Compression of electrocardiogram signals based on deep neural networks

Electrocardiogram (ECG) signals are signals that indicate the electrical activity of the heart, one of the important organs of the human body. These signals, which bandwidth in the range of 0.05-150 Hz, are produced as a result of the relaxation and contraction of the heart muscles. ECG signals are frequently used in the determination of appropriate diagnosis-treatment, following the treatment applied, evaluation of the treatment applied and detection of potential problems. For this reason, in modern medical applications, ECG signals should be recorded, processed and transmitted in real time. The area covered of ECG data of recorded for a long time can be very large. This situation brings along speed and memory problems during the processing, storage and transmission of ECG signals. These problems can be solved by compressing ECG signals efficiently. In this thesis, compression studies of electrocardiogram signals based on convolutional autoencoders and discrete wavelet transform are discussed. In this thesis, compression studies of electrocardiogram signals based on convolutional autoencoders (CAE) and discrete wavelet transform (DWT) has been discussed. The success of the developed models has been tested on the MIT-BIH Arrhythmia dataset containing 48 ECG recordings using the compression ratio (CR) and mean-independent percentage mean square difference (MPRD) criteria. In the first study conducted (Proposed Method-1), compression and reconstruction of one-dimensional (1D) ECG signals given to the input of the developed system has been performed. The ECG signals has been passed through the encoder part of the 1D-CAE, the DWT-based encoder/decoder, the decoder parts of the convolutional autoencoder, respectively. In the encoder parts of the system, the size of the signals has been reduced, and in the decoder parts of the system, the size of the signals has been restored. In Proposed Method-1, experimental results have been obtained with an average compression ratio of 32.27: 1 and an average MPRD of 18.91%. In the second study (Recommended Method-2), the system that performs compression and reconstruction operations with 4 different models working in parallel with each other, with one-dimensional ECG signals given to its input, was developed. Thus, 4 different compression ratios and 4 independent percent average square difference values are obtained for the signal given to the input of the system. When the experimental results in the proposed method-2 are examined, the average CR values are; 16.22: 1, 22.88: 1, 32.22: 1, 49.82: 1 and average MPRD values respectively; It was obtained as 12.28%, 14.74%, 18.40%, 28.21%.