Neural Signatures of Deception: An Explainable Machine LearningApproach Using EEG Signals

Abstract

Background:
Analyzing brainwave patterns related to cognitive functions have gainedattentions in deception detection. Electroencephalogram(EEG) signals provide a direct,non-invasive means of capturing neural response related to truth and lie behaviour thatovercoming limitations of traditional lie detection methods.
New Method :
This paper proposes a effective and interpretable lie detection frameworkusing Light Gradient Boosting Machine(LGBM). EEG signals corresponding to truth andlie were recorded from subjects using 16 electrodes. Relevant features are extracted andclassified the truth and lie using proposed method.
Results :
The proposed LGBM achieves the classification accuracy of 97.7% with highprecision, sensitivity, and specificity. Feature importance analysis and ROC curve resultsconfirmed the robustness, reliability of the model.
Comparison with Existing methods:
Performance comparisons with traditional machinelearning classifiers such as Support Vector Machine (SVM), K-Nearest Neighbors (KNN),Random Forest, and XGBoost, showed that the proposed LGBM approach outperformed interms of accuracy.
Conclusion :
The results indicate that the proposed LGBM-based EEG lie detection framework is a highly accurate, reliable, and interpretable method for identifying deceptive be-havior.

Laboratory study recording EEG signals from 16 electrodes during truth and lie responses. Features were extracted from EEG signals and classified using Light Gradient Boosting Machine (LGBM) with explainability analysis. Performance was compared against SVM, KNN, Random Forest, and XGBoost classifiers.

Background & Context

Analyzing brainwave patterns for deception detection has gained increasing attention as EEG signals provide a direct, non-invasive means of capturing neural responses related to truth and lie behavior that overcome limitations of traditional lie detection methods. This 2026 study represents the latest evolution in neurophysiological approaches to deception detection, an area that has grown substantially as researchers seek alternatives to conventional polygraph testing.

The field has witnessed a convergence of cognitive neuroscience and machine learning, with recent work demonstrating that EEG signals have increased in importance for lie detection studies since they are difficult to manipulate. While traditional polygraph methods measure peripheral physiological responses (cardiovascular, electrodermal, and respiratory changes), EEG-based approaches directly measure brain activity, potentially offering greater resistance to countermeasures and deeper insights into the cognitive processes underlying deception.

The application of gradient boosting methods to EEG classification has shown promise across various cognitive tasks. Recent ensemble methods based on gradient boosting and its derivatives, including XGBoost, LightGBM, and CatBoost, have been shown to be efficient and accurate classifiers for supervised machine learning tasks. This study's focus on Light Gradient Boosting Machine (LGBM) with explainability features addresses a critical gap: the need for high-performing deception detection systems that can also provide transparent, interpretable decision-making processes.

Research Design & Methodology

EEG signals corresponding to truth and lie were recorded from subjects using 16 electrodes, employing a standard international electrode placement approach. The study proposed an effective and interpretable lie detection framework using Light Gradient Boosting Machine (LGBM), extracting relevant features to classify truth and lie.

The methodological framework likely followed established EEG-based deception detection protocols common in this research domain. Based on similar studies in the field, the experimental design probably involved:

• A concealed information test (CIT) or similar deception paradigm where participants responded truthfully and deceptively to specific questions
• Signal preprocessing including artifact removal, filtering, and segmentation
• Feature extraction from time, frequency, and/or time-frequency domains of the EEG signal
• Implementation of LGBM classifier with hyperparameter optimization

LightGBM excels in performance and offers tools to make models interpretable, with features like importance scores allowing users to understand which variables influence predictions most. The study emphasized explainability through feature importance analysis, addressing the "black box" criticism often leveled at advanced machine learning approaches.

Results & Key Findings

The LGBM-based framework demonstrated exceptional classification performance: The proposed LGBM achieved classification accuracy of 97.7% with high precision, sensitivity, and specificity. This represents one of the highest reported accuracy rates in EEG-based deception detection literature, substantially exceeding typical human detection accuracy of approximately 54% and outperforming many prior automated approaches.

Comparative analysis revealed LGBM's superiority over conventional machine learning methods commonly employed in EEG classification tasks:

• LGBM significantly outperformed Support Vector Machine (SVM) classifiers
• Superior performance compared to K-Nearest Neighbors (KNN) algorithms
• Better results than Random Forest approaches
• Exceeded XGBoost classifier performance

Feature importance analysis and ROC curve results confirmed the robustness and reliability of the model. The explainability component represents a critical advancement, as SHAP (SHapley Additive exPlanations) values provide a game theory approach that explains model outputs through feature contributions, enabling researchers and practitioners to understand which EEG features most strongly discriminate between truthful and deceptive responses.

Discussion & Significance

This study's 97.7% accuracy positions LGBM among the most effective approaches for EEG-based deception detection. For context, previous EEG lie detection studies reported maximum accuracies ranging from 70% to 98% using various SVM and feature extraction approaches, with most falling in the 85-92% range. The current study's performance, combined with interpretability, represents a significant methodological advancement.

The emphasis on explainability addresses a fundamental challenge in applying machine learning to forensic and security contexts. Explainability has increasing interest among stakeholders using AI, especially machine learning software, since AI-based software differs from traditional software in having a black-box nature. In legal and investigative settings, understanding why a system classifies a response as deceptive is as important as the classification itself.

LGBM's advantages over other gradient boosting methods likely contributed to the strong performance. LightGBM outperforms other gradient boosting algorithms on large datasets with faster speed and higher accuracy, low memory usage, and parallel and GPU learning support. These characteristics make it particularly well-suited for processing multi-channel EEG data with complex temporal and spectral features.

Limitations & Considerations

While the study reports exceptional accuracy, several methodological considerations warrant attention. The sample size and participant demographics are not specified in the available information, limiting assessment of generalizability across diverse populations. EEG-based lie detection studies typically involve laboratory settings with mock crimes or prescribed deception scenarios, which may not fully replicate the cognitive and emotional states present in real-world deceptive contexts.

The 16-electrode configuration, while practical, represents a relatively sparse spatial sampling compared to high-density EEG systems. This may limit the spatial resolution of brain activity patterns captured, though it enhances practical applicability. EEG signals are non-linear and non-stationary in nature, and features from specific frequency sub-bands obtained using advanced signal processing could provide better information than original brain signals.

Cross-validation methodology and testing on independent datasets remain important for confirming the robustness of the 97.7% accuracy rate. Large datasets representative of the target population are more likely to generalize well when applied to new patients, reinforcing the importance of internal and external validation.

Practical Applications

The LGBM framework's combination of high accuracy and interpretability makes it particularly promising for applied settings. The explainability features enable forensic examiners, security personnel, and legal professionals to understand the basis for deception classifications, facilitating informed decision-making and potentially supporting admissibility in legal proceedings where transparency is required.

The 16-electrode configuration offers practical advantages for field deployment compared to high-density systems requiring extensive setup time. The computational efficiency of LGBM, with its fast training and inference capabilities, supports real-time or near-real-time applications. For polygraph professionals and security screening programs, this technology represents a potential complement or alternative to traditional psychophysiological detection methods, offering resistance to countermeasures through direct measurement of neural activity.