Journal of NeuroEngineering and Rehabilitation

Classification error rates are frequently reported to characterize the performance of upper-limb pattern recognition control systems. Pre-trial classification error and TAC test data were available from six of the nine subjects: two subjects had prior experience con- trolling the virtual prosthesis so they did not complete pre-home trial virtual environment testing, and data from one subject was lost due to a computer malfunc- tion. Post-trial data were available from all nine subjects. After the home trial, average classification error across subjects dropped from 13.4 to 8.3%, which was significantly lower ( p = 0.03) (Fig. 2 ) . All TAC test per- formance metrics also improved significantly after the home trial: failure rate improved from 19.9 to 3.7% to Y ( p = 0.001), and completion time dropped from 7.5 to 5.5 s ( p = 0.007). All nine subjects completed the outcome measures using the physical prosthesis. All outcome measures for use of the physical prosthesis tended to improve after the home trial; however, only the SHAP ( p = 0.001) and the Blocks and Box test ( p = 0.03) showed statistically significant improvements. We performed Pearson correlation analyses to inves- tigate the relationship between TAC test outcome measures and physical outcome measures (Table 3 ) . We found strong and statistically significant correla- tions between TAC completion times and several of the physical outcome measures for the post home-trial outcomes (Fig. 3 ) . Correlations between TAC comple- tion time and the SHAP, Box and Blocks test, and ACMC were negative, i.e., faster completion times were associated higher test scores, indicating better performance in these physical measures. The correl- ation between completion time and the Jebsen-Taylor test was positive, i.e., faster completion times were as- sociated with faster test times, indicating better phys- ical task performance. We also found that the classification error rate, which is an offline measure of performance, did not show statistically significant cor- relations with any virtual measure, but did correlate strongly with performance in the Clothespin Relocation task ( p = 0.018). Discussion Pattern recognition – based myoelectric control systems have seldom been systematically evaluated outside of a controlled laboratory environment. Most studies are performed to evaluate control algorithm classification error rates, or to evaluate ways to make classification error rates more robust to environmental factors, pro- longed use, or non-ideal conditions [ 18 – 20 ] . However, initial clinical case series evaluating commercially avail- able pattern recognition systems have reported positive patient experiences [ 21 ] . Our results show that users are capable of using pattern recognition – controlled myoelectric limbs within their home environment. Al- though the amount of time the prosthesis was worn and the frequency with which the control system was recalibrated was variable, the results of outcome tests using the pattern recognition – controlled prosthesis were equivalent or superior to measures recorded after subjects used the same prosthesis during an equivalent home trial, using conventional control [ 22 ] . We observed statistically significant improvements in the SHAP and Blocks and Box test after a 6-week home trial. The Clothespin Relocation task and the Jebsen-Taylor test also showed trends toward improve- ment that were not statistically significant. We also found statistically significant improvements in classifi- cation error rate and all outcome metrics associated with the virtual TAC test after the trial. Clearly, sub- jects learned to control the device better during the home trial. Limited published data has supported the hypothesis that patients learn to perform more consistent, distinct contractions over time with practice, leading to specula- tion that this would lead to improved control of a Fig. 2 Outcome measures when using a virtual prosthesis (left) or a physical prosthesis (right). Measures were performed before and after a 6-week home trial. *Denotes statistical significance at p = 0.05 Hargrove et al. Journal of NeuroEngineering and Rehabilitation 2018, 15 (Suppl 1):60 Page 25 of 72