Decoding Spontaneous Emotional States in the Human Brain

Abstract

Pattern classification of human brain activity provides unique insight into the neural underpinnings of diverse mental states. These multivariate tools have recently been used within the field of affective neuroscience to classify distributed patterns of brain activation evoked during emotion induction procedures. Here we assess whether neural models developed to discriminate among distinct emotion categories exhibit predictive validity in the absence of exteroceptive emotional stimulation. In two experiments, we show that spontaneous fluctuations in human resting-state brain activity can be decoded into categories of experience delineating unique emotional states that exhibit spatiotemporal coherence, covary with individual differences in mood and personality traits, and predict on-line, self-reported feelings. These findings validate objective, brain-based models of emotion and show how emotional states dynamically emerge from the activity of separable neural systems.

Conflict of interest statement

The authors have declared that no competing interests exist.

Figures

Fig 1. Distributed patterns of brain activity predict the experience of discrete emotions.

(A) Parametric maps indicate brain regions in which increased fMRI signal informs the classification of emotional states. See [13] for details of the development and validation of these neural decoding models. (B) Sensitivity of the seven models. Error bars depict 95% confidence intervals. The data underlying this figure can be found in S1 Data.

Fig 2. Emotional states emerge spontaneously during resting-state scans.

(A) Procedure for classification of resting-state data. Scores are computed by taking the scalar product of preprocessed data and regression weights from decoding models. (B) Frequency distributions for the classification of all seven emotional states (n = 499). The mean, 25th, and 75th percentiles are indicated by black lines. The solid gray line indicates the number of trials that would occur from random guessing. The data underlying this figure can be found in S1 Data. The raw fMRI resting state data can be obtained from https://www.haririlab.com/projects.

Fig 3. Emotional states exhibit coherence during resting-state scans.

Gray circles reflect the sample mean classification scores for all seven emotions (n = 499). Thick colored lines display group-average predicted time course using smoothing splines (with bordering 95% confidence interval). Text overlay (rcv) indicates the average cross-validated correlation between splines fitted for each subject and tested on the average fit of other subjects. Dashed lines indicate linear fits over time. Solid black dots indicate time points at which a model has the highest proportion of classifications. Data are concatenated across two sessions of 256 s (solid vertical line). Note the early peak for fear scores and general increases in neutral scores over time. The data underlying this figure can be found in S1 Data. The raw fMRI resting state data can be obtained from https://www.haririlab.com/projects.

Fig 4. Individual differences in mood and personality modulate the occurrence of spontaneous emotional brain states.

(A) Differences in depressive and anxious mood are associated with increases in the frequency of sad and fear classifications during rest. (B) Emotional traits of Anxiety, Angry Hostility, and Depression track differences in the frequency of fear, anger, and sad classifications (n = 499, error bars reflect standard error, * indicates effects significant at Punc < .05). The data underlying this figure can be found in S1 Data. The raw fMRI resting state data can be obtained from https://www.haririlab.com/projects.

Fig 5. Spontaneous emotional brain states exhibit correspondence with self-report.

(A) Participants (n = 21) participated in an experience sampling task in which they reported their current emotional state at random intervals exceeding 30 s during fMRI scanning. The five samples of data (lasting 10 s, TR = 2 s) preceding each rating were used to compute predictions of emotional state using multivariate decoding models. (B) Scores for classification models congruent with self-report are greater than incongruent models (z = 2.311, Punc = 0.0208; Wilcoxon signed rank test). Classification scores are calculated based on the inner product of neural activity and classifier weights and indicate the relative evidence for the different emotion models. (C) The frequency of classifications from multivariate models significantly correlates with those made by participant self-report (r = .3876 ± 0.102 [s.e.m.], t20 = 2.537, Punc = .0196; one sample t test). Gray line indicates best-fitting least-squares line for group mean. In all panels, error bars reflect standard error of the mean. The data underlying this figure can be found in S1 Data.