Estimating visual corrective reaction times in double-step paradigms using machine learning: A proof of principle


The double-step paradigm has long been the gold-standard for estimating visual corrective reaction times (CRT) in motor control. Accordingly, many statistical approaches have been developed and are widely utilized. These traditional approaches usually either determine when spatio-temporally speaking, trajectories are statistically significantly different from one another, or use extrapolated linear fits to estimate the CRT. Here, I instead propose using a classifier based in machine learning. From a machine learning perspective, the question changes to: "When is there a sufficient pattern in the data to afford better classification of the individual trajectories into their associated conditions." Although the trajectories associated with movements completed towards single-step and double-step stimuli are characteristically different overall, it was hypothesized that the classification accuracy would quickly increase when including data beyond the CRT. To test this hypothesis, a publicly available data-set including 10 participants was re-processed and re-analyzed ( Reaches made towards single- and double-step stimuli were completed using a mouse, a stylus, or in an unconstrained 3D environment. A Random-Forest classifier was used to predict the condition (i.e., single- vs. double-step) using a 20% hold-out cross-validation sample within each of the three tasks. Shorter, and more reliable estimates of CRT were observed in the axis of the primary perturbation. Notably, these CRT estimates were comparable, and sometimes nominally shorter than the estimates reported in the original manuscript, which utilized a common, sensitive metric. Thus, a machine-learning based classification approach may provide an additional avenue to obtain estimates of time-sensitive measures of motor behaviour and should be investigated further.