Deception Detection from Five-Channel Wearable EEG on LieWaves: A Reproducible Baseline for Subject-Dependent and Subject-Independent Evaluation

Deception detection with low-channel wearable EEG requires protocols that generalize across people while remaining practical for portable devices. Using the public LieWaves dataset (27 subjects recorded with a five-channel Emotiv Insight headset), we evaluate to what extent five-channel head-mounted EEG can support lie–truth discrimination under both subject-independent and subject-dependent evaluations. For the subject-independent setting, we train a compact Residual Network with Squeeze-and-Excitation blocks (ResNet-SE) model on raw overlapping windows with focal loss, light data augmentation, and grouped cross-validation by subject; out-of-fold window probabilities are averaged per session and converted to labels using a single decision threshold estimated from the cross-validated session scores. For the subject-dependent setting, we adopt an overlapping short-window Residual Temporal Convolutional Network with Squeeze-and-Excitation and Attention (Res-TCN-SE-Attention) model that fuses raw EEG with discrete wavelet transform (DWT)-based spectral and handcrafted band-power and Hjorth features, using an 80/10/10 split at the recording/session level (stratified by session label), so that all windows from a given session are assigned to a single subset; because each subject contributes two sessions, the same subject may still appear across subsets via different sessions. The subject-independent model attains 66.70% session-level accuracy with an AUC of 0.58 on unseen subjects, underscoring the difficulty of person-independent generalization from low-channel wearable EEG. Because practical deployment requires generalization to previously unseen individuals, we treat the subject-independent evaluation as the primary estimate of real-world generalization. In contrast, the subject-dependent pipeline reaches 99.94% window-level accuracy under the overlapping sliding-window (OSW) setting with a session-disjoint split (no session contributes windows to more than one subset). This near-ceiling performance reflects the optimistic nature of subject-dependent evaluation with highly overlapping windows, even when avoiding within-session train–test overlap, and should not be interpreted as a meaningful indicator of deception-detection capability under realistic deployment constraints. These results suggest limited, above-chance separability between lie and truth sessions in LieWaves using a five-channel wearable EEG under the studied protocol; however, performance remains far from deployment-ready and is strongly shaped by evaluation design. Explicit reporting of both protocols, together with clear rules for windowing, aggregation, and threshold selection, supports more reproducible and comparable benchmarking.

Background & Context

The promise of portable, wearable electroencephalography (EEG) for deception detection has generated considerable interest as an alternative to traditional polygraph methods. Low-channel head-mounted systems with 5-8 electrodes sampling from frontal and temporal sites reduce setup time and cost while avoiding the logistics of dense caps, making them attractive for field deployment. However, this accessibility comes with fundamental challenges.

Limited spatial sampling, reliance on dry or semi-dry electrodes more sensitive to motion and interface artifacts, and tighter constraints on preprocessing and model complexity characterize wearable EEG systems. The field has suffered from inconsistent evaluation protocols that make cross-study comparisons nearly impossible and often produce inflated performance estimates that collapse when tested on new individuals.

This study addresses a critical gap in the literature: the need for standardized, clearly reported evaluation protocols and reproducible baselines for five-channel wearable EEG deception detection. By systematically comparing subject-independent (cross-subject) and subject-dependent (within-subject) evaluation approaches on the publicly available LieWaves dataset, the researchers expose how evaluation design fundamentally shapes reported performance.

Research Design & Methodology

The study evaluated deception detection from five-channel wearable EEG on the public LieWaves dataset comprising 27 subjects recorded with an Emotiv Insight headset. The Emotiv Insight uses five channels positioned at AF3, T7, Pz, T8, and AF4, and subjects participated in two experiments, taking on the roles of deceivers and truth-tellers.

The researchers implemented two distinct evaluation frameworks to reveal how methodology affects results:

• Subject-Independent Model: A compact ResNet-SE (Residual Network with Squeeze-and-Excitation blocks) model trained on raw overlapping windows with focal loss, light data augmentation, and grouped cross-validation by subject
• Subject-Dependent Model: An overlapping short-window Res-TCN-SE-Attention (Residual Temporal Convolutional Network with Squeeze-and-Excitation and Attention) model fusing raw EEG with discrete wavelet transform (DWT)-based spectral and handcrafted band-power and Hjorth features
• Data Splitting: Subject-dependent evaluation used an 80/10/10 split at recording/session level with sessions stratified by label, while subject-independent evaluation ensured complete separation of individuals between training and testing
• Decision Aggregation: Out-of-fold window probabilities were averaged per session and converted to labels using a single decision threshold estimated from cross-validated session scores

Results & Key Findings

The study reveals a stark performance divide depending on evaluation methodology. The subject-independent model attained 66.70% session-level accuracy with an AUC of 57.80% (reported elsewhere as 0.58) on subjects not seen by the model, representing the realistic scenario where the system must generalize to new individuals.

Key performance metrics:

• Subject-independent (real-world scenario): 66.70% accuracy, AUC = 0.58
• Subject-dependent (optimistic scenario): 99.94% window-level accuracy
• Performance gap: Over 33 percentage points difference between evaluation methods

Because practical deployment requires generalization to previously unseen individuals, the researchers treat the subject-independent evaluation as the primary estimate of real-world generalization. The near-perfect subject-dependent performance is explicitly contextualized as misleading.

The subject-dependent pipeline reaches 99.94% window-level accuracy under the overlapping sliding-window (OSW) setting with a session-disjoint split, but this near-ceiling performance reflects the optimistic nature of subject-dependent evaluation with highly overlapping windows and should not be interpreted as a meaningful indicator of deception-detection capability under realistic deployment constraints.

Discussion & Significance

This study makes a crucial methodological contribution to deception detection research by explicitly demonstrating how evaluation design shapes outcomes. The pronounced generalization gap between subject-dependent and subject-independent models underscores that substantial additional work on robustness, calibration, and external validation is required before real-world deployment, with performance remaining far from deployment-ready.

The findings challenge optimistic claims in the EEG-based lie detection literature. An AUC of 0.58 indicates only marginally better than chance performance (0.50) when tested on new individuals—the scenario that matters for real-world applications. This aligns with longstanding skepticism about brain-based lie detection but provides empirical evidence through rigorous methodology.

The study's emphasis on reproducibility represents a significant advancement. By making subject-wise splits, windowing rules, decision thresholds, and session-level aggregation explicit, the researchers provide a concrete, reproducible benchmark for five-channel wearable EEG deception detection and highlight the need for standardized, clearly reported evaluation protocols. This transparent reporting enables direct replication and meaningful cross-study comparisons.

Limitations & Considerations

The study acknowledges several important constraints inherent to low-channel wearable EEG deception detection:

• Limited spatial resolution: Five-channel systems cannot capture the full complexity of neural activity associated with deception across distributed brain networks
• Dataset constraints: With only 27 subjects, generalization estimates remain uncertain; larger validation samples are needed
• Protocol specificity: Results apply specifically to the LieWaves task paradigm (bead theft scenario) and may not transfer to other deception contexts
• Sensor limitations: Dry electrodes in consumer-grade devices introduce noise and artifacts that may fundamentally limit signal quality for deception detection

The researchers are transparent that even their "realistic" subject-independent evaluation may overestimate real-world performance, as controlled laboratory conditions differ substantially from field deployment scenarios.

Practical Applications

For polygraph professionals and policymakers, this study delivers a sobering reality check. The 66.70% accuracy and near-chance AUC (0.58) on unseen subjects indicate that current five-channel wearable EEG technology cannot reliably detect deception when tested under conditions that approximate real-world deployment. This performance level falls far short of standards needed for forensic, security, or employment screening applications.

The study's most valuable contribution may be methodological rather than technological. By establishing transparent evaluation protocols and demonstrating how subject-dependent evaluation inflates performance estimates, the researchers provide a framework for future studies to report results honestly. Consumers, administrators, and researchers can now better evaluate claims about EEG-based lie detection systems by demanding subject-independent validation with explicit reporting of windowing, aggregation, and threshold selection procedures. Until substantially higher subject-independent performance is demonstrated on larger, more diverse samples, wearable EEG deception detection should be considered a research tool rather than a deployment-ready technology.