2. Department of Preventive Medicine, University of Tennessee Health Science Center, USA.
3. King Abdullah Petroleum Studies and Research Center, Riyadh, Saudi Arabia.
4. Risk Dynamics Consultancy, Istanbul, Turkey.
In this research article, we apply clustering within the Symbolic Pattern Recognition (SPR) framework to problems related to classifying different clinical categories of atrial fibrillation by modeling the changes in electrical activity of the heart. SPR characterizes a sequential dataset by modeling the transition behavior exhibited by patterns of symbols; clearly, this technique requires continuous data to be discretized into a set of defined symbols. With SPR, we were able to find hidden patterns in electrocardiograms (ECG) recorded during normal sinus rhythm that allowed us to classify patients as having paroxysmal atrial fibrillation (PAF) vs. those that did not. Even without extensive tuning of the model, our correct classification rate of 80% is inline with other published models. Additionally, we were able to identify normal sinus rhythm ECGs of PAF patients when a PAF episode was imminent. Finally, we used SPR clustering to distinguish between episodes of atrial fibrillation which would end within one minute (spontaneously-terminating) vs. those which needed intervention to stabilize (sustained). These are very important considerations for clinical practitioners for several reasons. The ability to screen for, and diagnose, PAF even with no known history or ongoing episode would be invaluable. This is especially true as related to elderly patients whom are at greater risk from atrial fibrillation, many of whom undergo regular ECG screenings anyway. Secondly, early warning that a PAF episode is imminent can give caregivers the chance to prepare an appropriate intervention in advance. For certain patients, this could mean the difference between life and death. Lastly, it is recognized that intervention to stabilize atrial fibrillation is not always in the best interest in the patient. One consideration is how long the episode is expected to last; in many cases, it may be better to allow an episode to spontaneously terminate.
Keywords: Symbolic Pattern Recognition, clustering of sequential data, ECG, atrial fibrillation, paroxysmal atrial fibrillation, cardiac arrhythmia, time series modeling