Investigation Of Paroxysmal Atrial Fibrillation Using Multitaper Spectral Analysis Techniques
H. Bokil, M. Jarvis, P. Mitra
California Institute of Technology
Pasadena, CA, USA
With a view to identifying subjects suffering from Paroxysmal Atrial Fibrillation (PAF), we analyse the ECG time-series data provided for the "Computers in Cardiology Challenge 2001" using multitaper spectral analysis techniques. We observed that subjects suffering from PAF can be distinguished from normal subjects by the presence of enhanced power at low frequency (0-0.15 Hz) in the continuous process. Based on this observation we performed a discrimination of patients from normal subjects using linear discriminant and nearest neighbour classification techniques. Cross validation applied to the training data showed that the linear discriminant performs slightly better than the nearest neighbour classification: linear discrimination correctly classified 23 out of 25 patient records and 23 out of 24 normal records, while the nearest neighbour scheme correctly classified 22 patient records and 21 normal records.Performing linear discriminant on the test data we find that we correctly predict 29 out of the 50 records. Further investigations of the robustness of the scheme are in progress.