PPG Signal Acquisition and Preprocessing: The PPG signals are acquired from the smartphones by using their camera and flash. In contrast to other studies that simply average 50 x 50 pixels of each video frame, we propose a novel method to better capture the variations of the pulsatile signal while reducing the computational cost and memory requirements.

Statistical Feature Extraction: Similar to other studies, we calculate statistical features from the PPG signal to distinguish between AF and normal sinus rhythm (NSR). In addition, we also make use of the features to detect strong hand movement during the measurement and label them as vibration (Vib). Implemented in a smartphone application, the user can be immediately informed that the measurement should be repeated. The statistical features are divided into two categories that are associated with the time-domain and the frequency domain of the PPG signal: Time-Domain Features and Frequency-Domain Features.

Feature Selection and Classification: We divide the classification procedure into two different stages: wrapper type feature selection and classification via support vector machines (SVM). Both stages are computational demanding but are not intended to be implemented in smartphone applications, as they only need the resulting classifications rules. First, the feature selection aims at finding the best feature combination of the presented set of features to distinguish between the classes. Feature selection reduces the computational cost and reveals the different levels of the features’ significance. Then, the SVM is applied to find the optimal decision equation to separate the classes. Finally, only the features of the best combination need to be calculated in the smartphone application and classification can be performed based on decision lines for pairs of features or hyperplanes in case of multiple features.