The electrocardiogram (ECG) is one of the most important physiological markers utilised in arrhythmia diagnosis. The ECG arrhythmia classification devices, which include multi-module sensors, clustering algorithms and neural networks, have established themselves as vital monitoring and detection systems for cardiovascular disorders during the last few years. The current ECG arrhythmia classification algorithms encounter two main difficulties because they have complicated models and they need extended periods to complete their operations. The identification of arrhythmias in ECG data needs precise evaluation to perform successful cardiac monitoring and receive correct medical care. The main problem with ECG dataset analysis through deep learning methods stems from the unbalanced distribution of arrhythmia types, which makes some arrhythmia types less common than others. This research presents a new method to enhance arrhythmia detection from ECG data by integrating deep residual networks (ResNets) with the synthetic minority over-sampling technique - edited nearest neighbours (SMOTE-ENN). The ResNet architecture serves to extract deep hierarchical features from ECG data, which enables the model to learn complex arrhythmia patterns with high accuracy. The system uses SMOTE-ENN to create artificial minority class examples while it eliminates unimportant and borderline data points, which produces a dataset that is both clean and balanced. The proposed framework demonstrates superior performance to conventional methods according to experimental results, which were conducted on ECG benchmark datasets. The combination of ResNet with SMOTE-ENN produces better model generalisation, which results in improved accuracy, F1-score, and sensitivity performance, especially for the minority arrhythmia classes. The research shows that deep learning systems, which use improved data balancing techniques, can successfully detect arrhythmias in ECG monitoring systems which function in actual clinical settings.