SparseMambaNet: A Novel Architecture Integrating Bi-Mamba and a Mixture of Experts for Efficient EEG-Based Lie Detection

Traditional lie detection technologies, such as the polygraph and event-related potential (ERP)-based approaches, often face limitations in real-world applicability due to their sensitivity to psychological states and the complex, nonlinear nature of electroencephalogram (EEG) signals. In this study, we propose SparseMambaNet, a novel neural architecture that integrates the recently developed Bi-Mamba model with a Sparsely Activated Mixture of Experts (MoE) structure to effectively model the intricate spatio-temporal dynamics of EEG data. By leveraging the near-linear computational complexity of Mamba and the bidirectional contextual modeling of Bi-Mamba, the proposed framework efficiently processes long EEG sequences while maximizing representational power through the selective activation of expert networks tailored to diverse input characteristics. Experiments were conducted with 46 healthy subjects using a simulated criminal scenario based on the Comparison Question Technique (CQT) with monetary incentives to induce realistic psychological tension. We extracted nine statistical and neural complexity features, including Hjorth parameters, Sample Entropy, and Spectral Entropy. The results demonstrated that Sample entropy and Hjorth parameters achieved exceptional classification performance, recording F1 scores of 0.9963 and 0.9935, respectively. Statistical analyses further revealed that the post-response “answer” interval provided significantly higher discriminative power compared to the “question” interval. Furthermore, channel-level analysis identified core neural loci for deception in the frontal and fronto-central regions, specifically at channels E54 and E63. These findings suggest that SparseMambaNet offers a highly efficient and precise solution for EEG-based lie detection, providing a robust foundation for the development of personalized brain–computer interface (BCI) systems in forensic and clinical settings.

Background & Context

Traditional lie detection technologies—including the polygraph and event-related potential (ERP)-based approaches—face critical limitations in real-world settings due to their sensitivity to psychological states and the complex, nonlinear characteristics of electroencephalogram (EEG) signals. This study addresses a fundamental challenge in forensic neuroscience: how to reliably decode deceptive behavior from brain signals despite their low signal-to-noise ratio and inherent temporal complexity.

The research fills a significant gap by introducing advanced deep learning architectures originally developed for sequence modeling—Structured State Space Models (SSMs) and Mixture of Experts (MoE)—to the domain of EEG-based lie detection. The SparseMambaNet architecture integrates the Mamba model, a Structured State Space Model offering near-linear computational complexity, which is particularly adept at capturing global dynamic patterns in nonlinear and temporally complex EEG signals. This represents a paradigm shift from traditional approaches that struggle with long-range temporal dependencies in EEG data.

The work is positioned at the intersection of cognitive neuroscience, machine learning, and forensic science, offering potential applications not only in criminal investigations but also in developing personalized brain-computer interface (BCI) systems for clinical and security contexts.

Research Design & Methodology

Experiments were conducted with 46 healthy subjects using a simulated criminal scenario based on the Comparison Question Technique (CQT) with monetary incentives to induce realistic psychological tension. The CQT protocol created ecological validity by motivating participants to deceive convincingly, mimicking the stakes of real-world polygraph examinations.

Nine statistical and neural complexity features were extracted from EEG signals, including statistical features (Skewness, Kurtosis, Zero Crossing, Mean, and Variance) and neural complexity measures (Sample Entropy, Spectral Entropy, Higuchi Fractal Dimension, and Hjorth parameters—consisting of Activity, Mobility, and Complexity components). EEG data were segmented using a sliding window approach to capture temporal dynamics across distinct phases of the questioning protocol.

Key methodological features include:

• Temporal segmentation: Analysis of both "question" and "answer" intervals to determine optimal detection windows
• Advanced architecture: Bi-Mamba for bidirectional temporal modeling combined with Sparsely Activated Mixture of Experts to enhance representational power while reducing computational costs through individualized learning
• Multi-level analysis: Interval classification, subject identification, feature comparison, and channel-level statistical mapping
• Statistical validation: ANOVA testing to determine whether performance differences between temporal intervals were statistically significant, with p-values below α = 0.05 for most features indicating significant differences between temporal segments

Results & Key Findings

Sample entropy and Hjorth parameters achieved exceptional classification performance, recording F1 scores of 0.9963 and 0.9935, respectively. These remarkably high accuracy rates demonstrate the model's capacity to distinguish deceptive from truthful responses with near-perfect precision.

Major findings across the four research questions include:

• Temporal dynamics: The post-response "answer" interval provided significantly higher discriminative power compared to the "question" interval, suggesting that cognitive and physiological signatures of deception are most pronounced during the response phase
• Neuroanatomical specificity: Channel-level analysis identified core neural loci for deception in frontal and fronto-central regions, specifically at channels E54 and E63, with E54 displaying statistically significant differences across most EEG features
• Subject identification: Sample Entropy achieved the highest performance in subject classification experiments (0.9963 ± 0.0023), demonstrating that EEG signals reflect unique neural activity patterns for each individual, providing a strong neural foundation for personalized BCI technologies
• Feature performance: Nonlinear and complexity-based features, including Hjorth parameters, Sample Entropy, and Spectral Entropy, successfully captured cognitive complexity and signal irregularity during deception

The frontal and fronto-central regions are well-known cortical areas associated with executive control, inhibition, and working memory—all critically engaged when a person intentionally suppresses truth to generate a lie, supporting the neural interpretation that deceptive behavior elicits consistent modulation within specific cortical lobes.

Discussion & Significance

The findings provide compelling evidence that EEG-based lie detection can achieve exceptional accuracy when sophisticated deep learning architectures are combined with principled feature engineering. By integrating Bi-Mamba, SparseMambaNet overcomes the quadratic computational complexity of traditional Transformers, enabling efficient capture of long-term temporal dependencies with near-linear complexity. This computational efficiency is critical for real-time forensic applications where processing speed matters.

The identification of specific cortical regions (E54, E63) provides neurobiological validation for the model's classifications. The consistent significance of channel E54 across nine different features provides robust evidence that the model captures universal neuro-cognitive mechanisms, and the localized activity in frontal and fronto-central regions suggests SparseMambaNet is decoding the intrinsic cognitive load of deception rather than relying on arbitrary subject-specific identifiers. This addresses a longstanding concern in lie detection research: whether models detect actual deception or merely individual differences.

The integration of the Sparsely Activated Mixture of Experts module plays a pivotal role by selectively activating specialized expert networks, effectively partitioning individualized neural signatures from broader cognitive features of lying, ensuring the model does not merely memorize individual traits but isolates and refines detection of deceptive patterns within heterogeneous datasets. This modularity represents a significant architectural innovation for handling individual differences in neurophysiological responses.

Limitations & Considerations

Although conducted with a limited number of subjects and not employing subject-independent validation methods such as leave-one-out cross-validation (LOOCV), the study serves as a robust proof-of-concept, with future research needing to incorporate larger, more diverse samples and rigorous validation techniques to further enhance reliability and applicability. The lack of LOOCV is a significant methodological limitation that affects generalizability claims.

Additional considerations include:

• Laboratory setting with simulated crime scenarios may not fully replicate the psychological states of actual suspects in criminal investigations
• Monetary incentives, while enhancing motivation, differ qualitatively from the stakes in real forensic contexts (criminal prosecution, imprisonment)
• The 46-subject sample, while adequate for proof-of-concept, is modest for establishing population-level norms
• High accuracy rates may partially reflect model overfitting without more rigorous cross-validation protocols

Practical Applications

SparseMambaNet offers a foundation for developing next-generation forensic BCI systems that could supplement or eventually replace traditional polygraph testing. The near-linear computational complexity makes real-time processing feasible in field settings, while the exceptionally high F1 scores suggest potential for deployment in high-stakes contexts. The identification of specific EEG channels (E54, E63) provides practical guidance for developing simplified electrode arrays optimized for deception detection, potentially reducing setup time and equipment costs.

For polygraph examiners and security professionals, this research demonstrates that focusing on post-response intervals yields superior detection accuracy—a finding that could inform questioning protocols and timing strategies. The model's ability to achieve individual subject identification while maintaining deception detection accuracy suggests applications in personalized screening systems for counterintelligence, pre-employment security vetting, and clinical assessment of credibility in forensic psychiatry contexts.