EEG Signals Classification Using Novel Acquisition Protocol for Lie Detection System

Background & Context

EEG-based lie detection represents a promising alternative to traditional polygraph testing, which relies on peripheral physiological measures that skilled individuals can potentially manipulate. Traditional polygraph methods measuring heart rate and respiratory rate are influenced by fear or tension and can be manipulated by skilled criminals. The emergence of deep learning and brain-computer interface technologies has opened new avenues for automated deception detection directly from neural signals.

This 2025 study by Hamza and colleagues addresses critical gaps in the EEG lie detection literature. Prior research showed a lack of studies on real-life settings using eyewitness investigation in EEG analysis for deceit detection, along with accuracy and efficiency issues in EEG data processing. Most existing protocols relied on simple visual stimuli like images, rather than realistic crime scenarios that better approximate forensic applications.

The researchers developed a novel experimental protocol featuring a video-based crime scenario—a robbery from a store—creating a more ecologically valid paradigm for testing deception detection. The protocol utilizes investigation with eyewitness for a crime robbery from a store seen in a video, making it more realistic. This approach represents an important step toward bridging laboratory research and real-world forensic applications.

Research Design & Methodology

The experiment utilized the OpenBCI Ultracortex "Mark IV" EEG Headset, which acquired 14 channel of EEG data from ten participants. The study created a novel dataset called "EXP" specifically for this investigation, alongside testing on the publicly available Dryad dataset for comparison purposes.

The experimental protocol employed a video-based stimulus depicting a store robbery, where participants acted as eyewitnesses. The EXP used video and Dryad used pictures protocol. The experiment dataset underwent pre-processing steps, which included applying Fast Fourier Transform (FFT) filter with a range of 0.1–30 Hz, capturing the primary frequency bands relevant to cognitive deception processes.

The study employed three distinct deep learning architectures for classification:

• Convolutional Neural Network (CNN) for spatial feature extraction from EEG patterns
• Long Short-Term Memory (LSTM) network for temporal sequence modeling
• Multilayer Perceptron (MLP) as a baseline fully connected architecture

The proposed method was implemented on Python 3.11.5 using 5-fold cross-validation, with fold accuracies ranging from 99.52% to 99.93%, dividing the dataset into five equal-sized folds with the model training on four and testing on the remaining fold five times. This rigorous validation approach helped ensure model generalizability and reduced overfitting concerns.

Results & Key Findings

The CNN technique achieved the highest incredible accuracy of 99.96% on the EEG data set in the experimental dataset and 99.36% on the Dryad dataset. This represents exceptional performance that surpasses most prior EEG-based lie detection research.

Performance metrics across all three deep learning models demonstrated remarkably high classification accuracy:

• CNN: 99.96% on the experimental dataset, 99.36% on the Dryad dataset
• LSTM: 99.87% in the EEG dataset, 99.64% in the Dryad dataset
• MLP: 98.99% on the experimental EEG dataset, 97.67% on the Dryad dataset

All models exhibited high performance achieving exceptional levels of precision, recall, F1 score, and accuracy, with the CNN model consistently achieving the highest results across all criteria and the LSTM model closely following. The confusion matrix analysis revealed exceptional discriminative power, with 30,000 samples total in testing (15,000 samples per category), with only 12 samples misclassified.

The proposed EXP dataset consistently demonstrated superior accuracy across all models compared to the publicly accessible Dryad dataset, illustrating the effectiveness of the suggested method in catching more subtle signal patterns. This suggests that the novel video-based crime scenario protocol elicited stronger and more distinguishable neural signatures of deception.

Discussion & Significance

This study makes several important contributions to the field of neurophysiological deception detection. The exceptionally high accuracy rates—particularly the CNN's 99.96% performance—place this work among the most successful EEG-based lie detection studies to date. These results significantly exceed typical polygraph accuracy rates reported in field studies and rival the best laboratory-based deception detection systems.

The superiority of CNN architecture aligns with broader trends in EEG signal processing. The CNN achieved an astounding 99.96% accuracy, demonstrating how well CNNs capture complex patterns in the data, which makes them indispensable tools for tasks requiring delicate analysis. CNNs' ability to automatically learn hierarchical spatial features from multi-channel EEG data appears particularly well-suited for detecting the distributed neural patterns associated with deceptive cognition.

The video-based protocol represents a meaningful methodological advancement. Unlike static image protocols used in most prior research, presenting participants with a dynamic crime scenario creates cognitive demands more similar to real forensic contexts. The consistently superior performance on the EXP dataset compared to the image-based Dryad dataset suggests this ecological validity translates into stronger, more detectable neural signatures of deception.

The study's cross-validation approach and testing on an independent public dataset (Dryad) strengthen confidence in the findings. The models maintained extremely high accuracy even when tested on data collected under different protocols and with different participants, suggesting reasonable generalizability of the deep learning approach.

Limitations & Considerations

The study's small sample size of ten participants represents a significant limitation. More participants are required to enhance system dependability and enable real-time execution. While the deep learning models achieved exceptional within-sample performance, the limited participant pool raises questions about how well these models would generalize to broader populations with diverse demographic characteristics, cognitive abilities, and cultural backgrounds.

The laboratory-based "mock crime" paradigm, while more realistic than simple image protocols, still differs substantially from actual forensic interrogation contexts. Participants in this study lacked the genuine stakes, stress, and consequences present in real criminal investigations. Whether these exceptionally high accuracy rates would be maintained in genuine high-stakes scenarios with motivated deceivers remains an empirical question.

The study does not appear to have tested resistance to countermeasures—deliberate strategies that examinees might employ to defeat the detection system. The vulnerability of EEG-based methods to mental countermeasures requires investigation before operational deployment. Additionally, their suitability for other situations could be restricted, suggesting the need for validation across diverse deception scenarios beyond eyewitness identification.

Practical Applications

The indicated manner is non-surgical, effective, and powerful, with least time complication, consequently appropriate for real-time applications. The use of affordable, portable OpenBCI hardware combined with automated deep learning classification could enable deployment in field settings where traditional neuroimaging methods (fMRI, MEG) are impractical.

The video-based protocol could be adapted for specific forensic applications, such as eyewitness verification in criminal investigations or security screening contexts. The ability to achieve near-perfect classification with relatively brief EEG recordings and automated analysis addresses longstanding concerns about the practical feasibility of brain-based deception detection. However, substantial additional validation research with larger samples, diverse populations, real-world stakes, and countermeasure testing would be essential before considering operational use in consequential decision-making contexts such as criminal justice or security screening.