Enhanced electrocardiogram classification using Gramian angular field transformation with multi-lead analysis and segmentation techniques

Abstract

Conventional manual or feature-based ECG analysis methods are limited by time inefficiencies and human error. This study explores the potential of transforming 1D signals into 2D Gramian Angular Field (GAF) images for improved classification of four ECG categories: Atrial Fibrillation (AFib), Left Ventricular Hypertrophy (LVH), Right Ventricular Hypertrophy (RVH), and Normal ECG.

• •
  The study employed GAF transformations to convert 1D ECG signals into 2D representations at three resolutions: 5000 × 5000, 512 × 512, and 256 × 256 pixels.
• •
  Segmentation methods were applied to enhance feature localization.
• •
  The ConvNext deep learning model, optimized for image classification, was used to evaluate the transformed ECG images, with performance assessed through accuracy, precision, recall, and F1-score metrics.

The 512 × 512 resolution achieved the optimal balance between computational efficiency and accuracy. F1-score for AFib, LVH, RVH and Normal ECG were 0.781, 0.71, 0.521 and 0.792 respectively. Segmentation methods improved classification performance, especially in detecting conditions like LVH and RVH. The 5000 × 5000 resolution offered the highest accuracy but was computationally intensive, whereas the 256 × 256 resolution showed reduced accuracy due to loss details.