Session S32.3

A Statistical Approach for Accurate Detection of Atrial Fibrillation and Flutter

S Dash*, E Raeder, S Merchant, K Chon

Stony Brook University
Stony Brook, NY, USA

Atrial fibrillation (AF) is the most common clinical arrhythmia afflicting 2-3 million Americans. It is a major risk factor for ischemic stroke and therefore early detection of AF must be a public health priority. Atrial fibrillation is generally considered to be a random sequence of heart beat intervals with markedly increased beat-to-beat variability. We have developed an algorithm for real-time detection of AF and atrial flutter (AFL) which combines four statistical techniques to exploit these characteristics, namely the Root Mean Square of Successive RR interval differences to quantify variability, the Turning Points Ratio to test for randomness of the time series, Shannon entropy to characterize its complexity and a high-resolution time-frequency spectral method to find the number of high frequency spectral peaks in a given 128-beat RR interval segment.
In an analysis of long-term recordings in the MIT Atrial Fibrillation database, we have achieved a beat-to-beat sensitivity of 95% and a specificity of 96.7% in detecting AF. For clinical applications, the most relevant objective is to detect the presence of AF in a given recording but not necessarily every single AF beat. Using this latter criterion, we achieved episode detection accuracy of 100% for the MIT-BIH AF database. In a more recent analysis of 36 Holter recordings provided by the Scottcare Corporation (www.scottcare.com), we correctly identified the presence of AF episodes in all subjects with a beat-to-beat sensitivity of 95% and specificity of 69%. The algorithm performed well even when tested against AF mixed with several other potentially confounding arrhythmias in the MIT-BIH Arrhythmia Database (Sensitivity = 90.2%, Specificity = 91.2%). The flutter detection algorithm has undergone preliminary testing on 2 files of the MIT AFIB database which contained around 80 minutes of AFL. High sensitivity (97%) and specificity (95%) have been obtained. Due to the simplicity of our algorithm, the computational speed is higher, thus making it easier to implement and requiring less memory than competing methods which require storage of more training data.

(Abstract Control Number: 44)