Implementing Deep Learning for Atrial Fibrillation Diagnosis

Atrial fibrillation (AF) is a critical yet often underdiagnosed cardiac arrhythmia, particularly in low-resource environments where digital ECG systems and cardiology specialists may be inaccessible. This study presents a deep learning-based diagnostic tool designed to classify AF and normal sinus rhythm (NSR) from real-world, clinic-verified ECG images. Leveraging EfficientNet-B0, a convolutional neural network optimized for efficiency and accuracy, the model was trained on preprocessed images of 12-lead ECG printouts captured via mobile phone. Preprocessing included grayscale conversion, denoising, sharpening, and resizing, ensuring consistent input quality across variable capture conditions. The model achieved a 93% accuracy, balanced F1-scores of 0.93 for both classes, and an ROC-AUC of 0.98 on a validation dataset. Beyond validation, the model was integrated into a web-based interface and tested on newly acquired ECG images under varied real-world conditions, maintaining strong classification performance with cautious predictions. By combining accessible imaging workflows with deep learning inference, this system offers strong potential to serve as an accessible diagnostic aid in under-resourced healthcare environments.