Wavelet Decomposition and Learning Architectures for EEG-Based Lie Detection: A LOSOCV-Based Comparative Evaluation of DWT Feature Representations on the LieWaves Dataset

This study systematically evaluates discrete wavelet transform (DWT)-based EEG feature representations and machine/deep learning classifiers for subject-independent deception detection using the publicly available LieWaves dataset. EEG signals were bandpass filtered, processed with Automatic and Tunable Artifact Removal (ATAR), segmented with the Overlapping Sliding Window (OSW) method, and decomposed using DWT. Four configurations (db4-Level4, db4-Level5, db5-Level4, and db5-Level5) were compared within an outer Leave-One-Subject-Out Cross-Validation (LOSOCV) framework using a three-subject, subject-wise inner holdout procedure; db5-Level5 achieved the highest mean inner-validation performance. Using the fixed db5-Level5 representation, DWT-derived statistical features were evaluated with an EEGNet-inspired compact convolutional model, CNN, LSTM, BiLSTM, GRU, Random Forest, KNN, and SVM under LOSOCV. The EEGNet-inspired model achieved the highest average performance, with an accuracy of 0.5779±0.0903 and a macro F1-score of 0.5295±0.1359. Nevertheless, all models showed limited subject-independent performance. These results highlight the importance of subject-independent validation for highly overlapping EEG segments and indicate that the proposed pipeline should be regarded as an exploratory comparative framework rather than a ready-to-use practical or forensic lie detection system.