Date of Award

2026

Document Type

Thesis

Degree Name

Master of Science in Artificial Intelligence

Department

Digital Engineering

Committee Chair and Members

Kewei Li

Keywords

Affective computing, EEG emotion recognition, Hemispheric asymmetry, Mamba state space, SEED-IV, Temporal interpretability

Abstract

Electroencephalography (EEG)-based emotion recognition has emerged as a critical component of affective computing and clinical neuroscience. Existing approaches to this problem primarily reduce the multi-dimensional EEG time series to a single averaged feature vector, thereby discarding the temporal structure of the emotional response. The present work addresses three identified gaps in the literature: the absence of temporal interpretability, the uniform use of frequency bands, and the use of single-scale temporal feature extraction. A deep learning architecture, MST-Mamba-Asym, is proposed, comprising four components: Asymmetry Attention, which encodes hemispheric asymmetry by computing signed left–right channel differences, FreqBandAttention, which learns differential weights across the five Differential Entropy frequency bands; MultiScaleConvPool, which extracts temporal features at 4-second, 12-second, and 20-second scales in parallel and two stacked Mamba Selective State Space Model layers, which selectively retain emotional information across the sequence while discarding uninformative background windows. Experiments are conducted on the SEED-IV dataset, comprising 1,080 trials from 15 subjects across four emotion classes. The proposed model achieves 74.54% accuracy on a global 80/20 split (mean 72.1% ± 1.8% across five runs) and 86.67% mean accuracy under a rigorous per-subject evaluation protocol. A temporal ablation explainability analysis reveals that Fear peaks at 8 seconds, Sad at 32 seconds, and Happy at 252 seconds into the recording, findings that are consistent with established neuroscientific accounts of emotional response latency. A frequency band ablation analysis further reveals that Fear is most discriminated by the Beta band, whereas the remaining three classes are distinguished primarily through the Delta band. These results represent a meaningful advance over published baselines by using SEED-IV dataset and demonstrate that domain knowledge can be encoded directly into deep learning architectures for neuroscientific benefit.

Download

Share

COinS