Journal of NeuroEngineering and Rehabilitation

contraction of that muscle — as the user attempts to close their missing hand — generates EMG signals that close the motorized hand, conversely, reinnervation of a seg- ment of triceps by the transferred distal radial nerve generates EMG signals that control hand open. A custom-fabricated prosthesis was created for each subject using a Boston Digital Elbow (Liberating Technologies Inc.), a Wrist Rotator (Motion Control Inc.), and a single degree-of-freedom (DOF) terminal device of their choice (Table 1 ) . The prosthesis was capable of the following powered movements: elbow flexion (EF), elbow extension (EE), wrist pronation (WP), wrist supination (WS), terminal device open (TDO), terminal device close (TDC), in addition to no movement (NM). All subjects, except P9, were fit with custom-fabricated thermoplastic elastomeric gel liners (Alps Inc.). P9 was fit with a custom rolled silicone liner to minimize length of the prosthesis. Stainless steel electrodes were embedded into the wall of the liners, and EMG signals were transmitted, through stretchable conductive fabric leads, to the electronics at the distal end of the liner. A grid of electrodes was used, as described in previous work [ 15 ] , rather than placing electrodes over specific muscles. Briefly, signals were amp- lified and digitized using a Texas Instruments ADS1299 chip, sampled at 1000 Hz, and transmitted to an embed- ded controller. The pattern recognition algorithm, de- scribed in detail in Kuiken et al. [ 16 ] , interpreted the signals and sent appropriate commands to the prosthesis. Amplifier gains were set on a subject-specific basis, with a typical value of 2000, and data were digitally filtered be- tween 70 and 450 Hz. A recalibration switch was lami- nated into the outer wall of each socket so that the users could recalibrate the pattern recognition system, using prosthesis-guided training [ 16 ] , whenever they desired. An example of the EMG signal patterns collected during a representative recalibration sequence are shown in Fig. 1 . Seven of the nine subjects were naïve to pattern recogni- tion control. While their prosthesis was being constructed, an occupational therapist taught these subjects the con- cepts of pattern recognition and instructed them how to make repeatable and distinct muscle contractions [ 17 ] . During this initial training phase, subjects were given visualization exercises to strengthen their muscles, which has been shown to improve users ’ ability to make repeat- able and distinct contractions for pattern recognition con- trol [ 18 ] , but received no real-time control feedback. Once all subjects could perform consistent contractions for each intended prosthesis movement, EMG data from four repeti- tions of each prosthesis movement, held for 3 s, were col- lected used to train the pattern recognition algorithm. A series of images presented on a computer monitor were used to guide subjects through the data collection proced- ure. Immediately after data collection, subjects performed three blocks of the TAC Test [ 11 ] . The user was required to move a flesh-colored virtual limb to match the posture of a translucent grey-colored virtual target limb in real time, within a 15-s time frame, essentially a timeout ceiling that limited the length of the trials. Each block comprised a set of 12 postures. Target limb postures were selected such that the subject had to control each DOF of the prosthesis through 75% of its range of movement, stop within the pos- ture location (±5° of each DOF), and maintain the target posture location for 2 s. Outcome metrics for this virtual outcomes test included (i) the number of postures success- fully acquired within their allotted 15-s time frames, and (ii) the median completion time required to match the set of postures in a block. Median completion time was used ra- ther than the mean, as the data were skewed by the 15-s timeout ceiling. Immediately after completion of the three blocks, data from four repetitions of each movement, held for 3 s, were collected and used to evaluate the classifica- tion error rate of the pattern recognition system. Subjects were then fit with the physical prosthesis and received occupational therapy over 3 – 4 consecutive days for approximately 6 h per day. Subjects then performed a set of outcome measures (pre-home trial testing) that Table 1 Patient demographics Patient Age (years) Time since amputation (years) Time since TMR (years) Side Gender Etiology Terminal device P1 35 4 3 R M Trauma (military) Hook-ETD P2 45 2 1 R M Trauma (train) Hand P3 54 6 < 1 L M Trauma (military) Hook-ETD P4 58 5 1 L M Sarcoma Hook-ETD P5 25 6 6 L M Trauma Hook-ETD P6 31 8 7 L M Trauma (military) Hook-Greifer P7 27 2 1 R M Trauma (crushing) Hook-Greifer P8 31 1 1 R M Trauma (MVA) Hook-ETD P9 44 1 < 1 R F Trauma (MVA) Hand PGT Prosthesis Guided Training Hargrove et al. Journal of NeuroEngineering and Rehabilitation 2018, 15 (Suppl 1):60 Page 23 of 72