Decoding Spatial Pain Anticipation from EEG Using Deep Learning

This thesis investigates whether electroencephalography (EEG) recorded while a participant waits for a possible painful stimulus contains information about where on the body that pain is expected. The study uses a Pavlovian threat-learning task in which visual cues signalled expected pain on the left arm, expected pain on the right arm, or safety. The classification problem was a three-class distinction between left-threat, right-threat, and safety during the cue period. The analysis focuses on Block 1, the acquisition phase in which cue meanings were stable. The Block 1 raw scan covered 30 recording files, while the final machine-learning export contained 29 retained recording identifiers and 1711 clean cue epochs, with almost balanced class counts: 571 left-threat epochs, 572 right-threat epochs, and 568 safety epochs. Three model families were compared using a subject-wise train/validation/test split and train-only normalization: a baseline temporal convolutional network (TCN), an attention-augmented TCN, and a compact convolutional neural network trained on short-time Fourier transform (STFT) spectrograms. In the full-band analysis, the attention TCN obtained the strongest held-out performance, with 39.93% test accuracy and macro-F1 0.3867. This is above the balanced three-class chance level of 33.33%, but the margin is modest. In the exploratory analyses, the alpha-restricted attention TCN achieved the highest test accuracy overall, reaching 40.27% and macro-F1 0.3975. Overall, the results suggest that cleaned cue-period EEG contains some information about the expected spatial content of pain, but that this information is weak and difficult to generalize across participants. The main contribution of the thesis is twofold: it provides a cautious empirical decoding result and a transparent preprocessing and evaluation pipeline for spatially specific EEG decoding in a technically challenging dataset.