A statistical feature based approach to predicting termination of atrial fibrillation

This paper presents a nonlinear signal classification approach to differentiating different atrial fibrillation terminalion stages; non-terminating (N), terminating in one minute (S), and terminating immediately (T) following the end of the recording. The nonlinear approach is based on Gaussian mixture models of reconstructed phase spaces of the last 2s of data in each recording. The removal of the ventricular component of the signal was removed by one of two methods: QRST averaging and subtraction and 4-9 Hz bandpass filtering of the recording. The accuracy of the approach is 63.3 and 66.7% for differentiaring N vs. T and 60 and 70% accuracy for differentiating S us. Z for QRST subtracted and filtered data respectively. When the training data was augmented with 150 more training cases, the results improved to 66.7 and 80% for N vs. T ond 70 and 70% for S us. T. An artifact was noted in the recordings that allowed a different set of criteria (the slope and percent of datu remaining after last Q point at the end of each of the recordings) to accurately classify N us. T at 80% and S vs. Tat 85%.