VivoOptiNet: A Gradient Update Adam Optimizer for ECG Arrhythmia Classification

The aim of this work is to develop a hybrid AI model, termed VivoOptiNet, to classify five arrhythmia types and provide their interpretations from ECG signal data. The RNN-based model uses an attention mechanism, a transformer encoder and the Adam optimizer to extract local and global ECG features and classify them. The Adam optimizer proposed in this work addresses the difficulties in non-stationary ECG signals by correcting the gradient norm behavior over time and adding Gaussian noise to achieve natural exploration and adaptation to local minima. The arrhythmia (MIT–BIH) dataset is used for preprocessing, and the SMOTE–ENN technique is employed to balance the classes and assess the model’s performance. The results indicate that VivoOptiNet achieved 99.51% accuracy in classifying five arrhythmia types from ECG signals. The addition of channel attention mechanism and a transformer encoder has inherent interpretability, highlighting the ECG segments that are most relevant, thereby leading to greater omniscience of the ECG sequence and supporting explainable AI in clinical decisions. The success of the VivoOptiNet model in extracting arrhythmias from ECG data represents a significant development in the diagnosis of cardiac patients, potentially reducing the complexity of, for example, multi-lead ECG systems involved and simplifying their use in clinical settings for arrhythmia detection. The model demonstrates its ability to handle imbalanced data using the SMOTE–ENN technique, providing evidence that such models can deliver satisfactory results and that these techniques may be helpful for real-world problems where certain types of arrhythmia are underrepresented in the training data.