Fig. 2: Classifier construction.
From: Representation of probabilistic outcomes during risky decision-making
a Distribution of the participant-specific set of 135 sensors that contained the fewest eye blinks. Warm-coloured sensors were retained more frequently. b Average balanced classification accuracy (thick black line) with its standard error of the mean (grey shade), obtained from n = 21 participants, as a function of time elapsed after outcome onset (t = 0). For selection of the post-outcome time bin, the Lasso coefficient λ was set arbitrarily to 0.025. At the resulting peak time bin (310 ms), λ was then optimised to build the classifier used for decoding, resulting in an average balanced accuracy of 0.70 ± 0.02 (mean ± s.e.m.; black dot). c Same as in b, but after dividing the training set according to loss probability. Classification accuracy is particularly low when negative outcomes are rare (low). Inset: comparison of the baseline-to-peak accuracy at 310 ms for each threat probability. d Source reconstruction of brain activity around the time bin used for building the classifier (310 ms), computed with a beamforming algorithm on an interval of 100 ms duration. Figure shows brain regions where broadband oscillatory power (1–50 Hz) was higher for the negative outcome (N, warm colours) or for the positive outcome (P, cold colours). Results are corrected for whole-brain family-wise error (FWE) at p < 0.05. Copyright (C) 1993–2004 Louis Collins, McConnell Brain Imaging Centre, Montreal Neurological Institute, McGill University.
