Accurate detection of arrhythmias is critical for the early diagnosis and prevention of cardiovascular diseases. However, the complex and highly variable nature of electrocardiogram (ECG) signals presents significant challenges in feature extraction and computational efficiency. This study explores a deep learning-based framework for arrhythmia classification, evaluating several neural network architectures, including Residual Neural Network (ResNet), Transformer-based Convolutional Neural Networks (CNN), Attention-based CNN, Gated Recurrent Units (GRU), and Long Short-Term Memory (LSTM) networks. The models were trained and evaluated using the MIT-BIH Arrhythmia Dataset, which includes ECG recordings (time-series data) categorized into five arrhythmia types. To address class imbalance and enhance model robustness, a hybrid loss function was employed, integrating categorical cross-entropy, mean absolute error (MAE), and mean squared error (MSE). This combination aims to improve class-wise learning while minimizing misclassification errors. Experimental results indicate that the custom convolutional ResNet architecture achieved the most consistent performance, with a classification accuracy of 99.64% (95% confidence interval 99.03–99.87%), sensitivity of 99.61% (95% CI 98.98–99.85%), specificity of 99.91% (95% CI 99.72–99.97%), and a loss of 0.07, with comprehensive measurement error analysis confirming the stability and reliability of these results. These results were benchmarked against existing methods reported in the literature and demonstrate notable improvements in classification metrics. The proposed approach offers a promising direction for reliable and efficient arrhythmia detection from ECG signals, potentially enhancing decision support in clinical settings.