Journal of NeuroEngineering and Rehabilitation

pattern recognition – controlled prosthesis [ 14 , 18 , 23 ] . Our data support this idea, as classification error rates improved significantly ( p < 0.05) after the home trial. However, classification error rate had a strong, statisti- cally significant correlation with the Clothespin Re- location task, but did not correlate significantly with any of the other physical outcome measures with any TAC outcome metric. In contrast, TAC test comple- tion time correlated strongly and significantly with all physical outcome measures except the Clothespin Re- location task. These results reinforce the growing body of literature supporting the importance of performing online testing, preferably with individuals with ampu- tations, rather than relying solely on classification error analyses during offline experiments to evaluate control [ 10 , 24 – 26 ] . This work is important because it demonstrates a correlation between virtual test measures and physical performance. However, the study has several limita- tions. As with any correlation analysis, correlation does not imply causation. Without further study, we cannot say that working within a virtual environment will transfer to better functional outcomes, although the work of van Dijk et al. suggests that this may be true under certain situations [ 13 ] . Furthermore, developing validated and reliable outcome measures in the field of upper-limb prosthetics is a challenging problem. As for physical outcome measures, virtual measures must be thoroughly described to ensure con- sistent administration, analysis, and interpretation. For example, the ACMC has an established test-retest, inter-rater, and intra-rater reliability and clinical inter- pretation guidelines. The TAC test may be made easier or more difficult by changing the length of time allowed to acquire postures, the number of DOFs needed to acquire the posture, the distance the virtual limb must be moved, and the tolerances required for matching the target posture. Finally, the results of this study are only applicable to transhumeral amputees. Future investigations will need to be performed to de- velop appropriate relationships between virtual and physical outcome measures for individuals with other levels of upper limb amputation. Conclusions Providing users with an opportunity to use a pattern recognition – controlled prostheses in their home-en- vironment for at least 6 weeks resulted in improved functional control, as measured by a set of outcome measures. This highlights the need for practice, in addition to comprehensive occupational therapy, before assessing outcomes. Improvements were seen in both the offline performance metric of classification error rate and in real time control outcome measures when controlling a virtual or physical prosthesis. Finally, we found that some outcome measures, particularly the TAC test completion time, correlated strongly with physical outcome measures. Future work will further investigate these relationships using a more standardized test configuration, and with a broader population of subjects. Fig. 3 Statistically significant relationships between virtual and physical outcome measures. Each relationship was strong, with a Pearson correlation coefficient |R| > 0.75 Table 3 Pearson correlation coefficients, R, between virtual and physical outcome measures Predictor SHAP CRT Box and blocks Jebsen-Taylor ACMC Completion Time − 0.86* 0.30 − 0.82* 0.87* − 0.85* Failure Rate 0.19 0.52 − 0.54 0.27 − 0.37 Classification Error − 0.13 0.76* − 0.46 0.39 − 0.31 *Denote statistical significance at the p = 0.05 level Hargrove et al. Journal of NeuroEngineering and Rehabilitation 2018, 15 (Suppl 1):60 Page 26 of 72