P-Wave Spatial Variability in 12-Lead ECG Predicts Temporal Changes in Atrial Fibrillation

Atrial fibrillation (AF), the most common sustained cardiac arrhythmia, is closely linked to abnormalities in the atrial substrate, including structrural and electrical remodeling. P-wave indices have been proposed for the early identification of patients prone to AF, with the beat-to-beat variability of the P-wave morphology in long vectorcardiographic (VCG) recordings revealing that AF patients present a more variable P-wave morphology compared with healthy subjects. In the current work, we use unsupervised learning techniques to categorize AF patients based on this temporal variability of the P-waves. Subsequently, supervised machine learning techniques are used for the classification of the patients using P-wave characteristics extracted from a standard 12-lead ECG. Thirty AF patients were included, for whom both VCG and 12-lead ECG recordings were acquired. k-means clustering, based on the silhouette value, revealed the existence of two main categories of patients using the P-wave temporal variability from the X lead of the VCG. P-wave characteristics on a beat-to-beat basis were calculated for all the leads of the 12-lead ECG and the standard deviation for each patient was calculated in order to quantify the spatial variability of the P-waves. Statistically insignificant features were excluded, and the Boruta feature selection algorithm, together with a voting method, identified the most important features, which were ultimately used for training two models: a Support Vector Machine and a Random Forest. A class balancing technique was also applied. Due to the small dataset, the entire process was repeated 100 times, revealing that the RF model achieved an accuracy of 80.2%, suggesting that the temporal variation of the P-wave can be predicted from the analysis of its spatial characteristics in a short recording. A larger dataset is required to improve the trustworthiness and reproducibility of the model.Clinical Relevance- This work can be used for the stratification of AF patients during sinus rhythm using a standard 12-lead ECG, reducing time and cost, and leading to a better allocation of resources.