Fig. 2: Predicting key cognitive scores using ReAl-LiFE connection weights.
a, Schematic of the SVR-RFE prediction model. For each participant (n = 200), the 68 × 68 whole-brain connectivity matrix was vectorized to give 1,122 connectivity features. Feature vectors from all participants were collated to form a feature matrix of size 200 × 1,122, which was used to predict 60 different behavioral and cognitive scores. Data were divided into training and testing folds, and the prediction model was trained on the train fold using SVR (dashed box). Feature selection was implemented using RFE. b, Top: number of scores significantly predicted as a function of the uncorrected P-value threshold for predictions based on the number of fibers in the unpruned connectome (red circles) and connection weights in the ReAl-LiFE-pruned connectome (purple circles). Bottom: average correlations between the observed and predicted scores as a function of the uncorrected P-value threshold. Other conventions are the same as in the top panel. c–e, As in b, but for scores from the cognition (c), emotion (d) and personality (e) categories. Other conventions are the same as in b. f, Word clouds showing the different behavioral scores from the cognition (left), emotion (middle) and personality (right) categories, sized based on their prediction accuracy values using ReAl-LiFE connection weights. Larger words indicate better predicted scores. g, Proportion of ReAl-LiFE features chosen by the SVR-RFE model for predicting each of the 60 behavioral and cognitive scores using a combined feature set, including both the number of fibers in the unpruned connectome (1,122 features), as well as ReAl-LiFE connection weights (1,122 features). The proportions of features are shown in descending order, separately, for each of the three categories of scores—cognition (blue, n = 13), emotion (magenta, n = 23) and personality (green, n = 24).
