BioTWavePat: An EEG-Based Mental Performance Detection Model

Background/Objectives: EEG-based mental performance classification requires participant-separated validation and traceable feature identities. Methods: BioTWavePat-XFE constructs five single-rhythm and 15 signed cross-rhythm channel-rank transition layers, uses four cumulative-weighted selectors and tkNN/IMV classification, and decodes retained identities with Directed Lobish. We compared Protocol A (whole-dataset feature and pooled-test outcome selection), Protocol B (training-fold outcome selection with Protocol A features), and Protocol C (all data-dependent steps restricted to outer training data). STEW and MAT used participant-wise LOSO. TMPD included all 55 participants in 54 identifier groups; two same-class participants sharing an identifier were withheld together. Results: Protocol A segment accuracies were 83.08% (STEW), 93.14% (TMPD), and 91.15% (MAT). Protocol C accuracies were 80.46%, 82.50%, and 74.65%, with balanced accuracies of 80.46%, 75.17%, and 64.15%, respectively. These are the primary estimates without the identified selection leakage under the stated participant-separated validation design. Conclusions: Whole-dataset selection inflated the original estimates, particularly for TMPD and MAT. MAT Protocol C performance was close to majority prediction. Directed Lobish exposes selected feature identities without establishing physiological connectivity, causality or biomarker validity.