Multimodal Emotion-Constrained Memory Generation Using EEG-Derived Features and Large Language Models

Emotion recognition from brain signals has attracted growing interest in affective computing andhuman-computer interaction. In this work, we present a proof-of-concept system that combines EEG-derivedDifferential Entropy (DE) features with large language model (LLM) text generation to produce emotionallyconditioned memory narratives. Using the publicly available FACED dataset — which contains 32-channel EEGrecordings from 123 subjects across 28 trials each, spanning nine emotion classes — we trained a Random Forestclassifier that achieved approximately 35.05% cross-validated accuracy on the nine-class task, significantly abovethe 11.11% random baseline. The decoded emotion probabilities and EEG frequency-band summaries were thenused as conditioning signals for LLM-based narrative generation via a custom prompt engineering pipeline,producing outputs we hypothesize are more emotionally aligned than cue-only baselines. This paper describes ourfull methodology including pipeline architecture, preprocessing workflow, module interactions, visualizationcomponents, and experiment logging. This is a proof-of-concept developed during my first year as an undergraduatestudent. Keywords — EEG, emotion recognition, dif erential entropy, FACED dataset, af ective computing, large languagemodels, memory generation, Random Forest, brain-computer interface, prompt engineering.