Deep Learning with Convolutional Neural Networks Applied to Electromyography Data: A Resource for the Classification of Movements for Prosthetic Hands

This paper represents an ongoing investigation of dexterous and natural control of upper extremity prostheses using the myoelectric signal (MES). The scheme described within uses pattern recognition to process four channels of MES, with the task of discriminating multiple classes of limb movement. The method does not require segmentation of the MES data, allowing a continuous stream of class decisions to be delivered to a prosthetic device. It is shown in this paper that, by exploiting the processing power inherent in current computing systems, substantial gains in classifier accuracy and response time are possible. Other important characteristics for prosthetic control systems are met as well. Due to the fact that the classifier learns the muscle activation patterns for each desired class for each individual, a natural control actuation results. The continuous decision stream allows complex sequences of manipulation involving multiple joints to be performed without interruption. Finally, minimal storage capacity is required, which is an important factor in embedded control systems.

Improving the function of prosthetic arms remains a challenge, because access to the neural-control information for the arm is lost during amputation. A surgical technique called targeted muscle reinnervation (TMR) transfers residual arm nerves to alternative muscle sites. After reinnervation, these target muscles produce electromyogram (EMG) signals on the surface of the skin that can be measured and used to control prosthetic arms. To assess the performance of patients with upper-limb amputation who had undergone TMR surgery, using a pattern-recognition algorithm to decode EMG signals and control prosthetic-arm motions. Study conducted between January 2007 and January 2008 at the Rehabilitation Institute of Chicago among 5 patients with shoulder-disarticulation or transhumeral amputations who underwent TMR surgery between February 2002 and October 2006 and 5 control participants without amputation. Surface EMG signals were recorded from all participants and decoded using a pattern-recognition algorithm. The decoding program controlled the movement of a virtual prosthetic arm. All participants were instructed to perform various arm movements, and their abilities to control the virtual prosthetic arm were measured. In addition, TMR patients used the same control system to operate advanced arm prosthesis prototypes. Performance metrics measured during virtual arm movements included motion selection time, motion completion time, and motion completion ("success") rate. The TMR patients were able to repeatedly perform 10 different elbow, wrist, and hand motions with the virtual prosthetic arm. For these patients, the mean motion selection and motion completion times for elbow and wrist movements were 0.22 seconds (SD, 0.06) and 1.29 seconds (SD, 0.15), respectively. These times were 0.06 seconds and 0.21 seconds longer than the mean times for control participants. For TMR patients, the mean motion selection and motion completion times for hand-grasp patterns were 0.38 seconds (SD, 0.12) and 1.54 seconds (SD, 0.27), respectively. These patients successfully completed a mean of 96.3% (SD, 3.8) of elbow and wrist movements and 86.9% (SD, 13.9) of hand movements within 5 seconds, compared with 100% (SD, 0) and 96.7% (SD, 4.7) completed by controls. Three of the patients were able to demonstrate the use of this control system in advanced prostheses, including motorized shoulders, elbows, wrists, and hands. These results suggest that reinnervated muscles can produce sufficient EMG information for real-time control of advanced artificial arms.

We describe the approach that won the final phase of the German traffic sign recognition benchmark. Our method is the only one that achieved a better-than-human recognition rate of 99.46%. We use a fast, fully parameterizable GPU implementation of a Deep Neural Network (DNN) that does not require careful design of pre-wired feature extractors, which are rather learned in a supervised way. Combining various DNNs trained on differently preprocessed data into a Multi-Column DNN (MCDNN) further boosts recognition performance, making the system insensitive also to variations in contrast and illumination.