Multivariate neural biomarkers of emotional states are categorically distinct

Abstract

Understanding how emotions are represented neurally is a central aim of affective neuroscience. Despite decades of neuroimaging efforts addressing this question, it remains unclear whether emotions are represented as distinct entities, as predicted by categorical theories, or are constructed from a smaller set of underlying factors, as predicted by dimensional accounts. Here, we capitalize on multivariate statistical approaches and computational modeling to directly evaluate these theoretical perspectives. We elicited discrete emotional states using music and films during functional magnetic resonance imaging scanning. Distinct patterns of neural activation predicted the emotion category of stimuli and tracked subjective experience. Bayesian model comparison revealed that combining dimensional and categorical models of emotion best characterized the information content of activation patterns. Surprisingly, categorical and dimensional aspects of emotion experience captured unique and opposing sources of neural information. These results indicate that diverse emotional states are poorly differentiated by simple models of valence and arousal, and that activity within separable neural systems can be mapped to unique emotion categories.

Keywords: Bayesian model comparison; affect; emotion; functional magnetic resonance imaging; multi-voxel pattern analysis; pattern classification.

© The Author (2015). Published by Oxford University Press. For Permissions, please email: journals.permissions@oup.com.

Figures

Fig. 1

Dimensional sampling from self-report of emotion. Scatterplot shows the distribution of self-reported emotion along dimensions of arousal and valence (standardized within subjects). The x-axis is the average of valence-related self-report items (‘good’, ‘positive’ and ‘pleasant’ minus the scores for ‘bad’, ‘negative’ and ‘unpleasant’) and the y-axis is the average of arousal-related items (‘agitated’, ‘active’ and ‘excited’ minus ‘calm’, ‘passive’ and ‘relaxed’). Marginal histograms depict distributions of valence and arousal for all emotions. The distributions generally conform to dimensional models of emotion: contentment and amusement oppose fear, anger and sadness along the valence dimension whereas contentment and sadness oppose fear, anger, surprise and amusement in terms of arousal.

Fig. 2

Performance measures of multi-voxel pattern classification. (A) Confusion matrix for classifying all seven emotional states (top) and receiver-operating characteristic (ROC) curves for one-vs-all classifications (bottom). Increasing classification performance would yield a confusion matrix with more predictions falling along the diagonal and push ROC curves toward the upper left corner, away from chance. (B) Binomial regression predicting the number of instances assigned to each emotion category based on the self-reported experience of the targeted emotion. The x-axis reflects standardized scores of self-report for the target emotion and the y-axis indicates the number of trials labeled as the target emotion. Note the curves approach 18 predictions (near random assignment of 128 trials into 7 classes) as self-report scores near zero.

Fig. 3

Distributed patterns of neural activity predict discrete emotional states. Each map depicts voxels whose activation consistently led to predictions of the related emotion across all subjects and stimulus modalities (P < 0.05, FWE corrected; see Materials and Methods), overlaid on the group-average anatomical image.

Fig. 4

Stacked bar plot showing the coverage of all regions in the AAL atlas by each emotion-predictive pattern.

Fig. 5

Categorical and dimensional models of emotion. (A) Hierarchical clustering of self-report (left) and radar plots of categorical and dimensional terms (right). Note the increased distance between positive and negative emotions in the dimensional model and the reduced overlap between emotions in the categorical model. (B) BIC weights suggest that a combined model with both categorical and dimensional terms was most likely to produce the observed distribution of classification errors (using subject-independent cross-validation). (C) Parameter estimates from the combined model show that errors are more frequent as distance increases along valence and arousal, and that errors are less frequent as categorical distance increases. Error bars reflect 95% confidence intervals.