Improving accuracy of atrial fibrillation detection in lossy ECG streams

Continuous physiological monitoring is often the best available tool for detecting and treating asymptomatic, intermittent pathologies like Atrial Fibrillation. A particularly effective algorithm is based on the variance of inter-beat intervals. This algorithms relies on the detection of the QRS complex and is thus fairly robust to noise. Unfortunately, we find that the algorithm is very susceptible to lost data and can quickly degrade even when small parts of the ECG stream are missing. For home-based environments with small devices and wireless data transmission, data loss and noise are inevitable and as such an algorithm that is both robust to noise and lost data becomes necessary. In this paper we present a new Atrial Fibrillation detection algorithm that has the above stated desired qualities. We have run the original and the modified algorithms on a collection of patients from the Physionet database exhibiting Atrial Fibrillation. Even with data loss as little as 10% the original algorithm degrades rapidly and its output is only 2-3% similar to the no-loss case. The loss-conscious algorithm continues to provide output that is more than 90% similar to the no-loss case even for data loss rates as high as 30%