Multi-frame event dependent locomotion mode classification with FIRNNs
University of New Brunswick
Using electromyography (EMG) data obtained from four transfemoral amputees, Finite Impulse Response Neural Networks (FIRNNs) were investigated for the task of locomotion mode identification. Classification accuracy is found to improve as the duration of the observation interval, presented to the FIRNN, increases. Improvement in classifier accuracy is found to depend on the associated gait event; the more locomotion-mode transitions associated with a gait event, the higher the improvement in the classifier accuracy. Overall, the average classification accuracy on transitions improves by 15.9%. FIRNNs prove much more tolerant to increasing input dimensionality when compared with Linear Discriminant Analysis (LDA) classifiers. When Principal Component Analysis (PCA) is used to reduce input dimensionality, LDA performance is nearly equivalent to FIRNN performance without PCA. A confidence based rejection system is implemented from scaled FIRNN outputs and found to increase classification accuracy and average confidence for the nonrejected patterns. Increasing the observation interval also leads to improved confidence and reduced rejection ratios for fixed decision thresholds.