MFG-RNN: A multi-feedback gated recurrent neural network with HJB-based weight update for EEG-based concealed information detection

Abstract

Reliable concealed information detection using electroencephalogram (EEG) signals has been a persistent challenge for decades. Researchers have utilized the P300 test, which examines how the brain reacts to crime-related stimuli, to distinguish between guilty and innocent individuals. However, EEG’s inherent variability across sessions and susceptibility to noise necessitate more robust approaches for accurately discerning brain activity. This work presents a novel deep learning model, a multi-feedback gated recurrent neural network (MFG-RNN), designed to classify these brain responses. Unlike conventional recurrent neural networks (RNN) that rely solely on the immediate past state, the proposed MFG-RNN integrates multiple delayed hidden states through gated feedback paths, enhancing the model’s capacity to capture long-term dependencies. This multi-feedback mechanism allows the network to dynamically regulate memory contributions over time, improving temporal representation and mitigating gradient-vanishing issues. Additionally, a Hamilton-Jacobi-Bellman (HJB) based optimization framework is employed to guide weight updates through long-term performance criteria, transforming training into an optimal control problem. We evaluated the model on an in-house dataset of 31 participants, in which the guilty group committed a mock crime, and the innocent group refrained from any involvement. Our method surpassed traditional RNN approaches and state-of-the-art methods for concealed information detection, achieving an average accuracy of over 92 % using a second-order MFG-RNN. The model also achieved 97 % accuracy on a comparable, non-public external dataset, demonstrating strong generalizability. The proposed framework enables single-trial classification of EEG probe trials for concealed information detection, contributing to intelligent human-computer interaction. © 2025 Elsevier Inc.