A ground reaction force artificial neural network classifier for the diagnosis of parkinson’s disease

Abstract

Parkinson's disease (PD) is a common neurodegenerative disease with symptoms of bradykinesia, rest tremor, rigidity, and postural instability. Currently there is no definitive diagnosis of PD. The disease is diagnosed by a clinician who qualitatively evaluates a patient’s visible symptoms during a physical exam. Post-mortem histology has shown that the accuracy of clinical diagnoses can be low, ranging from 74% to 90%. We have developed an artificial neural network (ANN) which classifies subjects as healthy or PD based on vertical GRF features. Data from a total of 40 PD subjects and 40 healthy controls (COs) was gathered from two previously published studies via a public online database. Eight vertical GRF features were measured and used as the input into the ANN. The average PD subject’s vertical GRF was found to having less high frequency power, smaller first and second peak amplitudes, and a delayed occurrence of the first peak. Detrended fluctuation analysis (DFA) determined the PD subjects had on average more long term correlation in their swing time intervals as measured over 70 strides. The ANN successfully diagnosed 10 out of 10 PD patients (sensitivity of 100%) and 9 out of 10 healthy COs (specificity of 90.0%).