We present techniques to predict various types of terminating and non-terminating atrial fibrillation (AF) as required by the Computers in Cardiology Challenge 2004. First, we describe an automatic technique to distinguish non-terminating atrial fibrillation (AF) from terminating AF. Our method models R-R intervals using mixtures of Gaussians and achieves an accuracy of 100% on the training set and 76.7% on the challenge test set. Second we describe a semi-automatic technique to distinguish immediately terminating AF from AF which terminates one minute later. Our method first uses a novel automatic technique to determine which two examples are recorded from the same patient. Specifically, we convert each ECG record to a series of frames and convert each frame to ceptral features. Such features are a well-studied compact representation of the amplitude frequency spectrum. We then model the features for each patient by a Gaussian model and use the Kullback Liebler (KL) distance to determine the distance between each pair of Gaussians and hence each pair of records. The closest records are hypothesized as belonging to the same patient. This technique achieves 100% accuracy on the training set and partitions the test set into 10 unique record pairs. To complete the second task of the challenge, we then examine the R-R series for each patient and determine by hand the likely time ordering of the records in each pair, thus distinguishing which record terminates immediately. This technique achieves an accuracy of 90% on the challenge test set.
Supported by the National Institute of Biomedical Imaging and Bioengineering (NIBIB) under NIH grant number R01EB030362.