Pattern Classification Of Force Signals Underneath The Feet Based On BP Neural Network

Parkinson's disease (PD) is a hackneyed and hazardous neural functional disorder disease, people often suffer from it after 50.Withing the elderly population growing, and there is a trend that the incidence, morbidity and the deformity rate are growing. But so far, the pathogeny of PD has not yet been know, the diagnosis of PD mainly depends on clinical experience and finally confirmed by autopsy. The rate of misdiagnosis is very high. Gait of PD is a research focus. However, few studies explored the gait features from the force signals underneath the feet. After extracting feature parameters from the foot-pressure signals in PD patients and control subjects, we use T test to find the identified differences as the inputs of the pattern classifier, which choosing the widely used BP neural network.The data from three research groups:Ga,Ju,and Si were analyzed. The database include 7,9,9 control subjects and 13,11,13PD patients respectively. In order to analyze the impact of Parkinson's disease on the gait of the patient quantitively, eight force sensors were mounted underneath each foot to measure the vertical ground reaction force as a function of time, then five feature points within each stride were localized from the left feet force record. Not only some timing measures including the stride time and each slice of the stride time and also its percentage over the whole stride time were computed, but also the force changing rate over each time slice was calculated. After T test and compared with their countparts, PD patients need more support time(t1+t2+t3+t4) from normal subjects, but their force changing rates sl s2, s3 over some time slices t1,t2,t3 are smaller and step length is shorter in the stride time. Using the five characteristic parameters as the inputs of the three-layer BP neural network, which having 8 nodes in hidden layer. There are 10 training samples and 10 testing samples. Finally, the testing error is acceptable, and the design meets the initial requirements. However, due to BP neural network has its own limitations, and the sets of the samples are too concentrated, this research should be studied more deeply.