Deep Learning Framework for Biometric Identification from Wrist-Worn PPG With Acceleration Signals

The photoplethysmography (PPG) signal has advantages in terms of accessibility and portability, which makes its usage in many applications such as user access control very attractive. In this paper, we propose a novel deep learning-based biometric identification framework (BIDNET), which uses photoplethysmography (PPG) and 3-axis acceleration signal data collected from wrist-worn sensors in an ambulatory environment. We developed a completely personalized data-driven method using eight layers of deep neural networks, which uses five convolutional neural network (CNN) layers and two bidirectional long short-term memory (Bi-LSTM) layers, followed by one dense output layer. It is used to model the time series inherent in the beating signal representing the heart activity. The proposed network structure was evaluated on the ISPC dataset, which was collected from 20 subjects (including the test set) participating in sports activities. The best performance achieved in terms of average accuracy, F1 score, recall, and precision are 0.98, 0.99, 1.00 and 0.99 respectively over 20 subjects for five-fold cross-validation. The proposed model outperforms the state-of-the-art methods for PPG-based biometric identification.