Patch-Embedded Transformer Model For ECG Noise Removal

Abstract

Cardiovascular disease ranks as one of the most dangerous medical conditions that exists throughout the globe, which doctors use electrocardiogram (ECG) signals to monitor.The analysis and diagnosis process becomes challenging because different types of noise disrupt ECG signal at various strength levels. The medical field requires proper denoising methods because they help establish better signal-to-noise ratios, which enable accurate cardiovascular patient monitoring. The research presents a deep learning system for ECG signal denoising, which combines one-dimensional convolutional neural networks with Transformer architecture. The 1D convolutional layers extract features from the ECG signal using multiple kernel sizes, creating a multi-scale patch embedding representation. The system processes this embedding through a Transformer-based network, which strengthens contextual learning while enhancing overall denoising results. The proposed model effectively removes noise while preserving important morphological features of the ECG signal.In additional convolutional block is incorporated within the architecture to enhance local feature preservation and improve reconstruction quality. By combining multi-scale local feature extraction global contextual modeling through self-attention, and enhanced reconstruction via the encoder–decoder structure the proposed method achieves effective noise suppression while preserving important morphological features of the ECG waveform