Freezing of gait (FoG) is a general and disabling indicator during the severe level of Parkinson's disease (PD) that affects millions of PD patients worldwide. Under episodic condition that cannot be predicted, FoG influences gait in term of delay and sudden inability to perform walking. FoG does not respond well to treatment through medication; therefore effective non-medication assistance is necessary. A wearable assistance system for FoG detection has been developed to assist patients in order to proceed walking. However, current objective evaluations for automated FoG detection are not sufficient as only the standard time-based and frequency-based features were extracted and there are still spaces for improvement in term of FoG detection performance. In this paper, we first attempt to adapt and extend current robust feature extraction and inference techniques in order to include additional features compared to the currently existing features. Then we go a step further by applying feature selection with the purpose of obtaining the maximum recognition results using the current available DAPHNet dataset. This dataset was collected using a wearable health assistive system that consists of 3-axes accelerometer to measure patient's movement. Ten PD patients were chosen to perform several walking tasks under laboratory environment. The overall performance was evaluated via subject-dependent and subject independent using the proposed feature extraction, feature selection and classification algorithms. The outcomes showed that the suggested machine learning methods had the ability in detecting FoG with maximum mean accuracy, sensitivity, specificity and area under curve (AUC) of approximately 99%.
