Research And System Implementation Of Cardiovascular Disease Intelligent Classification For Multi-source Data

Cardiovascular disease is one of the major diseases that threaten human life and health.The use of information technology such as artificial intelligence to enhance the treatment level of cardiovascular disease in grassroots institutions is the main direction of our country’s medical development and an important means related to the medical health of the people.Compared with doctors’ diagnosis,intelligent diagnosis of cardiovascular diseases has natural advantages in work intensity and work efficiency.However,most of the existing studies are based on public data sets,and their scale of ECG data is small and the disease types are limited,it is difficult to train a cardiovascular disease classification model that meets the needs of the real-life medical environment.In order to expand the scale of medical data and disease types,the current engineering practice usually integrates the ECG data of multiple class A tertiary hospitals,and trains a global classification model through federal learning for grass-roots hospitals.However,after evaluation,it is found that the global model has the problem of low classification accuracy in medical institutions other than the class A tertiary hospitals.The reason is that grass-roots hospitals are not the same in terms of medical characteristics,equipment configuration,doctors’ level,etc.,which makes it difficult for the global model to adapt to personalized medical needs.The data characteristics and case distribution of different levels of medical institutions are different,which affects the construction of the classification model.In the feature extraction stage,the difference in the feature distribution of different devices makes it impossible to extract the ECG features uniformly;in the disease classification stage,the difference in the distribution of cases in different hospitals makes the classification model difficult to generalize;in the model deployment stage,the portable device model is difficult to deployed due to the insufficient model efficiency.The existing methods cannot meet the personalized needs in the construction stage of cardiovascular disease classification model.Therefore,adapting to the three heterogeneous issues of feature matching,model generalization,and performance assurance in various medical institutions is an important issue addressed in this paper.This dissertation is oriented to the personalized needs of assisted diagnosis of cardiovascular diseases,and focuses on the three stages of classification model construction and optimization technology to solve the three heterogeneous problems of various medical institutions in feature matching,model generalization,and performance assurance.It mainly includes the following research contents and innovation results.(1)Aiming at the problem of extracting ECG feature mismatch,a transfer learning method based on feature alignment is proposed.The feature distance relationship between multiple samples is formed into a graph representation structure,and the constraint relationship of the graph structure is constructed based on this.In addition,in the local feature alignment module,the idea of metric learning is adopted,and local data is used to construct the constraint relationship of the distance between the intra class and the inter class distance,and the mapping relationship of the feature alignment is constructed to achieve the effect that different collection devices can be seamlessly connected to match the ECG features extracted by different devices.The accuracy of local cardiovascular disease recognition is improved from 71.33% to 88.15%.(2)Aiming at the problem that the disease classification model is difficult to generalize,a multi-task ECG data training mechanism based on dynamic loss is proposed.According to the proportion of each symptom of the hospital node,the loss weight of different tasks is adjusted through an adaptive method.The model will be more sensitive to samples of unbalanced tasks.To achieve the effect of accurately predicting different symptoms,so as to improve the learning ability of low-proportion symptoms,the recognition accuracy value of cardiovascular disease in characteristic medical institutions has increased from 72.23% to 80.35%.(3)Aiming at the problem of low performance of portable ECG devices,an adaptive compression mechanism based on knowledge distillation is proposed.The feature expansion module is used to match the input dimension changes of the acquisition device,so that the model can be deployed in different device types,achieving the effect that portable devices can also be effectively deployed,so as to meet the performance requirements of local medical nodes.The accuracy of cardiovascular disease recognition for portable devices has been increased from 72.32% to 82.33%.Based on theoretical results,this research designs and implements a multi-source data-oriented cardiovascular disease rational classification system,which is deployed on the Jiutian artificial intelligence cloud platform and medical platforms at all levels,provides effective cardiovascular disease classification solutions for grassroots institutions,provides high-performance,personalized secondary diagnosis capabilities for grassroots clinicians,and gives full play to the application and social value of artificial intelligence technology in the field of auxiliary medical care.