A GRU-LSTM Network Bank for Efficient Detection of Freezing of Gait for Parkinson's Disease Patients

Freezing of Gait (FoG) is a serious complication of Parkinson's disease (PD) that increases the likelihood and frequency of falls and severe injuries. Deep learning (DL) has been proven to be successful in detecting PD and its complications allowing for a timely intervention and a prompt initiation of effective therapeutic treatments. In this paper, we introduce a novel solution for the detection of FoG using a hybrid gated recurrent unit (GRU) - long short term memory (LSTM) network bank applied on three-axial acceleration time-series data collected for patients experiencing FoG. The proposed framework achieved a high subject independent accuracy up to 88 % for classifying the data into FoG or no FoG. This represents a reliable approach where the agreement among multiple machine learning models applied on different components of the acceleration data was deployed to predict the occurrence of FoG.