Predicting the end of an atrial fibrillation episode: the physionet challenge

The PhysioNet Challenge 2004 addresses two differenr goals: to separate the persistent atrial $brillation (A F) from the paroxysmal AF (event I ) and, in case of paroxysmal AF, to identify the one-minute ECG strip just before the termination of the AF episode (event 2). Both events were approached through the separation of the atrial activity by the ventricular one in the ECG recordings (I-minute, two leads, 128 Hz). This separation was obtained through two different methods: a) QRST cancellation through cross-channel adaptive filtering; b) beat classification and class avernged beat subtraction. For event I , the averaged RR (index of ventricular uch'vity) was put into relationship with the Dominant Atrial Frequency (DAF) (index of atrial activity). A linear classifier was evaluated separating the RWDAF plane into the N-type and T-type regions. The best score was 95% on learning sets and 27/30 on testing set A. For event 2, once the S-type and T-type signals were joined for each patient using a QRST correlation method, sign9cative parameters were singled out in the DAFs during the penultimate and last two seconds of the S-rype and T-type recordings. Criteria based on rhe DAF trend of each signal in its last seconds and criteria based on the DAF comparison between S-rype and T-type signals were jointly used. The best score was 80% on learning sets and iBR0 on testing set B. 1.