Abstract: . . .  sufficiently different R-R statistics from terminating AF to build a useful classifier. Further study of the physiological reasons behind this is warranted. Second, spectral analysis 5 Page 7 Figure 4: The last beat of records b01 and b03. b03 is chosen as the terminating record because of an absence of p-waves during ventricular recovery. can be used to automatically find records belonging to the same patient. This could have wide application to electronic record systems. References [1] R. Duda, P. Hart, and D. Stork. Pattern Classification . John Wiley & Sons, 2000. [2] A.L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. Ch. Ivanov, R. . . .
. . .  8 [5] George Moody. Computers in cardiology challenge 2004. URL=http://www.physionet.org/challenge/2004/, October 2003. [6] L. R. Rabiner and B. H. Juang. Fundamentals of Speech Recognition . PrenticeHall, 1993. [7] I Savelieva and A. John Camm. Atrial pacing for the prevention and termination of atrial fibrillation . Am J Geriatr Cardiol , 11(6):380398, 2002. 7 Page 9 Record Closest Second Third s01 t01 0.1118 s07 3.5054 t08 3.6134 s02 t02 0.0329 t05 0.4165 s05 0.4851 s03 t03 0.1488 s09 0.7625 t09 0.8337 s04 t04 0.0164 t10 0.7218 s10 1.0065 s05 t05 0.0330 s02 0.4851 . . .
. . .  Page 8 [5] George Moody. Computers in cardiology challenge 2004. URL=http://www.physionet.org/challenge/2004/, October 2003. [6] L. R. Rabiner and B. H. Juang. Fundamentals of Speech Recognition . PrenticeHall, 1993. [7] I Savelieva and A. John Camm. Atrial pacing for the prevention and termination of atrial fibrillation . Am J Geriatr Cardiol , 11(6):380398, 2002. 7 Page 9 Record Closest Second Third s01 t01 0.1118 s07 3.5054 t08 3.6134 s02 t02 0.0329 t05 0.4165 s05 0.4851 s03 t03 0.1488 s09 0.7625 t09 0.8337 s04 t04 0.0164 t10 0.7218 s10 1.0065 s05 t05 0.0330 . . .
. . .  a testing set. We use the supplied QRS annotations to compute the R-R interval series for each record and train Gaussian mixture models for Group N and Group T using the data in each groups training set. For each record in the test set, we calculate the likelihood that each R-R interval fits the model for Group N and Group T. Each R-R interval then generates a vote for the model with the greater likelihood. We then assign the entire record to Group N or Group T according to which model received the most votes. Table 1 shows the results of this experiment for models with 1, 2 and 4 mixture components. We see that the 1 mixture system, corresponding to simply using a Gaussian model for each group, confuses . . .
--3000,4,375,3303,19141