Signiﬁcance: Movement intent decoding algorithms can interpret human bioelectrical signals to control prosthetic limbs with many degrees of freedom (DOFs). This work involves decoding volitional movement intent from surface electromyogram (sEMG) signals to control prosthetic arms. To train these algorithms, patients ﬂex their muscles to “follow” a movement prompt, and the machine learns to associate the resulting sEMG signals with the desired movements. In the homologous case in which the patient has functioning forearm muscles, they can follow the prompt using the same muscles originally used to control the corresponding DOFs of their own hand. However, in the more common heterologous case in which some or all of the forearm muscles cannot be used, a new, unintuitive muscle action-to-DoF mapping must be created. This thesis presents three approaches to tackle this additional motor learning challenge eﬃciently.
Objective: This thesis presents three action selection (AS) algorithms that were developed to address this map creation problem. Given a prosthesis with m DOFs and a set of n available muscle actions A, the AS algorithms search for the size-m subset (n/m) of A with the smallest expected decoding error among all m-element subsets of A. Evaluating every possible subset is likely intractable because the number of subsets is typically large. The three algorithms trade accuracy for simplicity to varying degrees.
Methods: Three AS algorithms are presented. The Greedy algorithm selects the m actions with the strongest sEMG signals. The Second-Order algorithm uses oﬄine decoder cross-validation to evaluate the individual performance of each action and the decoder confusion between pairs of actions. A cost is calculated for each size-m action subset based on this information and the subset with the lowest cost is chosen. Finally, the Action Screening algorithm uses cross-validation to remove poorly performing or indistinguishable actions from the action set, then evaluates every remaining size-m action subset.
The algorithms were validated on eight intact-arm human subjects. A small amount of sEMG data from 24 actions was collected for each subject. This data was input into the three algorithms, each of which output a size-5 action subset. Two reference subsets were used for comparison: MedianRef, the median subset as predicted by the Second-Order algorithm, and EasyRef, the homologous mapping consisting of ﬂexion of the ﬁve ﬁngers. For each of these ﬁve subsets (identity hidden and tested in random order), a larger training set was collected for the ﬁve actions in the subset. Finally, the subjects tested the trained decoder online.
Results: Test subjects were able to control the trained system to match prompted targets with signiﬁcantly lower RMSE using the action-to-DoF mappings chosen by the Action Screening and Second-Order algorithms relative to the MedianRef map-ping (p = 0.0017, p = 0.0497, respectively). The Action Screening algorithm also outperformed the Greedy approach (p = 0.0174). Unintended movements in the DoFs that were supposed to be at rest were signiﬁcantly lower for the Action Screen-ing algorithm’s chosen mapping than for the MedianRef mapping (p = 0.0193). The Action Screening and Second-Order algorithms’ performances were not signiﬁcantly diﬀerent from the homologous mapping EasyRef.
Conclusions: The Action Screening and Second-Order algorithms, but not the Greedy algorithm, choose action-to-DoF mappings that prosthesis users can control more accurately relative to the MedianRef mapping. Also, the selected heterolo-gous mappings did not perform signiﬁcantly diﬀerent from the homologous mapping EasyRef. This means that prosthesis users that require heterologous control may use these algorithms to overcome the additional motor learning challenge of learning a new action-to-DoF mapping. For all users, The reduction in decoding error provided by these algorithms will reduce frustration during training and usage. This will increase prosthesis utility and patient satisfaction and will lower rejection rates in clinical practice. The Action Screening algorithm performed the best, but the Second-Order algorithm may also be used if computational capability is limited. More data must be collected to justify the selection of either of these algorithms over the homologous mapping, where applicable.