Atrial fibrillation detection using machine learning algorithm from single lead electrocardiograms

Aim . Atrial fibrillation (AF) represents one of the most critical cardiac arrhythmias, as it significantly increases the risk of stroke. Its detection is particularly challenging due to the unpredictable nature of its episodes. Methods . This study proposes a low-complexity algorithm, enabling integration into embedded devices for realtime AF episode detection. The proposed method integrates non-linear, time-domain and frequency-domain features extracted from electrocardiogram signals with The LightGBM algorithm (an extension of decision tree algorithm) is used to classify and detect AF. Results . The model was trained using the MIT-BIH AF Database (MIT-AFDB), achieving sensitivity (Se), specificity (Sp), accuracy rates (Acc), precision (PPV), F1-score and AUC of 0.9838, 0.9690, 0.9748, 0.9543, 0.9688 and 0.9957, respectively. We also performed 10‑fold cross‑validation on this dataset. The obtained values for Se, Sp, Acc, PPV, F1-score, and AUC were, respectively, 0.9837 ± 0.0020, 0.9701 ± 0.0021, 0.9755 ± 0.0007, 0.9559 ± 0.0029, 0.9696 ± 0.0008, and 0.9959 ± 0.0002. This indicates that the model achieves good performance compared to current studies in AF recognition and detection. Conclusions . The experimental results demonstrate that the model achieves high performance in the classification and detection of AF episodes. Furthermore, the model is suitable for integration into real-time arrhythmia detection systems.