Analysis of ECG signals using deep recurrent neural networks
Tipus de documentProjecte Final de Màster Oficial
Data2019-10-09
Condicions d'accésAccés obert
Llevat que s'hi indiqui el contrari, els
continguts d'aquesta obra estan subjectes a la llicència de Creative Commons
:
Reconeixement-NoComercial-CompartirIgual 3.0 Espanya
Abstract
The study of the ECG signalsis essential to detect several diseases. The present projectaims to analyze the ECG signals in order to detect different types of heart beat sassociated with arrhythmia, using data from the MIT-BIH Arrhythmia database. The proposed pipeline is the following: (a) design and program a data visualizer for the MIT-BIH Arrhythmia database; (b) reduce the dimensionality of the data using a Principal Component Analysis; and (c) compress thebeats using a deep autoencoder.The first step of the analysis consists of an approach to the MIT-BIH Arrhythmia databasefor better understanding on how the data is presented. For this purpose, an ECG data visualizer is designed. This application allows the user to select several graph settings and records of the database and plots the selected signals.Diverse graphic and GUI libraries are considered, such as WFDB, TkInter and Bokeh. But the package that best fits the requirements in this case is Dash, which includes both the graph and the user interface and updates the changed settings in real time.The second step of the analysis consists of reducing the dimensionality of the data by means of a Principal Component Analysis. PCA has been performed for each one of the following types of heartbeats: normal beat (N); left bundle branch block beat (L); right bundle branch block beat (R); premature ventricular contraction (V); and atrial premature beat (A). The importance of every principal component for the shape of each beat is discussed in this section.The final step is to compress the data using a deep auto encoder. The encoder pretends to represent an input heart beat in a reduced number of neurons. The decoder reconstructs the initial beat using the same configuration of layers but arranged in reverse order. The firstproposal is a convolutional autoencoder with 10 neurons. The obtained loss is0.228 (22.8%). The correlation coefficient between the input vector and the output of the autoencoder is 0.99; 0.89; 0.96; 0.93; and 0.92 for beats N; L; R; V; and A, respectively. For further compression, the same autoencoder but using only 5 neurons has been trained as well. The loss results are also around 0.228, but the correlation coefficients are, respectively, 0.95; 0.86; 0.83; 0.89; and 0.71.The performance is worse in this case. Hence,the autoencoder with 10 neurons is better for the data compression.A recurrent LSTM autoencoder is proposed for future work. It would allow both to use input data of variable length and to predict next steps in ECG sequences, which might improve the performance of the autoencoder and thus, the ECG beats classification for arrhythmia detection
MatèriesSignal processing, Neural networks (Computer science), Arrhythmia, Tractament del senyal, Xarxes neuronals (Informàtica), Arítmia -- Diagnòstic
TitulacióMÀSTER UNIVERSITARI EN ENGINYERIA INDUSTRIAL (Pla 2014)
Col·leccions
Fitxers | Descripció | Mida | Format | Visualitza |
---|---|---|---|---|
analysis-of-ecg ... s-miguel-rueda-sotorra.pdf | 2,530Mb | Visualitza/Obre |