Wavelet-Based Denoising and Classification of ECG Signals Using Hybrid LSTM-CNN Models

Abstract

Electrocardiogram (ECG) signals are essential as means to diagnose and monitor cardiovascular diseases, in particular arrhythmias. Nevertheless different noise sources like baseline wander, powerline interference and muscle artifacts make an accurate interpretation of ECG signals difficult as these noise components tend to cover up morphological features needed for diagnosis. To overcome these limitations, this work proposes an integrated signal processing and deep learning framework combining denoising using Discrete Wavelet Transform (DWT) and robust classification using a hybrid Long Short-Term Memory-Convolutional Neural Network (LSTM-CNN) architecture. The ECG signals are effectively decomposed into multi resolution components within the DWT and the selective suppression of noise can be achieved without loss of clinically relevant features like QRS complexes, P waves and T waves. After denoising, a hybrid LSTM CNN model is used, in which convolutional layers get spatial patterns and the LSTM layers capture temporal dependencies embedded in sequential ECG data. We train and validate the model on the MIT-BIH Arrhythmia Database, a well known benchmark for ECG analysis. Experimental results show that proposed framework substantially outperforms existing methods (entropy based selection of prevalent methods), achieving its accuracy 98.6%, sensitivity 98.2%, and specificity 99.1%. As such, these results verify the effectiveness of using wavelet based signal enhancement coupled with deep hybrid modeling to perform ECG interpretation more accurately and in real time for use in clinical and wearable health care applications.