A Deep Learning-Based Retrospective Evaluation Prediction System for Emotional Experiences: Temporal Dynamic Feature Extraction and ERP Neural Mechanisms of the Peak-End Effect

Abstract

Retrospective evaluations of emotional experiences are significantly influenced by the peak-end effect, yet existing electroencephalography (EEG)-based emotion recognition studies have focused on classifying immediate emotional states, leaving integration of the peak-end effect into predictive modeling of retrospective evaluations underexplored. This study proposes a theory-driven deep learning framework extracting peak-end neural features from event-related potential (ERP) signals to predict retrospective overall ratings. Thirty participants completed a sequential emotion induction paradigm based on the International Affective Picture System (IAPS), with peak position and endpoint intensity systematically manipulated. Three ERP components—Early Posterior Negativity (EPN), P300, Late Positive Potential (LPP)—were extracted from 64-channel EEG to construct the TCN-Attention-PeakEnd Gate (TAPE) model, which combines a temporal convolutional network (TCN) encoder with multi-head self-attention and a peak-end feature gating module that embeds the cognitive theory into the network as differentiable operations. Under leave-one-subject-out (LOSO) cross-validation, TAPE achieved MAE = 1.038, Pearson \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r$$\end{document} = 0.654, and three-level classification accuracy of 70.4%, significantly outperforming eight baselines under Holm-Bonferroni-corrected paired-sample \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t$$\end{document}-tests. Across 50 independent fivefold cross-validation evaluations, TAPE simultaneously attained the highest median \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r$$\end{document} (0.682) and smallest interquartile range (0.054). External validation on SEED replicated TAPE's ranking above the strongest deep-learning baselines (84.7% vs 81.9% for Transformer). Ablation confirmed independent contributions of each module, and attention weight visualization revealed high consistency between learned temporal patterns and peak-end theoretical predictions. This study provides a "cognitive theory and data-driven" dual-track fusion modeling paradigm for EEG affective computing.