Research On Small Sample Electrocardiogram Time Series Data Classification Based On Transfer Learning

In recent years,the deep learning model has achieved remarkable results in many fields.The key to deep learning is the large-scale,annotated dataset.However,in reality,large-scale dataset is difficult to be annotated completely.Thus,how to use small sample data to fully training models is significant.Actually,medical data is usually hard to be collected,and it contains few negative samples and few lables of samples.Therefore,large-scale labeled dataset in the medical field is scarce.The ECG uses the electrocardiograph to record the changes in point activity generated by each cardiac cycle from patients' body surface.It is an important aid and reference for doctors in diagnosing heart disease.Nevertheless,there is high demand for diagnosising ECG,which means that achieving real-time diagnosis is difficult.At present,there are few studies in small sample electrocardiogram signals classification,and the results still have a large room for improvement in accuracy and efficiency.Based on the deep learning,this paper proposes a cost-sensitive classification model based on stacked denoising autoencoders and bidirectional long and short term memory neural network.This model effectively improves accuracy and efficiency and solves the problem of imbalance dataset.Simultaneously,the idea of migration learning is utilized.The classification model is improved with adaption,and the migration of small samples from the source domain to the target domain is accomplished.This paper first proposes a time series data classification model which can solve the problem of imbalanced ECG dataset.In this model,the stacked denoising autoencoders act as an encoder,automatically learning the semantic features in the time series data rather than manually extract features in ECG like other complex methods.Subsequently,the bidirectional long and short emory neural network classifier realizes classification by the features extracted by the stacked denoising autoencoders.The stacked denosing autoencoders not only compress dimensions of data,but also achieves noise reduction,while the long and short term neural memory network classifier makes full use of the temporal information in the time series data.At the same time,the model solves the problem of imbalanced dataset by using a cost-sensitive loss function.The classification model is based on ECG signals and there are some experiments on the MIT-BIH arrhythmia database,SVDB and NSTDB.The final experimental results prove that the basic time series classification model not only has higher accuracy,but also improves efficiency.Secondly,the classification model combines the adaptive batch standardization mechanism to realize the classification of small sample ECG signals.Since the data distribution in source domain is different from data distribution in target domain,direct migration of model will results in bad classification and even negative migration.Compared with other migration learning methods,the principle of adaptive batch standardization is very simple.It does not require complicated calculations and does not need to add new parameters to achieve domain adaptation.The improved model was tested on selected UCR ECG datasets.The experimental results show that the improved model based on adaptive batch normalization mechanism is a solution to small sample ECG time series classification.