Signal Processing and Deep Learning Based Smartwatch Photoplethysmography Data Classification of Atrial Fibrillation, Premature Atrial and Ventricular Contraction

Atrial Fibrillation (AF) with high mortality rate needs to be monitored and detected accurately. AF is indicated as varying pulse-to-pulse intervals in a PPG signal. To record Photoplethysmography (PPG) signal, wrist-watches are used. AF detection is made using features, discriminating AF from Normal Sinus Rhythm (NSR). The presence of Premature Atrial and Ventricular Contraction
$(\text{PAC}/\text{PVC})$
in subjects due to its randomness may lead to false AF detection. The proposed methodology utilizes Convolutional Neural Network (CNN) on Time-Frequency spectra (TFS) along with signal processing for classification of PPG signal into AF, NSR and
$\text{PAC}/\text{PVC}$
, The Poincare plot based
$\text{PAC}/\text{PVC}$
detection implemented in this paper not only separates
$\text{PAC}/\text{PVC}$
from NSR and AF but also improves the accuracy of AF and NSR detection. The results for training and testing are validated on UMass Simband dataset (Smartwatch PPG Data with AF, NSR, PAC, PVC) and MIMIC III dataset (Finger Tips Pulse Oximetry PPG data with AF, NSR,
$\text{PAC}/\text{PVC})$
., The experimental results have shown that the proposed system gives higher accuracy, sensitivity and specificity values for both datasets.