EEG-Based Single Trial Classification Emotion Recognition: A Comparative Analysis in Individuals with and without Autism Spectrum Disorder

Background: Recent studies support the use of novel statistical techniques, including EEG-based single trial classification, to better understand individual differences in neural responses to emotion reappraisal tasks (Singh & Singh, 2017) a core deficit in Autism Spectrum Disorder (ASD). The Late-Positive-Potential (LPP) is an event-related potential (ERP) that occurs 400-1500ms (Ferri et al., 2012) post-stimulus, is related to emotion reappraisal, and is attenuated in ASD (Benning et al., 2016). Research in typically developing (TD) populations (Mehmood, & Lee, 2016) has evaluated EEG-based single trial classification using LPP features to predict neural response to emotion categories. However, little is known about how well these pipelines can predict emotion states in individuals with and without ASD. Exploring the performance of a novel EEG single trial emotion classification pipeline in TD and ASD groups can advance understanding of both functional network-activation and of emotion reappraisal in individuals with ASD relative to TD peers.

Objectives: Evaluate the performance of a novel EEG single trial classification pipeline for predicting pleasant/unpleasant emotions elicited by DANVA-2 (Nowicki, 2004) stimuli in TD and ASD groups, as well as the capacity of models in one group to predict pleasant/unpleasant emotions in the other.

Methods: 17 participants (M=13.15, SD=1.36) including 9 TD and 8 ASD with ADOS-2 confirmed diagnosis (Lord et al., 2012). For each participant we ran a 3-fold cross-validation (Figure 1) including a complete dataset composed of 48 affective faces (examples) times the corresponding LPP time-points as features. We evaluated three LPP windows per-channel: early (400-800ms), middle (800-1200ms), and late (1200-1500ms). We then used the top channels modality (Kushaba, et al., 2011) to automatically select the two most predictively-sound channels per fold. Subsequently, we compared the predictive power of the top two predictively-sound channels in the TD model to the top two channels in the ASD model and vice versa to examine generalizability of these models across diagnostic groups (different group channel modality).

Results: The model, with high accuracy, predicted concurrent emotion in both TD and ASD groups using the identified top channels (Figure 2A). The average Precision (Pr) and Recall (Re) differed between modalities, and – crucially – between TD/ASD groups for the LPP middle window features (Figure 2B & 2C); These values differed between modalities across all three windows (all F(1,30) > 178, all p < .001). For instance, in the middle window the degradation in predictive Precision was 35.2% for TD and 34.2% for ASD between modalities (Figure 2C).

Conclusions: Our results suggest that LPP response represents an important set of features for pleasant/unpleasant face-elicited emotion prediction across/between TD/ASD groups. However, when the top channels for each group were used to predict the emotion state of the facial stimulus in the opposite group, the recall and precision metrics degraded significantly. Moreover, the cross-group model results lend support to the proposition that individuals with ASD demonstrate more diffuse functional neural-networks during emotional reappraisal than TD individuals, which is consistent with structural connectivity findings (Rudie et al., 2013).