RESEMBLING THE MORPHOLOGIES OF ECG SIGNALS USING REGULARIZED DENOISING AUTOENCODER

ECG recording often requires an effective method of denoising to provide a clean signal for an accurate and valid diagnosis. Denoising autoencoder (DAE) has shown optimistic results in denoising ECG signals, especially a simple DAE consisting of input, hidden, and output layers. However, to obtain a good and efficient denoising, the optimal number of hidden neurons needs to be estimated. If the number of neurons in the hidden layer is less than those of the input or output layer, a dimension reduction occurs, which is known as the ‘bottleneck effect’. This forces the DAE network to learn the relevant feature map of the input during training. Here, we propose a framework to denoise the ECG segments using a regularized denoising autoencoder (RDAE), with one hidden layer only, where the bottleneck effect is introduced by applying a sparsity penalty or regularizations to the weights to learn sparse feature maps instead of the redundant information in the input signals. In this work, 𝑳 𝟏 and 𝑳 𝟐 Kernel regularizations are evaluated in terms of denoising ECG signals. The optimal regularization parameter is evaluated using a statistical method known as the Gini index, to find the optimal trained decoding weights which resemble the morphologies of the ECG signal efficiently. In conclusion, the 𝑳 𝟏 - and 𝑳 𝟐 -RDAE models with a suitable regularization parameter can effectively capture features that resemble the morphologies of ECG signals from its noisy version with an average signal-to-noise ratio improvement of 13.60 dB and 10 dB, respectively.