Cardiovascular diseases remain a leading cause of mortality worldwide, with arrhythmias posing significant diagnostic challenges. This survey presents a comprehensive review of Deep learning approaches applied to electrocardiogram-based arrhythmia detection, covering research from 2022 to 2024. It categorizes the literature by model architecture-including CNNs, RNNs, hybrid designs, and GCNs. It analyses performance across key benchmarks such as accuracy, precision, and real-time feasibility. Comparative insights into datasets, feature extraction strategies, and model generalization are also provided. The review highlights challenges such as interpretability, data imbalance, and clinical deployment limitations, while proposing future directions including Transformer-based architectures, XAI techniques, and IoT integration for real-time monitoring. This structured survey would support researchers and clinicians in selecting and developing effective AI tools for cardiac health diagnostics.