This study introduces a new EMG-based muscle fatigue index that combines EMG signal features in spatial, temporal, and spectral domains. This index incorporates a novel spatial EMG feature: Eigen Ratio Based Fatigue Index (ERFI) that can capture the variations in the active motor unit distribution during dynamic muscle fatiguing exercises. The new fatigue index maps the ERFI, a wavelet-based feature, a spectral feature, and an amplitude-based feature to the reduction in maximum voluntary contraction, which is considered the direct measure of muscle fatigue. The mapping function was implemented as a Multi-layer Perceptron (MLP). To evaluate the fatigue index, several fatigue tests under various speed and load conditions were conducted on four subjects. ERFI showed a significant variation (p < 0.01) over time for more than 85 % of the tests. Several MLP input configurations to predict muscle fatigue were compared in this study based on various combinations of EMG features. An input configuration that used the novel spatial EMG feature among five other features performed the best and was able to predict muscle fatigue with a mean coefficient of determination over 65 %. It was also noticed that ERFI’s relationship with muscle fatigue is less dependent on the load and speed of the cyclic exercise when compared with other EMG features that were proposed in previous studies. Thus it can be considered a better alternative to use in EMG-based control of active prosthetic and orthotic devices to compensate for the effect of muscle fatigue.