Abstract 18248: Development and Evaluation of a Novel Algorithm for Diagnosing Atrial Fibrillation Using Wristband Technology
Introduction: Atrial fibrillation (AF) increases the risk of ischemic stroke by 5-fold, yet is often asymptomatic and undiagnosed. Wristband technologies using photoplethysmography (PPG) for pulse detection may capture undiagnosed, asymptomatic AF episodes. Signal processing methods which exclude and/or filter noisy segments may also be advantageous to clinical interpretations.
Hypothesis: A computer-based algorithm using median wavelet entropy and noise detection may accurately detect AF in short-term wristband PPG recordings with accuracy superior to human expert readers.
Methods: We used the Samsung Simband to record short-term PPG signals from 46 adult inpatient subjects on telemetry monitoring with (n=15) and without (n=31) AF. Multiple channels of PPG were recorded for approximately 5 minutes at 128 Hz. After noise filtering and automatic elimination of artifact, 10 pre-specified complexity (entropy) features were employed in an Elastic Net logistic regression algorithm. Ten-fold cross-validation was performed to test the model in 36 subjects, and independent validation was performed in 10. A blinded cardiologist’s interpretations were used as a basis of comparison.
Results: All 10 pre-specified features of complexity were significant predictors of AF. The area under the curve from the 10-fold cross-validation was 0.99, with a sensitivity of 0.97 and specificity of 0.94. The algorithm successfully predicted AF status in all 10 subjects of the independent cohort. When a blinded cardiologist reviewed all 46 PPG rhythms, 8 of 15 cases of AF were diagnosed correctly, and 28 of 31 cases of non-AF rhythms were diagnosed correctly. This was inferior to computer reads with chi-square p=0.01.
Conclusion: We have demonstrated accurate classification of rhythm into AF (vs. non-AF) based on PPG from a wristband which is superior to clinical interpretations, which likely misclassified rhythms because of motion-related noise. Further research is warranted.
Author Disclosures: N. Isakadze: None. O. Levantsevych: None. G. Clifford: None. S. Nemati: None. A. Shah: None.
- © 2016 by American Heart Association, Inc.