Freezing of Gait Detection Using Discrete Wavelet Transform and Hybrid Deep Learning Architecture

Abstract

Freezing of gait (FoG) detection using wearable sensors plays an important role in both online and offline monitoring of Parkinson's disease patients. In a FoG detector, feature extraction is commonly considered as a critical part for distilling the sensor signals before the FoG classification. Manually extracted features with domain knowledge are widely used in conventional machine learning methods while recent deep learning algorithms introduce the automatic feature learning approach. In this paper, we propose a FoG detection framework, in which hand-crafted features are used as input to a hybrid deep learning model for further feature learning and classification task. The hand-crafted features with time-frequency representation are extracted from the raw sensor signal by using a multi-level discrete wavelet transform (DWT). A hybrid deep learning architecture constructed from two algorithms: convolutional neural network (CNN) and bidirectional long short-term memory network is then deployed to extract deep features and classify FoG events. For performance comparison purposes, experiments on different input data types and machine learning methods are carried out on the Daphnet public dataset.