Learning Spatial-Spectral-Temporal EEG Representations with Deep Attentive-Recurrent-Convolutional Neural Networks for Pain Intensity Assessment

Electroencephalogram (EEG)-based quantitative pain measurement is valuable in the field of clinical pain treatment, providing objective pain intensity assessment especially for nonverbal patients who are unable to self-report. At present, a key challenge in modeling pain events from EEG is to find invariant representations for intra- and inter-subject variations, where current methods based on hand-crafted features cannot provide satisfactory results. Hence, we propose a novel method based on deep learning to learn such invariant representations from multi-channel EEG signals and demonstrate its great advantages in EEG-based pain classification tasks. To begin, instead of using typical EEG analysis techniques that ignore spatial information of EEG, we convert raw EEG signals into a sequence of multi-spectral topography maps (topology-preserving EEG images). Next, inspired by various deep learning techniques applied in neuroimaging domain, a deep Attentive-Recurrent-Convolutional Neural Network (ARCNN) is proposed here to learn spatial-spectral-temporal representations from EEG images. The proposed method aims to jointly preserve the spatial-spectral-temporal structures of EEG, for learning representations with high robustness against intra-subject and inter-subject variations, making it more conducive to multi-class and subject-independent scenarios. Empirical evaluation on 4-level pain intensity assessment within the subject-independent scenario demonstrated significant improvement over baseline and state-of-the-art methods in this field. Our approach applies deep neural networks (DNNs) to pain intensity assessment for the first time and demonstrates its potential advantages in modeling pain events from EEG.