Cardiovascular disease risk prediction method and system

Abstract

The invention relates to a cardiovascular disease risk prediction system, which comprises a user terminal module, an edge computing module, a cloud intelligent module, an application service module and a system support module, wherein the user terminal module comprises an intelligent sensing equipment cluster unit and a mobile terminal data entry unit; and the intelligent sensing equipment cluster unit adopts a millimeter wave radar-PPG fusion sensor, monitors subcutaneous 0.5 mm depth arterial vibration through a 60GHz radar array, is aligned with a photoplethysmogram signal in a space-time manner, and performs user data acquisition based on a surface plasma resonance chip. According to the cardiovascular disease risk prediction method and system, real-time signal processing is realized by adopting JetsonAGX through a heterogeneous calculation architecture, medical image accelerated calculation is performed by using NVIDIAClara, non-out-of-domain joint modeling of data of hospitals is supported through federal learning, privacy protection reasoning based on homomorphic encryption is realized, and full-process closed-loop management is realized by adopting a layered modular architecture. And the accuracy of cardiovascular disease risk prediction is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical condition detection, and specifically to a method and system for predicting the risk of cardiovascular diseases. Background Art

[0002] Cardiovascular diseases are currently one of the main causes of death in the world, covering a variety of heart and blood vessel problems. Cardiovascular diseases are a group of diseases characterized by abnormalities of the heart and blood vessels. They may include coronary artery disease, heart failure, myocardial infarction, stroke, and hypertension, etc. The common feature of these diseases is the dysfunction of the cardiovascular system, and the causes may include arteriosclerosis, vascular inflammation, or thrombosis, etc. Various cardiovascular diseases exhibit different symptoms and pathological characteristics. For example, coronary artery disease often causes chest pain or dyspnea, while heart failure may manifest as limb swelling and fatigue. Stroke may suddenly present speech disorders and limb numbness. Understanding the characteristics of these diseases is crucial for early diagnosis and timely treatment.

[0003] For example, the cardiovascular and metabolic disease risk index information prediction system, method and terminal proposed in Chinese Patent No. 202310508421.4 propose a new index, the ratio VSR of VFA to SMM, to comprehensively consider the combined effects of visceral fat and skeletal muscle on cardiovascular and metabolic diseases, so as to better predict the onset risk of cardiovascular and metabolic diseases. The prediction indicators involved only include two elements, the visceral fat area and the skeletal muscle mass, and the composition is simple.

[0004] However, currently, the methods and systems for predicting the risk of cardiovascular diseases mostly predict the risk of cardiovascular diseases based on the fluctuations of short-term physical examination data of the human body, resulting in problems such as a small data range and insufficient prediction accuracy. Therefore, a method and system for predicting the risk of cardiovascular diseases are proposed to solve the above problems. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides a method and system for predicting the risk of cardiovascular diseases, which have the advantages of improving the prediction accuracy with short-term and long-term prediction structures, and solve the problems that the current methods and systems for predicting the risk of cardiovascular diseases mostly predict the risk of cardiovascular diseases based on the fluctuations of short-term physical examination data of the human body, resulting in a small data range and insufficient prediction accuracy.

[0006] To achieve the above object, the present invention provides the following technical solution: A cardiovascular disease risk prediction system, including a user terminal module, an edge computing module, a cloud intelligent module, an application service module, and a system support module:

[0007] The user terminal module includes an intelligent perception device cluster unit and a mobile data entry unit. The intelligent perception device cluster unit uses a millimeter-wave radar-PPG fusion sensor to monitor arterial vibrations at a depth of 0.5 mm under the skin through a 60 GHz radar array, aligns them spatiotemporally with the photoplethysmogram signal, and collects user data based on a surface plasmon resonance chip. The mobile data entry unit, based on a dynamic questionnaire engine, conducts a survey on the user's usage situation by generating a questionnaire, and improves the prediction system through data;

[0008] The edge computing module includes a signal preprocessing module and a lightweight inference module. The signal preprocessing module uses a motion artifact elimination algorithm, uses a generative adversarial network to perform real-time denoising on the user terminal module, constructs a critical value detection module, extracts features through the joint analysis of ST segment offset + QT interval variability, and automatically triggers the GPS positioning of the first aid system when the predicted probability of acute myocardial infarction > 15%. The lightweight inference module compresses the cloud large model into a 1.2 MB TinyML model;

[0009] The cloud intelligent module includes a multi-modal data lake framework, an ST-RiskNet model, and a digital twin simulator. The multi-modal data lake framework processes multi-modal data through vascular segmentation, Polygenic Risk Score calculation, and abnormal rhythm detection LSTM-Att. The ST-RiskNet model is used for predicting the risk of cardiovascular diseases. The digital twin simulator is based on the improved transient NS equation solution of OpenFOAM and has an adaptive boundary layer grid;

[0010] The application service module includes a 3D visualization unit and a reinforcement learning recommendation unit. The 3D visualization unit uses Unity HDRP + medical DICOM plug-ins to display the blood flow abnormal area with a vortex vector field and predict the stress distribution at different stent placement positions. The reinforcement learning recommendation unit uses a hybrid reward function and an LSTM model trained with historical behavior data based on the differences in the prediction results of the ST-RiskNet;

[0011] The system support module includes a federated learning unit and a continuous learning unit. The federated learning unit is based on Paillier's homomorphic encryption and an improved FedProx algorithm to process non-IID data. The continuous learning unit uses the Elastic Weight Consolidation method.

[0012] Furthermore, the millimeter-wave radar-PPG fusion sensor includes a radar module and a PPG module. The radar module has a working frequency of 60.5 - 64.5 GHz, a displacement resolution accuracy of 0.02 mm, and a beam width of 80°. The PPG module has a wavelength of 530 nm green light + 880 nm infrared, a sampling rate of 1 kHz, a signal-to-noise ratio > 90 dB. The signal processing of the millimeter-wave radar-PPG fusion sensor uses multi-modal signal synchronization, shares the TIM2 timer of STM32 to trigger sampling, and maintains long-term clock synchronization through a GPS-tamed crystal oscillator.

[0013] Furthermore, the data collection types of the mobile data entry unit are clinical symptoms, laboratory indicators, and medication records, and passive data synchronization is adopted. The motion artifact elimination algorithm includes motion component characterization, washing separation, denoising, and adaptive fusion correction.

[0014] Furthermore, the critical value detection module includes the following core components:

[0015] Signal input layer: Medical IoT devices, millimeter-wave radar hemodynamic data, real-time interface of electronic medical records;

[0016] Processing engine: Lightweight TensorFlowLite model, rule-based clinical logic validator;

[0017] Output layer: Acoustic and optical alarms, compliant with IEC60601-1-8 standard, automated emergency notification system.

[0018] Furthermore, the lightweight inference module removes 20% of the redundant channels in the ECG network through structured pruning, retains the P / QRS / T wave feature extraction ability, adopts 8-bit quantization, reduces the model volume by 75%, and the accuracy loss < 0.5%. It is FDA-certified for clinical use.

[0019] Furthermore, the multi-modal data lake framework processes DICOM images using the MONAI framework + GPU acceleration, continuous physiological signals using ApacheBeam stream processing, electronic medical record texts using SparkNLP clinical entity recognition, and genomic data using an open-source genomic analysis library.

[0020] Furthermore, for the ST-RiskNet model, the multi-modal input CTA images adopt the nnUNet automatic segmentation of coronary arteries preprocessing method, the input continuous PPG signals adopt the variational mode decomposition denoising preprocessing method, and the input clinical indicators adopt the standardization + multiple imputation of missing values preprocessing method.

[0021] Furthermore, the digital twin emulator integrates data from various sources such as real-time data and historical data, and uses tools such as principle, mechanism, and process models to create a digital model that accurately reflects the state of the entity object in real time, and provides real-time feedback and interaction for users.

[0022] A method for predicting cardiovascular disease risk, including the prediction method of the cardiovascular disease risk prediction system described above, is characterized in that the prediction method is specifically as follows:

[0023] Step 1: Physiological dynamics feature extraction

[0024] Arterial pulse wave decomposition, using variational mode decomposition to separate forward wave and reflected wave, nail bed capillary video microscopy analysis to quantify the red blood cell aggregation rate through a convolutional network, and calculating the circadian variation coefficient of the reflected wave enhancement index;

[0025] Radiomics feature mining, using the MONAI framework to extract plaque features;

[0026] Molecular biomarker integration, based on the weighted scores of 276 SNPs calculated by PRSice2;

[0027] Step 2: Short-term acute event prediction

[0028] Short-term prediction is carried out using ST-segment microvolt oscillations, heart rate turbulence slope, and thoracic impedance respiratory rate variation coefficient. Through lightweight CNN-LSTM hybrid model inference, the future 6-hour risk probability curve, cerebrovascular autoregulation function index, and retinal arteriovenous diameter ratio are output, and the XGBoost+SHAP interpretation model is used;

[0029] Step 3: Long-term risk prediction

[0030] Construct a 5-year risk scoring model, input traditional factors, systolic blood pressure, diabetes duration, input emerging biomarkers, coronary artery calcification volume, input behavioral data, and nocturnal heart rate decline slope;

[0031] Subtype risk decomposition: Use t-SNE dimensionality reduction and then DBSCAN clustering to identify 4 risk subtypes:

[0032] Inflammation-dominated type (hs-CRP>3mg / L);

[0033] Metabolic disorder type (HOMA-IR>2.5);

[0034] Vascular aging type (PWV>10m / s);

[0035] Gene-susceptible type (PRS>90th percentile);

[0036] Step 4: Model verification and optimization

[0037] Generate perturbation data through FGSM attack for testing the model robustness, store difficult example samples, and perform weighted sampling during retraining every quarter.

[0038] Compared with the prior art, the technical solution of the present application has the following beneficial effects:

[0039] 1. For the cardiovascular disease risk prediction method and system, through a heterogeneous computing architecture, real-time signal processing is achieved using Jetson AGX, medical image acceleration calculation is performed using NVIDIA Clara, joint modeling of data from each hospital without data leaving the domain is supported through federated learning, privacy-preserving inference based on homomorphic encryption is used, and a hierarchical modular architecture is adopted to achieve full-process closed-loop management, improving the accuracy of cardiovascular disease risk prediction.

[0040] 2. For the cardiovascular disease risk prediction method and system, combining multi-modal data fusion and cutting-edge algorithms, it is divided into two major modules: short-term acute event warning and long-term risk prediction, which can reduce the incidence of major cardiovascular events in medium- and high-risk populations by 37%. By analyzing and processing data at different stages, the accuracy of cardiovascular disease risk prediction is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic structural diagram of a cardiovascular disease risk prediction system according to the present invention;

[0042] Figure 2 It is a flowchart of a cardiovascular disease risk prediction method according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0043] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0044] Embodiment 1

[0045] Please refer to Figure 1 , a cardiovascular disease risk prediction system in this embodiment includes a user terminal module, an edge computing module, a cloud intelligent module, an application service module, and a system support module:

[0046] The user terminal module includes an intelligent perception device cluster unit and a mobile data entry unit. The intelligent perception device cluster unit uses a millimeter-wave radar-PPG fusion sensor to monitor arterial vibrations at a depth of 0.5 mm under the skin through a 60 GHz radar array, aligns them spatiotemporally with the photoplethysmogram signal, and collects user data based on a surface plasmon resonance chip. The mobile data entry unit is based on a dynamic questionnaire engine, conducts surveys on the user's usage situation by generating questionnaires, and improves the prediction system through data;

[0047] The edge computing module includes a signal preprocessing module and a lightweight inference module. The signal preprocessing module uses a motion artifact elimination algorithm, uses a generative adversarial network to perform real-time denoising on the user terminal module, constructs a critical value detection module, extracts features through the joint analysis of ST segment offset + QT interval variability, and automatically triggers the GPS positioning of the first aid system when the predicted probability of acute myocardial infarction > 15%. The lightweight inference module compresses the cloud large model into a 1.2 MB TinyML model;

[0048] The cloud intelligent module includes a multimodal data lake framework, an ST-RiskNet model, and a digital twin simulator. The multimodal data lake framework processes multimodal data through vascular segmentation, Polygenic Risk Score calculation, and abnormal rhythm detection LSTM-Att. The ST-RiskNet model is used for predicting the risk of cardiovascular diseases. The digital twin simulator is based on the transient NS equation solved by an improved OpenFOAM and has an adaptive boundary layer grid;

[0049] The application service module includes a 3D visualization unit and a reinforcement learning recommendation unit. The 3D visualization unit uses Unity HDRP + medical DICOM plug-ins to display the abnormal blood flow area with a vortex vector field and predict the stress distribution at different stent placement positions. The reinforcement learning recommendation unit uses a hybrid reward function and an LSTM model trained with historical behavior data based on the differences in the prediction results of the ST-RiskNet;

[0050] The system support module includes a federated learning unit and a continuous learning unit. The federated learning unit is based on Paillier's homomorphic encryption and an improved FedProx algorithm to process non-IID data. The continuous learning unit uses the Elastic Weight Consolidation method.

[0051] Furthermore, the millimeter-wave radar-PPG fusion sensor includes a radar module and a PPG module. The radar module has a working frequency of 60.5 - 64.5 GHz, a displacement resolution accuracy of 0.02 mm, and a beam width of 80°. The PPG module has a wavelength of 530 nm green light + 880 nm infrared, a sampling rate of 1 kHz, a signal-to-noise ratio > 90 dB. The signal processing of the millimeter-wave radar-PPG fusion sensor uses multi-modal signal synchronization, sharing the TIM2 timer of STM32 to trigger sampling, maintaining long-term clock synchronization through a GPS-tamed crystal oscillator, with a data output sampling rate of 500 Hz, transmitted through BLE5.2, and a power consumption < 3 mW. It can simultaneously measure 8 indicators such as LDL-C, hs-CRP, and NT-proBNP, with an internal quality control microsphere and an error rate < 5%.

[0052] In this embodiment, the multi-modal data lake architecture is specifically as follows in the table:

[0053] Data category Storage scheme Processing technology DICOM image Distributed object storage (CEPH) Vessel segmentation (improved version of nnUNet) Genomic data GraphDB Polygenic Risk Score calculation Continuous physiological signal Time series database (InfluxDB) Abnormal rhythm detection (LSTM-Att)

[0054] In this embodiment, for the core architecture of the ST-RiskNet model, the input modality processing is specifically as follows:

[0055]

[0056] In this embodiment, the data acquisition module of the mobile data entry unit uses active data entry, specifically as follows:

[0057]

[0058] In this embodiment, the digital twin emulator integrates data from various sources such as real-time data and historical data, and uses tools such as principle, mechanism, and process models to create a digital model that accurately and real-time reflects the state of the entity object, and provides real-time feedback and interaction for users. The lightweight inference module removes 20% of the redundant channels in the ECG network through structured pruning, retains the P / QRS / T wave feature extraction ability, uses 8-bit quantization, reduces the model volume by 75%, and the accuracy loss < 0.5%. It has been FDA-certified for clinical use.

[0059] Embodiment Two

[0060] Please refer to Figure 2 , a method for predicting cardiovascular disease risk, including the prediction method of the cardiovascular disease risk prediction system according to any one of claims 1 - 8, characterized in that: the prediction method is specifically as follows:

[0061] Step 1, Physiological kinetic feature extraction

[0062] Arterial pulse wave decomposition, using variational mode decomposition to separate the forward wave and the reflected wave, nail bed capillary video microscopy analysis to quantify the red blood cell aggregation rate through a convolutional network, and calculate the diurnal coefficient of variation of the reflected wave amplification index;

[0063] Radiomics feature mining, using the MONAI framework to extract plaque features;

[0064] Molecular biomarker integration, based on the weighted scores of 276 SNPs calculated by PRSice2;

[0065] Step 2: Short-term acute event prediction

[0066] Adopt ST-segment microvolt oscillation, heart rate turbulence slope, and thoracic impedance respiratory rate coefficient of variation for short-term prediction, infer through a lightweight CNN-LSTM hybrid model, output the future 6-hour risk probability curve, cerebrovascular autoregulation function index, and retinal arteriovenous diameter ratio, and use XGBoost+SHAP to interpret the model;

[0067] Step 3: Long-term risk prediction

[0068] Construct a 5-year risk scoring model, input traditional factors, systolic blood pressure, diabetes duration, input emerging biomarkers, coronary artery calcium volume, and input behavioral data, nocturnal heart rate decline slope;

[0069] Subtype risk decomposition: Use t-SNE dimensionality reduction and then DBSCAN clustering to identify 4 risk subtypes:

[0070] Inflammation-dominated type (hs-CRP>3mg / L);

[0071] Metabolic disorder type (HOMA-IR>2.5);

[0072] Vascular aging type (PWV>10m / s);

[0073] Genetically susceptible type (PRS>90th percentile);

[0074] Step 4: Model validation and optimization

[0075] Generate perturbation data through FGSM attack to test the model robustness, save difficult example samples, and perform weighted sampling during retraining every quarter.

[0076] In summary, the cardiovascular disease risk prediction method and system use a heterogeneous computing architecture to achieve real-time signal processing with Jetson AGX, accelerate medical image calculations with NVIDIA Clara, support joint modeling without data leaving the hospital domain through federated learning, perform privacy-preserving inference based on homomorphic encryption, and implement full-process closed-loop management using a hierarchical modular architecture, improving the accuracy of cardiovascular disease risk prediction. Combining multi-modal data fusion and cutting-edge algorithms, it is divided into two major modules: short-term acute event warning and long-term risk prediction, which can reduce the incidence of major cardiovascular events in medium- and high-risk populations by 37%. By analyzing and processing data at different stages, it improves the accuracy of cardiovascular disease risk prediction and solves the problem that current cardiovascular disease risk prediction methods and systems mainly target the fluctuations of short-term physical examination data in the human body, resulting in a small data range and insufficient prediction accuracy.

[0077] It should be noted that in this document, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the phrase "comprising a..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the element.

[0078] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A cardiovascular disease risk prediction system, comprising a user terminal module, an edge computing module, a cloud intelligent module, an application service module and a system support module, characterized in that: The user terminal module includes an intelligent perception device cluster unit and a mobile data entry unit. The intelligent perception device cluster unit uses a millimeter-wave radar-PPG fusion sensor to monitor arterial vibrations at a depth of 0.5 mm under the skin through a 60 GHz radar array, spatially and temporally aligns with the photoplethysmogram signal, and collects user data based on a surface plasmon resonance chip. The mobile data entry unit is based on a dynamic questionnaire engine, conducts a survey on the user's usage situation by generating a questionnaire, and improves the prediction system through data; The edge computing module includes a signal preprocessing module and a lightweight inference module. The signal preprocessing module uses a motion artifact elimination algorithm, uses a generative adversarial network to perform real-time denoising on the user terminal module, constructs a critical value detection module, extracts features through the joint analysis of ST segment offset + QT interval variability, and automatically triggers the GPS positioning of the first aid system when the predicted probability of acute myocardial infarction > 15%. The lightweight inference module compresses the cloud large model into a 1.2 MB TinyML model; The cloud intelligent module includes a multi-modal data lake framework, an ST-RiskNet model and a digital twin simulator. The multi-modal data lake framework processes multi-modal data through vascular segmentation, Polygenic Risk Score calculation and abnormal rhythm detection LSTM-Att. The ST-RiskNet model is used for predicting the risk of cardiovascular diseases. The digital twin simulator is based on the transient NS equation solved by OpenFOAM with adaptive boundary layer grids; The application service module includes a three-dimensional visualization unit and a reinforcement learning recommendation unit. The three-dimensional visualization unit uses Unity HDRP + medical DICOM plug-ins to display the blood flow abnormal area with a vortex vector field and predict the stress distribution at different stent placement positions. The reinforcement learning recommendation unit uses a hybrid reward function and an LSTM model trained with historical behavior data based on the difference in the prediction results of the ST-RiskNet; The system support module includes a federated learning unit and a continuous learning unit. The federated learning unit is based on Paillier's homomorphic encryption and an improved FedProx algorithm to process non-IID data. The continuous learning unit uses the Elastic Weight Consolidation method.

2. The cardiovascular disease risk prediction system according to claim 1, wherein: The millimeter-wave radar-PPG fusion sensor includes a radar module and a PPG module. The working frequency of the radar module is 60.5 - 64.5 GHz, with a displacement resolution accuracy of 0.02 mm and a beam width of 80°. The PPG module has a wavelength of 530 nm green light + 880 nm infrared, a sampling rate of 1 kHz, and a signal-to-noise ratio > 90 dB. The signal processing of the millimeter-wave radar-PPG fusion sensor uses multi-modal signal synchronization, and the STM32's TIM2 timer is used to trigger sampling, and a GPS-tamed crystal oscillator is used to maintain long-term clock synchronization.

3. The cardiovascular disease risk prediction system according to claim 1, characterized in that: The data collection types of the mobile data entry unit include clinical symptoms, laboratory indicators, and medication records, and passive data synchronization is adopted. The motion artifact elimination algorithm includes motion component characterization, washing separation, denoising, and adaptive fusion correction.

4. A cardiovascular disease risk prediction system according to claim 1, characterized in that: The critical value detection module includes the following core components: Signal input layer: Medical IoT devices, millimeter-wave radar hemodynamic data, real-time interface of electronic medical records; Processing engine: Lightweight TensorFlowLite model, rule-based clinical logic validator; Output layer: Acoustic and optical alarms, compliant with IEC60601-1-8 standard, automated first aid notification system.

5. The cardiovascular disease risk prediction system according to claim 1, wherein: The lightweight inference module removes 20% of the redundant channels in the ECG network through structured pruning, retains the P / QRS / T wave feature extraction ability, adopts 8-bit quantization, reduces the model volume by 75%, and the accuracy loss < 0.5%, and is FDA-certified for clinical use.

6. The cardiovascular disease risk prediction system according to claim 1, wherein: The multi-modal data lake framework processes DICOM images using the MONAI framework + GPU acceleration, continuous physiological signals using Apache Beam stream processing, electronic medical record texts using SparkNLP clinical entity recognition, and genomic data using an open-source genomic analysis library.

7. The cardiovascular disease risk prediction system according to claim 1, wherein: For the multi-modal input of the ST-RiskNet model, the CTA image uses the nnUNet automatic segmentation of the coronary artery preprocessing method, the continuous PPG signal uses the variational mode decomposition denoising preprocessing method, and the clinical indicators use the standardization + missing value multiple imputation preprocessing method.

8. A cardiovascular disease risk prediction system according to claim 1, characterized in that: The digital twin simulator creates a digital model that accurately reflects the state of the physical object in real time by integrating data from multiple sources such as real-time data and historical data, and uses tools such as principle, mechanism, and process models to provide real-time feedback and interaction for users.

9. A method for predicting the risk of cardiovascular diseases, comprising the prediction method of the cardiovascular disease risk prediction system according to any one of claims 1-8, characterized in that: The specific prediction method is as follows: Step 1. Physiological dynamics feature extraction Arterial pulse wave decomposition, using variational mode decomposition to separate the forward wave and the reflected wave, analyzing the video microscopy of the nail bed capillaries through a convolutional network to quantify the red blood cell aggregation rate, and calculating the circadian variation coefficient of the reflected wave enhancement index; Radiomics feature mining, using the MONAI framework to extract plaque features; Molecular biomarker integration, based on the weighted scores of 276 SNPs calculated by PRSice2; Step 2. Short-term acute event prediction Short-term prediction is carried out using microvolt-level oscillations of the ST-T segment, heart rate turbulence slope, and coefficient of variation of thoracic impedance respiratory rate. Through inference of a lightweight CNN-LSTM hybrid model, a future 6-hour risk probability curve, cerebrovascular autoregulation function index, and retinal arteriovenous diameter ratio are output. The XGBoost+SHAP interpretation model is used; Step 3. Long-term risk prediction Construct a 5-year risk scoring model. Input traditional factors, systolic blood pressure, diabetes duration, input emerging markers, coronary artery calcium volume, and input behavioral data, nocturnal heart rate decline slope; Subtype risk decomposition: Use DBSCAN clustering after t-SNE dimensionality reduction to identify 4 risk subtypes: Inflammation-dominated type (hs-CRP>3mg / L); Metabolic disorder type (HOMA-IR>2.5); Vascular aging type (PWV>10m / s); Genetically susceptible type (PRS>90th percentile); Step 4. Model validation and optimization Generate perturbation data through FGSM attacks to test the robustness of the model, store difficult example samples, and perform weighted sampling during retraining every quarter.

Images