<![CDATA[Automated arrhythmia classification from electrocardiogram (ECG) signals remains challenging because public datasets are highly imbalanced and fine-grained multiclass performance may degrade when labels are mapped to the clinically standardized AAMI EC57 grouping scheme. This study proposes real-record dataset enrichment combined with a compact one-dimensional convolutional neural network (1D-CNN) for both fine-grained and AAMI-grouped beat classification. Fourteen records from the MIT-BIH Supraventricular Arrhythmia Database were inserted into the MIT-BIH Arrhythmia Database, adding 4,649 S beats, 4,530 V beats, and 47 Q beats without synthetic oversampling. Preprocessing included Christov R-peak segmentation, beat extraction, per-beat min-max normalization, and resampling to 180 Hz. The 1D-CNN was evaluated under 16-class, 17-class, and 5-class AAMI EC57 schemes. Using ASGD, the model achieved accuracies of 99.10%, 98.58%, and 99.38%, with macro F1-scores of 0.90, 0.87, and 0.97, respectively. Cross-database testing on INCARTDB reached 99.13% accuracy across four mappable classes (N, V, R, A), indicating limited 4-class transferability rather than full AAMI generalization. The approach preserves authentic ECG morphology while addressing minority-class scarcity. The findings show that real-beat enrichment can improve balanced ECG classification, although results are based on beat-level random splits and require future record-wise validation before clinical deployment.]]>
