Distributed mobile smart medical remote system based on active health model

Abstract

The invention discloses a distributed mobile smart medical remote system based on an active health model, and relates to the technical field of cerebral apoplexy health management, the system comprises a smart medical remote platform, the smart medical remote platform is in communication connection with the following modules: a multi-source data acquisition module, a multi-source data processing module and a multi-source data processing module; the cerebral apoplexy multi-source data collection module is used for collecting cerebral apoplexy multi-source data including clinical data and health data and preprocessing the collected data. The deep Q network and the GPT large model are utilized to construct the virtual patient generator, the synthetic data conforming to the cerebral apoplexy characteristics of the target area are generated, the defect of real data is overcome, the high-fidelity medical record text can be generated through the reinforcement learning algorithm, the medical knowledge graph is called to ensure the physiological rationality, and the method is suitable for popularization and application. And the synthetic data is used for training the cerebral apoplexy risk prediction model, so that the performance of the cerebral apoplexy risk prediction model on real data is remarkably improved, high-risk patients can be found in time in the early stage of diseases, and the disability rate and the death rate of cerebral apoplexy are reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of stroke health management, and in particular to a distributed mobile intelligent medical remote system based on an active health model. BACKGROUND

[0002] Stroke (stroke) is one of the important causes of disability and death worldwide, and is divided into two types: ischemic stroke (cerebral infarction) and hemorrhagic stroke. The disease is acute, the treatment window is short, and early intervention can significantly reduce mortality and disability rate. In remote areas and rural areas, medical resources are scarce, there is a lack of professional neurologists and modern medical equipment, and many patients have difficulty in obtaining effective diagnosis and treatment in time, which has greatly reduced the timeliness and effectiveness of stroke patient treatment. With the rapid development of information technology, mobile Internet, Internet of Things, artificial intelligence, cloud computing and other technologies have gradually matured in the medical field. The distributed mobile intelligent medical remote system can effectively combine these technologies to provide convenient health monitoring and real-time medical consultation services, which can assist doctors in diagnosing and guiding treatment for patients remotely.

[0003] In the prior art, stroke-related data in remote areas is scarce, and federated learning and transfer learning are often used to assist remote diagnosis and treatment. However, the federated learning method only aggregates parameters and cannot solve the problem of data distribution heterogeneity and patient differences in perception areas. Transfer learning requires target domain data, but rural areas often have zero-labeled data. Therefore, how to use a deep Q network to build a virtual patient generator and convert the actions of the deep Q network into high-fidelity medical record text using a GTP large model to call a medical knowledge graph to ensure physiological reasonableness and improve the accuracy of remote diagnosis and treatment of patients is a problem to be solved. Therefore, the present application proposes a distributed mobile intelligent medical remote system based on an active health model. SUMMARY

[0004] The present application aims to provide a distributed mobile intelligent medical remote system based on an active health model to solve the problems raised in the background art.

[0005] To solve the above technical problems, the technical solution adopted by the present application is: A distributed mobile intelligent medical remote system based on an active health model, comprising an intelligent medical remote platform, wherein the intelligent medical remote platform is communicatively connected with the following modules, wherein: A multi-source data acquisition module is used to collect stroke multi-source data including clinical data and health data, ensure the diversity and comprehensiveness of the data, and preprocess the collected data to form a multi-source data set; A data generation enhancement engine constructs a virtual patient generator based on a reinforcement learning algorithm, and converts actions generated by the virtual patient generator into high-fidelity medical record texts by using a GTP large model, while calling a medical knowledge graph to analyze physiological rationality. An intelligent risk prediction module constructs a stroke risk prediction model based on the generated high-fidelity medical record texts and actual health data of the patient, predicts and analyzes the stroke risk, and solves the adaptability problem of the model in remote areas. A risk early warning module analyzes the stroke risk status of the patient based on the prediction results of the stroke risk prediction model, integrates clinical data and end health data, and early warns and intervenes in high-risk patients to reduce the disability rate and death caused by stroke. A mobile remote collaboration module is used to connect primary hospitals with superior experts through 5G / IoT technology, establish a telemedicine collaboration network, integrate three-level medical resources of cities, counties and villages, empower primary doctors, and perform regional management of stroke prevention and control.

[0006] Further improvement of the technical scheme of the application is that the multi-source data acquisition module specifically comprises: Stroke multi-source data is widely collected from primary medical institutions in remote areas, community health centers and personal health devices (smart bracelets, sphygmomanometers, etc.), including clinical data and health data related to stroke, covering multi-dimensional data of patient basic information, medical records, physiological indicators and image data, to ensure the diversity of data sources and the comprehensiveness of coverage range. The collected stroke multi-source data is centrally preprocessed, repeated, incorrect or invalid records are removed through data cleaning, and standardized processing is performed to unify data format and coding specifications, so that the data structure is consistent through formatting operation. The preprocessed stroke multi-source data is detected, and missing values and abnormal values in the stroke multi-source data are processed, and the data set is perfected by interpolation, estimation or rejection method to form a structured and high-quality multi-source data set.

[0007] Further improvement of the technical scheme of the application is that the data generation enhancement engine comprises a synthetic data generation unit and a GPT large model text conversion unit. The synthetic data generation unit constructs a virtual patient generator based on a deep Q network reinforcement learning algorithm, generates synthetic data conforming to the characteristics of stroke in the target area, and makes up for the lack of real data. The GPT large model text conversion unit is used to convert the generated actions into medical record texts conforming to medical specifications by using a GTP large model, and call a medical knowledge graph to analyze physiological rationality, ensure the logical consistency of symptoms and image description, and generate high-fidelity medical record texts directly input to the next process to solve the problem of real data scarcity.

[0008] A further improvement to the technical solution of the present invention is that the synthetic data generation unit specifically includes: Collect stroke epidemiological statistics for the target region, including incidence rate, age distribution, characteristics of high-risk groups, and distribution of common symptoms. Then, organize and encode the statistical data and input it into a reinforcement learning algorithm based on a deep Q-network as the initial state. Based on the initial input state, the deep Q network makes decisions and generates actions. Specifically, it creates a synthetic electronic medical record containing symptom descriptions and imaging feature descriptions of virtual patients. Then, by adjusting the network parameters, the generated actions gradually conform to the real characteristics of stroke patients in the target area, thus obtaining the constructed virtual patient generator and producing synthetic data that conforms to the characteristics of stroke in the target area. The generated synthetic data is used to train a logistic regression-based simulation evaluation model, which is then applied to real remote area data for validation. The AUC (area under the curve) improvement of the simulation evaluation model is calculated as a reward, where the AUC improvement reflects the improvement effect of the synthetic data on the model performance. The parameters of the deep Q network are adjusted according to the reward to optimize the action generation strategy. This process is repeated to optimize the action generation process, generate high-quality synthetic data, and further improve the model's performance on real data.

[0009] A further improvement to the technical solution of this invention is that the GPT large model text conversion unit specifically includes: It receives action information generated by a virtual patient generator, including symptom descriptions and imaging features of virtual patients. Utilizing the text generation capabilities of the GPT large model, and based on the common structure and expression habits of medical texts, it initially converts the action information into a preliminary medical record text that conforms to medical standards. The medical knowledge graph in the field of stroke is invoked. The preliminary medical record text is input into the invoked medical knowledge graph. Based on the medical knowledge and logical relationships in the medical knowledge graph, the physiological rationality analysis of the symptom description and imaging features in the preliminary medical record text is performed to check whether there are any logical contradictions between the two, and to ensure that the content of the medical record text conforms to medical reality and clinical rules. After medical knowledge graph analysis, medical record texts with unreasonable aspects are corrected and optimized to generate high-fidelity medical record texts that conform to medical standards. These high-fidelity medical record texts are then output and used directly as input data for the next process, namely the intelligent risk prediction module, effectively solving the problem of scarcity of real data.

[0010] A further improvement to the technical solution of the present invention is that the intelligent risk prediction module includes a multi-model ensemble trainer and a transfer learning adapter; The multi-model ensemble trainer integrates multiple heterogeneous machine learning models for the stroke risk assessment task, and constructs a stroke risk prediction model through weighted fusion. It takes high-fidelity medical record text and the patient's actual health data as input, and outputs a stroke risk prediction score, reducing the bias of a single model, improving prediction accuracy, and adapting to the differences in data distribution in different regions. The transfer learning adapter is used to fine-tune the stroke risk prediction model using adversarial training domain adaptation techniques on unlabeled target domain data (rural areas), minimizing the difference in feature distribution between the source domain and the target domain. The source domain is labeled data from urban hospitals, and the target domain is synthetic data from rural areas. This allows the stroke risk prediction model pre-trained on urban data to adapt to the distribution of rural data, thereby enabling the prediction and analysis of stroke risk for patients in the target domain.

[0011] A further improvement to the technical solution of the present invention is that the multi-model ensemble trainer specifically includes: The input high-fidelity medical record text is structured and parsed to extract the medical characteristics of stroke patients. Simultaneously, patient health data is integrated, and data cleaning, missing value imputation, and standardization are performed. Through feature selection and dimensionality reduction techniques, risk features highly correlated with stroke risk are screened out, and a unified feature vector set is constructed. The medical features include blood pressure, blood glucose level, blood lipid levels, age, body mass index, duration of hypertension, duration of atrial fibrillation, and homocysteine ​​(Hcy) level; the risk features include blood pressure, blood glucose level, blood lipid levels, body mass index, and homocysteine ​​(Hcy) level. For the task of stroke risk assessment, the multi-model ensemble trainer selects a variety of heterogeneous machine learning models, including logistic regression, random forest, support vector machine and neural network. Each model is trained independently based on the risk features in the feature vector set. The performance of each model on the local dataset is evaluated by cross-validation. The heterogeneous machine learning models are ranked according to the accuracy and recall metrics to ensure that each model has reliable predictive ability. For each heterogeneous machine learning model, a corresponding allowable value is set based on the extracted risk features. If the risk feature deviates from the allowable value, it is considered abnormal. Then, by combining each risk feature and its corresponding allowable value, the risk score of each heterogeneous machine learning model is calculated. Weights are assigned based on the ranking results of each heterogeneous machine learning model, with weights set to 0.4, 0.3, 0.2, and 0.1 respectively from highest to lowest. A weighted fusion method is used to construct a stroke risk prediction model. By combining the risk scores and assigned weights from each heterogeneous machine learning model, a stroke risk prediction score is output, which is then combined with a pre-set risk threshold. To assess the risk of stroke in patients and improve the generalization ability of the model.

[0012] A further improvement to the technical solution of the present invention is that the transfer learning adapter specifically includes: The transfer learning adapter collects data from the source domain and the target domain. The source domain is labeled data from urban hospitals, and the target domain is synthetic data from rural areas. Then, it uses a stroke risk prediction model pre-trained on the source domain to extract shared risk feature representations from the source and target domains. A domain discriminator is built through a gradient inversion layer (GRL) to initialize the feature extractor in an adversarial training manner, so that the risk feature distributions of the source and target domains are initially aligned. Adversarial training is performed between the feature extractor and the domain discriminator. The feature extractor generates confusing features through a gradient inversion layer, making it impossible for the discriminator to distinguish whether the risk features come from the source domain or the target domain. At the same time, the main loss of the risk prediction task is minimized to ensure the discriminative power of the risk features for stroke prediction. Through iterative optimization, the difference in the distribution of risk features between the two domains is gradually reduced. The feature extractor parameters are fixed after the adversarial training converges, and the pre-trained stroke risk prediction model is fine-tuned to adapt to the target domain data distribution. Finally, the stroke risk prediction model is directly applied to rural patient data to output a stroke risk prediction score. No target domain labeled data is required for training, thus achieving cross-domain adaptive prediction.

[0013] A further improvement to the technical solution of the present invention is that the risk warning module specifically includes: By integrating the prediction results of the stroke risk prediction model, the patient's clinical data, and end-point health data, the stroke risk prediction score and corresponding risk characteristics of the patient are extracted. Based on the set risk thresholds, the stroke risk prediction score is analyzed to determine whether the patient's risk of stroke is high, and abnormal risk characteristics are marked in combination with the allowable values ​​of each risk characteristic. If the patient is at high risk, the Level 1 early warning mechanism is triggered, alerting medical staff and the patient to the need for timely intervention. If the patient is not at high risk but has more than two abnormal risk characteristics, the Level 2 early warning mechanism is triggered, alerting medical staff and the patient and generating a risk report.

[0014] A further improvement to the technical solution of the present invention is that the mobile remote collaboration module specifically includes: Primary hospitals collect patient clinical data through 5G / IoT terminals (portable ultrasound, mobile electrocardiographs, smart medical kits, etc.). After preprocessing by edge computing devices, the patient clinical data is uploaded to the smart healthcare remote platform in real time via the 5G network. The platform automatically matches the HIS / EMR system interface to complete the structured data conversion and encrypted storage. At the same time, primary care physicians enter unstructured information such as patient medical history and medication records through a mobile APP, use NLP technology to extract key fields, and generate standardized electronic medical record templates to ensure data integrity and standardization. The smart healthcare remote platform allocates consultation resources based on the patient's condition: village-level patients are initially seen by neurologists at county-level hospitals. If a level-one early warning mechanism is triggered, a collaborative consultation with municipal-level experts can be initiated. Higher-level experts can access grassroots data in real time through a high-definition video interaction system, analyze the data, and form consultation opinions. Primary care physicians implement intervention measures based on consultation opinions and regularly upload follow-up data via mobile devices. At the same time, they use digital twin technology to build a regional stroke prevention and control map, which displays the number of inpatients and resource utilization rates in each primary care institution in real time and triggers resource allocation mechanisms.

[0015] Due to the adoption of the above technical solution, the technical progress achieved by this invention compared to the prior art is as follows: 1. This invention provides a distributed mobile smart medical remote system based on an active health model. It utilizes a deep Q-network and a GPT large model to construct a virtual patient generator, generating synthetic data that conforms to the characteristics of stroke in the target region, thus compensating for the lack of real data. Through reinforcement learning algorithms, it can generate high-fidelity medical record text and call upon medical knowledge graphs to ensure physiological rationality. Furthermore, the synthetic data is used to train a stroke risk prediction model, significantly improving the model's performance on real data. This enhances the accuracy and reliability of early stroke diagnosis, helping to identify high-risk patients in the early stages of the disease and reducing the disability and mortality rates of stroke.

[0016] 2. This invention provides a distributed mobile smart healthcare remote system based on an active health model. It employs a multi-model ensemble trainer and a transfer learning adapter. By integrating multiple heterogeneous machine learning models and using a weighted fusion approach, a stroke risk prediction model is constructed. The transfer learning adapter utilizes domain adaptation technology through adversarial training, enabling the model to adapt to differences in data distribution across different regions. Even without labeled target domain data, it maintains stable performance, significantly enhancing the model's generalization ability. This allows the model to accurately predict stroke risk in different regions and with different data distributions, thereby improving the model's reliability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a schematic diagram illustrating the workflow of a distributed mobile smart medical remote system based on an active health model according to the present invention. Figure 2 This is a data flow diagram of a distributed mobile smart medical remote system based on an active health model according to the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1, such as Figure 1 , Figure 2 As shown, this invention provides a distributed mobile smart healthcare remote system based on an active health model, including a smart healthcare remote platform. The smart healthcare remote platform has the following communication connections: The multi-source data acquisition module is used to collect multi-source stroke data, including clinical and health data, from primary healthcare institutions, community health centers, and patients' personal health devices (smart bracelets, blood pressure monitors, etc.) in remote areas. This ensures the diversity and comprehensiveness of the data. The module also preprocesses the collected data, including cleaning, standardization, and formatting, handling missing and outlier values ​​to form a multi-source dataset. This allows for the extensive collection of multi-source stroke data, including stroke-related clinical and health data, from primary healthcare institutions, community health centers, and patients' personal health devices (smart bracelets, blood pressure monitors, etc.) in remote areas. The dataset includes multi-dimensional data covering patients' basic information, medical records, physiological indicators, and imaging data to ensure the diversity of data sources and the comprehensiveness of coverage. The collected multi-source stroke data undergoes centralized preprocessing, including data cleaning to remove duplicate, erroneous, or invalid records, and standardization to unify data formats and coding standards. Formatting operations ensure consistent data structures. The preprocessed multi-source stroke data is then inspected, and missing and outlier values ​​are addressed. Interpolation, estimation, or elimination methods are used to improve the dataset, resulting in a structured, high-quality multi-source dataset. The specific tasks of the multi-source data acquisition module are as follows: Firstly, it involves extensively collecting stroke-related data from multiple channels. This includes establishing partnerships with primary healthcare institutions and community health centers in remote areas to obtain clinical data on patients receiving treatment at these institutions, including symptom presentation, preliminary diagnosis, and treatment measures, reflecting the patient's initial disease condition and the status of primary care. Secondly, it integrates data generated by patients' personal health devices, such as real-time physiological data from smart bracelets and blood pressure monitors, collected through multiple channels. This data covers basic patient information (age, gender, and occupation), medical records (past medical history, family history, and treatment process), physiological indicators (blood pressure, blood sugar, and blood lipids), and imaging data (CT, MRI). Multi-dimensional data (such as...) supplements the deficiencies of traditional medical data, ensuring diverse data sources and covering different scenarios and levels to fully demonstrate the overall health status and disease progression characteristics of stroke patients. After collecting massive amounts of multi-source stroke data, it enters a centralized preprocessing stage. This includes data cleaning to screen for duplicate records, avoiding information redundancy that could affect analysis results; accurately identifying and removing erroneous or invalid records to ensure data accuracy and reliability; and standardizing the data format to ensure structural consistency across different sources. Simultaneously, standardized coding ensures that all types of data are classified and labeled according to unified standards, eliminating data comprehension barriers caused by coding differences. This is achieved through formatting operations... This process ensures that data adheres to unified standards during storage and processing, improving data quality and usability. After preprocessing, the multi-source stroke data is inspected, with a focus on missing and outlier values. Missing values ​​arise from oversights during data collection or equipment malfunctions, while outliers reflect data recording errors or patients' unique physiological conditions. For missing values, interpolation methods are used to infer the missing data based on data characteristics and actual needs, or estimations are made to fill in the missing data reasonably based on existing data. For outliers, their causes are analyzed in depth. If they are erroneous records, they are decisively removed; if they are indeed real special cases, they are retained and properly labeled, thus improving the dataset and forming a well-structured, high-quality multi-source dataset. The data generation enhancement engine constructs a virtual patient generator based on reinforcement learning algorithms and uses the GPT large model to convert the actions generated by the virtual patient generator into high-fidelity medical record text. At the same time, it calls medical knowledge graph to analyze physiological rationality. The data generation enhancement engine includes a synthetic data generation unit and a GPT large model text conversion unit. The synthetic data generation unit constructs a virtual patient generator based on a reinforcement learning algorithm using a deep Q-network. This generator produces synthetic data that matches the stroke characteristics of the target region, compensating for the lack of real data. The state is regional stroke statistics, and the action is generating a synthetic electronic medical record containing symptom and image descriptions. The reward is the improvement in the AUC (area under the curve) of the model trained on the synthetic data on real, remote regional data. The unit collects stroke epidemiological statistics for the target region, including incidence rate, age distribution, high-risk population characteristics, and common symptom distribution. This statistical data is then organized, encoded, and input into the deep Q-network-based reinforcement learning algorithm as the initial state. The deep Q-network makes decisions and generates actions based on the input initial state, specifically creating symptom descriptions of virtual patients. The system synthesizes electronic medical records based on imaging features. By adjusting network parameters, the generated actions are gradually made to match the real characteristics of stroke patients in the target area, resulting in a virtual patient generator. This generator produces synthetic data that matches the stroke characteristics of the target area. The generated synthetic data is used to train a logistic regression-based simulation evaluation model. The trained simulation evaluation model is then applied to real remote area data for validation. The AUC (area under the curve) improvement value of the simulation evaluation model is calculated as a reward. The AUC improvement value reflects the improvement effect of the synthetic data on the model performance. The parameters of the deep Q network are adjusted according to the reward to optimize the action generation strategy. This process is repeated to optimize the action generation process, generate high-quality synthetic data, and further improve the model's performance on real data. The specific tasks of the synthetic data generation unit are as follows: Collecting stroke epidemiological statistics for the target region, comprehensively covering key information such as incidence rate, age distribution, high-risk population characteristics, and common symptom distribution, reflecting the prevalence and population characteristics of stroke in the target region. After collection, systematic organization and encoding are performed. Organization ensures the integrity and accuracy of the data, while encoding converts different types of data into a computer-recognizable and processable format. The organized and encoded statistical data is then input into a reinforcement learning algorithm based on a deep Q-network as the initial state, providing the deep Q-network with a basic feature framework of stroke in the target region. After receiving the input initial state, the deep Q-network makes decisions and generates actions. By learning from a large amount of data and experience, the deep Q-network evaluates the value of different actions based on the current state, thereby selecting the optimal action. Specifically, the generated action involves creating a synthetic electronic medical record containing symptom descriptions and imaging feature descriptions of a virtual patient. During the generation process, the deep Q-network continuously adjusts its parameters to gradually align the generated actions with the real characteristics of stroke patients in the target region, resulting in a completed virtual patient generator. Through multiple iterations and... The virtual patient generator produces synthetic data that matches the characteristics of stroke in the target region, exhibiting high similarity to real patient data in terms of symptoms and imaging features. This generated synthetic data is proportionally divided into training and testing sets. The training set is used to train a logistic regression-based simulation assessment model, enabling the model to learn the characteristics and patterns of stroke. The loss function is minimized by continuously adjusting the model parameters. After training, the performance of the simulation assessment model is initially evaluated using the testing set. The trained model is then validated on real, remote region data. The AUC (area under the curve) improvement of the simulation assessment model is calculated as a reward. The AUC improvement directly reflects the improvement effect of the synthetic data on model performance. Based on the reward, the parameters of the deep Q-network are adjusted to optimize the action generation strategy. A high reward value indicates that the generated actions are beneficial to improving model performance, and the action strategy can be strengthened. A low reward value leads to adjustments and improvements to the strategy. This process is repeated to continuously optimize the action generation process, generating high-quality synthetic data and further improving the model's performance on real data. The expression for the AUC boost value is as follows: ; In the formula: The AUC improvement value is used to measure the effect of synthetic data on improving model performance; The AUC value is obtained by validating the model on real remote area data without using synthetic data for training. It represents the area under the curve of the logistic regression-based simulation evaluation model when it is validated on real remote area data without including the generated synthetic data that conforms to the characteristics of stroke in the target area in the training process. It reflects the model's ability to predict or classify stroke in real remote areas when trained using only raw data (which may have limited data volume or insufficient feature coverage). The AUC value, obtained by validating a model trained on synthetic data and then validating it on real-world remote area data, refers to the area under the curve (AUC) obtained by using synthetic data that matches the stroke characteristics of the target area, dividing the training and test sets according to a certain ratio, training a logistic regression-based simulation evaluation model using the training set, and then validating the model on real-world remote area data. It reflects the improvement in the model's ability to predict or classify stroke in real-world remote areas after training with synthetic data. A positive improvement occurs when the synthetic data provides the model with more useful information, enabling the model to better learn the characteristics and patterns of stroke. Greater than ,at this time Furthermore, as the quality of synthetic data continues to improve and its consistency with real data characteristics becomes increasingly better, Gradually increase This will also increase, indicating that the synthetic data has an increasingly significant effect on improving model performance; in the case of negative improvement, if the generated synthetic data has problems such as noise, large differences in distribution from the real data, or irrelevance, it may cause interference to the model during the learning process, making it... Less than ,Right now As the problems with synthetic data intensify, Further reduce, It will also become smaller (more negative), indicating that synthetic data has a negative impact on model performance; without improvement, when synthetic data does not play a substantial role in the model's learning, ,at this time This indicates that the introduction of synthetic data did not change the model's performance; The GPT large-scale model text conversion unit is used to convert generated actions into medical record text conforming to medical standards using the GPT large-scale model. It also calls upon a medical knowledge graph to analyze physiological rationality, ensuring logical consistency between symptom and imaging descriptions. The generated high-fidelity medical record text is directly input into the next process to address the scarcity of real data. It receives action information generated by a virtual patient generator, including symptom descriptions and imaging features of the virtual patient. Utilizing the text generation capabilities of the GPT large-scale model, and based on common medical text structures and expression habits, it initially converts the action information into a preliminary medical record text conforming to medical standards, invoking medical knowledge in the field of stroke. The medical knowledge graph inputs the initially generated medical record text prototype into the invoked medical knowledge graph. Based on the medical knowledge and logical relationships in the medical knowledge graph, it performs a physiological rationality analysis on the symptom descriptions and imaging features in the medical record text prototype, checks for any logical contradictions between the two, and ensures that the content of the medical record text conforms to medical reality and clinical rules. After the medical knowledge graph analysis, the medical record text with unreasonable parts is corrected and optimized to generate a high-fidelity medical record text that conforms to medical standards. This high-fidelity medical record text is then output and directly used as the input data for the next process, namely the input data for the intelligent risk prediction module, effectively solving the problem of scarcity of real data. The specific tasks of the GPT large-scale model text conversion unit are as follows: The GPT large-scale model text conversion unit receives action information generated by the virtual patient generator, accurately covering the virtual patient's symptom description and imaging features. Utilizing the text generation capabilities of the GPT large-scale model, it works according to common medical text structures and expression habits, initially converting the action information into a preliminary medical record text conforming to medical standards. This preliminary medical record text includes basic patient information, chief complaint, present illness, past medical history, physical examination, and auxiliary examination sections. Following this structure, the symptom description is refined to conform to clinical expression logic; imaging features are paraphrased to ensure the preliminary text conforms to medical standards in terms of format and basic content; and publicly available medical knowledge in the field of stroke is utilized. A medical knowledge graph is constructed by inputting the initially generated medical record text prototype into the medical knowledge graph. The medical knowledge graph contains medical knowledge and logical relationships in the field of stroke, including the relationship between symptoms and diseases, and the correspondence between imaging features and diseases. The GPT large model text conversion unit uses the medical knowledge graph to perform physiological rationality analysis on the symptom descriptions and imaging features in the medical record text prototype. This includes checking the matching between symptoms and imaging manifestations, as well as the logical rationality of disease development. It checks whether symptoms and imaging manifestations match each other and whether there are any logical contradictions. Some symptoms are only reasonable to appear under specific imaging features. For example, patients with cerebral hemorrhage usually show high-density shadows on head CT scans. If the medical record describes the patient as having symptoms of cerebral hemorrhage, but the imaging... If the imaging examination does not show a corresponding high-density shadow, there is a logical contradiction. Based on the disease development patterns in the medical knowledge graph, the symptom descriptions and imaging features in the medical record are checked to see if they conform to the natural progression of the disease. For example, a stroke patient may only present with mild limb weakness in the early stages, and the symptoms will gradually worsen as the disease progresses. If the medical record describes a patient with severe hemiplegia and altered consciousness in the early stages without other specific causes, it may not conform to the logical progression of the disease. If discrepancies are found, the medical knowledge graph is used for precise identification to ensure that the content of the medical record closely aligns with medical reality and clinical patterns, avoiding errors that violate medical common sense, and improving the quality and credibility of the medical record. Furthermore, medical knowledge in the field of stroke... The Stroke Atlas integrates and visualizes medical knowledge about stroke and its interrelationships in a structured manner. Its core content includes: disease classification, distinguishing between ischemic stroke (such as cerebral infarction) and hemorrhagic stroke (such as cerebral hemorrhage), and clarifying the pathological mechanisms of different types of stroke; symptoms and signs, listing typical symptoms of stroke (such as hemiplegia, hemianopsia, aphasia) and signs (such as facial drooping, limb weakness), and associating them with specific damaged brain regions; etiology and risk factors, integrating risk factors such as hypertension, diabetes, atherosclerosis, and atrial fibrillation, as well as the direct cause of vascular rupture or blockage; and imaging characteristics, describing the manifestations of stroke in imaging examinations such as CT and MRI (such as high-density shadows and low-density lesions), and associating them with disease type and severity.Treatment methods include thrombolysis, thrombectomy, medication, and rehabilitation training, clearly defining the applicable scenarios and effects of different treatments. The medical knowledge graph is constructed by extracting stroke-related knowledge from publicly available medical literature, electronic medical records, clinical guidelines, and authoritative databases (PubMed, CNKI), and extracting entities and relationships. Entities include diseases, symptoms, causes, examination methods, and treatment plans; relationships are "disease-symptom," "disease-cause," and "disease-treatment method." A tool (Gephi) is used to display entities and relationships as nodes and edges, forming an intuitive knowledge network. After analyzing the medical knowledge graph, if any inconsistencies are found in the medical record text, the GPT large-scale model text conversion unit performs correction and optimization. Based on the correct logic and knowledge in the medical knowledge graph, the symptom descriptions and imaging features are adjusted to ensure consistency. During the correction process, clinical realities and medical common sense are considered to ensure the accuracy of the modified content. After optimization, a high-fidelity medical record text conforming to medical standards is generated, which is directly used as input data for the next process's intelligent risk prediction module, effectively solving the problem of scarce real data. Furthermore, the medical knowledge graph includes: Disease-symptom relationship: Ischemic stroke → Symptoms: hemiplegia, hemianopsia, aphasia, sensory disturbance; Hemorrhagic stroke → Symptoms: severe headache, vomiting, altered consciousness, hemiplegia; Disease-cause relationship: Ischemic stroke → Causes: Atherosclerosis, cardioembolism, small vessel disease; Hemorrhagic stroke → Causes: Hypertension, aneurysm rupture, vascular malformation; Disease-Imaging Feature Relationship: Ischemic stroke (cerebral infarction) → CT findings: Low-density lesions can be seen 24 hours after onset; MRI findings: High signal on DWI sequence; Hemorrhagic stroke (cerebral hemorrhage) → CT findings: High-density shadows appear immediately after onset; MRI findings: High signal on T1WI, low signal on T2WI. Disease-treatment relationship: Ischemic stroke (within 4.5 hours of onset) → Treatment: Intravenous thrombolysis (e.g., alteplase); Hemorrhagic stroke (large hematoma) → Treatment: Surgical removal of hematoma or puncture and drainage of hematoma cavity; The intelligent risk prediction module, based on the generated high-fidelity medical record text and the patient's actual health data, constructs a stroke risk prediction model to predict and analyze stroke risk, and solves the problem of model adaptability in remote areas. The risk warning module, based on the prediction results of the stroke risk prediction model, integrates clinical data and end-point health data to analyze the stroke risk status of patients, provide early warning intervention for high-risk patients, and reduce the disability and mortality rates caused by stroke. The mobile remote collaboration module is used to connect primary hospitals with higher-level experts through 5G / IoT technology, establish a remote medical collaboration network, integrate medical resources at the city, county, township and village levels, empower primary doctors, and carry out regional management of stroke prevention and control.

[0021] Example 2, as Figure 1 , Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: the intelligent risk prediction module includes a multi-model ensemble trainer and a transfer learning adapter; The multi-model ensemble trainer, designed for stroke risk assessment, integrates multiple heterogeneous machine learning models and constructs a stroke risk prediction model through weighted fusion. It takes high-fidelity medical record text and the patient's actual health data as input and outputs a stroke risk prediction score, reducing the bias of single models, improving prediction accuracy, and adapting to data distribution differences across regions. It performs structured parsing of the input high-fidelity medical record text, extracting medical characteristics of stroke patients, and integrates patient health data, performing data cleaning, missing value imputation, and standardization. Through feature selection and dimensionality reduction techniques, it filters out risk features highly correlated with stroke risk and constructs a unified feature vector set. Medical features include blood pressure, blood glucose level, blood lipid levels, age, body mass index, duration of hypertension, duration of atrial fibrillation, and homocysteine ​​(Hcy) level. Blood pressure refers to systolic blood pressure (140 mmHg) and diastolic blood pressure (90 mmHg), a core risk factor for stroke. Hypertension damages vascular endothelium, increasing the risk of atherosclerosis and vascular rupture. Blood glucose level refers to fasting blood glucose (7.2...). Blood glucose levels (mmol / L) or glycated hemoglobin (6.8%) reflect long-term blood glucose control. Abnormal blood glucose levels in diabetic patients can accelerate vascular and nerve damage and affect cerebral blood circulation. Blood lipid indicators include total cholesterol (6.5 mmol / L), triglycerides (2.8 mmol / L), and low-density lipoprotein (LDL, 4.5 mmol / L).A blood pressure level of 2 mmol / L is considered high, as excessively high levels can lead to atherosclerosis and thrombosis, potentially blocking cerebral blood vessels. The risk of stroke increases significantly with age, especially in individuals over 65. Body mass index (BMI) is calculated by dividing weight (kg) by the square of height (m). A BMI ≥ 24 indicates overweight, and ≥ 28 indicates obesity. Overweight and obesity increase the risk of complications such as hypertension and diabetes, indirectly increasing the probability of stroke. The duration of a history of hypertension is recorded as the time since a patient was diagnosed with hypertension; poorly controlled blood pressure over a long period can lead to stroke. Continued damage to blood vessels significantly increases the risk of stroke; the duration of atrial fibrillation (AF) is the recorded duration of atrial fibrillation, which can lead to thrombus formation in the atria. A thrombus can break off and block cerebral blood vessels, causing ischemic stroke; homocysteine ​​(Hcy) level is a blood test value. Hcy ≥ 15 μmol / L indicates hyperhomocysteinemia, which damages the vascular endothelium, promotes thrombus formation, and increases the risk of stroke by 2-3 times; risk characteristics include blood pressure, blood glucose levels, blood lipid levels, body mass index, and homocysteine ​​levels. At the (Hcy) level, for stroke risk assessment tasks, a multi-model ensemble trainer selects various heterogeneous machine learning models, covering four types: logistic regression, random forest, support vector machine, and neural network. Each model is independently trained based on risk features in the feature vector set. Cross-validation is used to evaluate the performance of each model on the local dataset. The heterogeneous machine learning models are ranked according to accuracy and recall to ensure reliable predictive ability for each individual model. For each heterogeneous machine learning model, corresponding allowable values ​​are set based on the extracted risk features; deviations from the allowable values ​​are considered abnormal. The risk score of each heterogeneous machine learning model is then calculated by combining the risk features and their corresponding allowable values. Weights are assigned according to the ranking results of the heterogeneous machine learning models, with weights set to 0.4, 0.3, 0.2, and 0.1 from high to low. A weighted fusion method is used to construct a stroke risk prediction model. Combining the risk scores from each heterogeneous machine learning model and the assigned weights, a stroke risk prediction score is output, along with a pre-set risk threshold. To assess the risk of stroke in patients and improve the generalization ability of the model; The expression for the stroke risk prediction score is as follows: ; ; In the formula: A stroke risk prediction score. For risk scoring of logistic regression models, For the risk scoring of the random forest model, For risk scoring of support vector machine models, Risk scoring for neural network models, For the weights of the logistic regression model, The weights of the random forest model, For the weights of the support vector machine model, These are the weights of the neural network model; The risk score characterizing each heterogeneous machine learning model, i.e. , , or , The number of risk characteristics, For the first The actual value of each risk characteristic For the first The minimum permissible value for each risk characteristic. For the first The maximum permissible value for each risk characteristic; A value close to 0 indicates that the patient's risk is low, and most of the risk characteristics are within the acceptable range. A value close to 1 indicates that the patient has a higher risk, with most risk characteristics exceeding the allowable range. Greater than the risk threshold This indicates that the patient is in a high-risk state; A transfer learning adapter is used to fine-tune a stroke risk prediction model on unlabeled target domain data (rural areas) using adversarial training and domain adaptation techniques. This minimizes the difference in feature distribution between the source and target domains, where the source domain is labeled data from urban hospitals and the target domain is synthetic rural data. The adapter adapts the stroke risk prediction model pre-trained on urban data to the rural data distribution, enabling stroke risk prediction analysis for patients in the target domain. The transfer learning adapter collects data from both the source and target domains (labeled data from urban hospitals and synthetic rural data from rural areas). It then uses the stroke risk prediction model pre-trained on the source domain to extract shared risk features between the source and target domains and constructs a domain discriminator through a gradient inversion layer (GRL) for initial adversarial training. An initial feature extractor is used to initially align the risk feature distributions of the source and target domains. Adversarial training is then conducted between the feature extractor and the domain discriminator. The feature extractor generates confusing features through a gradient inversion layer, making it impossible for the discriminator to distinguish whether the risk features come from the source or target domain. At the same time, the main loss of the risk prediction task is minimized to ensure the discriminative power of the risk features for stroke prediction. Through iterative optimization, the difference in risk feature distributions between the two domains is gradually reduced. The parameters of the feature extractor after the adversarial training converges are fixed, and the pre-trained stroke risk prediction model is fine-tuned to adapt to the data distribution of the target domain. Finally, the stroke risk prediction model is directly applied to rural patient data to output a stroke risk prediction score without the need for target domain labeled data to participate in the training, thus achieving cross-domain adaptive prediction. The specific work of the transfer learning adapter is as follows: It integrates the data distributions of the source domain (annotated data from urban hospitals) and the target domain (synthetic data from rural areas). The transfer learning adapter extracts shared risk feature representations in an unsupervised manner. Specifically, it uses a stroke risk prediction model pre-trained on the source domain as the basic feature extractor, mapping the data from both domains to the same feature space. The adapter introduces a gradient inversion layer (GRL) to construct a domain discriminator, inserted between the feature extractor and the domain discriminator. This discriminator backpropagates the gradient inversion signal, directly transmitting risk features during forward propagation and inverting the gradient during backpropagation, forcing the feature extractor to generate confusing features, making it impossible for the discriminator to distinguish the feature source. This is achieved by minimizing the domain discriminant loss, ensuring the comparability of the two domain features in the initial stage. Based on the initial alignment, the transfer learning adapter initiates an adversarial training loop between the feature extractor and the domain discriminator. The feature extractor generates confusing features through the gradient inversion layer, attempting to "deceive" the domain discriminator. The domain discriminator improves its discriminative ability by optimizing the classification loss. The two form a zero-sum game, ultimately making the features generated by the feature extractor more discriminant. The algorithm approximates the intermediate state between the two domains, meaning it cannot be accurately classified by the domain discriminator. Simultaneously, the transfer learning adapter minimizes the main loss (cross-entropy loss) of the risk prediction task, ensuring that the extracted risk features still possess stroke discriminative power. Through iterative optimization, the transfer learning adapter gradually reduces the difference in risk feature distribution between the two domains, enabling the stroke risk prediction model to learn cross-domain universal feature representations even without labeled target domain data, avoiding performance degradation due to domain shift. After adversarial training converges, the feature extractor of the transfer learning adapter already possesses cross-domain feature alignment capabilities. With fixed feature extractor parameters, only the classification layer of the pre-trained stroke risk prediction model is fine-tuned. During the fine-tuning phase, only source domain labeled data is used, and the classification layer weights are adjusted through backpropagation to adapt to the target domain data distribution. Since the feature extractor has generated domain-invariant features, the classification layer only needs to learn a small number of domain-specific parameters to complete the adaptation. Finally, the transfer learning adapter directly applies the fine-tuned stroke risk prediction model to rural patient data, outputting a stroke risk prediction score without requiring target domain labeled data for training, significantly improving the generalization ability of the stroke risk prediction model in resource-limited scenarios. The risk warning module specifically includes: integrating the prediction results of the stroke risk prediction model, the patient's clinical data, and end-point health data; extracting the patient's stroke risk prediction score and corresponding risk characteristics; analyzing the stroke risk prediction score based on the set risk threshold to determine whether the patient's stroke risk is high; and marking abnormal risk characteristics based on the allowable values ​​of each risk characteristic. If the patient is high-risk, a level-one warning mechanism is triggered, issuing an alert to medical personnel and the patient, indicating the need for timely intervention. If the patient is not high-risk but has more than two abnormal risk characteristics, a level-two warning mechanism is triggered, issuing a warning to medical personnel and the patient, and generating a risk report. The specific functions of the risk warning module are as follows: It connects to the hospital's electronic medical record (EMR) system, wearable devices, and home medical terminals to collect patients' clinical and end-point health data. This data is then fused with the output of the stroke risk prediction model. The data is cleaned, normalized, and dimensionality reduced to extract each patient's stroke risk prediction score and related risk characteristics. Based on preset risk thresholds, the risk warning module analyzes the stroke risk prediction score to determine if the patient is in a high-risk state. Simultaneously, it marks abnormal risk characteristics based on the allowable values ​​for each risk feature. If a patient's stroke risk prediction score exceeds the set risk threshold, a level-one warning mechanism is triggered, immediately alerting medical personnel and the patient, indicating the need for timely medical intervention. For non-high-risk patients, if... If at least two abnormal risk characteristics are detected, a Level 2 warning mechanism is triggered, issuing warning alerts to relevant personnel and generating a detailed risk report. This ensures that patients at different risk levels receive appropriate warnings and interventions, improving the timeliness and effectiveness of stroke prevention. After the warning mechanism is triggered, the risk warning module issues corresponding alerts to medical personnel and patients. Level 1 warnings target high-risk patients, emphasizing urgency and prompting medical personnel to take swift action. Level 2 warnings target patients with some risk but not meeting the high-risk criteria, reminding them to pay attention to potential health problems. At the same time, a detailed risk report is generated to provide decision support for medical personnel and health guidance for patients, including the patient's stroke risk prediction score, abnormal risk characteristics, and corresponding medical advice to help patients manage their health status. The mobile remote collaboration module specifically includes: Primary hospitals collect patient clinical data, including vital signs (blood pressure, blood glucose, blood oxygen), imaging data (preliminary head CT / MRI images), and laboratory test results (coagulation function, blood lipid profile), via 5G / IoT terminals (portable ultrasound, mobile electrocardiographs, smart diagnostic kits, etc.). After preprocessing by edge computing devices, the patient clinical data is uploaded in real-time to the smart healthcare remote platform via the 5G network, automatically matching the HIS / EMR system interface to complete structured data conversion and encrypted storage. Simultaneously, primary care physicians input unstructured information such as patient medical history and medication records via a mobile app, extracting key fields using NLP technology to generate standardized electronic medical record templates, ensuring data integrity and standardization. The smart healthcare remote platform allocates consultation resources based on the patient's condition: village-level patients are initially seen by neurologists at county-level hospitals; if a level-one warning mechanism is triggered, a city-level expert collaborative consultation can be initiated, with higher-level experts accessing the primary care data in real-time via a high-definition video interaction system. The system analyzes multidisciplinary data to generate consultation opinions and supports online collaboration among teams. Experts from neurology, radiology, and rehabilitation departments jointly develop intervention plans, including medication adjustments, referral recommendations, or remote rehabilitation guidance. All consultation records are automatically archived in the patient's electronic health record (EHR), enabling full traceability. Primary care physicians implement intervention measures based on consultation opinions and regularly upload follow-up data (medication adherence, blood pressure control) via mobile devices. Simultaneously, a regional stroke prevention and control map is constructed using digital twin technology, displaying in real-time the number of inpatients and resource utilization rates (CT equipment idle rate, ambulance dispatch status) at each primary care institution. When an institution experiences resource shortages (insufficient emergency beds) or a public health emergency (mass hypertensive crisis), a resource allocation mechanism is triggered: coordinating with neighboring institutions to divert patients, allocating mobile stroke units (ambulances with CT and laboratory equipment) for support, or initiating remote surgical guidance. This ensures efficient collaboration of regional medical resources and improves the speed and coverage of stroke prevention and control response.

[0022] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A distributed mobile intelligent medical remote system based on an active health model, comprising an intelligent medical remote platform, characterized in that, The intelligent medical remote platform is communicatively connected with the following modules, wherein: A multi-source data acquisition module is configured to collect stroke multi-source data including clinical data and health data, and to pre-process the collected data to form a multi-source data set; A data generation enhancement engine is configured to construct a virtual patient generator based on a reinforcement learning algorithm, and to convert actions generated by the virtual patient generator into high-fidelity medical record texts by using a GTP large model, while calling a medical knowledge graph to analyze physiological rationality; An intelligent risk prediction module is configured to construct a stroke risk prediction model based on the generated high-fidelity medical record texts and actual health data of the patient, and to predict and analyze the stroke risk; A risk early warning module is configured to analyze the stroke risk status of the patient based on the prediction results of the stroke risk prediction model, and to perform early warning intervention on high-risk patients by fusing clinical data and end health data; A mobile remote collaboration module is configured to establish a remote medical collaboration network by using 5G / IoT technology, and to integrate three-level medical resources of city, county and village to empower primary doctors and perform regional management of stroke prevention and control. 2.The distributed mobile intelligent medical remote system based on the active health model of claim 1, wherein: The multi-source data acquisition module specifically includes: Stroke multi-source data is collected from primary medical institutions in remote areas, community health centers and personal health devices of patients, including clinical data and health data related to stroke, covering multi-dimensional data such as basic information, medical records, physiological indicators and image data of patients; The collected stroke multi-source data is centrally pre-processed, repeated, incorrect or invalid records are removed through data cleaning, and standardized processing is performed to unify data format and coding standards; The pre-processed stroke multi-source data is detected, and missing values and abnormal values in the stroke multi-source data are processed, and the data set is perfected by using interpolation, estimation or rejection method to form a structured multi-source data set. 3.The distributed mobile intelligent medical remote system based on the active health model of claim 2, wherein: The data generation enhancement engine includes a synthetic data generation unit and a GPT large model text conversion unit; The synthetic data generation unit constructs a virtual patient generator based on a reinforcement learning algorithm of a deep Q network to generate synthetic data conforming to the characteristics of stroke in the target area; The GPT large model text conversion unit is configured to convert the generated actions into medical record texts conforming to medical specifications by using a GTP large model, and to analyze the physiological rationality by calling a medical knowledge graph, and the generated high-fidelity medical record texts are directly input to the next process. 4.The distributed mobile intelligent medical remote system based on the active health model of claim 3, wherein: The synthetic data generation unit specifically includes: Stroke epidemiological data of the target area are collected, including incidence rate, age distribution, high-risk population characteristics and common symptom distribution, and the statistical data are sorted and coded and then input into the reinforcement learning algorithm based on the deep Q network as the initial state; The deep Q network makes decisions and generates actions based on the input initial state, specifically creating a synthetic electronic medical record containing symptom descriptions and image feature descriptions of virtual patients, and then adjusting network parameters to gradually adapt the generated actions to the real characteristics of stroke patients in the target area to obtain the constructed virtual patient generator; The generated synthetic data is used to train a simulation evaluation model based on logistic regression, and the trained simulation evaluation model is applied to real remote area data for verification, the AUC promotion value of the simulation evaluation model is calculated as a reward, and the parameters of the deep Q network are adjusted according to the reward, the action generation strategy is optimized, and the action generation process is optimized, so as to generate high-quality synthetic data. 5.The distributed mobile intelligent medical remote system based on the active health model of claim 3, wherein: The GPT large model text conversion unit specifically comprises: Receive the action information generated by the virtual patient generator, including the symptom description and imaging features of the virtual patient, use the text generation capability of the GPT large model, and convert the action information into a preliminary medical record text form according to the common structure and expression habits of medical texts; Call the medical knowledge graph in the field of stroke, input the preliminary generated medical record text form into the called medical knowledge graph, and analyze the physiological rationality of the symptom description and imaging features in the medical record text form based on the medical knowledge and logical relationship in the medical knowledge graph to check whether there is a logical contradiction between them; After the medical knowledge graph analysis, the medical record text with unreasonable places is corrected and optimized to generate a high-fidelity medical record text that meets the medical specifications, and then the high-fidelity medical record text is output as the input data of the next process, i.e., the input data of the intelligent risk prediction module. 6.The distributed mobile intelligent medical remote system based on the active health model of claim 3, wherein: The intelligent risk prediction module includes a multi-model integrated trainer and a transfer learning adapter. The multi-model integrated trainer integrates multiple heterogeneous machine learning models for stroke risk assessment tasks, and constructs a stroke risk prediction model through weighted fusion, inputs the high-fidelity medical record text and the actual health data of the patient, and outputs the stroke risk prediction score. The transfer learning adapter is used to fine-tune the stroke risk prediction model through the adversarial training domain adaptation technology to minimize the feature distribution difference between the source domain and the target domain, wherein the source domain is the urban hospital labeled data, and the target domain is the rural synthetic data, and then the stroke risk of the target domain patient is predicted and analyzed. 7.The distributed mobile smart medical remote system based on the active health model of claim 6, wherein: The multi-model integrated trainer specifically comprises: The input high-fidelity medical record text is structurally analyzed to extract the medical features of the stroke patient, and the patient's health data is integrated for data cleaning, missing value filling and standardization processing, and the risk features highly related to the stroke risk are selected through feature selection and dimension reduction technology to construct a unified format feature vector set, wherein the medical features include blood pressure value, blood sugar level, blood lipid index, age, body mass index, hypertension history length, atrial fibrillation duration and homocysteine level; the risk features include blood pressure value, blood sugar level, blood lipid index, body mass index and homocysteine level; For the stroke risk assessment task, the multi-model ensemble trainer selects multiple heterogeneous machine learning models, covering four types of logistic regression, random forest, support vector machine and neural network, independently trains based on the risk features in the feature vector set, evaluates the performance of each model on the local data set through cross-validation, and sorts the heterogeneous machine learning models according to the accuracy and recall rate indicators; For each heterogeneous machine learning model, set the corresponding allowed value based on the extracted risk features, that is, if the allowed value deviates, it is considered that the risk feature is abnormal, and then the risk score of each heterogeneous machine learning model is calculated by comprehensively considering each risk feature and the corresponding allowed value; According to the ranking results of the isomorphic machine learning models, weights are allocated, the ranking from high to low weights is set as 0.4, 0.3, 0.2 and 0.1, a weighted fusion method is used to construct a stroke risk prediction model, the stroke risk prediction score is output in combination with the risk scores in the isomorphic machine learning models and the allocated weights, and the risk threshold set in advance is combined to judge the risk of the patient suffering from stroke. , judge the risk of the patient suffering from stroke. 8.The distributed mobile smart medical remote system based on the active health model of claim 7, wherein: The transfer learning adapter specifically includes: The transfer learning adapter collects data of the source domain and the target domain, the source domain is the city hospital labeled data, and the target domain is the rural synthetic data, and then uses the stroke risk prediction model pre-trained on the source domain to extract the shared risk feature representation of the source domain and the target domain, and constructs a domain discriminator through a gradient reversal layer to initialize the feature extractor in an adversarial training manner, so that the risk feature distribution of the source domain and the target domain is preliminarily aligned; Adversarial training is performed between the feature extractor and the domain discriminator, the feature extractor generates confused features through the gradient reversal layer, so that the discriminator cannot distinguish whether the risk features come from the source domain or the target domain, while keeping the main loss of the risk prediction task minimized, and through iterative optimization, the difference between the risk feature distributions of the two domains is gradually reduced. Fix the parameters of the feature extractor after the convergence of the adversarial training, fine-tune the pre-trained stroke risk prediction model to adapt to the data distribution of the target domain, and finally apply the stroke risk prediction model directly to the rural patient data to output the stroke risk prediction score. 9.The distributed mobile smart medical remote system based on the active health model of claim 8, wherein: The risk warning module specifically includes: Integrate the prediction results of the stroke risk prediction model, the clinical data of the patient and the end health data, extract the stroke risk prediction score of the patient, and the corresponding risk features; According to the set risk threshold, analyze the stroke risk prediction score to determine whether the patient's stroke risk is high, and mark the abnormal risk features combined with the allowed value of each risk feature; If the patient is high-risk, trigger the first-level warning mechanism, and send an alarm to the medical staff and the patient, prompting timely intervention; if the patient is not high-risk, and there are at least two abnormal risk features, trigger the second-level warning mechanism, send a warning prompt to the medical staff and the patient, and generate a risk report. 10.The distributed mobile smart medical remote system based on the active health model of claim 1, wherein: The mobile remote collaboration module specifically includes: The primary hospital collects patient clinical data through a 5G / IoT terminal, the patient clinical data is preprocessed by an edge computing device, and then uploaded to the intelligent medical remote platform in real time through a 5G network, automatically matched with the HIS / EMR system interface, completed structured data conversion and encrypted storage, at the same time, the primary doctor enters the patient's medical history and non-structured information of drug record through the mobile terminal APP, and generates a standardized electronic medical record template; The smart medical remote platform allocates consultation resources according to the patient's condition. If the first level early warning mechanism is triggered, municipal expert collaborative consultation can be triggered. The superior experts can real-time access to basic data and form consultation opinions after analysis; The primary doctors implement intervention measures according to the consultation opinions and upload follow-up data regularly through mobile terminals. At the same time, the digital twin technology is used to build a regional stroke prevention and control map to real-time display the number of patients in hospital and resource utilization rate of each primary institution, and trigger the resource allocation mechanism.

Images