A stroke recurrence risk assessment and personalized management method and system

Abstract

A stroke recurrence risk assessment and personalized management method and system, the system comprises an application server, the application server comprises a database module, a recurrence risk dynamic prediction model module and a personalized management strategy generation module. The recurrence risk dynamic prediction model module is used to determine a long-term static risk index reflecting the inherent risk of the patient based on the clinical data stored in the database module, determine a short-term dynamic risk adjustment index reflecting the instantaneous risk contribution based on the dynamic feature sequence in the self-management data; by executing a risk coupling algorithm, the long-term static risk index and the short-term dynamic risk adjustment index are fused and calculated by using a weight coefficient to obtain a total risk index; the personalized management strategy generation module is used to convert the total risk index into a query vector, retrieve a matched guideline segment in a stroke management knowledge base; input the guideline segment as a constraint condition into a large language reasoning model to generate a personalized intervention scheme.

Description

Technical Field

[0001] This invention relates to the fields of smart healthcare and artificial intelligence technology, and in particular to a method and system for assessing and personally managing the risk of stroke recurrence. Background Technology

[0002] Stroke is a leading cause of death and disability in adults worldwide, and patients still face a very high risk of recurrence after their first rehabilitation and discharge. Clinical studies have shown that approximately 80% of recurrent events can be prevented through proactive intervention of risk factors. Therefore, establishing a precise system for predicting stroke recurrence risk and managing risk factors is of profound practical significance for secondary stroke prevention.

[0003] In current clinical practice, the assessment of stroke recurrence risk mainly relies on two types of methods: First, traditional risk assessment scales such as ABCD2, RRE-90, ESSEN, and SPI-II. These scales are typically used for one-time assessments upon hospitalization or discharge. While they provide references based on static indicators such as medical history and pathological classification, their static limitations mean they cannot reflect changes in the patient's health status over time after discharge, resulting in significant predictive lag. Second, monitoring models based on multiple regression or machine learning predict risk by analyzing patient physiological indicators, facial recognition, or imaging data. However, these methods often focus on capturing immediate fluctuations or single-dimensional characteristics, lacking in-depth consideration of the patient's medium- to long-term medical history and underlying inherent risks.

[0004] As mentioned above, existing technologies still face the following key challenges in the management of stroke recurrence: First, there is a serious disconnect between static assessment and dynamic monitoring. Static assessment alone is not timely, while dynamic monitoring lacks medium- to long-term baseline depth, making it difficult for the system to accurately characterize the complex evolutionary process of stroke recurrence, which is driven by both long-term medical history and short-term behavior.

[0005] Secondly, a closed-loop mechanism for generating prediction results and interventions has not yet been established. Existing systems mostly stop at outputting risk scores and fail to effectively map prediction results into specific clinical management pathways, resulting in patients lacking actionable guidance when facing warnings.

[0006] Furthermore, the medical authority of health management strategies is difficult to guarantee. When attempting to introduce large language models to generate suggestions, a lack of deep integration with authoritative medical guidelines can easily lead to "illusion" phenomena, resulting in interventions that often lack specificity and rigor, failing to meet the professional requirements of vertical medical fields.

[0007] Finally, due to the failure to fully incorporate the factors influencing the risk of patients’ short-term behavior adjustment, and the lack of effective behavioral incentive and feedback mechanisms, patients’ self-management compliance is generally low.

[0008] For example, CN111430029A discloses a multi-dimensional stroke prevention screening method and system based on artificial intelligence. This method collects clinical consultation data, blood biochemical indicators, and daily monitoring data, and combines this with a domain-specific medical knowledge base to provide early warnings and suggestions. While this technical solution integrates multi-dimensional data, it has core limitations in generating personalized intervention plans. First, there is a lack of deep logical coupling between its prediction results and intervention measures, failing to address how risk scores are automatically mapped to precise clinical intervention pathways. Second, the knowledge base mentioned in this technical solution is mostly based on pre-defined rule matching, lacking deep reasoning capabilities for complex individual states. Directly introducing general large models to generate suggestions faces a serious risk of "illusion," failing to ensure the medical authority and rigor of personalized intervention plans.

[0009] In summary, existing technologies struggle to balance long-term risk benchmarks with short-term dynamic fluctuations, and cannot generate personalized closed-loop management plans while ensuring medical rigor. Therefore, constructing a management system capable of coupling two-layer risk indices and generating rigorous, personalized intervention plans based on authoritative medical guidelines has become a critical issue urgently needing resolution in this field.

[0010] This invention aims to provide a method and system for assessing and personally managing the risk of stroke recurrence, thereby addressing current technical problems.

[0011] Furthermore, on the one hand, there are differences in understanding among those skilled in the art; on the other hand, the applicant studied a large number of documents and patents when making this invention, but due to space limitations, not all details and contents were listed in detail. However, this does not mean that the present invention does not possess the features of these prior art. On the contrary, the present invention already possesses all the features of the prior art, and the applicant reserves the right to add relevant prior art to the background art. Summary of the Invention

[0012] In current clinical practice, there is a significant disconnect between static assessment and dynamic monitoring. Traditional risk assessment scales (such as ABCD2 and ESSEN) are typically used only for a one-time assessment at discharge, exhibiting clear static limitations and predictive lags, failing to reflect changes in a patient's health status over time after discharge. While existing monitoring models can capture immediate fluctuations in physiological indicators, they often overlook the patient's medium- to long-term medical history and underlying inherent risks, making it difficult for the system to accurately characterize the complex evolutionary process of stroke recurrence driven by both long-term medical history and short-term behavior.

[0013] Secondly, existing systems have not yet formed a closed-loop mechanism for predicting outcomes and generating interventions. Most existing technologies only output risk scores and fail to effectively map the predictive results into specific, actionable clinical management pathways, resulting in patients lacking clear guidance when facing risk warnings. Furthermore, the medical authority of personalized intervention plans is difficult to guarantee. If general, large-scale reasoning models are directly introduced to generate suggestions without deep integration with authoritative medical guidelines, it is highly likely to produce a "hallucination" effect, leading to interventions that lack rigor and fail to meet the professional requirements of specific medical fields.

[0014] Furthermore, existing technologies are ineffective in improving patient self-management adherence. Because the systems fail to adequately incorporate quantifiable factors regarding the risk impact of short-term behavioral adjustments and lack effective behavioral incentives and psychological feedback mechanisms, patients are prone to burnout during long-term home-based rehabilitation. Existing health management interfaces are typically serious and dry, failing to provide positive guidance for behavioral improvement through intuitive and emotional visual feedback, thus resulting in generally low patient self-management adherence.

[0015] To address the shortcomings of existing technologies, this invention provides a stroke recurrence risk assessment and personalized management system from a first aspect. The system includes an application server, which comprises a database module, a recurrence risk dynamic prediction model module, and a personalized management strategy generation module. The database module stores clinical data and self-management data. The recurrence risk dynamic prediction model module determines a long-term static risk index reflecting the patient's inherent risk based on the clinical data stored in the database module, and determines a short-term dynamic risk adjustment index reflecting the immediate risk contribution based on the dynamic feature sequence in the self-management data. By executing a risk coupling algorithm, the long-term static risk index and the short-term dynamic risk adjustment index are fused using weighting coefficients to obtain a total risk index. The personalized management strategy generation module converts the total risk index into a query vector and retrieves matching guideline fragments from a stroke management knowledge base. The guideline fragments are then input as constraints into a large language inference model to generate a personalized intervention plan.

[0016] This technical solution overcomes the pain point of traditional models being unable to capture home risk fluctuations in real time through a dual-axis assessment system of static and dynamic adjustments. At the same time, it uses RAG (Retrieval Enhanced Generation) technology to inject authoritative guidelines as constraints into the large model, solving the illusion problem of large models in the medical field and ensuring that the generated intervention plan has both individual accuracy and clinical authority.

[0017] According to a preferred embodiment, the application server is connected to the interactive terminal, which receives and displays the total risk index and personalized intervention plan output by the application server; wherein, the interactive terminal includes a medical client and a mobile application, and uses a dynamic risk index display component to provide risk warnings.

[0018] This technical solution establishes a closed-loop information system between doctors and patients. Through a dynamic risk index display component, it enables the visualization of medical data, allowing medical staff to remotely monitor high-risk patients and enabling patients to intuitively perceive their risk status, thereby improving the efficiency of early warning information dissemination.

[0019] According to a preferred embodiment, the recurrence risk dynamic prediction model module determines the long-term static risk index in the following ways: discretization assessment, wherein the preset risk assessment period is divided into several discrete time steps; survival probability modeling, wherein, based on the patient's static characteristics, the recurrence risk probability at each time step is calculated, and the cumulative survival probability of remaining relapse-free within the assessment period is derived; the cumulative survival probability is used to determine the long-term static risk index representing the patient's individual pathological basis, which is used to characterize the inherent risk.

[0020] This technical solution employs survival probability modeling and discretized step-size analysis to scientifically address the nonlinear impact of individual patient characteristics on recurrence time, providing a robust long-term pathological risk benchmark for the system. According to a preferred implementation, the recurrence risk dynamic prediction model module determines and performs fusion of short-term dynamic risk adjustment indices through the following methods: temporal feature extraction, where a recurrent neural network unit with memory is used to process the dynamic feature sequence within a sliding time window to identify the immediate contribution of physiological index fluctuations and self-management behaviors to risk; nonlinear mapping, where a gating mechanism and activation function are used to transform temporal features into a dynamic adjustment index reflecting short-term risk variables; and two-layer weighted fusion, where, according to preset weight coefficients, the long-term static risk index reflecting the static pathological basis is linearly or nonlinearly coupled with the short-term dynamic risk adjustment index reflecting real-time performance to generate a total risk index.

[0021] This technical solution utilizes recurrent neural networks (RNNs) and their memory properties to capture the trend fluctuations (rather than single-point abnormalities) of patients' blood pressure, blood sugar, and other indicators. Combined with a two-layer weighted fusion algorithm, the final total risk index can more objectively and sensitively reflect the surge in short- to medium-term risk caused by recent behavioral changes in patients.

[0022] According to a preferred embodiment, the personalized management strategy generation module retrieves guide fragments in the following manner: The patient's current risk level, abnormal physiological indicators, and historical baseline characteristics are transformed into a query vector in a high-dimensional space. In a knowledge base containing pre-stored fragment vectors of medical guidelines and instructions, a vector similarity algorithm is used to calculate the correlation index between the query vector and each fragment vector. Based on a preset similarity threshold, multiple guideline fragments that meet the conditions are automatically filtered and recalled from the knowledge base as authoritative basis for subsequent intervention logic.

[0023] This technical solution, based on semantic vector retrieval technology, surpasses traditional keyword matching. It can deeply analyze the inherent logical relationship between patients' conditions and medical literature, achieving minute-level accurate retrieval of massive guideline fragments and ensuring the evidence-based medicine foundation of personalized intervention plans.

[0024] According to a preferred embodiment, the specific steps of the personalized management strategy generation module in generating personalized intervention plans through a large language reasoning model include: assembling the recalled guideline fragments, the patient's total risk index, and abnormal physical sign information into structured prompts according to a preset medical logic template; injecting the structured prompts into the large language reasoning model and using the guideline fragments as knowledge constraints to enable the large language reasoning model to perform natural language reasoning within the scope of authoritative medical knowledge to generate personalized intervention plans; and distributing the generated personalized intervention plan, which includes at least one of health management tips, dietary recommendations, or exercise prescriptions, to the interactive terminal for visual feedback and behavioral guidance via a downlink.

[0025] This technical solution enables the automated transformation from complex data to natural language suggestions. The structured prompt word engineering effectively limits the scope of reasoning, enabling the system to provide standardized suggestions that conform to medical guidelines, as well as differentiated behavioral guidance based on individual patient circumstances.

[0026] According to a preferred embodiment, the system further includes a mobile application terminal that communicates with the application server. The mobile application terminal has a real-time feedback interface, which includes a dynamic risk index display component for synchronously displaying the achievement status of multiple health management dimensions obtained based on the total risk index decomposition. The dynamic risk index display component dynamically represents the real-time achievement status of each dimension in blood pressure, blood sugar, medication, exercise, and diet through changes in the geometric features of the graphic area. The dynamic risk index display component uses different color features to correspond to different risk levels based on risk warning coloring logic, so as to generate an incentive effect on the patient's behavior improvement through visual feedback.

[0027] This technical solution reduces the cost for patients to understand health indicators through a multi-dimensional visualization scheme using geometric features and color coding. It forms a closed loop of "behavior-outcome-perception" by using real-time visual feedback, which effectively improves patients' compliance with home management.

[0028] According to a preferred embodiment, the mobile application communicating with the application server achieves closed-loop management in the following ways: based on the collection frequency and numerical fluctuation of self-management data, an early warning indicator for the indicator to be retested or the timeliness of the data is automatically triggered in the real-time feedback interface; the total risk index is displayed in real time in the real-time feedback interface, and an incentive flag is generated based on the number of consecutive days the patient meets the target, so as to enhance the patient's self-management compliance through psychological feedback; the patient is guided to check in for health through the interactive entry in the real-time feedback interface, and the collected behavioral data is continuously fed back to the database module for continuous dynamic evaluation by the relapse risk dynamic prediction model module.

[0029] This technical solution constructs a self-driven mechanism for home management through data timeliness early warning and psychological incentive markers, ensuring a continuous inflow of data, thereby achieving mutual iteration and continuous optimization between model prediction and actual behavior.

[0030] According to a preferred embodiment, the mobile application communicating with the application server has a real-time feedback interface. The real-time feedback interface has a dynamic risk index display component at its center, used to visually present the progress of health management goals. The dynamic risk index display component consists of a central circular core and several surrounding geometric regions, which are defined as the five core dimensions of secondary stroke prevention: blood pressure, blood sugar, medication, exercise, and diet. The colored area within each geometric region is dynamically determined by the patient's completion rate of the corresponding indicator's check-in goals. When the completion rate of a certain indicator increases, the color fill ratio of the corresponding geometric region increases accordingly. The dynamic risk index display component provides risk warnings through dynamic adjustments of facial expressions within the central circular core and changes in the color saturation of the geometric regions. The facial expressions change dynamically with the total risk index, and the colors of the geometric regions use different hues to distinguish between the target achievement status and high-risk warnings, thereby forming positive incentive feedback for behavioral improvement.

[0031] The dynamic risk index display component transforms serious and dry secondary prevention indicators into vivid images of vital signs through advanced anthropomorphic and emotional design. By dynamically changing expressions and colors, it deeply connects with patients' emotions, greatly reducing patients' professional burnout caused by long-term disease management.

[0032] This invention provides a method for assessing and personally managing the risk of stroke recurrence from a second aspect. The method includes: determining a long-term static risk index reflecting the patient's inherent risk based on stored clinical data, and determining a short-term dynamic risk adjustment index reflecting the immediate risk contribution based on dynamic feature sequences in self-management data; calculating a total risk index by fusing the long-term static risk index and the short-term dynamic risk adjustment index using a risk coupling algorithm and weighting coefficients; converting the total risk index into a query vector and retrieving matching guideline fragments from a stroke management knowledge base; and inputting the guideline fragments as constraints into a large language reasoning model to generate a personalized intervention plan.

[0033] This method constructs a complete closed-loop management logic for outpatient health, organically combining clinical expertise with large-scale model generation capabilities, and achieves low-cost, high-efficiency, and medically credible personalized secondary prevention management of stroke. Attached Figure Description

[0034] Figure 1 This is a simplified structural diagram of the stroke recurrence risk assessment and personalized management system provided by the present invention; Figure 2 This is a flowchart illustrating the stroke recurrence risk assessment and personalized management method provided by the present invention. Figure 3 This is the dynamic risk index diagram of the five-petaled flower provided by the present invention; Figure 4 This is a schematic diagram illustrating the logical principle of the stroke recurrence risk assessment and personalized management system provided by the present invention.

[0035] List of reference numerals 101: Healthcare Client; 102: Mobile Application; 200: Application Server; 301: Database Module; 302: Dynamic Recurrence Risk Prediction Model Module; 303: Personalized Management Strategy Generation Module; 401: Clinical Data; 402: Self-Management Data; 501: BRS Model; 502: DRM Model; 503: Risk Fusion Calculation Model; 601: Stroke Management Knowledge Base; 602: Large Language Reasoning Model. Detailed Implementation

[0036] The following is a detailed explanation with reference to the accompanying drawings.

[0037] Static features (X) staticStatic characteristics refer to patient features that are determined at a specific assessment starting point (such as the date of discharge) and do not change over time or change very slowly over time. Static characteristics include the TOAST classification and the NIHSS score. The TOAST classification is an etiological classification method for ischemic stroke, used to differentiate ischemic stroke from a medical etiological perspective, such as large artery atherosclerosis type and cardioembolic type; the NIHSS score is the National Institutes of Health Stroke Scale, which assesses the severity of neurological impairment in patients through quantitative scoring and reflects the baseline condition level of patients at the assessment starting point.

[0038] Dynamic feature sequence (x) t ): Time-series data collected through wearable devices or home monitoring devices.

[0039] Input Tensor: The data form of physiological indicators (blood pressure, blood sugar, heart rate, blood lipids, medication behavior, etc.) within the sliding time window W after being matrixed into a multidimensional matrix.

[0040] The Baseline Risk Score (BRS) 501 model is a long-term static risk assessment model used to comprehensively characterize the risk of relapse in patients. This long-term static risk assessment model is built upon a survival analysis framework and mainly includes two core elements: the discrete-time hazard function and the cumulative survival probability. The discrete-time hazard function (h... t The conditional probability of a relapse occurring in month t is given that the patient has survived to month t-1. The cumulative survival probability S(τ) reflects the probability that the patient remains relapse-free for a continuous period of τ months. It is an important output indicator in survival analysis and is used to characterize the long-term risk level.

[0041] The DRM (Dynamic Risk Modification) model 502 is a short-term dynamic risk adjustment model designed to dynamically correct risk assessment results by incorporating time-series information. This model comprises two key components: LSTM recurrent units and nonlinear activation functions. An LSTM recurrent unit is a recurrent neural network unit with forget gates, input gates, and output gates, enabling effective modeling and memorization of long-term physiological characteristics. Nonlinear activation functions (such as Sigmoid or Tanh) are used to map neuron outputs to predetermined numerical ranges (e.g., [0,1]) to achieve gating control or generate probabilistic risk outputs.

[0042] BGE-M3 Embedding Model: A deep learning model for achieving high-dimensional vectorization of medical text.

[0043] Guideline Excerpt: To meet the accuracy requirements of vector retrieval, the medical guidelines are physically segmented according to a granularity of 256-512 tokens.

[0044] Cosine Similarity Retrieval: By calculating the angle between vectors in the vector space, it matches the most relevant medical intervention suggestion fragment to the patient's current risk status.

[0045] Example 1 In the current clinical application system, there is a lack of effective connection between risk assessment and continuous health monitoring. Traditional stroke risk assessment scales, such as ABCD2 and ESSEN, are mostly scored once at the patient's discharge point. Their assessment results are essentially static judgments, which are difficult to reflect the dynamic evolution of post-discharge health status in a timely manner and also suffer from predictive time lag. On the other hand, although some existing monitoring models can perceive short-term changes in physiological parameters in real time, they fail to fully integrate the long-term risk background formed by the patient's past medical history and ignore the coupling effect between individual inherent risk and recent behavioral changes. This makes it difficult for the system to comprehensively depict the complex evolutionary process of stroke recurrence driven by both long-term disease course and short-term factors.

[0046] Meanwhile, current technological solutions generally lack a closed-loop support mechanism from risk prediction to intervention implementation. Most systems only output quantitative risk scores, failing to further translate the prediction results into clear and actionable clinical management or intervention pathways, leaving patients without specific guidelines to follow after receiving risk warnings. More critically, the medical reliability of intervention recommendations is difficult to guarantee. If intervention plans are generated directly based on the general large language reasoning model 602 without deep alignment with authoritative clinical guidelines and evidence-based medicine systems, it is highly susceptible to model "illusion," resulting in unrigorous or even inapplicable recommendations that fail to meet the professionalism and safety requirements of the medical vertical field.

[0047] Furthermore, existing technologies have limited effectiveness in promoting long-term patient self-management. On the one hand, the system fails to effectively quantify the impact of short-term behavioral adjustments on risk changes; on the other hand, it lacks continuous behavioral incentives and psychological feedback mechanisms, making patients prone to burnout during home rehabilitation and long-term management. Current health management interfaces are mostly serious and rational, lacking intuitive and emotionally guiding visual feedback, making it difficult to positively reinforce patients' active behaviors, thus leading to generally low overall self-management adherence.

[0048] This embodiment provides a stroke recurrence risk assessment and personalized management system, whose physical hardware connections and model functions are as follows: Figure 1 As shown.

[0049] like Figure 1 As shown, the stroke recurrence risk assessment and personalized management system consists of a front-end interaction layer, a network transmission layer, and a back-end central server cluster. The physical hardware of the front-end interaction layer is the interaction terminal. For example... Figure 1 As shown, the interactive terminals include a medical client 101 (such as a medical workstation PC) used by medical staff and a mobile application terminal 102 (such as a smartphone or tablet) held by the patient. Figure 4 As shown, the interactive terminal uses a dynamic risk index display component to provide risk warnings.

[0050] On the data source side, the mobile application 102 physically connects to home monitoring devices (such as smart blood pressure monitors and heart rate recorders) via Bluetooth Low Energy (BLE) or WiFi protocol to acquire dynamic feature sequences x in real time. t Application server 200 establishes a communication connection with the interactive client. For example... Figure 1 As shown, the physical hardware of the aforementioned front-end interaction layer is connected to the back-end application server 200 via an encrypted internet link. The interaction terminal receives and displays the total risk index and personalized intervention plan output by the application server 200.

[0051] like Figure 1 As shown, the application server 200 is connected to the database module 301, which has independent storage space and is in a data read-write interaction relationship with the recurrence risk dynamic prediction model module 302, via a high-speed local area network bus. This enables the database module 301 to act as an independent data persistence node and exchange data with the application server 200 at high throughput.

[0052] During the clinical data acquisition phase, the medical client 101 writes to the clinical data storage area 401 of the database module 301 under the condition of encrypted transmission channel.

[0053] During the risk assessment initiation phase, upon receiving a trigger instruction, the application server 200 establishes a batch data reading relationship with the database module 301 to obtain the static and dynamic feature sequences to be processed.

[0054] During the intervention plan generation phase, the personalized management strategy generation module 303 establishes an instruction distribution relationship with the mobile application terminal 102 according to the downlink communication link protocol. The data flow within the stroke recurrence risk assessment and personalized management system forms a link of dual-track data input and closed-loop decision output. For example... Figure 1 As shown, the medical client 101 will display the static characteristics X of the patient at the time of discharge. staticData such as TOAST classification and NIHSS score are transmitted and stored in the clinical data storage area 401 of the database module 301. Simultaneously, daily data collected from the patient, such as blood pressure, blood sugar, diet, and medication, are uploaded in real-time via the mobile application terminal 102 to the self-management data storage area 402 of the database module 301. The application server 200 retrieves the clinical data 401 and self-management data 402 from the aforementioned clinical data storage area 401 and self-management data storage area, respectively, and transmits them to the relapse risk dynamic prediction model module 302.

[0055] like Figure 4 As shown, the relapse risk dynamic prediction model module 302 is used to determine a long-term static risk index that reflects the patient's inherent risk based on the clinical data 401 stored in the database module 301.

[0056] The recurrence risk dynamic prediction model module 302 determines a short-term dynamic risk adjustment index reflecting the immediate risk contribution based on the dynamic feature sequence in the self-managed data 402. The recurrence risk dynamic prediction model module 302 calculates the total risk index by executing a risk coupling algorithm and using weighting coefficients to fuse the long-term static risk index and the short-term dynamic risk adjustment index.

[0057] To address the technical issue of hallucinations in existing large-scale medical models, the personalized management strategy generation module 303 implements restricted reasoning through structured cue word engineering.

[0058] The personalized management strategy generation module 303 has a vector transformation unit, a knowledge retrieval unit, and / or a reasoning constraint unit. The vector transformation unit and the knowledge retrieval unit can establish a cosine similarity matching relationship based on high-dimensional semantic space when the patient risk profile is in an unstructured text state, so that the personalized management strategy generation module 303 is anchored to the authoritative guidelines in the stroke management knowledge base 601.

[0059] like Figure 4 As shown, the personalized management strategy generation module 303 is used to convert the total risk index into a query vector and retrieve matching guide fragments from the stroke management knowledge base 601; the guide fragments are then input as constraints into the large language reasoning model 602 to generate personalized intervention plans.

[0060] Specifically, the personalized management strategy generation module 303 achieves closed-loop generation of intervention plans through the large language reasoning model 602. This involves assembling the recalled guideline fragments, the patient's total risk index, and abnormal physical signs information into structured prompts according to a preset medical logic template; injecting these structured prompts into the large language reasoning model 602, and using the guideline fragments as knowledge constraints, enabling the large language reasoning model 602 to perform natural language reasoning within the scope of authoritative medical knowledge to generate personalized intervention plans; and distributing the generated personalized intervention plan, which includes at least one of health management tips, dietary recommendations, or exercise prescriptions, to the interactive terminal via a downlink for visual feedback and behavioral guidance. Preferably, the personalized intervention plan may also include a follow-up visit warning.

[0061] The specific processing flow of the personalized management strategy generation module 303 is as follows.

[0062] The Large Language Inference Model 602 is used to perform natural language generation tasks within a constrained knowledge domain. A vector retrieval engine is used to retrieve Top-K relevant fragments from massive amounts of medical literature, and structured cue word templates are used to constrain the inference boundaries.

[0063] When the patient's total risk index triggers the warning threshold, the personalized management strategy generation module 303, with the support of the BGE-M3 embedding model, establishes a retrieval relationship with the stroke management knowledge base 601 based on vector similarity, and recalls guideline fragments directly related to the current abnormal indicators.

[0064] When constructing input structured prompts, the structured prompt template is embedded and spliced ​​with the recalled guide segments under the preset logical framework (Role-Context-Constraint-Task), forcing the model to generate suggestions only based on the recalled content.

[0065] And / or, in cases where the generated recommendations involve sensitive medical behaviors (such as medication dosage), the large language reasoning model 602 associates with the disclaimer and medical advice library according to a preset safety alignment strategy to generate non-diagnostic health management recommendations rather than direct prescriptions.

[0066] Finally, the total risk index and intervention recommendations generated by the recurrence risk dynamic prediction model module 302 are distributed via a downlink. The application server 200 sends them to the healthcare client 101 for clinical review, and they are simultaneously transmitted to the mobile application terminal 102 for visual feedback.

[0067] Example 2 This embodiment is a further improvement on embodiment 1, and repeated content will not be described again.

[0068] This system uses Figure 1The physical architecture shown includes a backend application server 200 comprising a database module 301 (storing clinical data 401 and self-management data 402), a recurrence risk dynamic prediction model module 302 (including a BRS model 501 and a DRM model 502), and a personalized management strategy generation module 303 (including a stroke management knowledge base 601 and a large language reasoning model 602). Preferably, the personalized management strategy generation module 303 can be deployed locally or invoked as a remote server.

[0069] The specific processing steps of this system are as follows.

[0070] S101: The medical team enters patient clinical data 401.

[0071] like Figure 2 As shown, the process begins with the initialization of clinical data 401. Healthcare professionals collect the patient's static characteristics X at discharge using the healthcare client 101. static The data is stored in the clinical data area 401 of database module 301. This data includes demographic indicators, TOAST classification, NIHSS score, past medical history, and imaging indicators such as carotid artery stenosis. Application server 200 performs one-hot encoding preprocessing on the classification indicators to provide standard tensor input for subsequent modeling.

[0072] S102: Calculate the patient's long-term static risk index using the long-term risk assessment model for stroke recurrence (BRS model 501).

[0073] Preferably, the recurrence risk dynamic prediction model module 302 divides the preset risk assessment period into several discrete time steps; based on the patient's static characteristics, it calculates the recurrence risk probability at each time step and derives the cumulative survival probability of not recurring within the assessment period; and uses the cumulative survival probability to determine the long-term static risk index representing the patient's individual pathological basis.

[0074] like Figure 1 and Figure 2 As shown, the relapse risk dynamic prediction model module 302 calls the BRS model 501. The application server 200 uses a survival analysis algorithm to set the maximum assessment period of 36 months... The dataset is discretized into 36 time steps, with each month as the step size, to construct a "human-cycle" dataset. The system calculates the discrete-time risk function h for each time step. t And then according to the formula Obtain the cumulative survival probability S(τ) over τ consecutive months, and calculate the long-term static risk index. This step defines the inherent risk of recurrence in patients.

[0075] In the above formula, h represents the discrete-time risk function. In this context, j is a multiplicative step-counting variable, representing each specific time month that increments sequentially from 1 to τ, used to extract the risk value for that month. . This represents the survival probability (i.e., the probability of no recurrence in that month) at the j-th time step (month).

[0076] This represents the calculation function for long-term static risk scoring, where T represents the preset risk assessment period. Indicates the time step or observation point.

[0077] S103: Initial health management goals are generated.

[0078] like Figure 2 As shown, the system automatically matches and generates initial management goals in the database based on the BRS risk level calculated by S102. These goals cover five dimensions: blood pressure, blood glucose, medication adherence, exercise volume, and dietary structure, serving as the benchmark for subsequent dynamic monitoring. Specifically, the five core dimensions are defined as the five core dimensions of secondary stroke prevention: blood pressure, blood glucose, medication, exercise, and diet.

[0079] S104: Patient self-monitoring and data recording.

[0080] like Figure 1 As shown, patients use mobile application 102 to record their behavior at home. Physiological indicators (such as blood pressure and blood sugar) are synchronized in real time from the wearable device via Bluetooth protocol, while behavioral indicators (such as diet, exercise, and medication) are manually entered by the patient. All data is reported in real time to the self-management data 402 area of ​​database module 301.

[0081] S105: Dynamic risk calculation.

[0082] The recurrence risk dynamic prediction model module 302 determines the short-term dynamic risk adjustment index and performs fusion in the following manner.

[0083] By utilizing recurrent neural network units with memory capabilities, dynamic feature sequences within a sliding time window are processed to identify the immediate contribution of physiological index fluctuations and self-management behaviors to risk; through gating mechanisms and activation functions, time-series features are transformed into dynamic adjustment indices reflecting short-term risk variables.

[0084] like Figure 1 and Figure 2As shown, the relapse risk dynamic prediction model module 302 calls the DRM model 502. The DRM model 502 computation nodes extract the dynamic feature sequence within the sliding time window W, and receive input through a recurrent neural network unit with memory capabilities (e.g., an LSTM recurrent unit). The hidden layers of the LSTM recurrent unit operate through a gate mechanism: the forget gate and the input gate determine information retention through the Sigmoid activation function, and the cell state update uses the Tanh function to handle nonlinear mapping, thereby automatically learning the immediate contribution of missed medication or blood pressure fluctuations to the risk, and finally outputting a short-term dynamic risk adjustment index.

[0085] S106: Risk Integration and Decision Making.

[0086] The recurrence risk dynamic prediction model module 302, based on preset weighting coefficients, linearly or nonlinearly couples the long-term static risk index, which reflects the static pathological basis, with the short-term dynamic risk adjustment index, which reflects the real-time performance, to generate a total risk index.

[0087] The relapse risk dynamic prediction model module 302 does not employ the method disclosed in existing technologies (such as CN114203295A) that simply uses a deep neural network (CNN-LSTM) for full-data black-box fusion, but instead uses an interpretable two-layer architecture coupling algorithm. In this architecture, the risk fusion calculation model 503 is used to perform weighted logistic regression calculation on the long-term static risk index (BRS) representing the patient's inherent pathological basis and the short-term dynamic risk adjustment index (DRM) representing the patient's recent behavioral fluctuations.

[0088] like Figure 2 As shown, the system executes a dual-layer coupling algorithm. Dual-layer coupling refers to the mechanism of mathematically fusing "long-term inherent risk" based on historical statistics with "short-term adjusted risk" based on real-time fluctuations. Risk fusion calculation model 503 uses system weight coefficients α (ranging from 0.6 to 0.8) to perform risk fusion calculations on the long-term static risk index of BRS model 501 and the short-term dynamic risk adjustment index of DRM model 502, obtaining the total risk index R. total .

[0089] Specifically, the risk fusion calculation model 503 has a long-term static input port, a short-term dynamic input port, and a weight adjustment unit. The long-term static input port and the BRS model 501 can establish a baseline mapping relationship based on survival probability under the condition of an assessment period of T (such as 36 months), so that the risk fusion calculation model 503 can be fixed to the patient's pathological classification baseline.

[0090] In order to resolve the technical contradiction of the chaotic weight calculation logic in the prior art, in this embodiment, the risk fusion calculation model 503 follows the following linear or nonlinear mathematical logic when executing the risk coupling algorithm.

[0091] .

[0092] In the above formula, The weighting coefficient is an adjustable parameter used in clinical practice to balance the influence of individual constitution (static) and recent performance (dynamic) on the final conclusion. The overall risk index represents the patient's overall probability of recurrence at the current moment. This represents a long-term static risk score. This represents the short-term dynamic risk adjustment index.

[0093] In weighting coefficients (Value range 0.6-0.8) is set to favor long-term pathological basis, and the risk fusion calculation model 503 establishes a dominant benchmark anchoring relationship with the BRS model 501 under the condition of updated clinical data 401, so as to prevent noise from short-term data from interfering with the overall assessment.

[0094] In weighting coefficients When configured to respond to immediate behavioral fluctuations, the DRM model 502 dynamically adjusts its relationship with the risk fusion calculation model 503 under the condition that a continuous abnormal sequence appears in the self-management data 402, so as to improve the overall risk index R. total It can keenly reflect immediate risk shifts such as medication interruption or sudden changes in blood pressure.

[0095] And / or, in Under the condition of dynamic decay over time, the risk fusion calculation model 503 and the DRM model 502 exhibit a progressively enhanced coupling relationship according to the time step progression, thus objectively reflecting the medical fact that the longer the hospital stay, the greater the weight of self-management behavior on the risk of relapse. In this way, the present invention effectively overcomes the deficiency of the prior art CN114203295A in being unable to distinguish the contribution of physical condition and behavior to risk, and achieves interpretable decomposition of risk sources.

[0096] In this step, the personalized management strategy generation module 303 simultaneously extracts query features. The system extracts the patient's current abnormal features (such as abnormal systolic blood pressure) and risk level as query vectors, and performs vector retrieval in the stroke management knowledge base 601, which stores professional literature such as the "Guidelines for Secondary Prevention of Stroke," using the BGE-M3 model.

[0097] The personalized management strategy generation module 303 retrieves guide fragments in the following ways.

[0098] The patient's current risk level, abnormal physiological indicators, and historical baseline characteristics are transformed into a query vector in a high-dimensional space. In a knowledge base containing pre-stored fragment vectors of medical guidelines and instructions, a vector similarity algorithm is used to calculate the correlation index between the query vector and each fragment vector. Based on a preset similarity threshold, multiple guideline fragments that meet the conditions are automatically filtered and recalled from the knowledge base as authoritative basis for subsequent intervention logic.

[0099] Specifically, the personalized management strategy generation module 303 extracts the patient's current total risk index R from the database. total Key abnormal physiological indicators (such as systolic blood pressure higher than 150 mmHg for 3 consecutive days) and historical baseline characteristics are converted into query vectors.

[0100] The personalized management strategy generation module 303 calls the built-in BGE-M3 embedding model to perform high-dimensional spatial mapping on the query vector, and then performs a retrieval in the stroke management knowledge base 601 (vector database). The stroke management knowledge base 601 uses the cosine similarity algorithm to calculate the cosine value of the angle between the query vector and the pre-stored medical guideline fragment vectors (256-512 tokens) in the database.

[0101] The stroke management knowledge base 601 automatically filters and recalls the top K (e.g., Top-5) most relevant text fragments from the "Guidelines for Secondary Prevention of Stroke" or drug instructions as guideline fragments based on a preset similarity threshold (>0.75), and feeds them back to the personalized management strategy generation module 303. This process ensures that the subsequently generated strategies are not only based on algorithmic predictions, but also supported by authoritative medical literature.

[0102] Specifically, the personalized management strategy generation module 303 inputs the unstructured patient risk profile into the BGE-M3 model and outputs a standardized query vector Q. The stroke management knowledge base 601 calculates Q in the vector space and compares it with all pre-stored fragmented vectors D in the database. i The cosine value.

[0103] The stroke management knowledge base 601 calculates the query vector Q and the medical guideline fragment vector D using the following mathematical formula. i The cosine of the angle between them. The formula for calculating the cosine of the angle is: .

[0104] Q represents the query vector, which is a 1024-dimensional high-dimensional semantic vector generated by the personalized management strategy generation module 303 after extracting the patient's current total risk index and key abnormal indicators (such as systolic blood pressure higher than 150 mmHg for 3 consecutive days). D represents the query vector. iThe segmented vector is a 1024-dimensional segmented vector generated by pre-segmenting medical guidelines and drug instructions into 256-512 token size segments using the stroke management knowledge base 601, and then transforming them using the same model (BGE-M3). This represents the dot product of two vectors, which is the sum of the numerical products of their corresponding dimensions. and These represent the L2 norm (modulus) of the query vector and the slice vector, respectively, and are calculated as the square root of the sum of the squares of the vector components.

[0105] The stroke management knowledge base 601 limits the calculation results to the range of [-1, 1], with the closer the value is to 1, the higher the semantic relevance. The stroke management knowledge base 601 outputs the search results: The stroke management knowledge base 601 selects the Top-5 guide fragments with a similarity score greater than 0.75 and feeds them back to the personalized management strategy generation module 303.

[0106] Subsequently, the personalized management strategy generation module 303 will generate the total risk index R. total The number of consecutive days of compliance with treatment guidelines and abnormal physiological indicators are converted into query vectors, which are then matched using a cosine similarity algorithm in the stroke management knowledge base 601. The recurrence risk dynamic prediction model module 302 assembles the retrieved Top-K medical guideline fragments with the patient profile into structured prompts of approximately 512 tokens.

[0107] Specifically, the personalized management strategy generation module 303 will retrieve the medical knowledge fragments and the patient's current total risk index R from step S106. total The number of consecutive days meeting the standards and information on abnormal vital signs are assembled into structured prompts according to a preset medical logic template.

[0108] The personalized management strategy generation module 303 injects structured cue words into the Large Language Inference Model (LLM) 602. The Large Language Inference Model (LLM) 602 receives the structured cue words and performs natural language inference under the constraints of recalled knowledge to generate a personalized intervention plan that includes specific health management tips, dietary recommendations, exercise prescriptions, and follow-up visit warnings.

[0109] The Large Language Reasoning Model (LLM) 602 reduces common-sense medical errors and generated illusions by limiting the reasoning process to the scope of recalled authoritative medical guidelines, ensuring the rigor and authority of the output solution.

[0110] S107: Generation of personalized management strategies.

[0111] like Figure 1 and Figure 2As shown, the Large Language Inference Model (LLM) 602 receives the aforementioned prompt words from the Relapse Risk Dynamic Prediction Model Module 302 and performs natural language inference. The Relapse Risk Dynamic Prediction Model Module 302 not only analyzes the overall risk index but also combines the recalled knowledge to generate targeted dietary recommendations, medication adjustment reminders, and psychological encouragement. The dietary recommendations generated by the Relapse Risk Dynamic Prediction Model Module 302 are sent from the application server 200 to the mobile application terminal 102.

[0112] S108: Determination of whether a follow-up visit is required.

[0113] like Figure 2 As shown, the system automatically assesses the current total risk index R. total Does the system trigger a follow-up clinical visit? If the overall risk index continues to rise or serious abnormalities occur, the system determines it as yes and guides the patient to schedule a follow-up visit through the medical client 101 and return to step one to update clinical data 401; if the risk is under control, the system determines it as no and the process proceeds to the next step.

[0114] S109: Real-time feedback and incentive mechanism.

[0115] The mobile application 102 features a real-time feedback interface, including a dynamic risk index display component. This component synchronously displays the achievement status of multiple health management dimensions derived from the overall risk index decomposition. This achievement status can be updated periodically or in real-time. The dynamic risk index display component dynamically represents the real-time achievement status of each dimension—blood pressure, blood sugar, medication, exercise, and diet—through geometric changes in the graphical area. Based on a risk warning coloring logic, the component uses different color features to correspond to different risk levels, aiming to motivate patients to improve their behavior through visual feedback.

[0116] The real-time feedback interface of mobile application 102 adopts a layered layout strategy. The dynamic risk index display component is connected to the main area of ​​the real-time feedback interface, which has data monitoring cards and is distributed in a hierarchical relationship with the data monitoring cards, in a floating or embedded manner. This allows the dynamic risk index display component to have a visually guided and data-linked interactive relationship with the functional cards (medication, exercise, diet) located at the bottom of the interface.

[0117] like Figure 2 and Figure 3As shown, after receiving data, the mobile application 102 uses a dynamic risk index display component to visually represent multi-dimensional data. This component has a central circular core area, surrounding polygonal geometric areas (specifically represented as a five-petaled flower in this embodiment, but not limited to this), and / or a dynamic coloring layer. The central circular core area and the polygonal geometric areas can establish a visual mapping relationship based on facial expression feedback and color saturation while the total risk index is dynamically changing, thus fixing the dynamic risk index display component to the upper-middle visual core of the real-time feedback interface. This five-petaled flower dynamic risk index chart provides color-coded warnings (green - meeting standards, red - high risk) for five dimensions: blood pressure, blood sugar, medication, exercise, and diet. The five-petaled flower dynamic risk index chart simultaneously displays personalized encouragement messages generated by LLM and an incentive mark indicating "goal achieved for X consecutive days," completing the closed-loop management of behavioral intervention.

[0118] The specific interaction logic of the dynamic risk index display component is as follows.

[0119] The dynamic risk index display component transforms the multi-dimensional compliance rates calculated in the background into a geometrically filled state. The central circular core area displays anthropomorphic emotional feedback expressions. The bottom function card area contains detailed records of specific values ​​(such as blood pressure and medication status).

[0120] If the pass rate for a certain dimension (such as blood pressure) is lower than a preset threshold, the corresponding sub-region (such as petals) in the polygonal geometric region will be filled with the red high-risk color spectrum under the driving condition of the color rendering engine, so that the region will be in a genus or warning color state.

[0121] When the overall risk index is declining, the central circular region, under the action of the animation controller, calls the positive incentive expression library.

[0122] And / or, when the user clicks the bottom function card, the bottom function card area retrieves historical data from the database module 301 according to the data backtracking logic, and causes the dynamic risk index display component to be partially highlighted.

[0123] Example 3 This embodiment is a further improvement on embodiment 1, and repeated content will not be described again.

[0124] This embodiment discloses a real-time feedback interface and its interaction logic for a stroke recurrence risk management system, aiming to achieve closed-loop management of "assessment-intervention-feedback" through intuitive visual components. The real-time feedback interface of this stroke recurrence risk management system serves as the core interaction layer of the mobile application terminal 102, and its display content and technical logic are as follows: Figure 3 As shown.

[0125] The dynamic risk index display component consists of a central circular core and several geometric regions surrounding it. These geometric regions are defined as the five core dimensions of secondary stroke prevention: blood pressure, blood sugar, medication, exercise, and diet. The colored area within each geometric region is dynamically determined by the patient's achievement of the corresponding indicator's goals; as the achievement of a certain indicator increases, the color fill ratio of the corresponding geometric region also increases.

[0126] The dynamic risk index display component provides risk warnings by dynamically adjusting facial expressions within the central circle and changing the color saturation of geometric areas. The facial expressions change dynamically with the total risk index, and the colors of the geometric areas use different hues to distinguish between acceptable states and high-risk warnings, thereby creating positive incentive feedback for behavioral improvement.

[0127] For example, the geometric area is designed in the shape of a petal. Figure 3 As shown, the core visual component of the real-time feedback interface of this stroke recurrence risk management system is a dynamic risk index display component in the shape of a five-petaled flower located on the left side, used to visually demonstrate the progress of various health management goals. Physically, the dynamic risk index display component consists of a central circular core and five symmetrical petals surrounding it. A cartoon avatar is displayed within the central circle, and its expression (such as the "striving" expression shown in the display) dynamically adjusts according to the total risk index calculated by the system. Preferably, the expression changes dynamically with the total risk index, and different hues of the petals distinguish between the target status and high-risk warnings, thereby forming a positive incentive feedback for behavioral improvement.

[0128] Preferably, the five petals are defined as the five core dimensions of secondary stroke prevention: blood pressure (systolic and diastolic), blood sugar, medication, exercise, and diet. The colored area within each petal is dynamically determined by the patient's achievement of the corresponding indicator's goals; as the achievement of a certain indicator (such as exercise) increases, the color filling ratio of the corresponding petal increases accordingly. Furthermore, the real-time feedback interface of this stroke recurrence risk management system uses color saturation for risk warnings: green represents achievement of goals, while yellow or red represents risk warnings. This mechanism provides positive incentives for improved health behaviors, forming a feedback logic of behavior improvement and risk reduction.

[0129] The mobile application client 102, which communicates with the application server 200, implements closed-loop management in the following ways.

[0130] Based on the collection frequency and numerical fluctuations of self-management data 402, early warning indicators for indicators to be retested or the timeliness of data are automatically triggered in the real-time feedback interface. The total risk index is displayed in real time in the real-time feedback interface, and incentive markers are generated based on the number of consecutive days the patient meets the target, in order to enhance the patient's self-management compliance through psychological feedback. Patients are guided to check in for health through the interactive entry in the real-time feedback interface, and the collected behavioral data is continuously fed back to the database module 301 for continuous dynamic evaluation by the relapse risk dynamic prediction model module 302.

[0131] like Figure 3 As shown, the right side of the real-time feedback interface of the stroke recurrence risk management system displays detailed real-time values ​​and statuses of various specific health physiological indicators. The blood pressure monitoring section displays the text "Blood Pressure" and the specific value "XXX / XX mmHg," with a red bubble in the upper right corner indicating "Pending Retest." This function allows the system to remind users of monitoring frequency; an alert will be triggered when data is overdue or fluctuates excessively. The blood glucose monitoring section displays the text "Blood Glucose" and the specific value "XX.X mmol / L," with the "2 days ago" label in the upper right corner reflecting the dynamic characteristic sequence x. t The collection frequency indicates the timeliness of patient data updates.

[0132] The real-time feedback interface of the stroke recurrence risk management system features three function cards at the bottom to display details of daily health behaviors. The medication card displays an icon, the text "Medication," and the numerical value "XXX," recording the patient's medication status for the day to assist the DRM model 502 in assessing medication adherence. The exercise card displays an icon, the text "Exercise," and the numerical value "XXX," representing the duration or intensity of exercise. The diet card displays the text "Diet" and the numerical value "XXX," reflecting the patient's effectiveness in adhering to intervention recommendations such as low-salt and low-fat diets.

[0133] The real-time feedback interface of the stroke recurrence risk management system displays evaluation text corresponding to the total risk index at the top, such as "Health Risk: Low". This result is the total risk index R calculated by the application server 200 by integrating the long-term BRS index and the short-term DRM score. total This is mapped from the actual data. The real-time feedback interface of the stroke recurrence risk management system clearly displays the encouraging feedback text at the bottom: "Goal achieved for 3 consecutive days." This highlighted numerical display affirms the patient's continued self-discipline and uses an incentive mechanism to enhance the patient's self-management motivation. The "Health Check-in" title and calendar entry at the bottom are functionally designed to continuously feed daily check-in data back to the self-management data storage area 402 in database module 301, thereby achieving continuous dynamic evaluation and intervention optimization.

[0134] It should be noted that the specific embodiments described above are exemplary. Those skilled in the art can devise various solutions inspired by the disclosure of this invention, and these solutions all fall within the scope of this invention and its protection. Those skilled in the art should understand that this specification and its accompanying drawings are illustrative and not intended to limit the scope of the claims. The scope of protection of this invention is defined by the claims and their equivalents. This specification contains multiple inventive concepts; terms such as "preferredly," "according to a preferred embodiment," or "optionally" indicate that the corresponding paragraph discloses an independent concept. The applicant reserves the right to file divisional applications based on each inventive concept.

Claims

1. A stroke recurrence risk assessment and personalized management system, comprising an application server (200), characterized in that, The application server (200) includes: The database module (301) is used to store clinical data (401) and self-management data (402). The relapse risk dynamic prediction model module (302) is used to determine a long-term static risk index reflecting the patient's inherent risk based on the clinical data (401) stored in the database module (301), and to determine a short-term dynamic risk adjustment index reflecting the immediate risk contribution based on the dynamic feature sequence in the self-management data (402); by executing a risk coupling algorithm, the long-term static risk index and the short-term dynamic risk adjustment index are fused and calculated using weight coefficients to obtain the total risk index; The personalized management strategy generation module (303) is used to convert the total risk index into a query vector and retrieve matching guide fragments in the stroke management knowledge base (601); the guide fragments are input as constraints into the large language reasoning model (602) to generate personalized intervention plans.

2. The system according to claim 1, characterized in that, The application server (200) is connected to the interactive terminal for communication. The interactive terminal receives and displays the total risk index and the personalized intervention plan output by the application server (200); The interactive terminal includes a medical client (101) and a mobile application terminal (102), and uses a dynamic risk index display component to provide risk warnings.

3. The system according to claim 1 or 2, characterized in that, The recurrence risk dynamic prediction model module (302) determines the long-term static risk index in the following manner; The preset risk assessment period is divided into several discrete time steps; Based on the patient's static characteristics, the probability of recurrence risk at each time step is calculated, and the cumulative survival probability of remaining relapse-free during the assessment period is derived. The cumulative survival probability is used to determine a long-term static risk index that represents the individual pathological basis of the patient.

4. The system according to any one of claims 1 to 3, characterized in that, The recurrence risk dynamic prediction model module (302) determines the short-term dynamic risk adjustment index and performs fusion in the following manner: By utilizing recurrent neural network units with memory capabilities, dynamic feature sequences within a sliding time window are processed to identify the immediate contribution of physiological index fluctuations and self-management behaviors to risk. By using gating mechanisms and activation functions, time-series characteristics are transformed into a dynamically adjusted index that reflects short-term risk variables; Based on preset weighting coefficients, the long-term static risk index, which reflects the static pathological basis, is linearly or non-linearly coupled with the short-term dynamic risk adjustment index, which reflects the real-time performance, to generate the total risk index.

5. The system according to any one of claims 1 to 4, characterized in that, The personalized management strategy generation module (303) retrieves guide fragments in the following manner: Transform the patient’s current risk level, abnormal physiological indicators and historical baseline characteristics into a query vector in a high-dimensional space; In a knowledge base containing pre-stored fragmented vectors of medical guidelines and instructions, a vector similarity algorithm is used to calculate the relevance index between the query vector and each fragmented vector. Based on a preset similarity threshold, multiple guideline fragments that meet the criteria are automatically filtered and retrieved from the knowledge base to serve as authoritative evidence for subsequent intervention logic.

6. The system according to any one of claims 1 to 5, characterized in that, The specific steps by which the personalized management strategy generation module (303) generates a personalized intervention plan through the large language reasoning model (602) include: The recalled guideline fragments, patients' total risk index, and abnormal signs information are assembled into structured prompts according to a preset medical logic template; The structured prompts are injected into the large language reasoning model (602), and the guide fragments are used as knowledge constraints, so that the large language reasoning model (602) can perform natural language reasoning within the scope of authoritative medicine to generate personalized intervention plans; The generated personalized intervention plan, which includes at least one of the following: health management tips, dietary advice, or exercise prescription, will be distributed to the interactive terminal via a downlink for visual feedback and behavioral guidance.

7. The system according to any one of claims 1 to 6, characterized in that, The system also includes a mobile application (102) that communicates with the application server (200). The mobile application (102) is equipped with a real-time feedback interface, which includes a dynamic risk index display component, used to synchronously display the compliance status of multiple health management dimensions obtained based on the total risk index decomposition. The dynamic risk index display component dynamically represents the real-time compliance status of various dimensions such as blood pressure, blood sugar, medication, exercise, and diet through changes in the geometric features of the graphical area. The dynamic risk index display component uses different color features to correspond to different risk levels based on risk warning coloring logic, so as to motivate patients to improve their behavior through visual feedback.

8. The system according to any one of claims 1 to 7, characterized in that, The mobile application client (102) that communicates with the application server (200) implements closed-loop management in the following ways: Based on the collection frequency and numerical fluctuation of the self-management data (402), an early warning indicator for the indicator to be retested or the timeliness of the data is automatically triggered in the real-time feedback interface. The total risk index is displayed in real time in the real-time feedback interface, and incentive markers are generated based on the number of consecutive days that the patient has achieved the target, so as to enhance the patient's self-management compliance by using psychological feedback. Patients are guided to check in for health through the interactive entry in the real-time feedback interface, and the collected behavioral data is continuously fed back to the database module (301) for continuous dynamic evaluation by the recurrence risk dynamic prediction model module (302).

9. The system according to any one of claims 1 to 8, characterized in that, The mobile application (102) that communicates with the application server (200) has a real-time feedback interface; The real-time feedback interface has a dynamic risk index display component in the center, which is used to visually present the progress of health management goals. The dynamic risk index display component consists of a central circular core and several geometric regions surrounding it. These geometric regions are defined as the five core dimensions of secondary stroke prevention: blood pressure, blood sugar, medication, exercise, and diet. The colored area within each geometric region is dynamically determined by the patient's achievement of the corresponding indicator's check-in target. When the achievement of a certain indicator increases, the color filling ratio of the corresponding geometric region also increases. The dynamic risk index display component provides risk warnings by dynamically adjusting facial expressions within the central circular core and changing the color saturation of the geometric area. The facial expressions change dynamically with the total risk index, and the colors of the geometric area use different hues to distinguish between the acceptable state and the high-risk warning, thereby creating positive incentive feedback for behavior improvement.

10. A method for assessing and personally managing the risk of stroke recurrence, characterized in that, The method includes: Based on stored clinical data (401), a long-term static risk index reflecting the patient's inherent risk is determined, and a short-term dynamic risk adjustment index reflecting the immediate risk contribution is determined based on the dynamic feature sequence in self-management data (402); by executing a risk coupling algorithm, the long-term static risk index and the short-term dynamic risk adjustment index are fused and calculated using weighting coefficients to obtain the total risk index; The total risk index is converted into a query vector and a matching guideline fragment is retrieved from the stroke management knowledge base (601). The guideline fragment is then input as a constraint into the large language reasoning model (602) to generate a personalized intervention plan.

Images