DECODING MOTOR ADAPTATION FROM HUMAN MAGNETOENCEPHALOGRAPHY

Motor adaptation is a critical aspect of human motor control but its neural mechanisms are not fully understood. While a large body of research explored putative computational mechanisms of motor adaptation, direct validation of these theories using human brain data is scarce. In this study, we investigated how the brain represents the process of adaptation to visuomotor rotations using magnetoencephalography (MEG) collected from healthy participants performing center-out reaching movements with a joystick. The task involved both stable perturbation periods (when perturbations were kept constant over many trials) as well as random perturbation periods.

Traditionally motor adaptation in experiments like that is modeled by a simple one-dimensional state space model where the latent variable represents a motor memory. There are many different extensions of this idea, in particular some of them offer predictions for the adaptation rate dynamics as well. We compared the predictive power of different computational models of motor adaptation: behavior-only state space models and machine learning approaches using behavioral and brain data together. We also decoded state space models’ variables from MEG signals.

We found that machine learning approaches that use brain and behavioral together improve decoding of behavioral variables. We also found that latent states of some state space models can be decoded from the MEG data.

Overall this approach (decoding features of interpretable behavioral models and comparing predictive power of different models) offers a new way to distinguish between different existing motor adaptation theories and formulate new ones.