Decoding intention at sensorimotor timescales

Abstract

The ability to decode an individual's intentions in real time has long been a 'holy grail' of research on human volition. For example, a reliable method could be used to improve scientific study of voluntary action by allowing external probe stimuli to be delivered at different moments during development of intention and action. Several Brain Computer Interface applications have used motor imagery of repetitive actions to achieve this goal. These systems are relatively successful, but only if the intention is sustained over a period of several seconds; much longer than the timescales identified in psychophysiological studies for normal preparation for voluntary action. We have used a combination of sensorimotor rhythms and motor imagery training to decode intentions in a single-trial cued-response paradigm similar to those used in human and non-human primate motor control research. Decoding accuracy of over 0.83 was achieved with twelve participants. With this approach, we could decode intentions to move the left or right hand at sub-second timescales, both for instructed choices instructed by an external stimulus and for free choices generated intentionally by the participant. The implications for volition are considered.

Figure 1. The three protocols used.

(far left) Participant carried out real repetitive movements. (centre) Participant carried out repetitive motor imagery and received online feedback. (far right) Simple precueing task where the participant simply reacted as quickly as possible.

Figure 2. ROC AUC of model trained on real and imaginary movement data and tested on precue data.

Aligned to direction cue and using three different window widths (500 ms, 300 ms and 100 ms). The black line shows the decoding accuracy achieved when condition labels were randomly reshuffled.

Figure 3. ROC AUC of model trained on real and imaginary movement data and tested on precue data.

Aligned to Go cue and using three different window widths (500 ms, 300 ms and 100 ms). The black line shows the decoding accuracy achieved when condition labels were randomly reshuffled.

Figure 4. Comparison of decoding using EEG and EMG, when aligned to Go cue.

The black line shows the decoding accuracy achieved when condition labels were randomly reshuffled.

Figure 5. Comparison of decoding using EEG and EMG during imaginary movement task for the last three participants.

The black line shows the decoding accuracy achieved when condition labels were randomly reshuffled.

Figure 6. Comparison of decoding using EEG and EMG for the last three participants, when aligned to Go cue.

The black line shows the decoding accuracy achieved when condition labels were randomly reshuffled.

Figure 7. ROC AUC of Free and Instructed choices.

Aligned to Go cue using a 300 ms window. The black line shows the decoding accuracy achieved when condition labels were randomly reshuffled.

Figure 8. ROC AUC of Free and Instructed choices.

Aligned to direction cue using a 300 ms window. The black line shows the decoding accuracy achieved when condition labels were randomly reshuffled.

Figure 9. ROC AUC of Obey vs Disobey under the Free condition.

Aligned to Go cue using a 300 ms window. The black line shows the decoding accuracy achieved when condition labels were randomly reshuffled.