Machine learning-based decoding of emotional valence from electrophysiological signals in the monkey brain

Abstract

Understanding how the brain operates in naturalistic settings requires methods that go beyond conventional repeated-measurement approaches, necessitating the development of single-trial neural activity analysis. Recent advances in machine learning offer new opportunities for analyzing brain electrophysiological signals. Here, we recorded surface electrocorticography (ECoG) and intracranial local field potentials (LFPs) from emotion-related brain regions in a monkey performing a Pavlovian conditioning task, in which sensory cues predicting reward or punishment were presented randomly, followed by the actual unconditioned outcome. We evaluated the performance of two machine learning algorithms, a Convolutional neural network (CNN) model and a Transformer-based model (EEG-Conformer), in classifying raw ECoG/LFP traces. Both models successfully classified valence type during conditioned and unconditioned stimulus presentation. Furthermore, the Transformer achieved significantly superior classification performance compared to the CNN, particularly in multi-state classification including baseline periods. By optimizing the training dataset for the Transformer model, we could detect dynamic fluctuations in emotional valence consistent with task type from continuously evolving ECoG/LFP patterns recorded throughout the task. These results demonstrate the utility of Transformer-based models for decoding emotional valence from neurophysiological signals in non-human primates.