Integrating Multilevel Functional Characteristics Reveals Aberrant Neural Patterns during Audiovisual Emotional Processing in Depression

Abstract

Emotion dysregulation is one of the core features of major depressive disorder (MDD). However, most studies in depression have focused on unimodal emotion processing, whereas emotional perception in daily life is highly dependent on multimodal sensory inputs. Here, we proposed a novel multilevel discriminative framework to identify the altered neural patterns in processing audiovisual emotion in MDD. Seventy-four participants underwent an audiovisual emotional task functional magnetic resonance imaging scanning. Three levels of whole-brain functional features were extracted for each subject, including the task-evoked activation, task-modulated connectivity, combined activation and connectivity. Support vector machine classification and prediction models were built to identify MDD from controls and evaluate clinical relevance. We revealed that complex neural networks including the emotion regulation network (prefrontal areas and limbic-subcortical regions) and the multisensory integration network (lateral temporal cortex and motor areas) had the discriminative power. Moreover, by integrating comprehensive information of local and interactive processes, multilevel models could lead to a substantial increase in classification accuracy and depression severity prediction. Together, we highlight the high representational capacity of machine learning algorithms to characterize the complex network abnormalities associated with emotional regulation and multisensory integration in MDD. These findings provide novel evidence for the neural mechanisms underlying multimodal emotion dysregulation of depression.

Keywords: MDD; machine learning; multilevel characteristics; multisensory emotional processing; neural representation.