RESUMO
Automatic detection of atrial fibrillation (AF) is a challenging issue. In this study we proposed and validated a model to identify AF by using facial video recordings. We analyzed photoplethysmographic imaging (PPGi) signals, extracted from video of a subject's face. Sixty-eight patients were included: 30 in sinus rhythm (SR), 25 in AF and 13 presenting with atrial flutter or frequent ectopic beats (ARR). Twenty-six indexes were computed. The dataset was divided in three subsets: the training, validation, and test set, containing, respectively, 58, 29, and 13% of the data. Mean of inter-systolic interval series (M), Local Maxima Similarity (LMS), and pulse harmonic strength (PHS) indexes were significantly different among all groups. Variability and irregularity parameters had the lowest values in SR, the highest in AF, with intermediate values in ARR. The PHS was higher in SR than in ARR, and higher in ARR than in AF. The LMS index was the highest in SR, intermediate in ARR and the lowest in AF. Similarity indexes were higher in SR than in AF and ARR. A model with three features, namely M, Similarity1 and LMS was chosen. With this model, the accuracy for the validation set was 0.947±0.007 for SR, 0.954±0.004 for AF and 0.919±0.006 for ARR; for the test set (never-seen data), accuracy was 0.876±0.021 for SR, 0.870±0.030 for AF and 0.863±0.029 for ARR. A contactless video-based monitoring can be used to detect AF, differentiating it from SR and from frequent ectopies.
RESUMO
OBJECTIVE: Undiagnosed atrial fibrillation (AF) patients are at high risk of cardioembolic stroke or other complications. The aim of this study was to analyze the blood volume pulse (BVP) signals obtained from a wristband device and develop an algorithm for discriminating AF from normal sinus rhythm (NSR) or from other arrhythmias (ARR). APPROACH: Thirty patients with AF, 9 with ARR and 31 in NSR were included in the study. The recordings were obtained at rest from Empatica E4 wristband device and lasted 10 min. The analysis, on a 2 min segment, included spectral, variability and irregularity analysis performed on the inter-diastolic interval series, and similarity analysis performed on the BVP signal. Main results and Significance: Variability parameters were the highest in AF, the lowest in NSR and intermediate for ARR, as an example pNN50 values were, respectively, [Formula: see text], [Formula: see text], [Formula: see text] (p < 0.05). The similarity parameters were the highest in NSR, the lowest in AF and intermediate for ARR, as an example using a threshold for assessing similarity of [Formula: see text]: [Formula: see text], [Formula: see text], [Formula: see text], all p < 0.05. The rhythm classification was preceded by over-sampling (using synthetic minority over-sampling technique) the class of ARR, being it the smallest class. Then, the features selection was performed (using the sequential forward floating search algorithm) which identified two variability parameters (pNN70 and pNN40) as the best selection. The classification by the k-nearest neighbor classifier reached an accuracy of about 0.9 for NSR and AF, and 0.8 for ARR. Using pNN70 and pNN40, the specificity for the three rhythms was Spnsr = 0.928, Spaf = 0.963, Sparr = 0.768, while the sensitivity was Spnsr = 0.773, Spaf = 0.754, Sparr = 0.758.
