Adaptive biasing of action-selective cortical build-up activity by choice history

Reviewer #1 (Public Review):

This paper aims to study the effects of choice history on action-selective beta band signals in human MEG data during a sensory evidence accumulation task. It does so by placing participants in three different stochastic environments, where the outcome of each trial is either random, likely to repeat, or likely to alternate across trials. The authors provide good behavioural evidence that subjects have learnt these statistics (even though they are not explicitly told about them) and that they influence their decision-making, especially on the most difficult trials (low motion coherence). They then show that the primary effect of choice history on lateralised beta-band activity, which is well-established to be linked to evidence accumulation processes in decision-making, is on the slope of evidence accumulation rather than on the baseline level of lateralised beta.

The strengths of the paper are that it is: (i) very well analysed, with compelling evidence in support of its primary conclusions; (ii) a well-designed study, allowing the authors to investigate the effects of choice history in different stochastic environments.

There are no major weaknesses to the study. On the other hand, investigating the effects of choice/outcome history on evidence integration is a fairly well-established problem in the field. As such, I think that this provides a valuable contribution to the field, rather than being a landmark study that will transform our understanding of the problem.

The authors have achieved their primary aims and I think that the results support their main conclusions. One outstanding question in the analysis is the extent to which the source-reconstructed patches in Figure 2 are truly independent of one another (as often there is 'leakage' from one source location into another, and many of the different ROIs have quite similar overall patterns of synchronisation/desynchronisation.). A possible way to investigate this further would be to explore the correlation structure of the LCMV beamformer weights for these different patches, to ask how similar/dissimilar the spatial filters are for the different reconstructed patches.

Reviewer #2 (Public Review):

In this work, the authors use computational modeling and human neurophysiology (MEG) to uncover behavioral and neural signatures of choice history biases during sequential perceptual decision-making. In line with previous work, they see neural signatures reflecting choice planning during perceptual evidence accumulation in motor-related regions, and further show that the rate of accumulation responds to structured, predictable environments suggesting that statistical learning of environment structure in decision-making can adaptively bias the rate of perceptual evidence accumulation via neural signatures of action planning. The data and evidence show subtle but clear effects, and are consistent with a large body of work on decision-making and action planning.

Overall, the authors achieved what they set out to do in this nice study, and the results, while somewhat subtle in places, support the main conclusions. This work will have impact within the fields of decision-making and motor planning, linking statistical learning of structured sequential effects in sense data to evidence accumulation and action planning.

Strengths:
- The study is elegantly designed, and the methods are clear and generally state-of-the-art
- The background leading up to the study is well described, and the study itself conjoins two bodies of work - the dynamics of action-planning processes during perceptual evidence accumulation, and the statistical learning of sequential structure in incoming sense data
- Careful analyses effectively deal with potential confounds (e.g., baseline beta biases)

Weaknesses:
- Much of the study is primarily a verification of what was expected based on previous behavioral work, with the main difference (if I'm not mistaken) being that subjects learn actual latent structure rather than expressing sequential biases in uniform random environments. Whether this difference - between learning true structure or superstitiously applying it when it's not there - is significant at the behavioral or neural level is unclear. Did the authors have a hypothesis about this distinction? If the distinction is not relevant, is the main contribution here the neural effect?
- The key effects (Figure 4) are among the more statistically on-the-cusp effects in the paper, and the Alternating group in 4C did not reliably go in the expected direction. This is not a huge problem per se, but does make the key result seem less reliable given the clear reliability of the behavioral results
- The treatment of "awareness" of task structure in the study (via informal interviews in only a sub-sample of subjects) is wanting

Reviewer #3 (Public Review):

This study examines how the correlation structure of a perceptual decision making task influences history biases in responding. By manipulating whether stimuli were more likely to be repetitive or alternating, they found evidence from both behavior and a neural signal of decision formation that history biases are flexibly adapted to the environment. On the whole, these findings are supported across an impressive range of detailed behavioral and neural analyses. The methods and data from this study will likely be of interest to cognitive neuroscience and psychology researchers. The results provide new insights into the mechanisms of perceptual decision making.

The behavioral analyses are thorough and convincing, supported by a large number of experimental trials (~600 in each of 3 environmental contexts) in 38 participants. The psychometric curves provide clear evidence of adaptive history biases. The paper then goes on to model the effect of history biases at the single trial level, using an elegant cross-validation approach to perform model selection and fitting. The results support the idea that, with trial-by-trial accuracy feedback, the participants adjusted their history biases due to the previous stimulus category, depending on the task structure in a way that contributed to performance.

The paper then examines MEG signatures of decision formation, to try to identify neural signatures of these adaptive biases. Looking specifically at motor beta lateralization, they found no evidence that starting-level bias due to the previous trial differed depending on the task context. This suggests that the adaptive bias unfolds in the dynamic part of the decision process, rather than reflecting a starting level bias. The paper goes on to look at lateralization relative to the chosen hand as a proxy for a decision variable (DV), whose slope is shown to be influenced by these adaptive biases.

This analysis of the buildup of action-selective motor cortical activity would be easier to interpret if its connection with the DV was more explicitly stated. The motor beta is lateralized relative to the chosen hand, as opposed to the correct response which might often be the case. It is therefore not obvious how the DV behaves in correct and error trials, which are combined together here for many of the analyses.