ECG Arrhythmia Classification Using 1D Convolutional Neural Network with Data Augmentation: A Comparative Study

Electrocardiogram signals (ECG) are an essential component of automated detection of heart rhythm disorders. Despite the widespread use of 1D-CNN NSO models, most current studies evaluate performance using balanced datasets or clinically rare cardiac disorders. A clear methodological gap in understanding how deep learning models are acting under realism between groups. This study addresses this gap by providing a precise systemic comparison between a 1D-CNN core-trained model without any imbalance strategy, and a developed model that combines an increase in signal-level data (Added Gaussian noise, and displacement). Circular and amplitude change) with the use of the category weight loss function, using the highly unbalanced MIT-BIH ARRHYTHMIA database, which contains 40,937 heartbeats distributed over five classes according to Aami standard. A stratification was applied at 80/20 pulse-level to keep the distribution of categories between the training and testing groups. The 1D-CNN architecture was selected due to its computational efficiency and proven suitability for temporal physiological signals processing. Quantitative results showed a significant improvement in performance, with a total accuracy increased from 45.39% to 83.71%, and the Macro F1-Score coefficient improved from 0.3413 to 0.5379. However, it is important to note that the performance of rare clinical-important groups remains a major challenge; Class A (early atrial contractility) achieved a value of F1-Score close to 0.20. A very rare training example. Grad-Cam's analysis of idiomy's ability to learn meaningful and distinctive physiological representations of each category across all five categories of CPR has also been confirmed.