Atrial fibrillation (AF) is the most commonly experienced sustained arrhythmia, and it increases risks of stroke and congestive heart failure. Unobtrusive wearable solutions with photoplethysmography (PPG) have been proposed for AF detection and the performance has been mainly evaluated for short-term measurements in controlled measurement settings. In this study, we evaluate the predictive value of features from PPG for AF detection under both hospital and free-living conditions. PPG from the wrist was measured from 18 patients before and after cardioversion and from 16 patients (4 with 100% AF) for 24 hours. Single-lead ECG and 24-hour Holter were used respectively as gold standards. Six PPG-based inter-beat interval (IBI) variability and irregularity features were computed in three different sliding time windows. Thresholds for AF classification for every individual feature were determined with the data from the hospital conditions and tested with the measurements from free-living conditions. Overall, the best classification results were obtained by using a 120-s window, pNN40 resulting as the best feature. On average, the sensitivity was higher in the hospital conditions (92.3% vs. 71.6%) and the specificity higher in the free-living conditions (60.7% vs. 84.9%). In conclusion, testing the classification perfomance in free-living conditions is essential to properly evaluate AF detection models.
|Number of pages||4|
|Journal||Computing in Cardiology|
|Publication status||Published - 1 Jan 2017|
|Event||44th Computing in Cardiology Conference (CinC 2017) - Rennes, France|
Duration: 24 Sep 2017 → 27 Sep 2017
Conference number: 44