Objective. Powered lower-limb prostheses relying on decoding motor intentions from non-invasive sensors, like electromyographic (EMG) signals, can significantly improve the quality of life of amputee subjects. However, the optimal combination of high decoding performance and minimal set-up burden is yet to be determined. Here we propose an efficient decoding approach obtaining high decoding performance by observing only a fraction of the gait duration with a limited number of recording sites. Approach. Thirteen transfemoral amputee subjects performed five motor tasks while recording EMG signals from four muscles and inertial signals from the prosthesis. A support-vector-machine-based algorithm decoded the gait modality selected by the patient from a finite set. We investigated the trade-off between the robustness of the classifier's accuracy and the minimization of (i) the duration of the observation window, (ii) the number of EMG recording sites, (iii) the computational load of the procedure, measured the complexity of the algorithm. Main results. When including pre-foot-strike data in the decoding, the combination of three EMG recording sites and the inertial signals led to correct rates above 94% at the 20% of the gait cycle, showing the best trade-off between invasiveness of the setup and accuracy of the classifier. The complexity of the algorithm proved to be significantly higher when applying a polynomial kernel compared to a linear one, while the correct rate of the classifier generally showed no differences between the two approaches. The proposed algorithm led to high performance with a minimal EMG set-up and using only a fraction of the gait duration. Significance. These results pave the way for efficient control of powered lower-limb prostheses with minimal set-up burden and a rapid classification output.