NNERVE: neural network extraction of repetitive vectors for electromyography--Part II: Performance analysis

Abstract

We have presented a new method for the decomposition of clinical electromyographic signals, NNERVE, which utilizes a novel "pseudo-unsupervised" neural network approach to signal decomposition. In this paper, we present a detailed performance analysis. We present definitions for quantitative performance criteria. NNERVE is shown to be highly reliable over a wide range of neural network architectures. It is also minimally sensitive to learning parameters. The degradations of performance over a wide range of signals and parameters are shown to be gradual, slight and graceful. These characteristics are shown to translate directly into a high degree of robustness over widely varying signals. Real signals obtained from the entire range of patients encountered in clinical situations are shown to be correctly handled without any modifications or adjustments of any parameters. This neural network method is then directly compared to a prior traditional signal processing method and is shown quantitatively to have consistently superior performance on both simulated and real signals. Clinically acceptable performance over a wide range of signals, recorded using standard clinical methodology, and the lack of a need for user interaction, will facilitate the use of motor unit quantitation in routine clinical electromyography.