Research On ECG Signal Classification Method Based On Multi-task Learning

Cardiovascular disease is an increasingly serious threat to human life and health,and the number of people suffering from it is on the increase.Early detection and timely treatment of cardiovascular disease has been a major focus of research in the medical field.Electrocardiogram(ECG),as the main tool of detecting and diagnosing cardiovascular disease,requires a high level of competence and expertise in its analysis and interpretation,and the multitude of ECG data adds to the burden on doctors in diagnosing cardiovascular disease.Computer-aided diagnostic technology can analyze and process large amounts of ECG signal data,helping to reduce the burden on doctors while improving the accuracy of diagnosis of cardiovascular disease.Most existing single-task learning models used by ECG signal classification algorithms ignore information from related tasks that can help optimize task learning,whereas multi-task learning(MTL)models can learn multiple related tasks simultaneously,share features from related tasks,and thus help improve task performance.The thesis combined two technical approaches,deep learning and multi-task learning,to investigate ECG signal classification algorithms,and the specific research work carried out is as follows:(1)The thesis proposes a multi-task learning-based algorithm for ECG signal classification.The algorithm uses Residual Network(ResNet)to build a multi-task deep learning Res-MTL network model,and decomposes a single ECG multi-classification task into multiple binary classification tasks,by sharing the features of each task in the Res-MTL network model to learn and mine the information beneficial to tasks.The optimal choices of some parameters of the Res-MTL network model are determined by designing experiments for testing.Experimental results from the ten-fold cross-validation on the Physikalisch Technische Bundesanstalt Extra Large(PTB-XL)ECG database show that the proposed algorithm has good performance for ECG signal classification.(2)Aiming at the task learning bias problem of the Res-MTL network model,the thesis goes on to propose an improved model based on loss optimization,namely the Gradient Magnitude Direction Adjustment(GMDA)network model.The GMDA network model combines both the magnitude and direction of the task loss gradient to optimize the loss function of the Res-MTL network model.This network model sets the loss weights for each task based on its learning effect,so as to adjust the magnitude of the loss gradient of the task itself,and adjusts the direction of the loss gradient of conflicting tasks by the method of gradient projection,avoiding manual setting of task loss weights and the negative migration caused by task loss gradients cancelling each other,thus achieving better learning of ECG signal classification tasks.The optimal parameters of the GMDA network model are determined through experimental tests on the PTB-XL ECG database,and the experimental results demonstrate the effectiveness of the proposed multi-task learning algorithm for the ECG signal classification task,with classification accuracy of 0.965.The proposed multi-task learning based ECG signal classification algorithm in the thesis has good experimental results,but still needs more in-depth research to improve it.Further research work can attempt to combine more network modules and techniques,and to experiment with the generalization capabilities of the network model proposed here on other ECG databases.