Key result
Dual-input transformer model classifies arrhythmias with ~100% accuracy on ECG datasets.
Why the study?
Existing ECG classification approaches often rely on a single data modality and struggle to capture complementary temporal and morphological information while maintaining real-time performance in resource-constrained environments.
Does a multimodal transformer-based deep learning framework combining ECG signals and recurrence plot images improve arrhythmia classification accuracy?
Does a multimodal transformer-based deep learning framework combining ECG signals and recurrence plot images improve arrhythmia classification accuracy?
A novel multimodal deep learning framework combining 1D ECG signals with 2D recurrence plot images achieved >99% accuracy in classifying arrhythmias across standard open-access datasets.
Background: Accurate detection of cardiac arrhythmia from electrocardiogram (ECG) signals is essential for early diagnosis and effective patient management, particularly in Internet of Medical Things (IoMT) healthcare systems. Existing ECG classification approaches often rely on a single data modality and struggle to capture complementary temporal and morphological information while maintaining real-time performance in resource-constrained environments. Objective: This study proposes a transformer-based deep learning framework integrated within a hybrid Cloud-Edge IoMT architecture to improve arrhythmia detection. Cloud-Edge IoMT architecture aims to enhance the accuracy, robustness, and deployment efficiency of arrhythmia classification. Methods: The proposed framework combines a 1D ECG signal with its corresponding 2D Recurrence Plot (RP) image to capture complementary temporal and spatial features. A GRU network extracts temporal dependencies from ECG signals, while a CNN learns morphological features from RP images. A transformer-based fusion module integrates both feature representations through multimodal attention to enhance feature interaction and classification performance. Model training and updating are conducted in the cloud, while inference is executed on edge devices to reduce latency, enhance data security, and enable timely classification. The framework is evaluated in the MIT-BIH Arrhythmia Database and the PTB Diagnostic Electrocardiography Database, including cross-dataset validation to assess generalization. Results: The proposed framework achieves an overall accuracy of 99.632% on the MIT-BIH dataset for five-class arrhythmia classification and 99.052% on the PTB dataset for binary classification. The evaluation of the dataset across different evaluation settings further demonstrates strong generalization capability while maintaining high precision, recall, and F1-score. Conclusion: The experimental results demonstrate that integrating multimodal ECG representation with transformer-based feature fusion within a hybrid Cloud-Edge IoMT framework significantly improves arrhythmia classification performance while supporting scalable, low-latency deployment for intelligent remote cardiac monitoring.
No takes yet. Share an insight, caveat, or question.
Ibrahim et al. (2026) studied Cardiac arrhythmia. Transformer-based multimodal deep learning framework was evaluated on Arrhythmia classification accuracy. A transformer-based framework integrating 1D ECG signals and 2D recurrence plot images achieved 99.632% accuracy for 5-class arrhythmia classification on the MIT-BIH dataset.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: