Fig 1.
Single subject movement data with analysis of perturbation adaptation and trial-by-trial adaptation rates.
(a) Trial-by-trial angular error data for a block of 80 consecutive movements of a human subject controlling a cursor with electromyographic signals. Red points indicate the trial and magnitude of catch-trial perturbations. Data from Subject ID# 2016–127 from Shehata et al. [25]. (b) Linear regression of Errorn+1vs. Errorn for trials in a. Red line indicates line of best fit whose slope (-0.32) is defined as the perturbation adaptation rate. (c) Linear regression of ΔError vs. Error for each consecutive sequence of unperturbed trials between perturbations. The trial-by-trial adaptation rate (-1.20) is calculated as the mean of all regression coefficients.
Fig 2.
Perturbation adaptation and trial-by-trial adaptation rates are different.
(a) Simulated data generated using a Bayesian learner model shows a significant difference between predicted perturbation adaptation rate and trial-by-trial adaptation rate (*** = p<0.001, two-tailed paired t-test). (b) Empirical results from 10 human subjects [25] show the same significant difference between adaptation rate analysis results (*** = p<0.001, two-tailed paired t-test).
Fig 3.
Single subject movement data and adaptation rate quantification.
(a) Trial-by-trial endpoint error data (one degree-of-freedom) for a block of 70 consecutive movements of a human subject controlling an on-screen cursor with a computer mouse. (b) Linear regression of ΔError vs. Error for initial trials in a. Red line indicates line of best fit whose slope (-0.69) is defined as the adaptation rate. ΔError = Errori+1-Errori. (c) Linear regression of ΔError vs. Error for steady-state trials in a. The regression coefficient equals -1.11.
Fig 4.
Adaptation rate depends on trial-set analyzed.
Adaptation rate calculated from the last 30 trials of a 70-trial sequence of movements was significantly higher than the result from the first 30 trials for simulated experiments (a) and empirical data (b) (** = p<0.01, *** = p<0.001, two-tailed paired t-test).
Fig 5.
Novel protocol to identify steady state trials for adaptation analysis.
(a) The expanding window protocol can identify the steady-state trials from a set of human movement data. (b) The start of the steady state trial set (which always ends at the last trial) is calculated as the point where the range of the 95% confidence interval of the offset of a zero-order robust linear regression is minimized. The minimum window size is 10 trials, thus the last potential starting trial plotted is trial number 60.
Fig 6.
Steady state trial-by-trial adaptation rates show improved consistency.
(a) As the initial gain estimate of the Bayesian learner model is varied, the resulting trial-by-trial adaptation rate calculated using the identified steady state trials remains consistent. (b) As the initial gain estimate is varied, the trial-by-trial adaptation rate calculated across all trials varies.
Fig 7.
Steady state trial-by-trial adaptation rates show improved estimation confidence.
The range of the 95% confidence interval of the regression coefficient representing the trial-by-trial adaptation rate is narrower when analyzing the identified steady-state trials compared to analysis of a constant number of trials (30) at the end of the experiment.