A Feasibility Study on Biometric Identification Using Multimodal Signal Analysis of Photoplethysmogram and Phonocardiogram

This paper presents a study on multimodal signal fusion for biometric identification based on cardiac physiological signals from commercial-grade wearable devices. Firstly, photoplethysmogram (PPG) and phonocardiogram (PCG) signals are collected with a smartwatch and segmented into individual cardiac cycles, which are then fused as inputs for models. Secondly, two models with different architectures are employed to train and classify these signals. One is modified by BiometricNet based on the convolutional neural network and the long short-term memory structure. The other is ResNet18. The data from 59 subjects is used to evaluate the biometric identification task. The accuracy of using fused signals is higher than that of using individual signals in both the Modified-BiometricNet and the ResNet, reporting best results of 81.42% from the modified BiometricNet and 94.71% from the ResNet. These results show that the fusion of multimodal cardiac physiological signals from different modalities can enhance the performance of biometric identification. We infer that the fusion of multimodal cardiac physiological signals can add more useful features and suppress interference features.