A deep learning approach using transfer learning on ECG images validated with real-world clinical data for arrhythmia detection
Biomedical Signal Processing and Control, cilt.126, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 126
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.bspc.2026.110897
- Dergi Adı: Biomedical Signal Processing and Control
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, EMBASE
- Anahtar Kelimeler: Arrhythmia detection, Cardiology, ECG, Pre-trained model, Signal processing
- Sivas Cumhuriyet Üniversitesi Adresli: Evet
Özet
Background and Objective: The applications of artificial intelligence technologies in the health sector are revolutionizing the diagnosis and treatment processes of diseases. Arrhythmias are critical cardiac conditions that require accurate and timely diagnosis, yet traditional ECG signal analysis methods often struggle to capture rhythm-level patterns effectively. This study addresses this limitation by leveraging deep learning and transfer learning on two-dimensional ECG images, aiming to develop high-performance models that mimic cardiologists’ visual interpretation processes. Methods: Three datasets were utilized: the MIT-BIH Arrhythmia Dataset, segmented into 10-s rhythm-level annotations; the PTB-XL Dataset, containing diverse multi-label annotations; and the ADIRA Dataset, offering real-world ECG recordings annotated by cardiologists. Pre-trained models, including VGG16, DenseNet201, and EfficientNetV2L, were applied to binary, multiclass, and multiclass-without-SR classification tasks. Results: Results demonstrated that VGG16 outperformed other models, achieving 97% accuracy in multiclass classification, while binary models showed superior performance for specific conditions like sinus bradycardia. Removing the SR (normal) class in multiclass classification had negligible impact, and the ADIRA dataset validated the models’ practical applicability in clinical settings. Conclusions: These findings underscore the effectiveness of image-based deep learning techniques for rhythm classification, bridging the gap between traditional signal processing and modern methods. Future work will focus on optimizing lightweight models for mobile platforms and refining rhythm-level diagnostics to enhance accessibility and applicability.