Automatic Screening of Atrial Fibrillation in Thumb-ECG Recordings

The present study proposes a novel sorting algorithm for identification of patients with atrial fibrillation in large one-lead ECG repositories. Repeated measurements at home with automatic transmission of data to a central database is presently tested in the search for atrial fibrillation for the long-term purpose to reduce the incidence of stroke. Such screening rapidly generates large databases of signals waiting to be sorted and prioritized. The one-lead ECGs were first preprocessed to remove baseline wander followed by beat detection and beat classification. A rhythm analysis stage was employed to perform RR interval analysis with negligible influence of ectopic beats and disturbances. RR interval information in combination with a waveform clustering procedure applied to the expected P wave intervals were used to sort the database into a low priority group containing mainly sinus rhythm, a high priority group containing all ECGs with irregular beat patterns, and a third group showing an unreliable RR series. The outcome of the algorithm was compared to an annotated database containing 2837 one-lead ECG recordings from 103 patients where each recording was visually inspected by a physician. The proposed method was able to divide the database into a low-priority group containing 93% (n=2357) of the sinus rhythm cases and a high priority group containing 98% (n=55) of the atrial fibrillation cases. In addition, 3.7% were found to have an unreliable RR series. In conclusion, automatic analysis of one-lead ECG databases can quickly guide the physician to find recordings with high probability to contain atrial fibrillation and can automatically indicate if a recording needs to be remade due to quality problems.