Session S31.6

Predicting the End of an Atrial Fibrillation Episode: The PhysioNet Challenge

F Cantini, F Conforti, M Varanini, F Chiarugi,
G Vrouchos

CNR Institute of Clinical Physiology
Pisa, Italy

Atrial fibrillation (AF) is the most common arrhythmia. Paroxysmal (spontaneously terminated) atrial fibrillation (PAF) is, by evidence, antecedent to sustained atrial fibrillation that requires a pharmacological or external electrical intervention (cardioversion) to allow its termination. The risks of sustained atrial fibrillation are, nevertheless, serious because it predisposes to thromboembolism as a result of stasis and thrombus formation within the atria that can cause stroke or other thromboembolic events. Thus, the discrimination between paroxysmal and sustained atrial fibrillation and the prediction of PAF termination can be invaluable in order to avoid useless therapeutic interventions, to minimize the risks for the patient and to save money when the health care costs are strictly monitored.
Data from 3 different learning sets provided by PhysioNet were analysed for a total of 30 records. Each record is 1 minute length and was extracted by a two-channel Holter ECG with sampling rate 128 Hz. An algorithm, based on adaptive filtering, was developed to separate the atrial from the ventricular activity, then the atrial activity was analysed in the frequency domain. For each record the dominant atrial frequency (DAF) was evaluated and associated with the average ventricular heart rate. A fully automated linear classifier was developed to discriminate the cases of sustained AF from the cases of paroxysmal AF and, for the latter, to evaluate their proximity to the termination.
In the learning set for the event 1, 90% of the atrial fibrillations were correctly classified as sustained or paroxysmal. In the learning set for the event 2, 70% of the cases were correctly classified. The algorithm was then assessed on the test set A (30 cases, for the event 1) provided by PhysioNet for the CinC challenge with a score of 27/30 (for the time being, entry 20040416.045753, entrant 6). The algorithm for the event 2 was applied to the test set B (20 cases, event 2) with a score of 18/20 (for the time being, entry 20040426.022624, entrant 6).
Further improvement can be done on the filter technique for a better evaluation of the DAF trying to minimize the effects of signal artefacts and extrasystoles in the adaptive filtering used.