Offline experiment

An offline experiment was performed to determine the best possible method to train the control system for performing hand-grasp selection and to familiarize subjects with pattern recognition prosthesis control prior to the real-time experiment. Data were collected using the protocol from a previous study (Earley et al., 2014), and data from this previous study were combined with these new data to perform statistical analysis on interaction and simple main effects, which influenced the development of the real-time algorithms detailed in the next section. EMG data were collected under 10 conditions. Subjects first held the wrist in a neutral position and performed the eight hand postures (six grasps, hand open, and no movement) for 4 s each. These postures were then performed with the wrist statically held in flexion, extension, radial deviation, ulnar deviation, pronation, and supination. Each hand posture was performed 4 times in each wrist position. Data collected during these seven static wrist conditions were combined into the variable wrist position data set (see Figure ​Figure4A).4A). Subjects were then asked to perform hand postures four times while moving the wrist from flexion to extension, before returning to flexion, and four times while moving the wrist from extension to flexion, and back to extension. Wrist movements spanned 2 s in each direction, for a total of 4 s. This procedure was repeated for wrist radial and ulnar deviation, and pronation and supination; data collected during these three dynamic conditions were combined into the dynamic wrist movement data set (see Figure ​Figure4B).4B). Analyses were performed using two-fold cross-validation. Classification performance was quantified by classification error, which is the percentage of incorrect classifications.

Classifier training protocols. (A) Variable wrist position protocol. Grasp is initiated and maintained for 4 s in the desired wrist position. (B) Dynamic wrist motion protocol. Grasp is initiated with the wrist in the starting position. With the grasp maintained for the entire duration, the wrist moves to the opposite position for 2 s before moving back to the starting position for 2 s. (C) Hybrid wrist motion protocol. Grasp is initiated and maintained in the starting position. Wrist held in the starting position for 2 s, moves to the opposite position for 2 s, held in the opposite position for 2 s, and moves back to the starting position for 2 s.

Two statistical analyses were performed, both with subjects as a random variable: an ANOVA with electrode placement and training method as factors, and an ANOVA with window length and available grasps as factors; main effect and interaction effect terms were included in each model. If the interaction was not found to be significant, the analysis was rerun with a reduced model consisting only of main effects. When interaction was found to be significant, a subsequent analysis was performed to determine the simple main effects. Pairwise comparisons were made using Bonferroni correction factors, and significance levels were set at α = 0.05. Analyses were split into two categories: grasp selection and grasp maintenance performance. Grasp selection was only tested on various wrist positions data—from an application perspective, this equates to prepositioning the wrist prior to making a hand-grasp selection. Grasp maintenance was tested on both various wrist positions and dynamic wrist motions data—this equates to the user being able to move their wrist freely, after grasp selection, as they maintain the selected grasp.

EMG training data were obtained from (1) extrinsic muscles, (2) intrinsic muscles, or (3) extrinsic and intrinsic muscles. Classifiers were trained with one of four sets of training data: (1) the wrist only in the neutral position, (2) the wrist in seven variable wrist positions, (3) the wrist moving in each of the three degrees of freedom (dynamic wrist movement), and (4) all static and dynamic wrist data. All possible combinations of electrode position and wrist position yielded a total of 12 conditions. Grasp selection analyses were performed with a 500 ms EMG feature extraction window (Smith et al., 2011), trained on all data in one cross-validation fold, and tested against the first 600 ms of each static trial included in the other fold, effectively evaluating the performance of the classifier on only transient EMG generated by the onset of grasp selection (Hudgins et al., 1993). Grasp maintenance analyses were performed with a 200 ms EMG feature extraction window, and tested against all but the first 300 ms of each included data collection.

Analyses of window length and available grasps were performed on tests run with the classifiers trained with both dynamic and static wrist movements and with both extrinsic and intrinsic EMG channels. Pattern recognition classifier performance was evaluated for 100, 200, 300, 400, and 500 ms feature extraction window lengths; and for N (2, 4, or 6) grasps available to the classifier, where grasps were the N most useful hand-grasps for performing ADLs (see Figure ​Figure3).3). In order to ensure adequate data to train a classifier for grasp selection analyses with feature extraction window length L, the first L+100 ms of each data collection were used, capturing the transient EMG from each grasp initiation. The remaining durations of data collections were used for grasp maintenance analyses for all window lengths.