## Session S91.6

A Statistical Feature Based Approach to Predicting Termination of Atrial Fibrillation

FM Roberts, RJ Povinelli

Marquette University

Milwaukee, WI, USA

The electrophysiological characteristics of the termination of atrial fibrillation (AF) are poorly understood. Hence, the Computers in Cardiology 2004 Challenge is to better understand these characteristics. The challenge is two-fold. The first part is to identify features in ECG signals that distinguish non-terminating from immediately terminating AF. The second part is to differentiate between AF that terminates immediately and AF that terminates in one minute.

The challenge dataset consists of 30 one-minute ECG records. All records are characterized by AF with 10 records of non-terminating AF, 10 records of terminating in one minute AF, and 10 records of terminating immediately AF. This paper presents two approaches to this problem. The first is a feature based approach. The second uses a nonlinear signal classification method.

The feature based approach uses a statistical classifier to characterize the temporal evolution of the following features: mean and standard deviation of the R-R variability, power spectrum of the fibrillatory baseline after QRST averaging and subtraction, and standard deviation of the fiduciary point of normalized QRS complexes. Using leave-one-out cross validation on the training set, this approach has a 80% accuracy on the non-terminating vs. immediately terminating problem and a 70% accuracy on immediately terminating vs. one-minute terminating problem.

The second approach is based on the reconstructed phase spaces (RPSs) of frequency sub-banded ECG signals. Given sufficient dimension, a RPS provides a method for reconstructing a state space description of the generating system, which in this case is the cardiac system, from a signal sampled from that system, which in this case is the ECG signal. As the cardiac system in AF evolves through time towards termination, the state structure describing that system will reflect this evolution. This time evolution is captured through Gaussian Mixture Models of the RPSs. A Bayesian classifier is used to identify previously unseen ECG signals.