GDT-Net: Multi-level feature extraction network for precise diagnosis of atrial and ventricular fibrillation.

Atrial fibrillation (AFIB) and ventricular fibrillation (VFIB) are two critical cardiovascular diseases, where accurate diagnosis is essential for timely intervention. Although deep learning technology has greatly improved the diagnostic accuracy of AFIB and VFIB, its performance degrades significantly under compound noise conditions. To address this challenge, we propose GDT-Net, a neural network specifically designed for robust ECG classification under compound noise conditions. GDT-Net comprises three key modules: the G Module, D Module, and T Module, each contributing to a comprehensive analysis of cardiac electrical activity. The G Module employs a grouped convolution strategy to extract lead-specific features, ensuring the precise characterization of intra-lead electrophysiological information. The D Module utilizes a densely connected architecture to capture spatial correlation patterns and facilitate multi-lead collaborative diagnosis. Meanwhile, the T Module integrates a Transformer encoder to model temporal dependencies and long-range feature relationships within ECG signals, enhancing the model's capacity to recognize complex arrhythmic patterns. We construct six subsets for the experiments. Experimental results demonstrate that GDT-Net significantly improves classification performance for AFIB and VFIB on the MIT-BIH dataset, achieving an F1-score of 99.46% for intra-patient classification and 97.20% for inter-patient classification. Even under compound noise conditions, the method maintains an F1-score above 90% for most subsets. These results underscore GDT-Net's strong diagnostic capability under compound noise conditions, offering a more reliable and effective solution for the automatic diagnosis of AFIB and VFIB.