Motor-evoked potentials reveal a motor-cortical readout of evidence accumulation for sensorimotor decisions

Many everyday activities, such as driving and sports, require us to engage in time-pressured sensorimotor decision making in response to visual cues. The computational principle of continuous evidence accumulation is the dominant account underlying models of speeded decision making, but the nature and locus of the decision variable that triggers action is debated. Traditionally, cognitive stages such as perception, stimulus-response translation, and the generation of motor plans, have been considered to occur in series. However, this idea is challenged by neurophysiological work in animals, suggesting that cognitive operations are distributed across sensorimotor cortex. Here, we investigate whether a decision variable can be observed in the primary motor cortex (M1) of humans. Participants categorised faces as male or female, with task difficulty manipulated using natural or morphed stimuli. Transcranial magnetic stimulation, applied at random across the reaction-time interval, produced motor-evoked potentials (MEPs) in two hand muscles that were the major contributors when generating the correct and incorrect pinch/grip movements. MEP magnitudes reveal covert action preparation, even when no action is produced. Smoothing MEPs using a Gaussian kernel allowed us to recover a continuous time-varying MEP average, comparable to an EEG component, which permitted precise localisation of the time at which the motor plan for the responding muscle began to dominate over the non-responding action. This moment was calculated in both stimulus-locked and response-locked analyses, and was found to occur at the same time with stimulus locking, but earlier with response locking, when ambiguous stimuli made the decision more challenging. This pattern is consistent with M1 providing a continuous readout of evidence accumulation. We predicted the evidence accumulation profile from a drift diffusion model, using only behavioural data, and found a good qualitative match to the observed neurometric MEP profiles.