Citation

Karimi-Rouzbahani H, Woolgar A, Rich AN. Elife 2023; 12: e91529.

Copyright

(Copyright © 2023, dLife Sciences Plublications, Ltd)

DOI

10.7554/eLife.91529

PMID

37552225

Abstract

In the decoding distance to object analyses on correct trials, classifiers were trained and tested using data from across timepoints (since each distance was a bin spanning multiple timepoints) and trials. Signal samples from all trials and timepoints within each distance were randomly allocated to one of the 10 folds to be used in the 10-fold cross-validation process. This meant that, on any given iteration of cross-validation, a single sample was never in both the training and the testing set, but samples in the training and testing set could be drawn from different timepoints of the same experimental trial. Autocorrelation in the data over time could therefore have inflated the decoding accuracy. To correct this, we have re-run the analysis decoding distance to object ensuring that training and testing sets drew from independent trials, and re-ran all analyses that drew on these decoding results. Analyses that did not decode distance to object are not affected, and analyses decoding distance to object on error trials are also not affected.

We replicate our main finding of being able to use neural patterns to predict the occurrence of errors, along with all the effects of attention. The interaction indicating decreased distance decoding for late vs early blocks in monitoring but not active conditions has decreased in strength from being evident (BF >3) at two consecutive time-windows to one (BF = 6.7) and there is no longer a main effect of target frequency. The link we reported between behaviour and informational connectivity is no longer present.

Language: en

Keywords

human; neuroscience