Health scheme recommendation method, device and equipment based on large model and knowledge graph

Abstract

The invention provides a health scheme recommendation method, device and equipment based on a large model and a knowledge graph. The method comprises the following steps: acquiring multi-modal data of a cardiovascular disease patient; the multi-modal data is preprocessed, and fusion of the multi-modal data is achieved through time alignment and feature alignment; features of the preprocessed multi-modal data are extracted, a multi-dimensional feature matrix is constructed, and the multi-dimensional feature matrix is identified and labeled; inputting the multi-dimensional feature matrix into a pre-constructed fusion behavior recognition model, enhancing the model performance through time sequence modeling and multi-modal fusion, realizing the recognition of psychological disorders, and obtaining a recognition result; and integrating the patient data, the recognition result and a pre-constructed psychological disorder knowledge graph, and adopting collaborative filtering, content recommendation and reinforcement learning methods to generate a personalized psychological health management scheme. Precise recognition and personalized management of the psychological disorder of the patient with the cardiovascular disease are achieved, and the psychological health level and life quality of the patient can be improved.

Description

Technical Field

[0001] The present invention relates to the field of medical information, and in particular to a method, device and equipment for recommending health plans based on a large model and a knowledge graph. Background Art

[0002] Studies have shown that depression is closely associated with death from cardiovascular disease and injuries, and targeted intervention measures are needed.

[0003] The mental health of patients with cardiovascular disease has a significant impact on their recovery and quality of life, but existing methods for identifying psychological disorders often have problems such as insufficient accuracy and low personalization. How to effectively integrate multi-dimensional data of patients and accurately identify psychological disorders is a technical problem that needs to be solved urgently.

[0004] Traditional methods mainly rely on a single data source, which makes it difficult to fully reflect the patient's psychological state; and simply superimposing multimodal data is prone to information redundancy and noise interference. In addition, the manifestations of psychological disorders are complex and diverse, and there are large individual differences between different patients. How to extract effective features from massive heterogeneous data and establish a dynamically updated knowledge system is a key challenge to achieve personalized recognition.

[0005] At the same time, how to closely integrate the results of psychological disorder identification with clinical practice to provide strong support for doctors to formulate personalized treatment plans is also an important issue. These technical difficulties are interrelated and together constitute a systematic challenge for the identification and management of psychological disorders in cardiovascular patients. Innovative exploration is needed at multiple levels such as data fusion, model construction, and knowledge representation to achieve full-process intelligence from data to knowledge and then to personalized decision-making.

[0006] In view of this, this application is filed. Summary of the invention

[0007] The present invention provides a health plan recommendation method based on a large model and a knowledge graph, which is characterized by comprising:

[0008] Acquiring multimodal data of a patient with cardiovascular disease, wherein the multimodal data includes demographic data, physiological data, behavioral data, text data, image data, and audio data;

[0009] Preprocessing the multimodal data to achieve fusion of the multimodal data through time alignment and feature alignment;

[0010] Extracting features of the preprocessed multimodal data, constructing a multidimensional feature matrix, and identifying and annotating the multidimensional feature matrix; wherein each row of the feature matrix represents features of different time periods, and each column represents feature dimensions extracted from different modalities;

[0011] Input the multi-dimensional feature matrix into a pre-constructed fusion behavior recognition model, enhance the model performance through time series modeling and multi-modal fusion to achieve the recognition of mental disorders, and obtain the recognition result;

[0012] Integrate patient data, recognition results, and a pre-constructed mental disorder knowledge graph, and use collaborative filtering, content recommendation, and reinforcement learning methods to generate a personalized mental health management plan.

[0013] Preferably, the preprocessing of multi-modal data includes:

[0014] Perform standardization or normalization on different types of physiological data so that they can be compared on a unified scale;

[0015] Synchronize the timestamps of different data sources to ensure that the data is analyzed in the same time dimension;

[0016] Align the features of different modalities to ensure that they can be processed and analyzed in the same feature space, reducing information loss.

[0017] Preferably, the extraction of features from the preprocessed multi-modal data to construct a multi-dimensional feature matrix includes:

[0018] Extract physiological data features, including time-domain features and frequency-domain features of heart rate variability;

[0019] Extract behavior data features, including the activity count per minute in daily activities and the proportion of deep sleep in the sleep pattern;

[0020] Use natural language processing techniques to analyze the frequency of emotion words and emotion tendencies in text data;

[0021] Use a convolutional neural network to extract emotion features and areas of interest in image data;

[0022] Extract speech features from the patient's audio data, including fundamental frequency, formant frequency, and speech rate, for emotion recognition.

[0023] Preferably, the recognition and annotation of the multi-dimensional feature matrix include:

[0024] By analyzing the features of the fused feature matrix, identify important features related to emotional and behavioral changes;

[0025] Associate and annotate the identified important features with mental disorders.

[0026] Preferably, the fusion behavior recognition model uses a deep recurrent neural network model.

[0027] Preferably, the mental disorder knowledge graph is constructed by the following method:

[0028] Define the core entities in the knowledge graph, including patients, mental disorders, symptoms, and treatment methods, and classify the entities.

[0029] Adopt the BERT+BiLSTM-CRF model to identify entities in the text based on natural language processing technology and automatically classify them.

[0030] Define the relationships between entities, including the "suffering from" relationship between patients and mental disorders, and the "manifesting" relationship between mental disorders and symptoms.

[0031] Define attributes for entities, including the age, gender, medical history, etc. of patients, and the severity and duration of mental disorders.

[0032] Preferably, integrating patient data, recognition results, and a pre-constructed mental disorder knowledge graph, adopt collaborative filtering, content recommendation, and reinforcement learning methods to generate a personalized mental health management plan, specifically including:

[0033] Integrate patient data, recognition results, and a pre-constructed mental disorder knowledge graph, adopt a collaborative filtering algorithm, and use the patient-intervention measure matrix to recommend possible effective mental intervention plans.

[0034] Integrate patient data, recognition results, and a pre-constructed mental disorder knowledge graph, adopt a content recommendation algorithm, match according to patient characteristics and intervention measure characteristics, and recommend the most suitable mental intervention plan.

[0035] Introduce a reinforcement learning framework, continuously optimize the recommendation strategy through patient feedback, and improve the intervention effect.

[0036] Preferably, it further includes:

[0037] Real-time monitor the changes in the patient's health status, and use a machine learning model to learn the patient's personalized response pattern.

[0038] Regularly train and update the model parameters according to new data and feedback.

[0039] Dynamically adjust the recommendation strategy according to the patient's latest status to ensure the effectiveness and adaptability of the intervention plan.

[0040] The embodiment of the present invention also provides a health plan recommendation device based on a large model and a knowledge graph, which includes:

[0041] A multi-modal data acquisition unit for acquiring multi-modal data of cardiovascular disease patients, where the multi-modal data includes demographic data, physiological data, behavioral data, text data, image data, and audio data.

[0042] A preprocessing unit for preprocessing the multi-modal data and realizing the fusion of multi-modal data through time alignment and feature alignment;

[0043] A feature extraction unit for extracting the features of the preprocessed multi-modal data, constructing a multi-dimensional feature matrix, and performing identification and annotation on the multi-dimensional feature matrix; wherein, each row of the feature matrix represents the features in different time periods, and each column represents the feature dimensions extracted from different modalities;

[0044] An identification unit for inputting the multi-dimensional feature matrix into a pre-constructed fusion behavior recognition model, enhancing the model performance through time series modeling and multi-modal fusion, realizing the identification of mental disorders, and obtaining an identification result;

[0045] A health plan recommendation unit for comprehensively considering patient data, the identification result, and a pre-constructed mental disorder knowledge graph, and generating a personalized mental health management plan by using collaborative filtering, content recommendation, and reinforcement learning methods.

[0046] An embodiment of the present invention further provides a health plan recommendation device based on a large model and a knowledge graph, which includes a memory and a processor. A computer program is stored in the memory, and the computer program can be executed by the processor to implement the health plan recommendation method based on a large model and a knowledge graph as described above.

[0047] The present invention obtains multi-modal data of patients, and uses a deep recurrent neural network for fusion analysis and mental disorder identification. At the same time, a large language model is used to extract knowledge from medical literature to dynamically update the mental disorder knowledge graph. Based on the identification result and the updated knowledge graph, the present invention performs reasoning and matching, generates a personalized health management plan, and presents it visually to doctors and patients. The present invention innovatively combines multi-modal data analysis, deep learning, knowledge graph, and large language model, realizes the accurate identification and personalized management of mental disorders of cardiovascular disease patients, and helps to improve the mental health level and quality of life of patients. Brief Description of the Drawings

[0048] Figure 1 It is a flowchart of the health plan recommendation method based on a large model and a knowledge graph provided by the first embodiment of the present invention.

[0049] Figure 2 It is a structural diagram of the health plan recommendation device based on a large model and a knowledge graph provided by the second embodiment of the present invention. Detailed Description of the Embodiment

[0050] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative work belong to the scope of protection of the present invention.

[0051] Please refer to Figure 1 , the first embodiment of the present invention provides a health plan recommendation method based on a large model and a knowledge graph, which can be executed by a health plan recommendation device based on a large model and a knowledge graph (hereinafter referred to as the recommendation device). Specifically, it is executed by one or more processors in the recommendation device to implement the following method:

[0052] S101. Obtain multimodal data of cardiovascular disease patients, where the multimodal data includes demographic data, physiological data, behavioral data, text data, image data, and audio data.

[0053] In this embodiment, the physiological data of the patient can be collected through wearable devices and medical monitoring devices, including indicators such as heart rate, blood pressure, and sleep quality, and the collected data is transmitted to the system for storage and analysis. Use a smartphone application to record the patient's daily behavioral data, such as exercise volume and eating habits, and upload the behavioral data to the recommendation device through a data interface. Obtain text data such as the patient's electronic medical record and consultation record, and use natural language processing algorithms to extract key information such as symptoms and diagnoses, such as using the TF-IDF algorithm to extract keywords. Collect image data such as the patient's facial expressions and body postures, and use computer vision technology to analyze the images to identify the patient's emotional state. Record the patient's voice data, convert it to text through a speech recognition API, and use an emotion analysis algorithm to judge the emotional tendency, such as using an LSTM model for emotion classification.

[0054] S102. Preprocess the multimodal data to achieve the fusion of multimodal data through time alignment and feature alignment.

[0055] In this embodiment, after obtaining multimodal data such as the patient's electronic health record, physiological indicators, psychological assessment results, patient preferences, and lifestyle, through data cleaning, missing values, outliers, and noise data are removed to obtain a high-quality original data set. Then, according to the medical domain knowledge graph, the cleaned multimodal data is standardized, and data from different sources and in different formats is converted into a unified representation form to eliminate the differences between the data and obtain standardized multimodal data. Perform time alignment on the standardized multimodal data. According to the timestamp information, different modal data is synchronized in the time dimension to ensure that subsequent analysis is performed on a unified time axis.

[0056] For example, after obtaining multi-modal data such as the patient's electronic health record and physiological indicators, through data cleaning and processing, records with missing values exceeding 20% are removed, and the remaining missing values are filled with the median. At the same time, box plots are used to identify outliers, and data beyond 1.5 times the interquartile range is regarded as abnormal and excluded to obtain a high-quality original data set. Then, according to the standard terms and coding system in the medical knowledge graph, the cleaned multi-modal data is standardized. For example, disease names from different sources are mapped to ICD-10 codes, and drug names are mapped to ATC codes to ensure data consistency and comparability. Next, the standardized multi-modal data is time-aligned. With a 1-hour time window, the data of each modality is aggregated into the corresponding time period to form an aligned time series.

[0057] S103, extract the features of the preprocessed multi-modal data, construct a multi-dimensional feature matrix, and perform identification and annotation on the multi-dimensional feature matrix; wherein, each row of the feature matrix represents the features of different time periods, and each column represents the feature dimensions extracted from different modalities.

[0058] Specifically, feature engineering techniques are used to extract effective features from the standardized and time-aligned multi-modal data, including the fundamental frequency, formant frequency, speech rate, etc. of speech, facial expression features, and physiological features such as heart rate variability, to obtain independent feature sets of multiple modalities. Through feature alignment techniques, the features extracted from different modalities are mapped to the same feature space to eliminate the differences between features and obtain an aligned multi-modal feature set. According to the feature fusion strategy, methods such as weighted average or deep fusion network are used to fuse the aligned multi-modal features to generate a multi-dimensional feature matrix. Each row of the feature matrix d represents the feature dimensions of a time period, and each column represents the feature dimensions of different modalities.

[0059] In this embodiment, the fused multi-dimensional feature matrix can also be normalized to scale the feature values to a unified range, eliminate the dimensional differences between features, and obtain a normalized multi-modal feature matrix. Through feature selection techniques, the most discriminative and representative feature subset is selected from the normalized multi-modal feature matrix to reduce the feature dimension and obtain an optimized multi-modal feature set. The optimized multi-modal feature set is used as the input for the subsequent psychological disorder recognition model to provide high-quality feature data for model training and prediction.

[0060] For example, methods such as wavelet transform and Mel-frequency cepstral coefficients are used to extract features such as fundamental frequency, formant frequency, and speech rate from speech data. Facial expression features are extracted using a facial key point detection algorithm, and heart rate variability indicators are calculated from electrocardiogram data to obtain feature sets of different modalities. Subsequently, each modality feature is mapped to the interval [0, 1] through min-max normalization to achieve feature alignment. Finally, in a weighted average manner, combined with the importance weights of each modality feature, a fused multi-dimensional feature matrix is generated, and the Z-score normalization method is used to normalize the feature matrix to eliminate the influence of dimensions. Through a recursive feature elimination algorithm, the top 20 features with the highest contribution to mental disorder recognition are selected to obtain an optimized multi-modal feature set for subsequent model training and prediction tasks.

[0061] In this embodiment, after obtaining the multi-dimensional feature matrix, during model training, it is necessary to perform identification and annotation on it, including:

[0062] Behavior annotation: By analyzing the fused feature matrix, features related to emotional and behavioral changes are identified. For example, when the heart rate increases and is accompanied by a tense change in facial expression, it may represent an anxious emotion. This annotation process depends not only on single-modal features but also on the comprehensive analysis of multi-modal features to ensure the accuracy of annotation.

[0063] Association between mental disorders and physiological behaviors: During the annotation process, heart rate, blood pressure, facial expressions, speech features, etc. will all be used as important indicators to comprehensively judge the user's emotional state. For example, an abnormal increase in heart rate can be annotated as an "anxious" or "tense" state; while during deep sleep, the situation where the heart rate drops and is accompanied by a low activity level may be annotated as a "relaxed" or "normal" state.

[0064] S104, input the multi-dimensional feature matrix into a pre-constructed fusion behavior recognition model, enhance the model performance through time series modeling and multi-modal fusion, realize the recognition of mental disorders, and obtain the recognition result.

[0065] In this embodiment, the fusion behavior recognition model can use a deep recurrent neural network model. Specifically, it is an LSTM+GRU model. Among them, LSTM can process time series data and capture the dependencies that change over time in multi-modal data. The LSTM+GRU model can also solve the gradient vanishing problem in traditional RNNs and learn feature changes over a longer time span. Finally, the model performance is enhanced through multi-modal data fusion to understand the user's mental state from multiple perspectives and improve the recognition accuracy and reliability.

[0066] Specifically, the deep recurrent neural network based on GRU includes 3 hidden layers, with each layer containing 128 neurons. It is trained using the Adam optimizer and the cross-entropy loss function, with the learning rate set to 0.001 and the batch size to 32. During the training process, the attention mechanism can be used to adaptively adjust the weights of different modality features to achieve dynamic feature fusion. The fused features can be sparsified through L1 regularization, and the top 20% of the features are selected as the input to the fusion behavior recognition model. Finally, the fused features are classified in the output layer using the radial basis kernel function, with the regularization parameter C set to 1.0. The performance of the model is evaluated through 5-fold cross-validation to obtain metrics such as the average accuracy, precision, recall, and F1 value, which are used to guide the optimization and improvement of the model.

[0067] S105, integrating the patient data, recognition results, and the pre-constructed mental disorder knowledge graph, uses collaborative filtering, content recommendation, and reinforcement learning methods to generate personalized mental health management plans.

[0068] In this embodiment, when constructing the mental disorder knowledge graph, according to the defined entity categories and relationship types, natural language processing techniques can be used to extract the entities and relationships related to the mental disorders of cardiovascular patients from data sources such as evidence-based guidelines, clinical practice guidelines, medical databases, authoritative literature, online platforms, and clinical health manuals to obtain the original knowledge triples. Then, a knowledge fusion algorithm is used to deduplicate and disambiguate the extracted knowledge triples, determine the unique entity identifiers, judge the confidence of the triples, and obtain high-quality fused knowledge triples. A knowledge graph representation learning algorithm is used to map the entities and relationships in the knowledge triples to a low-dimensional semantic space to obtain the distributed representations of the entities and relationships, which is convenient for subsequent knowledge reasoning and application. According to the characteristics of the mental disorders of cardiovascular patients, rule-based and pattern-based knowledge reasoning rules are designed, and methods such as link prediction and path reasoning on the knowledge graph are used to mine the implicit entity relationships and enrich the knowledge graph. A fact-based knowledge graph quality assessment method is used to evaluate the quality of the knowledge graph by calculating metrics such as the confidence, coverage, and consistency of the triples in the knowledge graph, and identify the knowledge that may be incorrect or missing. The feedback from experts on the knowledge graph quality assessment results is obtained, and the knowledge marked as incorrect or unreasonable by the experts is manually verified and corrected to improve the accuracy and reliability of the knowledge graph. According to the latest research progress in cardiology and psychology, in an incremental update manner, high-quality newly published literature is regularly obtained from the literature database, and the entities, relationships, and attributes in the knowledge graph are updated through knowledge extraction and reasoning.

[0069] In addition, an intelligent question - answering and decision - making support system based on the knowledge graph can be further developed. Doctors and patients can obtain the required knowledge about psychological disorders of cardiovascular disease patients through natural language inquiries or interactive interfaces, which supports mental health assessment and management. A visualization display system of the knowledge graph is constructed to present the entities, relationships, and attributes in the knowledge graph in a graphical way, facilitating users to explore and understand the knowledge about psychological disorders of cardiovascular disease patients and discover new associations and rules.

[0070] In this embodiment, when generating psychological intervention suggestions based on patient data, recognition results, and a pre - constructed knowledge graph of psychological disorders, a knowledge - graph - based reasoning algorithm such as DeepPath can be used to obtain personalized suggestions by searching for paths between patient characteristics and potential intervention measures in the knowledge graph.

[0071] Specifically, according to the patient's multi - modal data (such as speech, facial expressions, movements, heart rate, and sleep monitoring, etc.), independent and cross - modal feature analysis methods are used to extract key features reflecting the patient's mental state, and the recognition result of the patient's psychological disorder is obtained. If a patient is identified as having a psychological disorder, entities such as symptoms, diseases, and mental states related to this psychological disorder are queried in the knowledge graph to obtain the association relationships between relevant entities. Through the reasoning rules in the knowledge graph, based on the recognition result of the patient's psychological disorder and the association relationships between relevant entities, other potential mental health risk factors of the patient are judged. Treatment methods and health management plans that match the patient's psychological disorder and risk factors are searched in the knowledge graph to obtain a candidate set of personalized health management plans. According to the patient's specific situation (such as age, gender, medical history, etc.), the candidate health management plans are screened and ranked to obtain the most suitable personalized health management plan for the patient. Through natural language generation technology, the selected personalized health management plan is transformed into a form that is easy for the patient to understand and execute, and a personalized health management advice report is generated.

[0072] For example, by extracting features from multimodal data such as the patient's voice, facial expressions, movements, heart rate, and sleep monitoring, and using the support vector machine (SVM) algorithm to classify the features, it is identified that the possibility of the patient having depression is 85%. Query the knowledge graph for symptoms related to depression, such as sleep disorders and decreased appetite, and find that the patient's sleep monitoring data shows that the average sleep duration is only 5 hours, inferring that the patient may have risk factors for insomnia. According to the inference rules in the knowledge graph, treatment methods suitable for this patient, such as cognitive behavioral therapy (CBT) and drug therapy (such as selective serotonin reuptake inhibitors, SSRIs), are matched. Considering that the patient is 65 years old and has a history of hypertension, a non-drug treatment plan mainly based on CBT is preferentially recommended. Using natural language generation technology, a personalized health management advice report is generated, suggesting that the patient undergo CBT treatment 2 times a week, each time for 50 minutes, and at the same time receive sleep hygiene education. According to the results of the latest Meta-analysis research, the efficacy evaluation of CBT in the knowledge graph is updated to increase its weight. After expert demonstration, the safety and effectiveness of this plan are confirmed.

[0073] In this embodiment, a machine learning algorithm can also be used to optimize the personalized health management plan. Combining the patient's characteristic preferences and needs, the practicality and effectiveness of the intervention measures are ensured, and an optimized personalized health management plan is obtained. Through visualization technology, the optimized personalized health management plan is transformed into forms such as easy-to-understand charts and images, and it is judged whether the visualization effect is clear and intuitive. If it is not intuitive enough, the visualization presentation method is further optimized until a clear and easy-to-understand visualized health management plan is obtained. The visualized personalized health management plan is synchronously presented to doctors and patients. Through an interactive interface, both doctors and patients can conveniently view and understand the content and implementation steps of the health plan, and it is judged whether there are difficulties in understanding. If there are difficulties, necessary explanations are provided. The feedback opinions of the patient on the health management plan are obtained, including satisfaction scores, symptom improvement situations, etc., and the feedback data is input into the machine learning model. Through a reinforcement learning framework, the recommendation strategy is continuously optimized. According to the patient's feedback, a reward function is used to set the reward value to guide the model to adjust and optimize the personalized health management plan, and more accurate and effective intervention measures are obtained.

[0074] Furthermore, it also includes:

[0075] Real-time monitor the changes in the patient's health status, and use a machine learning model to learn the patient's personalized response pattern;

[0076] Regularly train and update the model parameters according to new data and feedback;

[0077] Dynamically adjust the recommendation strategy according to the patient's latest status to ensure the effectiveness and adaptability of the intervention plan.

[0078] Among them, the health status changes of patients can also be monitored in real time. Machine learning models such as LSTM and Transformer are used to learn the personalized response patterns of patients, and it is judged whether it is necessary to dynamically adjust the health plan according to the changes in the patient's status. If so, the dynamic optimization process of the health plan is triggered. According to the health status changes of patients and feedback opinions, the personalized health management plan is dynamically adjusted. Through the continuous learning and optimization of the machine learning model, a more accurate and personalized health intervention strategy is obtained. The adjusted personalized health management plan is presented to doctors and patients in a visual manner again, and their evaluation feedback on the optimized plan is obtained to judge whether further improvement is needed. If so, the above steps of the dynamic optimization process are repeated. Regularly collect the health data and feedback data of patients, and use these new data to retrain the machine learning model and update the parameters, continuously improving the accuracy and effectiveness of the personalized health management plan, and realizing the dynamic optimization and continuous improvement of the health plan.

[0079] Please refer to Figure 2 , the second embodiment of the present invention also provides a health plan recommendation device based on a large model and a knowledge graph, which includes:

[0080] A multimodal data acquisition unit 210, configured to acquire multimodal data of cardiovascular disease patients, where the multimodal data includes demographic data, physiological data, behavioral data, text data, image data, and audio data;

[0081] A preprocessing unit 220, configured to preprocess the multimodal data and realize the fusion of multimodal data through time alignment and feature alignment;

[0082] A feature extraction unit 230, configured to extract the features of the preprocessed multimodal data, construct a multi-dimensional feature matrix, and perform identification and annotation on the multi-dimensional feature matrix; wherein, each row of the feature matrix represents the features of different time periods, and each column represents the feature dimensions extracted from different modalities;

[0083] An identification unit 240, configured to input the multi-dimensional feature matrix into a pre-constructed fusion behavior recognition model, enhance the model performance through time series modeling and multimodal fusion, realize the recognition of mental disorders, and obtain recognition results;

[0084] A health plan recommendation unit 250, configured to generate a personalized mental health management plan by integrating patient data, recognition results, and a pre-constructed mental disorder knowledge graph, and adopting collaborative filtering, content recommendation, and reinforcement learning methods.

[0085] The third embodiment of the present invention also provides a health solution recommendation device based on a large model and a knowledge graph, which includes a memory and a processor. A computer program is stored in the memory and can be executed by the processor to implement the health solution recommendation method based on a large model and a knowledge graph as described above.

[0086] In the embodiment of the present invention, by obtaining multi-modal data of patients, a deep recurrent neural network and a graph neural network are used for fusion analysis and psychological disorder recognition. At the same time, a large language model is used to extract knowledge from medical literature to dynamically update the psychological disorder knowledge graph. Based on the recognition results and the updated knowledge graph, the present invention performs reasoning and matching to generate a personalized health management plan, which is visually presented to doctors and patients. The present invention innovatively combines multi-modal data analysis, deep learning, knowledge graph and large language model to achieve accurate recognition and personalized management of psychological disorders in cardiovascular disease patients, which helps to improve the mental health level and quality of life of patients.

[0087] The above description is only the preferred embodiment of the present application and the description of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the present application is not limited to the technical solution formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the concept of the present application. For example, the technical solutions formed by mutually replacing the above features with the technical features (but not limited to) having similar functions disclosed in the present application.

Claims

1. A health plan recommendation method based on a large model and knowledge graph, characterized in that: include: Obtain multimodal data on patients with cardiovascular disease; The multimodal data includes demographic data, physiological data, behavioral data, text data, image data and audio data; Preprocessing the multimodal data to achieve fusion of the multimodal data through time alignment and feature alignment; Extracting features of the preprocessed multimodal data, constructing a multidimensional feature matrix, and identifying and annotating the multidimensional feature matrix; wherein each row of the feature matrix represents features of different time periods, and each column represents feature dimensions extracted from different modalities; Inputting the multi-dimensional feature matrix into a pre-built fusion behavior recognition model to realize the recognition of psychological disorders and obtain recognition results; By integrating patient data, identification results and pre-built knowledge graphs of mental disorders, collaborative filtering, content recommendation and reinforcement learning methods are used to generate personalized mental health management plans.

2. The health plan recommendation method based on a large model and a knowledge graph as claimed in claim 1, characterized in that: The preprocessing of the multimodal data includes: Standardize or normalize different types of physiological data so that they can be compared on a unified scale; Synchronize the timestamps of different data sources to ensure that the data is analyzed on the same time dimension.

3. The health plan recommendation method based on a large model and a knowledge graph as claimed in claim 1, characterized in that: The extracting the features of the preprocessed multimodal data and constructing a multidimensional feature matrix includes: Extract physiological data features, including time domain features and frequency domain features of heart rate variability; Extract behavioral data features, including activity counts per minute in daily activities and deep sleep ratio in sleep patterns; Use natural language processing technology to analyze the frequency and emotional tendency of emotional words in text data; Use convolutional neural networks to extract emotional features and focus points from image data; Extract speech features from patient audio data, including fundamental frequency, formant frequency, and speaking rate, for emotion recognition.

4. The health plan recommendation method based on a large model and a knowledge graph as claimed in claim 1, characterized in that: Identification and annotation of multi-dimensional feature matrices include: By analyzing the features of the fused feature matrix, important features related to emotion and behavior changes are identified; The identified important features were associated with psychological disorders and annotated.

5. The health plan recommendation method based on a large model and a knowledge graph as claimed in claim 1, characterized in that: The fusion behavior recognition model uses a deep recurrent neural network model.

6. The health plan recommendation method based on a large model and a knowledge graph as claimed in claim 1, characterized in that: The knowledge graph of psychological disorders is constructed by the following method: Define the core entities in the knowledge graph, including patients, psychological disorders, symptoms, treatments, and classify the entities; The BERT+BiLSTM-CRF model is used to identify entities in texts based on natural language processing technology and automatically classify them; Define relationships between entities, including the "suffering from" relationship between patients and mental disorders, and the "manifestation" relationship between mental disorders and symptoms; Define attributes for entities, including the patient's age, gender, medical history, severity and duration of psychological disorders.

7. The health plan recommendation method based on a large model and a knowledge graph as claimed in claim 1, characterized in that: Combining patient data, identification results and pre-built knowledge graphs of mental disorders, collaborative filtering, content recommendation and reinforcement learning methods are used to generate personalized mental health management plans, including: By integrating patient data, identification results and pre-built knowledge graphs of psychological disorders, a collaborative filtering algorithm is used to recommend potentially effective psychological intervention plans using a patient-intervention matrix; By integrating patient data, identification results, and pre-built knowledge graphs of psychological disorders, a content recommendation algorithm is used to match patient characteristics with intervention measures to recommend the most suitable psychological intervention plan; A reinforcement learning framework was introduced to continuously optimize the recommendation strategy through patient feedback and improve the intervention effect.

8. The health plan recommendation method based on a large model and a knowledge graph as claimed in claim 7, characterized in that: Also includes: Monitor changes in patients’ health status in real time and use machine learning models to learn patients’ personalized response patterns; Regularly train and update model parameters based on new data and feedback; Dynamically adjust the recommended strategy based on the patient's latest status to ensure the effectiveness and adaptability of the intervention plan.

9. A health plan recommendation device based on a large model and knowledge graph, characterized in that: include: A multimodal data acquisition unit, used to acquire multimodal data of a cardiovascular disease patient, wherein the multimodal data includes demographic data, physiological data, behavioral data, text data, image data and audio data; A preprocessing unit, used to preprocess the multimodal data and realize fusion of the multimodal data through time alignment and feature alignment; A feature extraction unit, used to extract features of the preprocessed multimodal data, construct a multidimensional feature matrix, and identify and annotate the multidimensional feature matrix; wherein each row of the feature matrix represents features of different time periods, and each column represents feature dimensions extracted from different modalities; An identification unit, used for inputting the multi-dimensional feature matrix into a pre-built fusion behavior identification model to realize the identification of psychological disorders and obtain an identification result; The health plan recommendation unit is used to integrate patient data, recognition results and pre-built knowledge graphs of psychological disorders, and adopt collaborative filtering, content recommendation and reinforcement learning methods to generate personalized mental health management plans.

10. A health plan recommendation device based on a large model and knowledge graph, characterized in that: It includes a memory and a processor, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement the health plan recommendation method based on a large model and a knowledge graph as described in any one of claims 1 to 8.

Images