Algorithm Development

Some of the most common control schemes for prosthetic hands include: Direct Control, State Machine, Pattern Recognition, and Postural Control.​

Direct Control is the paradigm where one EMG channel corresponds to one degree of motion.¹ For example, an electrode placed on the extensor muscles may correspond to an open motion while an electrode placed on the flexor muscles may correspond to a closed motion.

A State Machine expands upon the direct control paradigm and often uses a co-contraction switch to go between different grip patterns.¹ However, it is sequential in nature and a person cannot jump between grasp patterns that are out of sequence freely. This may make smooth and efficient operation tougher than other types of control algorithms.

In Pattern Recognition systems, variations of machine learning algorithms are used to train the system to recognize what different grasp patterns look like in the multi EMG domain each time they are performed. The math behind each of these algorithms, such as Linear Discriminant Analysis, Linear Regression, Fuzzy C-Means, Decision Trees, Neural Networks, etc. is slightly different; however, all of these algorithms look for features to identify unique characteristics for each grasp pattern.

Linear Regression maps and interpolates any position between open and close to 2 EMG channels. One channel will typically be the flexion and the other will be the extension. Postural Control expands upon Linear Regression by mapping to 2 axes instead of just one. This gives the user more access to switching between pre-programmed gestures based on combinations of EMG signals instead of having to cycle through all postures to get to the desired one.

Figure 1^2,3: Linear Regression (b) systems use antagonistic muscles  (left side of image) to operate to each degree of freedom. For example, wrist flexion and extension would operate hand open and close, while wrist abduction and adduction would operate wrist pronation and supination. In Pattern Recognition (a) algorithms, when the user recreates the posture with their phantom limb, the algorithm decodes these signals, classifies the posture, and actuates the prosthetic hand accordingly.

However, with all algorithms, it is necessary to preprocess the signals so that less noise is mixed in with the signal of interest. Otherwise differentiating between two similar grasping patterns becomes impossible and the classification accuracy crashes.​