A CNN using only P-waves achieved 92.86% accuracy and 0.9231 F1-score for AF classification, outperforming full ECG CNN by ~18% accuracy and 12% F1-score.
Does a convolutional neural network trained exclusively on extracted P-waves improve the accuracy of atrial fibrillation classification compared to using the full 12-lead ECG?
Training AI models exclusively on P-waves rather than full 12-lead ECGs significantly improves the accuracy and precision of automated atrial fibrillation detection.
Absolute Event Rate: 0% vs 0%
Abstract Background Atrial fibrillation (AF) affects about 2% of the global population, placing a significant burden on healthcare systems. AF is characterised by rapid and irregular electrical excitations in the atria, which manifest as abnormalities in electrocardiogram (ECG) morphologies, particularly the P-wave. Artificial intelligence (AI) has been employed for AF diagnosis to facilitate automated detection, reducing both time and human resource demands. However, existing AI models in both research and clinical practice typically utilise the full ECG, which can lead to downstream issues such as false positives and potential reliance on physiologically irrelevant features – thus compromising model transparency. Moreover, AI models trained on excessive features can introduce adverse effects, commonly referred to as the "curse of dimensionality." Aim This study aims to investigate whether using P-wave extracted from 12-lead ECG data can improve the accuracy of AI classification for AF compared to using the full 12-lead ECG. This will improve efficacy and transparency of automated AI4AF diagnosis. Methods The 12-lead ECG dataset used in this study was obtained from the UK Biobank and included 212 ECGs from control patients and 186 ECGs from AF patients. An interval-based algorithm was developed to automatically extract P-waves from each ECG waveform in each lead. The extracted P-wave signals were then used to train a convolutional neural network (CNN) for AF classification. The performance of this CNN model was compared with that of a CNN trained on the full 12-lead ECG (Figure 1). Model performance was assessed using the standard metrics of accuracy and F1-score. Results The CNN trained exclusively on extracted 12-lead P-waves achieved an accuracy of 0.9286 and F1-score of 0.9231 in classifying AF patients. In contrast, the CNN trained on the full 12-lead ECG signals demonstrated a lower accuracy of 0.7634 and F1-score of 0.8118. Therefore, the AI model utilising only P-waves exhibited greater precision and fewer false positives than the model leveraging the full ECGs, as reflected in an ~18% increase in accuracy and a 12% improvement in F1-score. Conclusion This study demonstrates that AI models trained exclusively on P-waves for AF patient classification achieve significantly higher accuracy than AI utilising the full ECG. These findings can guide the development of more effective AI-based automated AF diagnostic systems and potentially enhance AF detection on a large population.AF classification from full ECG & P-wave
Hernandez et al. (Sat,) reported a other. A CNN using only P-waves achieved 92.86% accuracy and 0.9231 F1-score for AF classification, outperforming full ECG CNN by ~18% accuracy and 12% F1-score.