Cerebral stroke early diagnosis and early warning system and method based on machine learning

Abstract

The invention relates to the technical field of medical information systems, in particular to a cerebral apoplexy early diagnosis and early warning system and method based on machine learning, and the system comprises a data management module, a deep learning model module, a big data processing and analysis module, a big data processing and analysis module, and a big data processing and analysis module. The system comprises a big data processing and analysis module, a real-time monitoring and early warning module in communication connection with the big data processing and analysis module, a man-machine interaction module in communication connection with the real-time monitoring and early warning module, and a cloud computing platform in communication connection with the man-machine interaction module and used for providing data acquisition, feature extraction, data analysis and decision support services. And remote access and update services are provided for the deep learning model module. And the effects of early diagnosis and early warning of cerebral apoplexy are comprehensively improved. The early discovery rate and the treatment effect of cerebral apoplexy can be improved, and a new thought is provided for optimal configuration of medical resources.

Description

Technical Field

[0001] The present invention relates to the technical field of medical information systems, in particular to a stroke early diagnosis and warning system and method based on machine learning. Background Art

[0002] Stroke is an acute cerebrovascular disease that seriously threatens human health, characterized by high incidence, high disability rate, and high mortality rate. With the aggravation of population aging, the incidence of stroke shows an increasing trend year by year. Therefore, early diagnosis and warning of stroke are of great significance for reducing mortality and improving the quality of life of patients.

[0003] Traditional stroke diagnosis methods mainly rely on clinical symptoms and imaging examinations. However, this method has many limitations. First, early symptoms are often atypical and easily overlooked or misdiagnosed. Second, although imaging examinations such as CT or MRI can provide accurate diagnoses, they often require professional radiologists to interpret them, which may lead to diagnostic delays in areas with tight medical resources. In addition, traditional methods are difficult to achieve early warning of stroke and cannot provide timely intervention suggestions for high-risk populations.

[0004] In recent years, with the development of artificial intelligence technology, machine learning methods have shown great potential in the field of medical image analysis. Some studies have tried to apply deep learning algorithms to the imaging diagnosis of stroke and achieved certain results. However, these methods still have some problems. First, most studies only focus on single-type medical image data, such as CT or MRI, ignoring the advantages of multi-modal data fusion. Second, existing models can often only give the diagnosis results of stroke, lacking the ability to evaluate the severity of the disease and predict the prognosis. Moreover, these models are usually static and cannot be continuously optimized with the accumulation of new data, making it difficult to adapt to the rapid update of medical knowledge.

[0005] In addition, most existing stroke auxiliary diagnosis systems operate independently and lack effective integration with hospital information systems, which limits their application in clinical practice. At the same time, the issues of patient data security and privacy protection have not been fully emphasized.

[0006] In view of the above problems, there is an urgent need for a stroke diagnosis and warning system that can comprehensively utilize multi-modal data, achieve early diagnosis and warning, have the ability of continuous learning, and ensure data security at the same time. The present invention is proposed in response to this need. Summary of the Invention

[0007] The machine learning-based stroke early diagnosis and early warning system and method proposed in the present invention are intended to solve many problems existing in the prior art. The system integrates multi-source heterogeneous data and builds a deep learning model to achieve early diagnosis, severity assessment and early warning of stroke. At the same time, the system has the ability to continuously learn and update, and can adapt to the rapid development of medical knowledge. In addition, the present invention also focuses on data security and privacy protection, which provides a guarantee for the clinical application of the system.

[0008] The present invention proposes an early diagnosis and early warning system for stroke based on machine learning, comprising:

[0009] Data management module for:

[0010] Collect patient data from the cloud computing platform and store it in the system;

[0011] Clean the collected data and remove invalid data;

[0012] Process data according to data standardization templates to improve data quality;

[0013] A deep learning model module is communicatively connected to the data management module and is used to:

[0014] receiving processed patient data sent by the data management module;

[0015] Based on the patient data, a machine learning model is constructed, and supervised learning, unsupervised learning, and semi-supervised learning are performed on the data to achieve model training;

[0016] The big data processing and analysis module is connected to the deep learning model module for:

[0017] Using big data processing and analysis technology, the analysis results output by the deep learning model module are analyzed in real time and dynamically monitored;

[0018] The real-time monitoring and early warning module is connected to the big data processing and analysis module for:

[0019] Based on the analysis results of the big data processing and analysis module, a stroke diagnosis and early warning decision model is generated;

[0020] Provide early stroke diagnosis results and early warning predictions for stroke diagnosis reports at different stages;

[0021] The human-computer interaction module is connected to the real-time monitoring and early warning module for:

[0022] Realize the real-time display and interaction of the data management module, deep learning model module, big data processing and analysis module and real-time monitoring and early warning module;

[0023] Realize remote access and real-time synchronous update based on the cloud computing platform;

[0024] The cloud computing platform, communicatively connected to the human-computer interaction module, is configured to:

[0025] Provide data collection, feature extraction, data analysis and decision support services;

[0026] Provide remote access and update services for the deep learning model module.

[0027] Preferably, the data management module includes:

[0028] A data acquisition sub-module, configured to:

[0029] Establish a stroke data center through the cloud computing platform;

[0030] Receive brain image data, personal information data, medical records and treatment result data of stroke patients;

[0031] Locally store the received data;

[0032] A data processing sub-module, configured to:

[0033] Preliminarily process the data stored by the data acquisition sub-module, including cleaning and eliminating data with blanks, inconsistencies and missing annotations;

[0034] Standardize the data using a standardization method;

[0035] Establish a machine learning model through the deep learning model module, and perform feature extraction, feature selection and feature screening on the processed data.

[0036] Preferably, the deep learning model module includes:

[0037] A feature extraction network, configured to:

[0038] Extract image features using a network structure including residual units, SE channels, convolutional kernels and pooling layers;

[0039] The convolutional kernel has a size of 7×7, a quantity of 3, and a stride of 2;

[0040] A target classification network, communicatively connected to the feature extraction network, is configured to:

[0041] Perform stroke type classification using a network structure including convolutional kernels, pooling layers, fully connected layers and Softmax output layers;

[0042] The convolutional kernel has a size of 3×3, a quantity of 3, and a stride of 2;

[0043] A severity classification network, communicatively connected to the target classification network, for:

[0044] Judging the severity of stroke by using a network structure including a convolutional kernel, a pooling layer, a fully connected layer, and a Softmax output layer;

[0045] The size of the convolutional kernel is 3×3, the number is 3, and the stride is 2.

[0046] Preferably, the real-time monitoring and early warning module includes:

[0047] A diagnosis and evaluation sub-module, for:

[0048] Based on the model trained by the deep learning model module, diagnosing stroke for the input medical image data;

[0049] Calculating the accuracy rate, recall rate, and F1 metric of the diagnosis result;

[0050] An early warning decision sub-module, communicatively connected to the diagnosis and evaluation sub-module, for:

[0051] Based on the output result of the diagnosis and evaluation sub-module, combining with the real-time monitoring data of the patient, generating a stroke early warning information;

[0052] Adopting an active early warning mechanism to trigger an early warning according to a preset threshold.

[0053] Preferably, the big data processing and analysis module includes:

[0054] A data analysis sub-module, for:

[0055] Analyzing the correlation between the incidence rate of stroke, prognosis effect and factors such as age, gender, occupation, living habits, and region;

[0056] Constructing a stroke disease feature vector including multi-dimensional information;

[0057] A result output sub-module, communicatively connected to the data analysis sub-module, for:

[0058] Generating an analysis report on the stroke disease feature vector, incidence rate of stroke, and prognosis effect;

[0059] Sending the analysis result to the human-computer interaction module for display.

[0060] Preferably, the human-computer interaction module includes:

[0061] A user interaction sub-module, for:

[0062] Providing a graphical interface to display the diagnosis result and early warning information generated by the real-time monitoring and early warning module;

[0063] Receive the query request and operation instructions input by the user;

[0064] The system management sub-module, which is communicatively connected to the user interaction sub-module, is used for:

[0065] Manage user permissions and system configurations;

[0066] Coordinate the operation and data exchange of each functional module;

[0067] Implement remote access and real-time synchronous update of the system.

[0068] Preferably, the cloud computing platform adopts a distributed computing architecture, including:

[0069] A data storage cluster for storing a large amount of stroke-related data;

[0070] A computing resource pool for providing high-performance computing resources for the deep learning model module;

[0071] A task scheduler for coordinating the allocation and execution of data processing and model training tasks;

[0072] An API gateway for providing standardized interfaces to implement data exchange and service calls with external systems.

[0073] Preferably, it further includes a data security module for:

[0074] Perform desensitization processing on the collected patient data;

[0075] Use encryption algorithms to protect the security of data transmission and storage;

[0076] Implement role-based access control to ensure the legality of data access.

[0077] Preferably, it further includes a model update module for:

[0078] Regularly evaluate the performance of the deep learning model;

[0079] Based on the newly added training data, automatically trigger the incremental learning of the model;

[0080] Implement the online update of the model without affecting the normal operation of the system.

[0081] A method for early diagnosis and warning of stroke based on machine learning, including the following steps:

[0082] (1) Collect the brain imaging data, personal information data, medical records, and diagnosis and treatment result data of patients through the cloud computing platform;

[0083] (2) Clean and standardize the collected data to improve data quality;

[0084] (3) Build a deep learning model, including a feature extraction network, a target classification network, and a severity classification network, where:

[0085] The feature extraction network includes residual units, SE channels, 7×7 convolutional kernels, and pooling layers;

[0086] Both the target classification network and the severity classification network include 3×3 convolutional kernels, pooling layers, fully connected layers, and Softmax output layers;

[0087] (4) Train the deep learning model using the processed data, adopting a combination of supervised learning, unsupervised learning, and semi-supervised learning;

[0088] (5) Use the trained model to diagnose stroke in new medical image data, and calculate the accuracy, recall rate, and F1 metric of the diagnosis results;

[0089] (6) Based on the diagnosis results and real-time patient monitoring data, generate stroke warning information using an active warning mechanism;

[0090] (7) Use big data processing and analysis techniques to analyze the correlation between the incidence of stroke, prognosis, and various factors, and construct a multi-dimensional stroke disease feature vector;

[0091] (8) Real-time display the diagnosis results, warning information, and data analysis reports through a human-computer interaction interface;

[0092] (9) Regularly evaluate the model performance, and perform incremental learning based on new added data to achieve online update of the model;

[0093] (10) Adopt measures such as data desensitization, transmission encryption, and role-based access control to ensure the data security of the entire system.

[0094] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0095] First, from the overall architecture, the present invention constructs a complete early stroke diagnosis and warning ecosystem. Through the organic combination of a data management module, a deep learning model module, a big data processing and analysis module, a real-time monitoring and warning module, a human-computer interaction module, and a cloud computing platform, it realizes the full-process automation from data collection, processing, analysis to warning. This system-level innovation greatly improves the efficiency and accuracy of diagnosis, and at the same time provides comprehensive decision-making support for doctors.

[0096] Secondly, in terms of data utilization, the present invention breaks through the limitation of traditional methods that only rely on a single data source. By integrating multi-modal information such as the patient's brain imaging data, personal information, medical records, and real-time monitoring data, the system can conduct a more comprehensive and accurate assessment of the patient's condition. This multi-dimensional data analysis not only improves the accuracy of diagnosis but also provides a basis for formulating personalized treatment plans.

[0097] In terms of model design, the present invention adopts an innovative deep learning architecture. The residual units and SE channels introduced in the feature extraction network can effectively extract the deep features of images, while the designs of the target classification network and the severity classification network achieve an accurate assessment of the type and severity of stroke. This multi-task learning framework not only improves the generalization ability of the model but also provides richer information for clinical decision-making.

[0098] In terms of the early warning mechanism, the active early warning mechanism proposed by the present invention is a major innovation point. By comprehensively considering the model diagnosis results, the patient's physiological indicators, and lifestyle data, the system can timely detect potential stroke risks and provide early intervention suggestions for high-risk groups. This function is of great significance for reducing the incidence of stroke and improving the treatment effect.

[0099] In terms of the system scalability and adaptability, the model update module of the present invention ensures that the system can continuously learn and optimize. By regularly evaluating the model performance and conducting incremental learning, the system can continuously adapt to new medical discoveries and clinical practices and maintain the advancement of its diagnostic capabilities. This dynamic update feature enables the system to remain highly efficient and reliable in the long term.

[0100] In terms of data security, the present invention takes multiple measures to protect patient privacy and data security. Through technologies such as data desensitization, transmission encryption, and role-based access control, the system ensures the security of patient data while providing efficient diagnostic services. This clears the obstacles for the wide application of the system in clinical practice.

[0101] Finally, the human-computer interaction design of the present invention greatly improves the usability of the system. The intuitive display of diagnostic results and the presentation method of early warning information enable doctors to quickly understand and utilize the information output by the system. At the same time, the remote access and real-time synchronous update functions of the system make remote diagnosis and multi-center collaboration possible.

[0102] In summary, through the innovation of the system architecture, the integration of multi-modal data, advanced deep learning models, active warning mechanisms, continuous learning capabilities, and strict data security measures, the present invention comprehensively improves the effect of early diagnosis and warning of stroke. This not only helps to improve the early detection rate and treatment effect of stroke, but also provides new ideas for the optimal allocation of medical resources. The implementation of the present invention will have a profound impact on reducing the disability rate and mortality of stroke, improving the quality of life of patients, and alleviating the social medical burden, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0103] Figure 1 It is a diagram of the data management and deep learning model of the present invention.

[0104] Figure 2 It is a diagram of big data processing and real-time monitoring and warning of the present invention.

[0105] Figure 3 It is a diagram of the human-computer interaction and cloud computing platform of the present invention.

[0106] Figure 4 It is a diagram of data security and model update of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0107] The present invention provides a system and method for early diagnosis and warning of stroke based on machine learning. The system includes a data management module 1, a deep learning model module 2, a big data processing and analysis module 3, a real-time monitoring and warning module 4, a human-computer interaction module 5, and a cloud computing platform 6. These modules work together to achieve the full-process automation from data collection, processing, analysis to warning, greatly improving the accuracy and efficiency of early diagnosis of stroke.

[0108] The data management module 1 is used to collect patient data from the cloud computing platform 6 and store it in the system. In a preferred embodiment of the present invention, these data include but are not limited to the patient's brain CT images, magnetic resonance images, personal basic information, past medical history, etc. The data management module 1 is also responsible for cleaning the collected data and removing invalid data. For example, a threshold can be set to determine that the image data with a signal-to-noise ratio lower than 20 dB is invalid data and be removed. In addition, the data management module 1 will also process the data according to a preset data standardization template to improve the data quality. Here, the standardization processing can include image size normalization, gray value standardization, etc.

[0109] The deep learning model module 2 is communicatively connected to the data management module 1, and is used to receive the processed patient data and construct a machine learning model based on this data. The present invention adopts a combined method of supervised learning, unsupervised learning, and semi-supervised learning for model training to make full use of the labeled and unlabeled data. In a specific embodiment, a basic model can be first trained using the supervised learning method, then pre-trained on a large amount of unlabeled data using the unsupervised learning method, and finally the model can be fine-tuned using the semi-supervised learning method to obtain better generalization ability.

[0110] The big data processing and analysis module 3 is communicatively connected to the deep learning model module 2, and is responsible for real-time analysis and dynamic monitoring of the analysis results output by the model. This module adopts a distributed computing framework, such as Apache Spark, and can efficiently process large-scale stroke-related data. In an embodiment, this module can calculate the stroke incidence rates of different age groups and different regions in real time and conduct trend analysis.

[0111] The real-time monitoring and warning module 4 is communicatively connected to the big data processing and analysis module 3, and generates a stroke diagnosis and warning decision model based on the analysis results. This module can give early diagnosis results and warning predictions for the diagnosis reports of strokes at different times. For example, for newly input patient data, the system can give a diagnosis result (such as whether the patient has a stroke and what type it belongs to), and at the same time predict the risk probability of a severe stroke occurring within a future period (such as within 3 months).

[0112] The human-computer interaction module 5 is communicatively connected to the real-time monitoring and warning module 4, and is used to realize the real-time display and interaction of each functional module. This module provides an intuitive graphical user interface through which doctors can view diagnosis results, warning information, etc. In addition, this module also realizes the functions of remote access and real-time synchronous update based on the cloud computing platform 6, enabling doctors to obtain the latest diagnosis information anytime and anywhere.

[0113] The cloud computing platform 6 is communicatively connected to the human-computer interaction module 5, and provides powerful computing and storage capabilities for the entire system. This platform not only supports data collection, feature extraction, data analysis, and decision support, but also provides remote access and update services for the deep learning model module 2. In an embodiment, public cloud platforms such as AWS or Alibaba Cloud can be used, or a private cloud can be built according to the needs of the hospital.

[0114] The data management module 1 of the present invention further includes a data acquisition sub-module 11 and a data processing sub-module 12. The data acquisition sub-module 11 establishes a stroke data center through the cloud computing platform 6, receives various types of patient data and performs local storage. The data processing sub-module 12 is responsible for the preliminary processing of the stored data, including data cleaning and standardization. During the data cleaning process, multiple rules can be set to identify and process abnormal data. For example, for numerical data, the 3σ principle can be used to identify outliers; for date data, it can be checked whether it is within a reasonable range (e.g., the date of birth should not be later than the current date).

[0115] In terms of the standardization process, the present invention adopts an improved Z-score method. Specifically, for a numerical feature x, its standardized value x ′ is calculated as follows:

[0116]

[0117] where μ is the mean of the feature, σ is the standard deviation, and ∈ is a small positive number (such as 1×10 -8 ), used to avoid the denominator being zero. This method can effectively transform features of different scales to the same scale, facilitating subsequent processing by machine learning models.

[0118] The deep learning model module 2 includes a feature extraction network 21, a target classification network 22, and a severity classification network 23. The feature extraction network 21 adopts a structure including residual units and SE (Squeeze-and-Excitation) channels. This design can effectively extract the deep features of the image while maintaining the stable propagation of gradients. In an embodiment of the present invention, the feature extraction network 21 uses 3 convolutional kernels of 7×7 with a stride of 2. This setting can reduce the computational amount while maintaining a large receptive field.

[0119] Both the target classification network 22 and the severity classification network 23 adopt a structure including convolutional kernels, pooling layers, fully connected layers, and a Softmax output layer. They use 3 convolutional kernels of 3×3 with a stride of 2. Such small-sized convolutional kernels can reduce the number of parameters, and at the same time, a large receptive field can be obtained by stacking multiple layers. The Softmax output layer is used to convert the output of the network into a probability distribution, facilitating interpretation and decision-making.

[0120] During the model training process, the present invention adopts a strategy of dynamically adjusting the learning rate. Specifically, the value of the learning rate η at the t-th iteration is calculated as follows:

[0121] η t = η 0 ·(1 + α·t) -β ,

[0122] where η 0 is the initial learning rate (such as 0.01), and α and β are hyperparameters (such as α = 0.001, β = 0.75). This strategy can maintain a relatively large learning rate at the beginning of training for rapid convergence, and gradually reduce the learning rate in the later stage to fine-tune the model parameters.

[0123] From the above description, it can be seen that the system structure of the present invention is reasonable, the functions of each module are clear, and it can effectively achieve the early diagnosis and warning of stroke. Especially in the design and training strategy of the deep learning model, the present invention has made a number of innovations, improving the performance and generalization ability of the model. This machine learning-based method can identify the early signs of stroke faster and more accurately than traditional diagnostic methods, providing important decision-making support for clinical treatment.

[0124] The real-time monitoring and warning module 4 of the present invention further includes a diagnostic evaluation sub-module 41 and a warning decision sub-module 42. The diagnostic evaluation sub-module 41 diagnoses stroke on the input medical image data based on the model trained by the deep learning model module 2. In a preferred embodiment of the present invention, the diagnostic evaluation sub-module 41 can not only determine whether a patient has a stroke, but also further identify the specific type of stroke, such as ischemic stroke or hemorrhagic stroke.

[0125] To evaluate the reliability of the diagnostic results, the diagnostic evaluation sub-module 41 calculates the accuracy, recall rate, and F1 metric of the diagnostic results. The calculation formulas for these metrics are as follows:

[0126]

[0127] where TP is the number of true positive samples, FP is the number of false positive samples, and FN is the number of false negative samples. During the implementation of the present invention, a threshold can be set. For example, when the F1 metric exceeds 0.9, the diagnostic results are considered to have a high level of credibility. The warning decision sub-module 42 is communicatively connected to the diagnostic evaluation sub-module 41 and generates a stroke warning message based on the output results of the diagnostic evaluation and in combination with the real-time monitoring data of the patient. The present invention adopts an innovative active warning mechanism that triggers a warning by presetting a threshold. Specifically, the warning decision sub-module 42 calculates a comprehensive risk score S:

[0128] S = w 1 ·P stroke + w 2 ·R factor + w 3 ·T trend ,

[0129] where P stroke is the probability of stroke predicted by the model, Rfactor is the risk factor score (considering factors such as age, blood pressure, blood sugar, etc.), T trend is the recent index change trend, w 1 , w 2 , w 3 is the weight coefficient (such as w 1 = 0.5, w 2 = 0.3, w 3 = 0.2). When S exceeds the preset threshold (such as 0.7), the system will trigger an alarm. The big data processing and analysis module 3 of the present invention includes a data analysis sub-module 31 and a result output sub-module 32. The data analysis sub-module 31 is responsible for analyzing the correlation between the incidence rate of stroke, prognosis effect and various factors (such as age, gender, occupation, living habits, region, etc.). In a specific embodiment, the data analysis sub-module 31 uses a multiple logistic regression model to analyze the influence of these factors on the risk of stroke. The form of the model is as follows:

[0130]

[0131] where p is the probability of having a stroke, x 1 , x 2 ,..., x n are the respective influencing factors, and β 0 , β 1 ,..., β n are the model parameters. By analyzing the magnitudes and signs of these parameters, the influence degree and direction of each factor on the risk of stroke can be obtained.

[0132] In addition, the data analysis sub-module 31 is also responsible for constructing a stroke disease feature vector containing multi-dimensional information. This feature vector includes data in multiple dimensions such as the patient's basic information, clinical indicators, and imaging features. In the preferred embodiment of the present invention, the principal component analysis (PCA) method is used to reduce the dimension of the feature vector and extract the most representative features.

[0133] The result output sub-module 32 is communicatively connected to the data analysis sub-module 31 and is responsible for generating an analysis report on the stroke disease feature vector and the incidence rate and prognosis effect of stroke. These analysis results will be sent to the human-computer interaction module 5 for display. During the display process, the present invention uses a variety of visualization techniques, such as heat maps, scatter plots, trend lines, etc., to intuitively present the analysis results.

[0134] The human-computer interaction module 5 of the present invention includes a user interaction sub-module 51 and a system management sub-module 52. The user interaction sub-module 51 provides a graphical interface for displaying the diagnostic results and warning information generated by the real-time monitoring and warning module 4. In an embodiment of the present invention, the display of the diagnostic results uses an interactive 3D brain model, which can intuitively display the location and scope of stroke. The warning information is displayed in the form of a dashboard, including the risk level, the change trend of key indicators, etc.

[0135] The user interaction sub-module 51 is also responsible for receiving the query requests and operation instructions input by the user. For example, doctors can view the historical diagnostic records of patients and adjust the warning threshold through this interface. In order to improve the operation efficiency, the present invention adopts the responsive design principle in the user interface design, which can adapt to display devices of different sizes.

[0136] The system management sub-module 52 is communicatively connected to the user interaction sub-module 51 and is responsible for managing user permissions and system configurations. In terms of permission management, the present invention adopts a role-based access control (RBAC) model, which can flexibly set the operation permissions of different users. In terms of system configuration, the system management sub-module 52 provides a number of adjustable parameters, such as data update frequency, model retraining period, etc., to meet the needs of different medical institutions.

[0137] In addition, the system management sub-module 52 is also responsible for coordinating the operation and data exchange of each functional module. During the data exchange process, an encrypted transmission mechanism is adopted to ensure the security of patients' private data. An innovation point of the present invention is to implement the remote access and real-time synchronization update functions of the system. This enables doctors to view the latest diagnostic results and warning information anywhere and at any time, greatly improving the utilization efficiency of medical resources.

[0138] The cloud computing platform 6 of the present invention adopts a distributed computing architecture, including a data storage cluster 61, a computing resource pool 62, a task scheduler 63, and an API gateway 64. The data storage cluster 61 is used to store a large amount of stroke-related data. In a preferred embodiment of the present invention, a storage solution combining a distributed file system (such as HDFS) and a distributed database (such as HBase) is adopted, which can not only efficiently store structured and unstructured data, but also ensure the high availability and scalability of the data.

[0139] The computing resource pool 62 provides high-performance computing resources for the deep learning model module 2. The present invention adopts a GPU cluster to accelerate the training and inference processes of the deep learning model. In a specific embodiment, NVIDIA Tesla V100 GPUs are used, and each GPU has 32GB of video memory, which can effectively process a large amount of medical image data.

[0140] The task scheduler 63 is responsible for coordinating the allocation and execution of data processing and model training tasks. The present invention adopts a priority-based dynamic task scheduling algorithm, which can dynamically adjust the execution order and resource allocation of tasks according to the urgency of tasks and resource requirements. This method can effectively improve the overall efficiency of the system, especially when dealing with emergency diagnosis tasks.

[0141] The API gateway 64 provides standardized interfaces for realizing data exchange and service calls with external systems. The present invention adopts the RESTful API design principle and provides detailed API documentation for the convenience of integration with other medical systems. To ensure the security of the interfaces, the API gateway 64 implements security mechanisms such as identity authentication, access control, and traffic control.

[0142] From the above description, it can be seen that the system of the present invention has made innovations in aspects such as real-time monitoring and early warning, big data analysis, human-computer interaction, and cloud computing platforms. These innovations not only improve the accuracy and efficiency of early stroke diagnosis but also provide doctors with more intuitive and comprehensive decision-making support tools. At the same time, the system of the present invention has good scalability and adaptability and can be flexibly configured and expanded according to the needs of different medical institutions.

[0143] The present invention also includes a data security module 7 for ensuring the data security of the entire system. In the medical field, the privacy protection of patient data is particularly important. Therefore, the data security module 7 of the present invention takes multiple measures to ensure data security.

[0144] First, the data security module 7 performs desensitization processing on the collected patient data. In a preferred embodiment of the present invention, a data desensitization method combining rules and machine learning is adopted. For structured data, such as patient names, ID numbers, etc., predefined rules are used for replacement or encryption. For example, the name can be replaced with a randomly generated code name, and the date of birth part in the ID number is retained, and the other parts are replaced with *. For unstructured data, such as text descriptions in medical records, named entity recognition (NER) technology is used to identify sensitive information and perform corresponding desensitization processing.

[0145] Secondly, the data security module 7 uses encryption algorithms to protect the security of data transmission and storage. During data transmission, the present invention adopts the TLS 1.3 protocol, which provides higher security and faster handshake speed. For the stored data, the AES-256 encryption algorithm is used for encryption. Key management adopts a distributed key management system to avoid the risk of single-point failure.

[0146] In addition, the data security module 7 also implements role-based access control (RBAC) to ensure the legality of data access. During the implementation of the present invention, different roles can be defined according to the organizational structure and work processes of the hospital, such as doctors, nurses, administrators, etc., and corresponding access permissions are assigned to each role. The system records all data access operations for subsequent auditing and tracking.

[0147] The present invention further includes a model update module 8 for ensuring the continuous optimization of the deep learning model. In the medical field, with the continuous accumulation of new data and the update of medical knowledge, machine learning models need to be updated regularly to maintain their accuracy and effectiveness.

[0148] The model update module 8 first regularly evaluates the performance of the deep learning model. In an embodiment of the present invention, the model is evaluated once a week using newly collected data. The evaluation metrics include accuracy, recall rate, F1 score, etc. If a certain metric is lower than a preset threshold (for example, the F1 score is lower than 0.9), the model update process is triggered.

[0149] The model update module 8 adopts the method of incremental learning to update the model based on the newly added training data. This method can significantly reduce the consumption of computing resources compared to completely retraining, while maintaining the stability of the model. Specifically, the present invention adopts an incremental learning algorithm based on Elastic Weight Consolidation (EWC). The objective function of this algorithm is as follows:

[0150]

[0151] where L new (θ) is the loss function on the new data, θ is the model parameter, is the parameter of the old model, F i is the diagonal element of the Fisher information matrix, and λ is a hyperparameter that weighs the importance of the new and old tasks. In this way, the model can learn new knowledge while trying to retain the original important features.

[0152] The model update module 8 also realizes the online update of the model without affecting the normal operation of the system. Specifically, the system first trains an updated model in the background using new data, and then gradually switches the traffic to the new model through A / B testing. This smooth transition method can reduce the impact on clinical diagnosis.

[0153] Finally, the present invention also provides a method for early diagnosis and warning of stroke based on machine learning. The method includes the following steps:

[0154] First, collect the brain imaging data, personal information data, medical records, and treatment outcome data of patients through the cloud computing platform 6. In the preferred embodiment of the present invention, these data sources may include the hospital's electronic medical record system, Picture Archiving and Communication System (PACS), etc. The system supports the import of data in multiple formats, such as DICOM-format medical images, HL7-format electronic medical records, etc.

[0155] Next, clean and standardize the collected data to improve data quality. During the data cleaning process, the system will detect and process missing values, outliers, and duplicate values. For example, for numerical data, the median can be used to fill in missing values; for categorical data, the mode can be used. Outlier detection can use statistical-based methods (such as the Z-score method) or density-based methods (such as the Local Outlier Factor algorithm). Standardization processing includes unifying data formats, converting units, unifying encodings, etc.

[0156] Then, construct a deep learning model, including a feature extraction network 21, a target classification network 22, and a severity classification network 23. The feature extraction network 21 uses ResNet50 as the backbone network and introduces a spatial attention mechanism to better capture the key regions of the image. The target classification network 22 and the severity classification network 23 use a Multi-Layer Perceptron (MLP) structure, combined with Dropout and batch normalization layers to prevent overfitting.

[0157] Use the processed data to train the deep learning model, adopting a combination of supervised learning, unsupervised learning, and semi-supervised learning. During the training process, a multi-task learning framework is used to optimize the three tasks of stroke detection, type classification, and severity assessment simultaneously. The loss function is designed as follows:

[0158] L total = αL detection + βL classification + γL severity ,

[0159] where α, β, and γ are the weight coefficients of each task, and the optimal values can be determined by methods such as grid search.

[0160] After training is completed, use the model to diagnose strokes for new medical imaging data, and calculate the accuracy, recall rate, and F1 metric of the diagnosis results. In practical applications, a confidence threshold (such as 0.8) can be set. Only when the output probability of the model exceeds this threshold will a clear diagnosis result be given; otherwise, it is recommended to conduct further examinations.

[0161] Based on the diagnostic results and real-time patient monitoring data, the system adopts an active warning mechanism to generate stroke warning information. The warning mechanism takes into account multiple factors, including the diagnostic results of the model, the patient's physiological indicators (such as blood pressure, blood sugar), lifestyle data, etc. The system calculates the overall risk score using the method of weighted summation and triggers different levels of warnings according to the risk levels.

[0162] Meanwhile, the system utilizes big data processing and analysis techniques to analyze the correlation between the incidence rate of stroke, prognosis effects and various factors, and constructs a multi-dimensional stroke disease feature vector. This analysis process adopts a variety of statistical and machine learning methods, such as multiple regression analysis, decision tree, random forest, etc., to comprehensively evaluate the impact of various factors on stroke risk.

[0163] The method of the present invention also displays the diagnostic results, warning information and data analysis reports in real time through a human-computer interaction interface. The interface design follows the basic principles of human-computer interaction, such as consistency, visualization, feedback and high efficiency. For example, the diagnostic results will be intuitively displayed in the form of a 3D brain model, and different colors and icons will be used for warning information to represent different risk levels.

[0164] To ensure the timeliness of the model, this method also includes regularly evaluating the model performance and performing incremental learning based on the newly added data to achieve online update of the model. During the evaluation process, in addition to common indicators such as accuracy and recall, the performance of the model on different sub-populations (such as different age groups, different types of stroke) will also be considered to ensure the fairness and generalization ability of the model.

[0165] Finally, this method adopts measures such as data desensitization, transmission encryption and role-based access control to ensure the data security of the entire system. Especially when dealing with sensitive medical data, the system will strictly comply with relevant laws, regulations and ethical guidelines, such as the requirements of HIPAA (Health Insurance Portability and Accountability Act).

[0166] Through the above steps, the method of the present invention realizes the early diagnosis and warning of stroke, provides a powerful decision-making support tool for clinicians, and helps to improve the early detection rate and treatment effect of stroke.

[0167] It should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. The early diagnosis and early warning system of stroke based on machine learning is characterized by: include: Data management module for: Collect patient data from the cloud computing platform and store it in the system; Clean the collected data and remove invalid data; Process data according to data standardization templates to improve data quality; A deep learning model module is communicatively connected to the data management module and is used to: receiving processed patient data sent by the data management module; Based on the patient data, a machine learning model is constructed, and supervised learning, unsupervised learning, and semi-supervised learning are performed on the data to achieve model training; The big data processing and analysis module is connected to the deep learning model module for: Using big data processing and analysis technology, the analysis results output by the deep learning model module are analyzed in real time and dynamically monitored; The real-time monitoring and early warning module is connected to the big data processing and analysis module for: Based on the analysis results of the big data processing and analysis module, a stroke diagnosis and early warning decision model is generated; Provide early stroke diagnosis results and early warning predictions for stroke diagnosis reports at different stages; The human-computer interaction module is connected to the real-time monitoring and early warning module for: Realize the real-time display and interaction of the data management module, deep learning model module, big data processing and analysis module and real-time monitoring and early warning module; Realize remote access and real-time synchronization updates based on cloud computing platform; The cloud computing platform is connected to the human-computer interaction module for: Provide data collection, feature extraction, data analysis and decision support services; Provide remote access and update services for the deep learning model module.

2. The system according to claim 1, characterized in that The data management module comprises: The data acquisition submodule is used to: Establishing a stroke data center through the cloud computing platform; Receive brain imaging data, personal information data, medical records and treatment results data of stroke patients; Locally store the received data; Data processing submodule, used for: Performing preliminary processing on the data stored in the data acquisition submodule, including cleaning and removing data that is blank, inconsistent, or lacks annotations; The data were standardized using standardized methods; A machine learning model is established through the deep learning model module to perform feature extraction, feature selection and feature screening on the processed data.

3. The system according to claim 1, characterized in that The deep learning model module includes: Feature extraction network for: Image features are extracted using a network structure consisting of residual units, SE channels, convolutional kernels, and pooling layers; The convolution kernel size is 7×7, the number is 3, and the step length is 2; The target classification network is in communication with the feature extraction network and is used to: The stroke type was classified using a network structure consisting of convolution kernels, pooling layers, fully connected layers, and Softmax output layers; The convolution kernel size is 3×3, the number is 3, and the step length is 2; A severity classification network, in communication with the target classification network, is configured to: The severity of stroke is determined using a network structure that includes convolution kernels, pooling layers, fully connected layers, and Softmax output layers; The convolution kernel size is 3×3, the number is 3, and the step length is 2.

4. The system according to claim 1, characterized in that The real-time monitoring and early warning module includes: The diagnostic assessment submodule is used to: Based on the model trained by the deep learning model module, diagnose stroke on the input medical imaging data; Calculate the accuracy, recall and F1 index of the diagnosis results; The early warning decision submodule is in communication with the diagnosis and evaluation submodule and is used to: Based on the output results of the diagnosis and evaluation submodule and in combination with the patient's real-time monitoring data, stroke warning information is generated; Adopt active early warning mechanism to trigger early warning according to preset thresholds.

5. The system according to claim 1, characterized in that The big data processing and analysis module includes: Data analysis submodule, used to: To analyze the correlation between stroke incidence, prognosis and age, gender, occupation, lifestyle, region and other factors; Construct a stroke disease feature vector containing multi-dimensional information; The result output submodule is connected to the data analysis submodule for: Generate analysis reports of stroke disease feature vectors, stroke incidence, and prognosis effects; The analysis results are sent to the human-computer interaction module for display.

6. The system according to claim 1, characterized in that The human-computer interaction module comprises: User interaction submodule, used to: Provide a graphical interface to display the diagnostic results and warning information generated by the real-time monitoring and warning module; Receive query requests and operation instructions input by users; The system management submodule is communicatively connected with the user interaction submodule and is used to: Manage user permissions and system configuration; Coordinate the operation and data exchange of each functional module; Realize remote access and real-time synchronous update of the system.

7. The system according to claim 1, characterized in that The cloud computing platform adopts a distributed computing architecture, including: Data storage cluster, used to store large-scale stroke-related data; A computing resource pool, used to provide high-performance computing resources for the deep learning model module; Task scheduler, which coordinates the allocation and execution of data processing and model training tasks; API Gateway is used to provide standardized interfaces to realize data exchange and service calls with external systems.

8. The system according to claim 1, characterized in that Also includes data security modules for: Desensitize the collected patient data; Use encryption algorithms to protect the security of data transmission and storage; Implement role-based access control to ensure the legitimacy of data access.

9. The system according to claim 1, characterized in that Also included is a model updating module for: Regularly evaluate the performance of deep learning models; Automatically trigger incremental learning of the model based on newly added training data; The model can be updated online without affecting the normal operation of the system.

10. A method for early diagnosis and early warning of stroke based on machine learning, characterized in that: The following steps are involved: (1) Collect patients’ brain imaging data, personal information data, medical records, and treatment results data through the cloud computing platform; (2) Clean and standardize the collected data to improve data quality; (3) Construct a deep learning model, including a feature extraction network, a target classification network, and a severity classification network, where: The feature extraction network contains residual units, SE channels, 7×7 convolution kernels and pooling layers; Both the target classification network and the severity classification network contain 3×3 convolution kernels, pooling layers, fully connected layers, and Softmax output layers; (4) Use the processed data to train a deep learning model, using a combination of supervised learning, unsupervised learning, and semi-supervised learning; (5) Use the trained model to diagnose stroke on new medical imaging data and calculate the accuracy, recall, and F1 index of the diagnosis results; (6) Based on the diagnosis results and real-time monitoring data of patients, an active early warning mechanism is used to generate stroke warning information; (7) Using big data processing and analysis technology, analyze the correlation between stroke incidence, prognosis and multiple factors, and construct a multi-dimensional stroke disease feature vector; (8) Real-time display of diagnostic results, warning information and data analysis reports through the human-computer interaction interface; (9) Regularly evaluate model performance and perform incremental learning based on new data to achieve online model updates; (10) Measures such as data desensitization, transmission encryption, and role-based access control are adopted to ensure data security of the entire system.

Images