WiFOG: Integrating deep learning and hybrid feature selection for accurate freezing of gait detection

Abstract

This study investigates the feasibility of utilizing non-invasive WiFi sensing using the 4.8 GHz operating frequency band of the 5 G spectrum, which is suitable for Internet of Things applications. We propose WiFOG: a WiFi CSI system for detecting FOG in PD leveraging deep learning and wireless channel characteristics collected by wireless devices such as a radio frequency signal generator, a network interface card, and dipole antennas. The raw data for several activities, including sitting, slow-walking, fast-walking, voluntary stopping, and FOG episodes, is collected. Regress feature engineering is performed in which discrete wavelet transforms is used for signal denoising and Hilbert-Huang transforms for feature extraction. Further, we propose hybrid feature selection techniques based on whale optimization, recursive feature elimination, and select form models for dimensionality reduction. Moreover, we propose a deep-gated recurrent network (DGRU) for activity classification and FOG detection and compared the results with the state-of-the-art approaches in the existing work. The results show our proposed scheme surpasses existing FOG detection with a total improvement of approximately 4% in accuracy and a 29% reduction in training time.