A Freezing of Gait Detection System Based on Stepwise Feature Extraction Method

As one of the most common gaits disorders in Parkinson's motor symptoms, freezing of gait FoG severely interferes with patients' daily activities and causes them to fall, which has significant clinical and social consequences. What's worse, there is neither clear pathogenesis nor effective therapeutic regiment at present, making it urgent to develop effective adjuvant treatment, intervention methods to alleviate motor symptoms and improve mobility of patients. As the basis for the implementation of intervention measures, the detection of FoG can also provide symptom information for subsequent disease assessment. Therefore, the detection of FoG has important research significance and application value. This paper proposes a FoG detection system based on motion signals collected by wearable devices and machine learning techniques. A stepwise feature extraction method is proposed to detect FoG by first separating acceleration signal into body acceleration, gravity and FoG acceleration, then calculating the jerk and Euclidean norm of each component, finally extracting a series of temporal and frequency domain features that can reflect the characteristics of FoG and using them to train the classifier. The work characterizes the performance system on user-dependent and user-independent experiment. Furthermore, different preprocessing window size, feature selection methods and classification algorithms are also evaluated. The best result obtained belongs to Light Gradient Boosting Machine LightGBM with 256 sample window size and the feature selection based on mutual information method. The best result achieved an accuracy of 88.62%, a sensitivity of 87.69%, a specificity of 89.29% and an AUC of 88.49%, respectively, enhancing the results of former methods and providing a more balanced rate of sensitivity and specificity.