Fusion of multiple classifiers for motor unit potential sorting

Abstract

To achieve improved classification performance, a multiclassifier fusion approach for motor unit potential (MUP) sorting during electromyographic (EMG) signal decomposition was investigated. A classifier fusion system was developed that aggregates, at the abstract and measurement levels, the outputs of an ensemble of heterogeneous base classifiers to reach a collective decision, and then uses an adaptive feedback control system that detects and processes classification errors by using motor unit firing pattern consistency statistics. Three types of base classifiers were used: certainty, adaptive certainty, and adaptive fuzzy k-NN. Performance of the developed system was evaluated using real and synthetic simulated EMG signals with known properties and compared with the performance of the constituent base classifiers. Across the sets of EMG signal data sets studied, the classifier fusion schemes had better average classification performance, especially in terms of improving correct classification rates. Relative to the average performance of base classifiers and based on the difference between correct classification rate CC r and error rate E r , the adaptive average rule classifier fusion scheme shows on average: for the set of real signals an improvement of 9.2%; for the set of simulated signals of varying intensity an improvement of 6%; and for the set of simulated signals of varying amounts of shape and/or firing pattern variability an improvement of 7.7%.