The problem of classifying short atrial fibrillatory segments in ambulatory ECG recordings as being either paroxysmal or persistent is addressed by investigating a robust approach to signal characterization. The method comprises preprocessing estimation of the dominant atrial frequency for the purpose of controlling the subbands of a filter bank, computation of the relative subband (harmonics) energy, and the subband sample entropy. Using minimum-error-rate classification of different feature vectors, a data set consisting of 24-h ambulatory recordings from 50 subjects with either paroxysmal (26) or persistent (24) atrial fibrillation (AF) was analyzed on a 10-s segment basis; a total of 212,196 segments were classified. The best performance in terms of area under the receiver operating characteristic curve was obtained for a feature vector defined by the subband sample entropy of the dominant atrial frequency and the relative harmonics energy, resulting in a value of 0.923, whereas that of the dominant atrial frequency was equal to 0.826. It is concluded that paroxysmal and persistent AFs can be discriminated from short segments with good accuracy at any time of an ambulatory recording.
© 2011 IEEE