Cleaning ECG Noisy Data Based On Empirical Mode Decomposition And Rule Mining

Automatic detection of electrocardiogram(ECG)abnormality is a typical multi-label classification problem.On the one hand,it is difficult to eliminate the noises of ECG signal,which will have negative effects on the subsequent waveform recognition and feature extraction.On the other hand,in order to ensure the generalization ability of classifier,sufficient training samples with high quality labels are necessary.But in fact,the labels of ECG data are often partial and incorrect.Therefore,how to effectively denoise ECG signal and clean noisy labels to get a clean ECG data set is a challenging problem.The research content of this paper mainly includes the following two aspects.First,the elimination of baseline wander is the most important among different kinds of noises in ECG signal.In normal ECG,the baseline is almost in accordance with the x-axis.However,when the baseline wander is too large,it leads to serious deviation of ECG signal from the x-axis,affecting the recognition of various waveforms.Existing methods of eliminating baseline wander are usually not adaptive,which need to set parameters manually and easily lead to signal distortion.Therefore,empirical mode decomposition(EMD)combined with Hilbert transform(HT)is proposed to remove baseline wander.First,EMD is used to decompose ECG signal into several intrinsic mode functions(IMF).Then the instantaneous frequency of IMF is obtained based on HT.Further,the IMF components which represent baseline wander are determined according to adaptive average instantaneous frequency(AAIF).Not only does this method extract baseline wander more accurately,but also it is more adaptive.After the ECG data is preprocessed,abnormality patterns are identified according to feature patterns.An ECG abnormality often presents different combination patterns of multiple features.However,the existing methods of cleaning noisy labels often ignore the complex combinations between abnormalities and features.Therefore,cleaning ECG noisy labels based on rule mining(RM)is proposed.With the help of standard data set with complete and accurate labels,feature patterns and shared labels in the data set with noisy label are identified.Further,cleaning rules are mined to expand the initial relevant labels.Finally,binary classifiers are built to clean the remaining noisy labels in order to obtain all correct abnormality labels.Experimental results show the effectiveness and superiority of this method.The main contributions of this paper are as follows:(1)Empirical mode decomposition combined with Hilbert transform is proposed to remove baseline wander,which not only has strong adaptability,but also can accurately determine the IMF components where baseline wander is located according to AAIF;(2)Cleaning ECG noisy labels based on rule mining is proposed to fully exploit the combination relationship between abnormality labels and feature patterns,so as to effectively identify noisy labels.The clean data set after cleaning can be used in various applications.(3)Experiments on simulated data sets and real data sets have proved the effectiveness and superiority of the method.