Early detection of cardiac arrhythmias from electrocardiogram (ECG) signals is crucial for effective automated cardiac diagnostic systems. A deep neural network-based system using convolutional neural networks (CNNs) was developed for arrhythmia detection. Experiments were conducted using the PTB-XL dataset, with initial comparative analyses performed on both raw time-domain ECG signals and their frequency-domain representations. The CNN model trained on time-domain signals achieved a classification accuracy of 77.38%, significantly outperforming its frequency-domain counterpart which attained 68.43% accuracy. This superior initial performance is attributed to the CNN's inherent ability to directly learn fine-grained temporal patterns and morphological features crucial for arrhythmia detection from the raw time-domain signals, which motivated the subsequent focus on enhancing this pipeline. To further enhance the performance of this CNN-based arrhythmia detection system, supervised clustering with discriminative training was integrated, leading to a notable improvement with a final accuracy of 79.17% and a macro F1-score of 0.72. These results demonstrate that supervised clustering significantly improves classification performance in CNN-based arrhythmia detection systems.