimal solution external archive. ; After completing the particle swarm initialization, generate two random numbers between 0 and 1 that satisfy a uniform distribution. Then the particle's velocity vector is updated, calculated using the following formula: ,in, For inertia weights, their update formula is: ,in, This represents the maximum value of the inertia weight. This represents the minimum inertial weight. For a dynamically adjusted term, in a non-dominated solution, The value is 0.1, used to enhance global search capabilities and guide the population to explore sparse regions. For dominated particles, The value is -0.1 to enhance local search capabilities, allowing the search to occur near the optimal solution. Indicates the maximum number of iterations. Represents individual learning factors. Represents social learning factors. This represents the optimal position of a historical individual. This represents the global optimal position; the updated velocity vector is obtained. Based on this, add the particle's current position vector Obtain the updated particle position vector For the updated particle position vector Then calculate the corresponding two fitness function values ​​and set the particle's current position. Best position of historical individuals According to the Pareto dominance rule, if Pareto Domination Then update ,like Pareto Domination Then keep After updating the individual optimal positions of all particles according to the above steps, the external archive of the current position vectors of all particles and the Pareto optimal solutions from the previous round is used. The external archive set of all non-dominated solutions is selected as the Pareto optimal solution for this round of optimization. Subsequently, the roulette wheel method was used to obtain the Pareto optimal solution from the external archive set. Choose one solution to update the global optimal position.

[0021] The maximum number of iterations has been reached in the current iteration round. Or Pareto optimal solution external archive set Compared to the previous round When the solution is completely consistent and remains unchanged for 10 consecutive iterations, the iteration ends, and the final Pareto optimal solution is output as an external archive set.

[0022] Based on the external archive set of the final Pareto optimal solution A decision matrix is ​​constructed based on the fitness function value corresponding to each non-dominated solution, and Z-score standardization is used to eliminate dimensional differences. Then, the maximum value of the objective function value for treatment fit and the objective function value for treatment resource load fit are taken from the decision matrix as the positive ideal solution, and the minimum value as the negative ideal solution. Finally, the Euclidean distance between each non-dominated solution and the positive ideal solution is calculated. Euclidean distance to the negative ideal solution ,based on and The proximity of each nondominated solution is calculated. The solution with the highest degree of similarity is selected as the final output list of recommended brain diagnostic and treatment resources.

[0023] Preferably, in the BERT model, the token sequence has a dimension of 768, the medical semantic embedding vector has a dimension of 768, and the linear layer of the NER decoder has a dimension of 768. 12. The fully connected layer in the RE decoder consists of two layers with dimensions of [dimensions to be filled in]. The Adam optimizer was used for training, with an initial learning rate of 0.0002. The learning rate was adjusted with a decay weight of 0.5 every 10 training epochs, and the maximum number of training epochs was set to 50. Preferably, in the medical Transformer model, the dimensions of the 3-layer MLP are as follows: The obtained auxiliary attribute features have a dimension of 768, the position encoding has a dimension of 768, the multi-head self-attention layer has 12 attention heads, each with a dimension of 64, and the FFN layer includes two fully connected layers with dimensions of [missing data]. The final output patient condition representation vector has a dimension of 128. The optimizer used is Adam, with an initial learning rate of 0.001. The learning rate is adjusted with a decay weight of 0.1 every 20 training epochs. The batch size is set to 64, and the maximum number of training epochs is set to 300. Preferably, in the embedding layer, the dimensions of the three MLP layers are respectively The output patient static preference feature vector has a dimension of 128; the dimension of the bidirectional LSTM layer is... The output patient temporal feature vector has a dimension of 128; the dimension of the fully connected layer is... The final output patient personalized preference representation vector has a dimension of 128. The optimizer used is Adam, with an initial learning rate of 0.001. The learning rate is adjusted with a decay weight of 0.1 every 20 training epochs. The batch size is set to 128, and the maximum number of training epochs is set to 300. Preferably, in the cross-modal fusion Transformer, the dimension of the linear layer is... Both the disease feature mapping and the preference feature mapping have a dimension of 128, and the multi-head self-attention layer in the 2-layer Transformer has a dimension of [missing value]. The attention head has 8 attention heads, each with a dimension of 32. The FFN within it also includes two fully connected layers with dimensions of [dimensions to be filled in]. The intramodal encoding output has a dimension of 256. In the cross-attention of disease symptoms and cross-attention of preferences, the Query, Key, and Value are all 128-dimensional, and the enhanced disease symptoms and enhanced preference features are also 128-dimensional, resulting in a cross-modal fusion output of 256 dimensions. The final output patient personalized profile representation vector is 256-dimensional. The Adam optimizer is used for training, with an initial learning rate of 0.001. The learning rate is adjusted with a decay weight of 0.1 every 20 training epochs. The batch size is set to 64, and the maximum number of training epochs is set to 300. Preferably, RGCN contains 3 layers of graph convolutions with dimensions of [missing information]. The output initial aggregated feature vector has a dimension of 128; the dimension of the linear layer in KGAT is... The output diagnostic resource representation vector has a dimension of 256. The optimizer is Adam, with an initial learning rate of 0.001. The learning rate is adjusted with a decay weight of 0.2 every 30 training epochs. The batch size is set to sampling by node batches, and the maximum number of training epochs is set to 400. Preferably, in the dual-tower deep matching model, the dimensions of the three fully connected layers of the patient tower sub-network are as follows: The output patient profile encoding feature vector has a dimension of 128; the resource pyramid subnetwork has dimensions of 128 after two fully connected layers. The output feature vector of the diagnosis and treatment resource encoding has a dimension of 128. The optimizer is Adam, with an initial learning rate of 0.001. The learning rate is adjusted with a decay weight of 0.1 every 20 training epochs. The batch size is set to 256, and the maximum number of training epochs is set to 300. Preferably, in the adaptive inertia weighted discrete particle swarm optimization algorithm, the relevant parameters are set as follows: individual learning factor The social learning factor is 1.5. The maximum number of iterations is 1.5. It is 500. The maximum value of the inertia weight is set to 0.9. The minimum inertia weight is set to 0.4.

[0024] Furthermore, such as Figure 5 As shown, this invention proposes a brain therapy resource recommendation system driven by patient desensitization profiles, including a data desensitization module, a profile construction module, a resource representation module, and a recommendation decision module; The data anonymization module collects medical text data and patient attribute data for each medical institution, and uses statistical anonymization methods to anonymize the patient attribute data to obtain anonymized patient attribute data. It then uses the BERT model to extract the relationships between medical entities from the medical text data and outputs medical semantic embedding vectors. The anonymized patient attribute data is cleaned using standardization algorithms and anomaly handling. Finally, privacy enhancement is performed to obtain enhanced medical semantic embedding vectors and enhanced patient attribute data. The patient profile construction module constructs a multimodal fusion personalized patient profile representation vector based on enhanced medical semantic embedding vectors and enhanced patient attribute data. It then inputs the enhanced medical semantic embedding vectors into a medical Transformer model to capture long-distance dependencies between different medical entities, obtaining a patient condition representation vector. A dual-branch structure is employed, using an embedding layer and a bidirectional LSTM layer to extract patient personalized preference representation vectors from enhanced patient attribute data and diagnostic text data. Finally, the patient condition representation vector and the patient personalized preference representation vector are input into a cross-modal fusion Transformer, utilizing a cross-attention mechanism to obtain the patient personalized profile representation vector. The resource representation module constructs a heterogeneous knowledge graph for brain diagnosis and treatment based on the medical entities and relationships. It uses RGCN to process the various types of nodes and relationships in the heterogeneous knowledge graph for brain diagnosis and treatment, and generates an initial aggregated feature vector for the nodes. Then, it uses KGAT to introduce an attention mechanism, and adaptively learns the importance of different neighbor nodes in the heterogeneous knowledge graph for the target node based on the initial aggregated feature vector, to obtain the diagnosis and treatment resource representation vector. The recommendation decision module constructs a dual-tower deep matching model, inputting the patient's personalized profile representation vector into the patient tower sub-network and the treatment resource representation vector into the resource tower sub-network to calculate the treatment fit and treatment resource load fit. Subsequently, an adaptive inertial weighted discrete particle swarm optimization algorithm is used, with the treatment fit objective function and the treatment resource load fit objective function as the fitness function, to generate a brain treatment resource recommendation list based on the Pareto principle.

[0025] Furthermore, the data anonymization module collects, for each medical institution, medical text data and patient attribute data that have been individually informed and authorized by the patient; among them, medical text data includes the patient's electronic medical record, outpatient record, brain imaging report, inpatient record, consultation record, and examination record; patient attribute data includes the patient's identity identifier, age, gender, geographical location, and year of treatment; For patient identity identifiers that can directly determine patient identity information, an irreversible desensitization process is performed using a hash algorithm. A random salt value is added to the patient identity identifier before hashing. For patient basic information data such as age, gender, geographical location, and year of treatment that can be indirectly combined to determine patient identity information, interval generalization is used for age and year of treatment, and regional generalization is used for geographical location. After the above two-level desensitization mechanism, desensitized patient attribute data is obtained. For medical text data, a pre-trained BERT model is used for processing. First, the medical text data is segmented into sentences and words to generate a token sequence. This token sequence is then input into the BERT model, where a bidirectional Transformer layer encodes the text semantics, obtaining a semantic vector for each token. This outputs a medical semantic embedding vector for the entire token sequence. This medical semantic embedding vector is then input into a NER decoder, which outputs a medical entity label for each token. Finally, the medical semantic embedding vector and the entity labels output from the NER decoder are input into a RE decoder, where a fully connected layer extracts the relationships between medical entities. For anonymized patient attribute data, a... The guidelines perform outlier detection and removal on the attribute data of desensitized patients, and use the median to fill out the outlier data. The data after outlier removal is standardized by Z-score to map the data to a standard normal distribution with a mean of 0 and a standard deviation of 1 in order to eliminate the dimensional differences between different data. For the preprocessed desensitized patient attribute data, controllable noise following a Laplace distribution is added to perform differential privacy enhancement to obtain enhanced patient attribute data. For the medical semantic embedding vector, small noise following a normal distribution is added, and privacy enhancement is performed through small perturbations in the feature space to obtain enhanced medical semantic embedding vector.

[0026] Furthermore, the profile construction module uses a 3-layer MLP to perform nonlinear mapping on the enhanced patient attribute data to obtain auxiliary attribute features. The enhanced medical semantic embedding vector and auxiliary attribute features are input into the medical Transformer model. A bidirectional Transformer structure is adopted, and long-distance dependencies in the diagnosis and treatment text data are captured through self-attention. At the same time, an attention bias matrix is ​​introduced to output the patient condition representation vector. A dual-branch structure is adopted, and patient static preference feature vector and patient temporal feature vector are extracted from enhanced patient attribute data and diagnosis and treatment text data respectively through Embedding layer and bidirectional LSTM layer. The two feature vectors are concatenated and fused through fully connected layer to obtain patient personalized preference representation vector. The patient's condition representation vector and the patient's personalized preference representation vector are input into the cross-modal fusion Transformer, and the patient's personalized profile representation vector is obtained through intra-modal self-attention encoding and cross-modal cross-attention fusion.

[0027] Furthermore, the resource representation module constructs a heterogeneous knowledge graph for brain diagnosis and treatment based on the relationships between the medical entities, with medical entities as nodes and the relationships between medical entities as edges. The node features of each node are the semantic vectors of the corresponding medical entity token. The relational graph convolutional network RGCN is adopted to complete the aggregation and updating of node features in the heterogeneous knowledge graph of brain diagnosis and treatment through message passing mechanism. After multiple layers of message passing, the initial aggregated feature vector of each node in the heterogeneous knowledge graph of brain diagnosis and treatment is output. We employ the Knowledge Graph Attention Network (KGAT) and introduce an attention mechanism. Based on the initial aggregated feature vectors of nodes, we adaptively learn the importance of different neighboring nodes in the heterogeneous knowledge graph of brain diagnosis and treatment, thereby obtaining a representation vector of diagnosis and treatment resources.

[0028] Furthermore, the recommendation decision module obtains the treatment resource representation vectors corresponding to each medical institution and the personalized patient profile representation vectors corresponding to all patients through the resource representation module and the profile construction module, respectively. Then, a dual-tower deep matching model is constructed. The personalized patient profile representation vector is input into the patient tower sub-network, and after three fully connected layers, it undergoes nonlinear transformation and dimensional mapping to output the patient profile encoding feature vector. The treatment resource representation vector is input into the resource tower sub-network, and after two fully connected layers, it undergoes nonlinear transformation and dimensional mapping to output the treatment resource encoding feature vector. The cosine similarity between the patient profile encoding feature vector and the treatment resource encoding feature vector is calculated as the treatment fit between the medical institution's treatment resources and the patient's treatment needs. At the same time, the absolute value of the medical institution's load rate after receiving the patient minus the medical institution's optimal load rate is calculated, and 1 is subtracted from this absolute value to obtain the treatment resource load fit between the medical institution's treatment resources and the patient's treatment needs. The medical institution's load rate is calculated by dividing the current total number of patients by its maximum number of patients it can receive. After calculating the matching degree between patient treatment needs and the treatment resources of various medical institutions using a dual-tower deep matching model, and considering the resource load problem of each medical institution, an adaptive inertial weighted discrete particle swarm optimization algorithm was designed to achieve a global optimal balance between the personalized treatment needs of patients and the resource load of medical institutions. The algorithm uses the treatment matching degree objective function and the treatment resource load matching degree objective function as fitness functions, and generates a brain treatment resource recommendation list based on the Pareto principle.

[0029] This invention discloses a method and system for recommending brain treatment resources driven by patient desensitization profiles. It employs a two-stage desensitization mechanism and privacy enhancement methods to desensitize data, ensuring patient privacy. Through a medical Transformer model and a bidirectional LSTM network, it decouples patient condition and preference features. A cross-modal fusion method is used for semantic alignment to generate personalized patient profile representation vectors. Subsequently, a heterogeneous knowledge graph for brain diagnosis and treatment is constructed. Using a two-level graph neural network (RGCN and KGAT), it accurately models the representation vectors of treatment resources for all medical institutions. A dual-tower deep matching model is used to calculate the patient's suitability for treatment resources and the medical institution's suitability for the patient's treatment resource load, based on the patient's personalized profile representation vector and treatment resource representation vector. Finally, an adaptive inertial weighted discrete particle swarm optimization algorithm, based on Pareto principles, is used to generate a recommended list of brain treatment resources.

[0030] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for recommending brain therapy resources driven by patient desensitization profiling, characterized in that, Includes the following steps: Collect medical text data and patient attribute data, and de-identify them to obtain de-identified patient attribute data; use the BERT model to extract the relationships between medical entities from the medical text data and output medical semantic embedding vectors. Privacy enhancements yield enhanced medical semantic embedding vectors and enhanced patient attribute data; The enhanced medical semantic embedding vector is input into the medical Transformer model to generate the patient condition representation vector; the patient personalized preference representation vector is extracted from the enhanced patient attribute data and diagnosis and treatment text data through the Embedding layer and bidirectional LSTM layer; The patient's condition representation vector and the patient's personalized preference representation vector are input into the cross-modal fusion Transformer model, and the patient's personalized profile representation vector is obtained by using the cross attention mechanism. Construct a heterogeneous knowledge graph for brain diagnosis and treatment, and input it into RGCN to generate the initial aggregated feature vector of the nodes; KGAT is used to introduce an attention mechanism to learn the importance of different neighbor nodes based on the initial aggregated feature vector, and obtain the medical resource representation vector. A dual-tower deep matching model is constructed to calculate the treatment fit and treatment resource load fit. An adaptive inertial weighted discrete particle swarm optimization algorithm is adopted, with the treatment fit objective function and the treatment resource load fit objective function as fitness functions, and a brain treatment resource recommendation list is generated based on the Pareto principle.

2. The method for recommending brain therapy resources driven by patient desensitization profiling as described in claim 1, characterized in that, The method for obtaining the patient's condition representation vector includes the following steps: The enhanced patient attribute data is processed by nonlinear mapping to obtain auxiliary attribute features. The position code is obtained by using sine and cosine coding algorithm based on the token sequence corresponding to the enhanced medical semantic embedding vector. The enhanced medical semantic embedding vector, the position code and the auxiliary attribute features are concatenated to obtain the encoder input. The encoder input is processed by the encoder layer of the medical Transformer model. The encoder layer is stacked with multiple bidirectional Transformer layers, which include a multi-head self-attention layer and a feedforward network. An attention bias matrix is ​​introduced in the multi-head self-attention layer based on the relationship between the medical entities. The output of the multi-head self-attention layer is processed by the feedforward network. The output of the previous bidirectional Transformer layer is used as the input of the next bidirectional Transformer layer to obtain the global output of the encoder layer as the patient's condition representation vector.

3. The method for recommending brain therapy resources driven by patient desensitization profiling as described in claim 2, characterized in that, The method for obtaining the patient's personalized preference representation vector includes the following steps: The enhanced patient attribute data is input into the Embedding layer, and the high-dimensional sparse enhanced patient attribute data is mapped into a low-dimensional dense vector through a 3-layer MLP network, outputting a patient static preference feature vector. The patient's historical medical treatment sequence is obtained by organizing the medical treatment text data and arranging it in chronological order to generate a medical treatment time sequence. This sequence is then input into a bidirectional LSTM layer, and the outputs of the forward LSTM and backward LSTM sides are concatenated to obtain the patient's time sequence feature vector. The patient's static preference feature vector and the patient's temporal feature vector are fused through a fully connected layer to obtain the patient's personalized preference representation vector.

4. The method for recommending brain therapy resources driven by patient desensitization profiling as described in claim 3, characterized in that, The method for obtaining the patient's personalized profile representation vector includes the following steps: A cross-modal fusion Transformer framework was built, and the patient's condition representation vector and the patient's personalized preference representation vector were respectively mapped by linear layer mapping to obtain the condition feature mapping and the preference feature mapping. Intramodal self-attention is employed. After concatenating the disease feature map and the preference feature map, a two-layer Transformer is used for self-attention encoding to obtain the intramodal encoded output. Cross-modal cross-attention is employed, with disease feature mapping as the query and preference feature mapping as the key and value. Enhanced disease features are obtained based on a self-attention mechanism. Enhanced preference features are obtained by using feature preference mapping as the query and disease feature mapping as the key and value, based on a self-attention mechanism. Subsequently, the enhanced disease features and enhanced preference features are concatenated and then residually connected with the intramodal encoding output to obtain the cross-modal fusion output. The intramodal encoded output and the cross-modal fusion output are concatenated, and then nonlinear transformation and normalization are performed using a linear layer and the GELU activation function to obtain the patient's personalized profile representation vector.

5. The method for recommending brain therapy resources driven by patient desensitization profiling as described in claim 1, characterized in that, The method for obtaining the representation vector of the diagnostic and treatment resources includes the following steps: Based on the relationships between the medical entities, and with the medical entities as nodes and the relationships between the medical entities as edges, a heterogeneous knowledge graph for brain diagnosis and treatment is constructed. The relational graph convolutional network RGCN is adopted to complete the aggregation and updating of node features in the heterogeneous knowledge graph of brain diagnosis and treatment through message passing mechanism. After multiple layers of message passing, the initial aggregated feature vector of each node in the heterogeneous knowledge graph of brain diagnosis and treatment is output. The KGAT knowledge graph attention network is adopted, and an attention mechanism is introduced. The initial aggregated feature vectors of two nodes are multiplied by the mapping matrix and then concatenated. This concatenation is then multiplied by an initialized vector and passed through the LeakyReLU activation function to obtain the attention score between the nodes. Subsequently, the attention scores of all neighboring nodes of a target node are normalized using the softmax function to obtain normalized attention weights. Based on the normalized attention weights, the features of the neighboring nodes of the target node are weighted and aggregated, and then the target node's own features are added to complete the update of the target node's features. After updating the features of all nodes in the heterogeneous knowledge graph for brain diagnosis and treatment, the node features are processed through a linear layer for dimensional mapping and normalization to obtain the final representation vector of diagnosis and treatment resources.

6. The method for recommending brain therapy resources driven by patient desensitization profiling as described in claim 1, characterized in that, The calculation method for the diagnostic and treatment fit and the diagnostic and treatment resource load fit includes the following steps: A dual-tower deep matching model is constructed. The patient's personalized profile representation vector is input into the patient tower sub-network, and nonlinear transformation and dimension mapping are performed through 3 fully connected layers to output the patient profile encoding feature vector. The medical resource representation vector is input into the resource tower sub-network, and nonlinear transformation and dimension mapping are performed through 2 fully connected layers to output the medical resource encoding feature vector. The cosine similarity between the patient profile encoding feature vector and the medical resource encoding feature vector is used as the medical resource matching degree between the medical institution and the patient's medical needs. The absolute value of the load rate of the medical institution after receiving the patient minus the optimal load rate of the medical institution is calculated, and the absolute value of 1 minus the absolute value is used to obtain the medical resource load matching degree between the medical institution and the patient's medical needs.

7. The method for recommending brain therapy resources driven by patient desensitization profiling as described in claim 1, characterized in that, The initialization method of the adaptive inertia weighted discrete particle swarm optimization algorithm includes the following steps: Initial particles that conform to the aforementioned encoding rules are randomly generated. Each particle corresponds to a medical institution sorting scheme. Each particle is configured with a corresponding initial particle position vector and initial particle velocity vector. Subsequently, the initial position of each particle is set as the optimal position of the initial individual. The objective function value of the diagnosis and treatment fit and the objective function value of the diagnosis and treatment resource load fit for each initial particle are calculated respectively. All non-dominated solutions are selected from all initial particles and added to the Pareto optimal solution external archive set.

8. The method for recommending brain therapy resources driven by patient desensitization profiling as described in claim 7, characterized in that, The iterative update method of the adaptive inertia weighted discrete particle swarm optimization algorithm includes the following steps: After the particle swarm initialization is completed, two random numbers between 0 and 1 that satisfy a uniform distribution are generated. Then, the velocity vector of the particles is updated according to the particle category using adaptive inertia weights. The current position vector of the particles is added to the updated velocity vector to obtain the updated particle position vector. The objective function values ​​of the diagnosis and treatment fit and the diagnosis and treatment resource load fit corresponding to the updated particles are then calculated. The current position of the particles is compared with the best position of the historical individuals according to the Pareto dominance rule, and the best position of the historical individuals is updated. After updating the individual optimal positions of all particles, select all non-dominated solutions from the union of the current position vectors of all particles and the external archive of Pareto optimal solutions from the previous round as the external archive of Pareto optimal solutions for this round of optimization; then use the roulette wheel method to select a solution from the external archive of Pareto optimal solutions to update the global optimal position. The iteration ends when the current iteration round reaches the maximum number of iterations or when the external archive of the Pareto optimal solution is completely consistent with the previous round and has not changed for 10 consecutive iterations, and the final external archive of the Pareto optimal solution is output.

9. The method for recommending brain therapy resources driven by patient desensitization profiling as described in claim 8, characterized in that, The method for generating the brain diagnosis and treatment resource recommendation list includes the following steps: A decision matrix is ​​constructed based on the fitness function value corresponding to each non-dominated solution in the external archive set of the final Pareto optimal solution, and Z-score is used for standardization. Then, the maximum value of the objective function value of treatment fit and the objective function value of treatment resource load fit are taken from the decision matrix as the positive ideal solution, and the minimum value is the negative ideal solution. Then, the Euclidean distance between each non-dominated solution and the positive ideal solution, and the Euclidean distance between each non-dominated solution and the negative ideal solution are calculated. The proximity of each non-dominated solution is calculated based on the two Euclidean distances, and the solution with the largest proximity is selected as the final output list of recommended brain treatment resources.

10. A brain therapy resource recommendation system driven by patient desensitization profiling, characterized in that, include: The data anonymization module is used to collect diagnostic and treatment text data and patient attribute data, and to perform anonymization and preprocessing. Finally, through privacy enhancement, enhanced medical semantic embedding vectors and enhanced patient attribute data are obtained. The profile building module is used to obtain patient condition representation vectors from augmented medical semantic embedding vectors and extract patient personalized preference representation vectors from augmented patient attribute data and diagnosis and treatment text data. The patient's condition representation vector and the patient's personalized preference representation vector are input into the cross-modal fusion Transformer, and the patient's personalized profile representation vector is obtained by using the cross attention mechanism. The resource representation module is used to construct a heterogeneous knowledge graph for brain diagnosis and treatment, and uses RGCN to generate the initial aggregated feature vectors of nodes; Subsequently, KGAT was used to introduce an attention mechanism to obtain a diagnostic resource representation vector based on the initial aggregated feature vector; The recommendation decision module constructs a dual-tower deep matching model, inputting the patient's personalized profile representation vector into the patient tower sub-network and the treatment resource representation vector into the resource tower sub-network to calculate the treatment fit and treatment resource load fit. An adaptive inertial weighted discrete particle swarm optimization algorithm is employed, using the treatment fit objective function and the treatment resource load fit objective function as fitness functions, to generate a brain treatment resource recommendation list based on the Pareto principle.

Images