Chronic kidney patient cardiovascular adverse event prediction and clinical treatment system based on deep learning

Abstract

The invention provides a chronic kidney patient cardiovascular adverse event prediction and clinical treatment system based on deep learning, and the system comprises a multi-source information integration module which obtains multi-modal fusion data; the risk prediction module is used for constructing a risk prediction model and outputting the risk prediction probability, the risk level and the risk report of each patient; the decision support module is used for generating a plurality of interpretable layers, calculating high-risk significant factors and automatically recommending clinical decision suggestions; the personalized diagnosis and treatment module is used for constructing a process knowledge graph, carrying out personalized treatment recommendation and distributing corresponding rehabilitation execution tasks for different roles; and the doctor-patient interaction module is used for establishing a multi-terminal collaborative micro-service architecture and carrying out multi-terminal collaborative management and early warning information real-time pushing. The cardiovascular adverse event prediction accuracy and efficiency of the chronic kidney disease patient can be effectively improved, personalized treatment schemes are pushed to different patients, and early diagnosis and early intervention of chronic kidney disease related complications are achieved.

Description

Technical Field

[0001] This invention relates to the field of cardiovascular adverse event prediction technology, and in particular to a deep learning-based system for predicting and treating cardiovascular adverse events in patients with chronic kidney disease. Background Technology

[0002] Chronic kidney disease (CKD) refers to a condition characterized by abnormalities in kidney structure or function lasting for more than three months. It can be caused by various factors, including diabetes, hypertension, glomerulonephritis, and polycystic kidney disease. When CKD progresses to end-stage renal disease (ESRD), it transforms into uremia, at which point more than 90% of kidney function has been lost, making it impossible to maintain metabolic balance. Furthermore, patients with CKD have a significantly higher risk of cardiovascular disease (CVD) than the general population. With the increasing incidence of cardiovascular disease among CKD patients, the prediction of adverse cardiovascular events is of paramount practical significance and promise for improving the overall prognosis of CKD patients and reducing the social burden.

[0003] Cardiovascular adverse event prediction utilizes various information, such as clinical, laboratory, or imaging data, to estimate the risk of CKD patients experiencing cardiovascular adverse events (myocardial infarction, heart failure, stroke, cardiac death, etc.) over a certain period. Currently, the most commonly used cardiovascular risk prediction tools and standard risk models are employed. However, existing technologies suffer from several limitations: 1) Poor population applicability: Most general-purpose cardiovascular risk models do not adequately incorporate CKD-specific risk factors, resulting in limited accuracy in predicting CKD populations and often underestimating the true risk; 2) Limited data or insufficient information utilization: Many models only include traditional risk factors, failing to adequately consider CKD-specific factors, biomarkers, or even genetic factors; 3) Inconvenient clinical application: The clinical operability of these models is poor, hindering widespread adoption. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a deep learning-based system for predicting and treating cardiovascular adverse events in patients with chronic kidney disease. This system can effectively improve the accuracy and efficiency of predicting cardiovascular adverse events in patients with chronic kidney disease, deliver personalized treatment plans to different patients, achieve early diagnosis and early intervention of chronic kidney disease, reduce the burden on doctors, improve patient compliance, and promote intelligent and precise management of chronic diseases.

[0005] To achieve the above objectives, the present invention provides the following solution: a deep learning-based system for predicting and treating cardiovascular adverse events in patients with chronic kidney disease, comprising:

[0006] The multi-source information integration module is used to connect to the hospital system, acquire structured data, unstructured data, and patients' real-time health data, and then fuse and store the acquired data to obtain multimodal fused data;

[0007] The risk prediction module is used to design and integrate structured branch networks, medical imaging branch networks, medical text branch networks and attention fusion layers to obtain a risk prediction model. The risk prediction model is used to output the risk prediction probability, risk level and risk report for each patient.

[0008] The decision support module is used to generate multiple interpretable layers based on the multimodal fusion data and the output of the risk prediction model, calculate high-risk significance factors, and automatically recommend clinical decision suggestions based on the high-risk significance factors.

[0009] The personalized treatment module is used to construct a process knowledge graph, make personalized treatment recommendations based on individual rules, and assign corresponding rehabilitation tasks to different roles based on the personalized treatment recommendations.

[0010] The doctor-patient interaction module is used to build a multi-terminal collaborative microservice architecture, manage multi-terminal collaboration, and push early warning information in real time according to the risk level and the rehabilitation execution task.

[0011] The multi-source information integration module, the risk prediction module, the decision support module, the personalized diagnosis and treatment module, and the doctor-patient interaction module are interconnected.

[0012] Optionally, the multi-source information integration module includes:

[0013] The structured data acquisition unit is used to build API-ETL that links with hospital information systems, laboratory information systems and electronic medical record systems, and to develop dynamic medical data adapters compatible with different databases to achieve standard docking, acquire structured data, and use blockchain technology for encrypted recording.

[0014] The unstructured data acquisition unit is used to interface with the medical image archiving and transmission system, automatically acquire and store DICOM images, interface with the electronic medical record system to acquire unstructured documents, and combine the DICOM images and the unstructured documents to obtain unstructured data.

[0015] The data conversion unit is used to automatically extract image features from the DICOM image using deep learning technology to obtain structured indicators, and to use a medical NLP model to perform named entity recognition, relation extraction and event tracing on the structured document to convert text information into structured fields.

[0016] A continuous monitoring unit is used to design a wireless gateway, which automatically senses and connects to the patient's wearable device to obtain real-time health data.

[0017] The data fusion unit is used to perform time alignment, missing data completion, and consistency checks based on the structured data, the unstructured data, and the real-time health data to obtain multimodal fused data, and to construct a patient data archive that stores the multimodal fused data; wherein the main index of the patient data archive is the patient ID and the format is columnar storage.

[0018] Optionally, the risk prediction module includes:

[0019] Modular network units are used to design corresponding network structures for the structured data and the unstructured data respectively, to obtain a structured branch network, a medical image branch network and a medical text branch network, and to integrate the three branch networks to obtain an initial model;

[0020] The fusion layer unit is used to design the attention fusion layer. Based on the input of the initial model, the attention fusion layer performs feature attention scoring to dynamically adjust the feature weights and obtain the attention distribution. The calculation expression for the dynamic feature weights is:

[0021]

[0022]

[0023] in, For attention score, This is the first layer weight matrix. This is the second-layer weight matrix vector. For the first Each modal feature vector These are all bias parameters. For the first Dynamic attention weights for each modality Modal features;

[0024] The model training unit is used to embed the attention fusion layer into the initial model, and then perform end-to-end training of the model to obtain a risk prediction model. The risk prediction model is used to output the risk prediction probability, risk level and risk report for each patient.

[0025] Optionally, the structured branch network includes a first input layer for inputting structured data sequences, a multilayer perceptron for extracting high-order features, and a gated recurrent unit for capturing temporal dynamics.

[0026] The medical imaging branch network includes a second input layer for inputting standard medical images and an image encoder for capturing global-local multi-scale features;

[0027] The medical text branching network includes a third input layer for inputting text, a Transformer base layer for extracting word sequence context embeddings, and a CLS task adaptation layer for outputting final features and linearly transforming them to a shared dimension.

[0028] Optionally, the model training unit includes:

[0029] The training subunit utilizes the multimodal fusion data from the input model to perform forward propagation, global loss calculation, and backpropagation-optimizer training, dynamically adjusting the learning rate and regularization coefficient to complete the end-to-end training of the initial model, thereby obtaining the risk prediction model; the expression for the global loss calculation is:

[0030]

[0031]

[0032]

[0033] in, Main loss function, For feature separation regularization terms, For the global loss function, For the first The true label of each sample The risk probability output by the model. For modality With mode cosine similarity, The regularization coefficient is used.

[0034] The output subunit is used to output the predicted probability of cardiovascular events for each patient within a certain time window using the risk prediction model, and to classify the risk level and output a risk report based on the predicted probability.

[0035] Optionally, the decision support module includes:

[0036] Interpretable units are used to construct modal-level attention heatmaps, image region visualization maps, and text attention visualization maps based on temporal structured feature attention, resulting in multiple interpretable layers;

[0037] The risk factor analysis unit is used to filter out high-risk significant factors based on the interpretable layer, calculate and sort the high-risk significant factors, and obtain an individualized risk contribution list; the calculation expression for the high-risk significant factors is:

[0038]

[0039] in, For factor normalization scores, For attribution strength, For model attention, The frequency of factor abnormalities For weighted parameters;

[0040] The decision support unit is used to connect the individualized risk contribution list with authoritative CKD cardiovascular guidelines or corresponding knowledge graphs to automatically recommend clinical decision-making suggestions.

[0041] Optionally, the interpretable unit includes:

[0042] A modal attention heatmap is used to input the attention distribution output by the attention fusion layer and the temporal dynamics output by the gated loop unit, calculate the feature importance at each time point, output a modal attention radar map for each patient, and highlight high-risk and sensitive moments in the course of the disease in the modal attention radar map to obtain a modal attention heatmap.

[0043] An image region visualization map is used to input the multi-scale features output by the medical imaging branch network, output a fine-grained heat map, and introduce anatomical structure labels into the fine-grained heat map to obtain the image region visualization map.

[0044] A text attention visualization is generated by taking the attention weights of each layer of tokens output from the Transformer base layer as input and highlighting the risk content related to keywords in the structured document.

[0045] Optionally, the personalized treatment module includes:

[0046] The knowledge base unit is used to construct a process knowledge graph by utilizing the acquired CKD-cardiovascular authoritative guidelines and the hospital's case database. Literature NLP is deployed in the process knowledge graph to automatically monitor relevant papers. A reinforcement learning mechanism is introduced into the process knowledge graph to statistically analyze the long-term practical benefits of different rules in similar populations in order to dynamically adjust the priority of knowledge recommendations.

[0047] A personalized recommendation unit is used to design individual rules and use the individual rules to make personalized treatment recommendations.

[0048] The intelligent rehabilitation unit is used to develop corresponding follow-up plans based on the risk level and to break down the personalized treatment recommendations into patient lists, doctor lists, nutritionist lists, and rehabilitation lists, so as to assign corresponding rehabilitation tasks to different roles.

[0049] Optionally, the individual rules are used to dynamically retrieve the process knowledge graph based on the rule input, and perform intelligent matching and conflict resolution to output personalized treatment recommendations; wherein, the rule input includes the risk prediction probability, the high-risk significant factors, the patient's clinical status, and comorbidities.

[0050] Optionally, the doctor-patient interaction module includes:

[0051] The multi-terminal platform unit is used to build a data service bus that supports parallel collaboration between App, Mini Program, Doctor Web Workbench and Backend Management Terminal, resulting in a multi-terminal collaborative microservice architecture. It also utilizes blockchain and zero-trust identity authentication to perform hierarchical encryption and full operation traceability on sensitive data in the collaborative microservice architecture to protect data privacy.

[0052] The early warning triggering unit is used to generate corresponding early warning information on the doctor's end and the patient's end respectively, based on the risk level and the rehabilitation task, so as to carry out risk intervention.

[0053] This invention discloses the following technical effects by providing a deep learning-based system for predicting and treating cardiovascular adverse events in patients with chronic kidney disease:

[0054] 1. Full Utilization of Information: 1) By combining structured and unstructured data from hospitals with real-time health data from wearable devices, seamless integration of multimodal and multidimensional data is achieved, laying a solid data foundation for subsequent intelligent modeling and personalized diagnosis and treatment. 2) API-ETL interfaces and dynamic medical data adapters are built to ensure cross-platform and heterogeneous system data collection compatibility, achieving deep and extensive information integration and providing data support for the analysis of complex diseases. 3) Blockchain technology is used for encrypted records, improving data security and traceability, preventing the leakage of sensitive medical and patient information, and meeting compliance requirements.

[0055] 2. Reliable and accurate risk prediction: Through an end-to-end multimodal deep learning model, it outputs individual future risk probabilities, risk levels, and risk reports. This not only makes risk prediction more accurate and comprehensive, dynamically adapting to data heterogeneity across different patients and follow-up periods, but also enhances the forward-looking and proactive intervention capabilities of disease management. It significantly improves model interpretability and clinical applicability, and substantially reduces disability / mortality rates in CKD patients.

[0056] 3. High Clinical Translation Efficiency: 1) Through multiple interpretable layers, the high-risk factors and time points that the AI ​​model focuses on are presented intuitively, helping clinicians judge the rationality of the model and assisting doctors in making decision-making more transparent. 2) By outputting an individualized risk contribution list and automatically connecting to authoritative clinical guideline knowledge graphs, precise personalized intervention suggestions are automatically recommended, greatly improving doctors' work efficiency and the accuracy of patient interventions. 3) Through intelligent breakdown and push of rehabilitation tasks, doctors, patients, nutritionists, and other roles can collaborate to carry out precise rehabilitation treatment, promoting the personalization and refinement of chronic disease management, achieving efficient closed-loop management of the medical team, and promoting the development of modern medicine.

[0057] 4. By building a microservice architecture that enables multi-terminal, multi-platform, and multi-role collaboration, doctor-patient communication becomes more convenient and secure, enabling early intervention and remote management, preventing information silos, better coordinating and managing patients with chronic diseases, and greatly improving the response efficiency of medical services.

[0058] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 This is a schematic diagram of the system architecture provided in an embodiment of the present invention;

[0061] Figure 2 A schematic diagram of the risk prediction process provided in an embodiment of the present invention;

[0062] Figure 3 This is a schematic diagram of the personalized treatment recommendation process provided in an embodiment of the present invention. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0065] like Figure 1 As shown, this invention provides a deep learning-based system for predicting and treating cardiovascular adverse events in patients with chronic kidney disease, including interconnected multi-source information integration modules, risk prediction modules, decision support modules, personalized diagnosis and treatment modules, and doctor-patient interaction modules.

[0066] 1. Multi-source information integration module

[0067] like Figure 2 As shown, this module is used to interface with hospital systems, acquire structured data, unstructured data, and real-time health data of patients, and then fuse and store the acquired data to obtain multimodal fused data. The multi-source information integration module includes:

[0068] 1.1 Structured Data Acquisition Unit

[0069] API-ETL is used to build links with hospital information systems, laboratory information systems, and electronic medical record systems, and to develop dynamic medical data adapters compatible with different databases to achieve standard interfaces, acquire structured data, and use blockchain technology for encrypted recording; the collection, access, modification, and deletion of each piece of structured data are encrypted to ensure data compliance and trustworthiness.

[0070] Structured data: patient demographic information, laboratory results, medical history and medication history, etc.

[0071] 1.2 Unstructured Data Acquisition Unit

[0072] It is used to interface with medical image archiving and transmission systems, automatically acquire and store DICOM images, interface with electronic medical record systems, obtain unstructured documents, and combine the DICOM images and the unstructured documents to obtain unstructured data.

[0073] DICOM imaging: echocardiography, CT, MRI, etc. Unstructured documents: free text, chief complaint, discharge summary, medical orders, clinical symptoms, family history, etc.

[0074] 1.3 Data Conversion Unit

[0075] This is used to automatically extract image features from the DICOM images using deep learning technology to obtain structured indicators, and to use a medical NLP model to perform named entity recognition, relation extraction, and event tracing on the structured documents, so as to transform text information into structured fields.

[0076] 1.4 Continuous Monitoring Unit

[0077] This design utilizes a wireless gateway to automatically detect and connect to patient wearable devices, obtaining real-time health data. Wearable devices include Apple Watch, Xiaomi Mi Band, and Bluetooth blood pressure monitors.

[0078] 1.5 Data Fusion Unit

[0079] Based on the structured data, the unstructured data, and the real-time health data, time alignment, missing data completion, and consistency checks are performed to obtain multimodal fusion data, and a patient data archive is constructed to store the multimodal fusion data; wherein, the main index of the patient data archive is the patient ID, and the format is columnar storage.

[0080] 2. Risk Prediction Module

[0081] like Figure 2 As shown, a risk prediction model is designed and integrated with a structured branch network, a medical imaging branch network, a medical text branch network, and an attention fusion layer. This model outputs the risk prediction probability, risk level, and risk report for each patient. The risk prediction module includes:

[0082] 2.1 Modular Network Unit

[0083] For the structured data and the unstructured data, corresponding network structures are designed respectively to obtain a structured branch network, a medical image branch network, and a medical text branch network. These three branch networks are then integrated to obtain an initial model; wherein:

[0084] The structured branch network consists of: a first input layer for inputting structured data sequences, a multilayer perceptron for extracting high-order features, and a gated recurrent unit for capturing temporal dynamics.

[0085] The medical imaging branch network consists of a second input layer for inputting standard medical images and an image encoder for capturing global-local multi-scale features. Standard medical images include, for example, 2D / 3D slices in DICOM format.

[0086] The medical text branch network consists of: a third input layer for input text, a Transformer base layer for extracting word sequence context embeddings, and a CLS task adaptation layer for outputting final features and linearly transforming them to a shared dimension.

[0087] 2.2 Fusion Layer Unit

[0088] This is used to design an attention fusion layer. Based on the input of the initial model, the attention fusion layer performs feature attention scoring to dynamically adjust feature weights and obtain the attention distribution. The expression for calculating the dynamic feature weights is:

[0089]

[0090]

[0091] in, For attention score, This is the first layer weight matrix. This is the second-layer weight matrix vector. For the first Each modal feature vector These are all bias parameters. For the first Dynamic attention weights for each modality The three modal features can be concatenated into an input matrix.

[0092] 2.3 Model Training Unit

[0093] The model training unit is used to embed the attention fusion layer into the initial model, perform end-to-end training, obtain a risk prediction model, and output the risk prediction probability, risk level, and risk report for each patient using the risk prediction model.

[0094] 2.3.1 Training Subunit

[0095] This method utilizes multimodal fusion data from the input model to perform forward propagation, global loss calculation, and backpropagation-optimizer training, dynamically adjusting the learning rate and regularization coefficient to complete end-to-end training of the initial model, thereby obtaining a risk prediction model. The expression for the global loss calculation is:

[0096]

[0097]

[0098]

[0099] in, Main loss function, For feature separation regularization terms, For the global loss function, For the first The true label of each sample The risk probability output by the model. For modality With mode cosine similarity, This is the regularization coefficient.

[0100] 2.3.2 Output Subunit

[0101] This system is used to output the predicted probability of cardiovascular events for each patient within a certain time window using the risk prediction model, and to classify and report the risk level based on the predicted probability. Risk levels are, for example: low risk, intermediate risk, and high risk.

[0102] 3. Decision Support Module

[0103] The decision support module is used to generate multiple interpretable layers based on the multimodal fusion data and the output of the risk prediction model, calculate high-risk significance factors, and automatically recommend clinical decision-making suggestions based on the high-risk significance factors.

[0104] 3.1 Interpretable Unit

[0105] This is used to construct modal-level attention heatmaps, image region visualizations, and text attention visualizations based on temporal structured feature attention, resulting in multiple interpretable layers. Interpretable units include:

[0106] 1) Modal-level attention heatmap:

[0107] The attention distribution output by the attention fusion layer and the temporal dynamics output by the gated loop unit are used as inputs to calculate the feature importance at each time point, output the modal attention radar map for each patient, and highlight high-risk and sensitive moments in the course of the disease in the modal attention radar map to obtain a modal-level attention heatmap.

[0108] Modal attention radar chart: Displays the model's current attention weight towards structured, image, and text data. Doctors can see at a glance which types of information the model's risk assessment primarily relies on, making it intuitive and interpretable.

[0109] 2) Image area visualization map:

[0110] The multi-scale features output by the medical imaging branch network are used as input to output a fine-grained heatmap, and anatomical structure labels are introduced into the fine-grained heatmap to obtain a visualization map of the image region, which facilitates spatial tracing by doctors; among them, anatomical structure labels include, for example, thinning of the left ventricular wall layer, coronary artery calcification area, etc.

[0111] 3) Text attention visualization

[0112] The attention weights of each layer of tokens output by the Transformer base layer are used as input, and the relevant risk content of important keywords ("palpitation", "nighttime breathing difficulties", etc.) in the structured document are highlighted to obtain a text attention visualization.

[0113] 3.2 Risk Factor Analysis Unit

[0114] Based on the interpretable layer, high-risk significant factors are screened out, such as "eGFR<35, duration 60 days", "left ventricular hypertrophy", "severe proteinuria", etc.

[0115] The high-risk significant factors are calculated and ranked to obtain an individualized risk contribution list; the calculation expression for the high-risk significant factors is as follows:

[0116]

[0117] in, For factor normalization scores, For attribution strength, For model attention, The frequency of factor abnormalities These are weighted parameters.

[0118] 3.3 Decision Support Unit

[0119] This is used to connect the individualized risk contribution list with authoritative CKD-cardiovascular guidelines or corresponding knowledge graphs to automatically recommend clinical decision-making suggestions.

[0120] 4. Personalized Treatment Module

[0121] like Figure 3 As shown, this is used to construct a process knowledge graph, make personalized treatment recommendations based on individual rules, and assign corresponding rehabilitation tasks to different roles based on the personalized treatment recommendations. The personalized treatment module includes:

[0122] 4.1 Knowledge Base Unit

[0123] This method utilizes the acquired CKD-cardiovascular authoritative guidelines and the hospital's case database to construct a process knowledge graph. Literature NLP is deployed in the process knowledge graph to automatically monitor relevant papers. A reinforcement learning mechanism is introduced into the process knowledge graph to statistically analyze the long-term practical benefits of different rules in similar populations, so as to dynamically adjust the priority of knowledge recommendations. Among them, long-term practical benefits include, for example, a 10% reduction in events.

[0124] 4.2 Personalized Recommendation Unit

[0125] This is used to design individual rules, which are then used to make personalized treatment recommendations. The individual rules are used to dynamically retrieve the process knowledge graph based on the rule input, and to perform intelligent matching and conflict resolution to output personalized treatment recommendations. The rule input includes the risk prediction probability, the high-risk significant factors, the patient's clinical status, and comorbidities.

[0126] For example:

[0127] Intelligent matching: eGFR<45+ hypertension+ hyperkalemia → ACEI adjustment, diuretics to avoid potassium;

[0128] Conflict resolution: The system scans for contradictions in activated suggestions, such as "water restriction" versus "dehydration risk".

[0129] Personalized treatment recommendations include multi-dimensional suggestions such as medication control, follow-up examinations, nutritional adjustments, and lifestyle suggestions. The algorithm incorporates patient preferences and can adaptively expand to reflect individual needs.

[0130] 4.3 Intelligent Rehabilitation Unit

[0131] like Figure 3 As shown, it is used to develop corresponding follow-up plans based on the risk level, and to break down the personalized treatment recommendations into patient lists, doctor lists, nutritionist lists, and rehabilitation lists, so as to assign corresponding rehabilitation tasks to different roles.

[0132] Follow-up plans, for example:

[0133] High risk: Renal function and echocardiography should be checked every 2 months, and a full laboratory test should be performed monthly;

[0134] Medium risk: Quarterly follow-up visits + special inspections of key risk factors;

[0135] Low risk: Extend the follow-up period every six months to one year to facilitate optimal resource allocation.

[0136] List splitting, for example:

[0137] Patient side: Record daily blood pressure / weight / medication adherence;

[0138] Doctor's side: Remote assessment every 2 weeks, and adjustment of the treatment plan as needed;

[0139] Nutritionist's side: Schedule dietary interventions and receive reminders about seasoning restrictions.

[0140] 5. A doctor-patient interaction module, used to build a multi-terminal collaborative microservice architecture, manage multi-terminal collaboration, and push early warning information in real time based on the risk level and the rehabilitation execution task. The doctor-patient interaction module includes:

[0141] 5.1 Multi-terminal platform unit

[0142] This is used to build a data service bus that supports parallel collaboration between App, Mini Program, Doctor Web Workbench, and Backend Management Terminal, resulting in a multi-terminal collaborative microservice architecture. By utilizing blockchain and zero-trust identity authentication, sensitive data in the collaborative microservice architecture is subjected to hierarchical encryption and full operation traceability to protect data privacy.

[0143] 5.2 Early Warning Triggering Unit

[0144] Based on the risk level and the rehabilitation task, the system generates corresponding early warning information on both the doctor's and patient's ends to intervene in the risk.

[0145] Warning information, for example:

[0146] Level 1 Low Risk - Reminder: Brief information is pushed to the patient's end, and the doctor's end is silently recorded. The doctor can handle or pay attention to the information independently.

[0147] Level 2 Intermediate Risk - Follow-up: Automatically generates follow-up / re-examination tasks for doctors, and provides patients with pop-up self-test forms / encrypted AI consultations; pushes collaborative tasks with family / rehabilitation therapists;

[0148] Level 3 High-Risk - Emergency: Simultaneous notifications to doctors, responsible department heads, and the emergency system; strong reminders for patients with one-click SOS; supports AI multimodal "quick Q&A" to collect symptom changes, and recommends automatic referral / remote consultation work orders when necessary.

[0149] Therefore, by providing a deep learning-based system for predicting and treating cardiovascular adverse events in patients with chronic kidney disease, this invention can effectively improve the accuracy and efficiency of predicting cardiovascular adverse events in patients with chronic kidney disease, push personalized treatment plans to different patients, achieve early diagnosis and early intervention of chronic kidney disease, reduce the burden on doctors, improve patient compliance, and promote intelligent and precise management of chronic diseases.

[0150] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0151] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A deep learning-based system for predicting and treating cardiovascular adverse events in patients with chronic kidney disease, characterized in that, include: The multi-source information integration module is used to connect to the hospital system, acquire structured data, unstructured data, and patients' real-time health data, and then fuse and store the acquired data to obtain multimodal fused data; The risk prediction module is used to design and integrate structured branch networks, medical imaging branch networks, medical text branch networks and attention fusion layers to obtain a risk prediction model. The risk prediction model is used to output the risk prediction probability, risk level and risk report for each patient. The decision support module is used to generate multiple interpretable layers based on the multimodal fusion data and the output of the risk prediction model, calculate high-risk significance factors, and automatically recommend clinical decision suggestions based on the high-risk significance factors. The personalized treatment module is used to construct a process knowledge graph, make personalized treatment recommendations based on individual rules, and assign corresponding rehabilitation tasks to different roles based on the personalized treatment recommendations. The doctor-patient interaction module is used to build a multi-terminal collaborative microservice architecture, manage multi-terminal collaboration, and push early warning information in real time according to the risk level and the rehabilitation execution task. The multi-source information integration module, the risk prediction module, the decision support module, the personalized diagnosis and treatment module, and the doctor-patient interaction module are interconnected.

2. The deep learning-based system for predicting and treating cardiovascular adverse events in patients with chronic kidney disease according to claim 1, characterized in that, The multi-source information integration module includes: The structured data acquisition unit is used to build API-ETL that links with hospital information systems, laboratory information systems and electronic medical record systems, and to develop dynamic medical data adapters compatible with different databases to achieve standard docking, acquire structured data, and use blockchain technology for encrypted recording. The unstructured data acquisition unit is used to interface with the medical image archiving and transmission system, automatically acquire and store DICOM images, interface with the electronic medical record system to acquire unstructured documents, and combine the DICOM images and the unstructured documents to obtain unstructured data. The data conversion unit is used to automatically extract image features from the DICOM image using deep learning technology to obtain structured indicators, and to use a medical NLP model to perform named entity recognition, relation extraction and event tracing on the structured document to convert text information into structured fields. A continuous monitoring unit is used to design a wireless gateway, which automatically senses and connects to the patient's wearable device to obtain real-time health data. The data fusion unit is used to perform time alignment, missing data completion, and consistency checks based on the structured data, the unstructured data, and the real-time health data to obtain multimodal fused data, and to construct a patient data archive that stores the multimodal fused data; wherein the main index of the patient data archive is the patient ID and the format is columnar storage.

3. The deep learning-based system for predicting and treating cardiovascular adverse events in patients with chronic kidney disease according to claim 2, characterized in that, The risk prediction module includes: Modular network units are used to design corresponding network structures for the structured data and the unstructured data respectively, to obtain a structured branch network, a medical image branch network and a medical text branch network, and to integrate the three branch networks to obtain an initial model; The fusion layer unit is used to design the attention fusion layer. Based on the input of the initial model, the attention fusion layer performs feature attention scoring to dynamically adjust the feature weights and obtain the attention distribution. The calculation expression for the dynamic feature weights is: ； ； in, For attention score, This is the first layer weight matrix. This is the second-layer weight matrix vector. For the first Each modal feature vector These are all bias parameters. For the first Dynamic attention weights for each modality Modal features; The model training unit is used to embed the attention fusion layer into the initial model, and then perform end-to-end training of the model to obtain a risk prediction model. The risk prediction model is used to output the risk prediction probability, risk level and risk report for each patient.

4. The deep learning-based system for predicting and treating cardiovascular adverse events in patients with chronic kidney disease according to claim 3, characterized in that: The structured branch network includes a first input layer for inputting structured data sequences, a multilayer perceptron for extracting high-order features, and a gated recurrent unit for capturing temporal dynamics. The medical imaging branch network includes a second input layer for inputting standard medical images and an image encoder for capturing global-local multi-scale features; The medical text branching network includes a third input layer for inputting text, a Transformer base layer for extracting word sequence context embeddings, and a CLS task adaptation layer for outputting final features and linearly transforming them to a shared dimension.

5. The deep learning-based system for predicting and treating cardiovascular adverse events in patients with chronic kidney disease according to claim 4, characterized in that, The model training unit includes: The training subunit utilizes the multimodal fusion data from the input model to perform forward propagation, global loss calculation, and backpropagation-optimizer training, dynamically adjusting the learning rate and regularization coefficient to complete the end-to-end training of the initial model, thereby obtaining the risk prediction model; the expression for the global loss calculation is: ； ； ； in, Main loss function, For feature separation regularization terms, For the global loss function, For the first The true label of each sample The risk probability output by the model. For modality With mode cosine similarity, The regularization coefficient is used. The output subunit is used to output the predicted probability of cardiovascular events for each patient within a certain time window using the risk prediction model, and to classify the risk level and output a risk report based on the predicted probability.

6. The deep learning-based system for predicting and treating cardiovascular adverse events in patients with chronic kidney disease according to claim 5, characterized in that, The decision support module includes: Interpretable units are used to construct modal-level attention heatmaps, image region visualization maps, and text attention visualization maps based on temporal structured feature attention, resulting in multiple interpretable layers; The risk factor analysis unit is used to filter out high-risk significant factors based on the interpretable layer, calculate and sort the high-risk significant factors, and obtain an individualized risk contribution list; the calculation expression for the high-risk significant factors is: ； in, For factor normalization scores, For attribution strength, For model attention, The frequency of factor abnormalities These are weighted parameters; The decision support unit is used to connect the individualized risk contribution list with authoritative CKD-cardiovascular guidelines or corresponding knowledge graphs to automatically recommend clinical decision suggestions.

7. The deep learning-based system for predicting and treating cardiovascular adverse events in patients with chronic kidney disease according to claim 6, characterized in that, The interpretable unit includes: A modal attention heatmap is used to input the attention distribution output by the attention fusion layer and the temporal dynamics output by the gated loop unit, calculate the feature importance at each time point, output a modal attention radar map for each patient, and highlight high-risk and sensitive moments in the course of the disease in the modal attention radar map to obtain a modal attention heatmap. An image region visualization map is used to input the multi-scale features output by the medical imaging branch network, output a fine-grained heat map, and introduce anatomical structure labels into the fine-grained heat map to obtain the image region visualization map. A text attention visualization is generated by taking the attention weights of each layer of tokens output from the Transformer base layer as input and highlighting the risk content related to keywords in the structured document.

8. The deep learning-based system for predicting and treating cardiovascular adverse events in patients with chronic kidney disease according to claim 7, characterized in that, The personalized diagnosis and treatment module includes: The knowledge base unit is used to construct a process knowledge graph by utilizing the acquired CKD-cardiovascular authoritative guidelines and the hospital's case database. Literature NLP is deployed in the process knowledge graph to automatically monitor relevant papers. A reinforcement learning mechanism is introduced into the process knowledge graph to statistically analyze the long-term practical benefits of different rules in similar populations in order to dynamically adjust the priority of knowledge recommendations. A personalized recommendation unit is used to design individual rules and use the individual rules to make personalized treatment recommendations. The intelligent rehabilitation unit is used to develop corresponding follow-up plans based on the risk level and to break down the personalized treatment recommendations into patient lists, doctor lists, nutritionist lists, and rehabilitation lists, so as to assign corresponding rehabilitation tasks to different roles.

9. The deep learning-based system for predicting and treating cardiovascular adverse events in patients with chronic kidney disease according to claim 8, characterized in that, The individual rules are used to dynamically retrieve the process knowledge graph based on the rule input, and perform intelligent matching and conflict resolution to output personalized treatment recommendations; wherein, the rule input includes the risk prediction probability, the high-risk significant factors, the patient's clinical status and comorbidities.

10. The deep learning-based system for predicting and treating cardiovascular adverse events in patients with chronic kidney disease according to claim 9, characterized in that, The doctor-patient interaction module includes: The multi-terminal platform unit is used to build a data service bus that supports parallel collaboration between App, Mini Program, Doctor Web Workbench and Backend Management Terminal, resulting in a multi-terminal collaborative microservice architecture. It also utilizes blockchain and zero-trust identity authentication to perform hierarchical encryption and full operation traceability on sensitive data in the collaborative microservice architecture to protect data privacy. The early warning triggering unit is used to generate corresponding early warning information on the doctor's end and the patient's end respectively, based on the risk level and the rehabilitation task, so as to carry out risk intervention.

Images