Affective Response Modeling Leveraging EEG and Deep Learning Toward Mental Health Assessment

1681 

Abstract Submission 

Woojae Jeong¹, Takfarinas Medani¹, Hamzeh Alturk², Colin McDaniel², Idan Blank³, Dani Byrd⁴, Assal Habibi², Baruch Cahn², Shrikanth Narayanan¹, Richard Leahy¹

¹Ming Hsieh Department of Electrical and Computer Engineering, University of Southern California, Los Angeles, CA, ²Brain and Creativity Institute, University of Southern California, Los Angeles, CA, ³Department of Psychology and Department of Linguistics, University of California, Los Angeles, Los Angeles, CA, ⁴Department of Linguistics, University of Southern California, Los Angeles, CA

Woojae Jeong  
Ming Hsieh Department of Electrical and Computer Engineering, University of Southern California
Los Angeles, CA

Takfarinas Medani  
Ming Hsieh Department of Electrical and Computer Engineering, University of Southern California
Los Angeles, CA

Hamzeh Alturk  
Brain and Creativity Institute, University of Southern California
Los Angeles, CA

Colin McDaniel  
Brain and Creativity Institute, University of Southern California
Los Angeles, CA

Idan Blank  
Department of Psychology and Department of Linguistics, University of California, Los Angeles
Los Angeles, CA

Dani Byrd  
Department of Linguistics, University of Southern California
Los Angeles, CA

Assal Habibi  
Brain and Creativity Institute, University of Southern California
Los Angeles, CA

Baruch Cahn  
Brain and Creativity Institute, University of Southern California
Los Angeles, CA

Shrikanth Narayanan  
Ming Hsieh Department of Electrical and Computer Engineering, University of Southern California
Los Angeles, CA

Richard Leahy  
Ming Hsieh Department of Electrical and Computer Engineering, University of Southern California
Los Angeles, CA

Mental health is critical in determining one's overall health, quality of life, and functional status. After the COVID-19 pandemic, we witnessed a significant increase in the global prevalence of depression and suicide rates [1]. Accurate and timely assessment of depression and suicide risk is one the most effective ways to support mental healthcare. Current methods for assessing individual mental health and potential risk factors predominantly rely on self-reports or behavioral interviews through questionnaires. However, these approaches are often incomplete and unreliable. With advances in neuroscience and machine learning, recent studies have shown promising results in emotion recognition with high accuracy by decoding distinctive features from emotionally evoked electroencephalogram (EEG) time-frequency representations using different convolutional neural network (CNN) architectures [2, 3]. Here, we seek to develop a reliable means to objectively assess and classify mental states by identifying biomarkers from EEG signals evoked by four different emotional stimuli in an Emotional Strop paradigm using deep neural networks and machine learning classifiers.

Participants were pre-screened and grouped into control, depressed, and suicide risk groups based on their performance on PHQ-9 [4] and SIS (Suicidal Ideation Scale) scores [5]. Participants were asked to perform an Emotional Stroop task [6] with four groups of 60 emotional stimuli words related to happy, neutral, sad, and suicidality designed and developed to trigger an emotional response. Participants were to identify the color of the word while instructed to ignore the meaning of the word. Differential interference in color by emotional versus neutral words implicates excess attention deployed to stimuli related to underlying emotional processing, taking attentional resources away from color identification. We collected EEG using a 64-channel active electrode system (Brain Products) sampled at 1 kHz. The EEG signal was preprocessed using the standard Brainstorm pipelines [7]. We performed a preliminary analysis of the behavioral data and event-related potential (ERP) on data from 31 participants (healthy control: 17, depressed: 8, suicide risk: 6). We computed the average reaction time (time from stimulus onset to the participant's response) for each emotional category in each group. The ERP was computed by averaging across trials for each emotional category and averaged across the 9 selected channels in the fronto-central area, where we expected to capture signals originating from the Anterior Cingulate Cortex (ACC) and Dorsolateral Prefrontal Cortex (DLPFC) [8].

The overall reaction time of the healthy control group showed a faster trend than the depressed and the suicide risk groups (Figure 1). Our preliminary analysis suggests that, in comparison to the healthy control group, which exhibits similar neural responses to both affective and non-affective stimuli, the neural processing in response to affective stimuli appears to differ in the groups with depression and suicidal tendencies (Figure 2).

Our initial findings indicate that the emotionally aroused neural responses exhibit distinct characteristics across different groups both on reaction time and the underlying neural mechanisms, although larger samples will be required to verify statistical significance. We aim to collect 150 participants (control: 74, depressed: 38, suicide risk 38) and extend our analysis of these subjects. Future analysis will include extracting discriminative underlying features from the EEG time-frequency representations at the source level [7, 9, 10] by a deep convolutional neural network (CNN) [2]. By employing this approach, we aim to deepen our understanding of the neural responses that underlie different mental states and elaborate our capability to assess mental health with greater objectivity.

Classification and Predictive Modeling ²

EEG/MEG Modeling and Analysis ¹

Computational Neuroscience

Computing

Consciousness

Data analysis

Design and Analysis

Electroencephaolography (EEG)

Emotions

Machine Learning