Machine Learning for Analyzing Gait in Parkinson's Patients Using Wearable Force Sensors

Gait impairments are the prerequisite for the diagnosis of Parkinson's disease (PD). The sole purpose of this study is to objectively and automatically classify between healthy subjects and Parkinson patients. In this research total, 16 different positioned force sensors were attached to the shoes of subjects that recorded the Multisignal Vertical Ground Reaction Force (VGRF). From all sensors signals using 1024 window size over the raw signals, using the Packet wavelet transform (PWT) five different features namely entropy, energy, variance, standard deviation and waveform length were extracted and support vector machine (SVM) is applied to distinguish between Parkinson patients and healthy subjects. SVM is trained on 85% of the dataset and validated on 15% dataset. The training cohort depends on 93 patients with idiopathic PD (mean age: 66.3 years; 63% men and 37% women), and 73 healthy controls (mean age: 66.3 years; 55% men and 45% women). Among 16 sensors, 8 force sensors were attached to the left foot of subject and the remaining 8 on the right foot. The results show that 5th sensor worn on a Medial aspect of the dorsum of right foot represented by R5 gives 90.3% accuracy. Hence this research gives the insight to use only single wearable force sensor. Therefore, this study concludes that a single sensor may serve for identification between Parkinson patient and healthy subject.