P-wave Analysis in Atrial Fibrillation Detection Using a Neural Network Clustering Algorithm

Abstract

Absence of P-waves in ECG records with irregular interbeat intervals (R-R) is a sign of Atrial Fibrillation (AF). Detection of P-waves in ECG beats or even average beats could be challenging if the artifact resembles a P-wave, or an actual P-wave is buried in the artifact. We developed a neural network algorithm to generate the ECG beat clusters in segments of the record. Beats with matching QRS complexes were clustered using Self-Organizing Map (SOM) technique and then cross-correlated to combine and generate the dominant clusters. This process helps to eliminate the abnormal or artifact-corrupted beats. Fiducial points of the dominant average beat were measured by morphological techniques. If the P-wave was detected in the average beat, a smaller search window was defined for individual beats to exclude the potentially false P-waves. A set of P-wave features determined the presence of P-wave throughout an ECG segment. Our algorithm was tested on several datasets with annotated intervals for some cardiac rhythms. A decision tree ensemble of bagged trees classifier was developed and applied to the P-wave and interbeat interval features, resulting in AF/non-AF classification with average F1 score of 96.0% in training subset and 95.6% in test subset of all records.