Towards Explainable Modelling of Sensory and Affective Dimensions of Pain

: A Graph Attention Network Based Approach

Abstract

Pain is subjective and depends on individual expression, making it challenging to assess in those who cannot communicate effectively. Understanding pain requires analyzing brain signals involved in pain perception and modulation. The objective of this thesis is to develop an explainable modelling framework for signals from brain recordings, such as Electroencephalography (EEG), to better characterize emotional and sensory pain experiences. It focuses on addressing three key challenges in interpreting these signals: identifying brain activations following a stimulus, analyzing inter-regional connectivity, and determining the time span of relevant events. We propose the Graph Attention-Inspired Spatial Temporal (GIST) network, which uses attention blocks to highlight relevant nodes, positional relationships, and time windows for pain-related events. Visualisation of attention scores reveal characteristics that help to explain the model behaviour in relation to the classification task at hand. The proposed framework for high-level feature representation is initially combined with low-level features extracted via conventional signal processing and machine learning. We also developed a simple feature reconstruction based channel selection algorithm, identifying channels AF4, F8, O1, P7, T7, and T8 as the most informative for emotion recognition. We extended the GIST network by adding a CNN-based feature extraction block and a transformer-based temporal attention block for end-to-end brain signal analysis. The performance and explainability of our approach were demonstrated through comprehensive benchmark studies across diverse tasks and public datasets of EEG as well as functional Near Infrared Spectroscopy (fNIRS). This thesis demonstrates that leveraging the GIST network representation of EEG/fNIRS signals holds great potential for interpretable models of pain perception, with potential use in neurology and psychology studies.

Date of Award	2025
Original language	English
Awarding Institution	Aberystwyth University
Supervisor	Chuan Lu (Supervisor)