Session S91.4

On Predicting the Spontaneous Termination of Atrial Fibrillation Episodes Using Linear and Non-Linear Parameters of ECG Signal and RR Series

L Mainardi, N Gatti, M Matteucci

Politecnico di Milano
Milano, Italy

A few human studies evidenced the occurrence of spontaneous re-organization in the electrical activities of different atrial sites in the period preceding the termination of atrial fibrillation (AF) episodes. It could be argued that some signs of this re-organization may also appear on surface ECG recordings. Aim of this study is to assess the presence of subtle changes in ECG leads, revealing the regularization of atrial sites, that could be used as a predictor of the termination of the AF episodes. In the past, the analysis of residual ECG signal (i.e. the ECG signal in which the ventricular activities, QRST complex, has been cancelled through beat averaging techniques) has been proposed to characterize atrial activities: it evidences marked changed during AF, after infusion of drug or in presence of various atrial rhythms. In addition, some authors reported the presence of different patterns in the series of atrial activations during organized and non-organised atrial rhythm. In this study, a set of features is therefore extracted from both residual ECG signal (16 features) and RR interval series (9 features). The computed metrics aimed at measuring glimpse of re-organization in atrial activities through the quantification of both linear and non-linear parameters. These parameters include: Entropy based measures (Approximated and spectral Entropy, Regularization and Synchronization indexes), spectral analysis parameters (frequency and amplitude of fibrillation waves, coherence function) and model based indexes (level of predictability). The features were computed on the data set of Computers in Cardiology Challenge 2004. The features extracted from the dataset have been reduced from a bigger set using principal component analysis and they have been normalized before training a feed forward neural network. We used leave one out cross-validation to select the number of neurons in the hidden layer and the final output is obtained from a committee of networks. Due to the small number of records we decided to apply the co-train framework to the two different sets of features (i.e., features from ECG signal and RR interval series) building two different classifiers and gaining information also from the test sets. Results comparing the co-training approach and classical classifiers using a mixed set of features are evaluated.