Combining Neural Classifiers in EMG Diagnosis

In the case of difficult pattern recognition problems, the combination of the outputs of multiple classifiers using for input multiple feature sets, can improve the overall classification performance. In this work a modular neural network decision support system was developed for the assessment of the electromyographic (EMG) signal recorded from normal subjects and subjects suffering with myopathy and motor neuron disease. Six different feature sets were computed from the motor unit action potentials (MUAPs) composing the EMG signal, as follows: (i) time domain parameters, (ii) frequency domain parameters, (iii) cepstral coefficients and (iv) three different wavelet coefficients (Daubechies, Chui, and Battle-Lemarie). The multiple feature sets were inputted into multiple self-organising feature map (SOFM) classifiers and the classification results were combined using majority voting. Furthermore, a confidence measure was computed which weighted the contribution of each feature set to the final diagnostic yield. The average diagnostic yield for the individual feature sets was 69.1% whereas when combining the six classification results was 76.9%. The use of the confidence measure enhanced further the diagnostic yield to 79.6%