This study investigated the characteristics of postural sway in cases of Parkinson's disease (PK). We previously reported sway patterns, its measurements, and factor analysis. This paper investigates learning and evaluation by neural networks.
Each subject stood on a stabilometer with their eyes open or closed, and the sway of the center of gravity was recorded for one minute. The learning and evaluation was performed using a neural network program developed by Anima corporation. Input data were 6 units (measurements values related to area and length) in analysis A and 20 units (measurements values related to area and length, power spectrum, vector of velocity, amplitude probability density distribution curve) in analysis B. Output consisted of 3 units (positive, negative a, and negative b symptoms) in both analyses. There were 16 subjects, of whom 13 were used for learning and 3 for evaluation. Learning was carried out until the number of learning trials reached 10000, or the square error margin decreased to 0.005.
Results
1) In analysis A, learning revealed a square error margin of 0.005 after 690 trials, while the evaluations revealed a square error margin of 0.263.
2) In analysis B, learning revealed a square error margin of 0.005 after 215 trials, while the evaluations revealed a square error margin of 0.267.
Three symptoms were classified clearly as results of learning in analysis A and B. The number of learning trials in analysis B was less than that in analysis A, indicating that increased input resulted in fewer learning trials. The square error margin for evaluation was not good in either analysis.
