Promise of embedded system with GPU in artificial leg control: Enabling time-frequency feature extraction from electromyography

Applying electromyographic (EMG) signal pattern recognition to artificial leg control is challenging because leg EMGs are non-stationary. Time-frequency features are suitable for representing non-stationary signals; however, the computational complexity to extract time-frequency features is too high and current embedded systems used for artificial limb control are inadequate for real-time computing. The aim of this study was to quantify the computational speed of a novel embedded system, the Graphic Processor Unit (GPU), on EMG time-frequency feature extraction. The computational time derived from a GPU was compared to that derived from a general purpose CPU. The results indicated that the GPU significantly increased the computational speed. When the size of EMG analysis window was set to 100 ms, the GPU extracted EMG time-frequency features over 50 times faster than the CPU setting. Therefore, high performance GPU shows a great promise for EMG-controlled artificial legs and other medical applications that need high-speed and real-time computation.